Upload 200 files

.gitattributes

@@ -41,3 +41,11 @@ face_anon_simple-main/my_dataset/train/celeb/real/01758_01758.png filter=lfs dif
 face_anon_simple-main/my_dataset/train/celeb/real/01758_09704.png filter=lfs diff=lfs merge=lfs -text
 face_anon_simple-main/my_dataset/train/celeb/real/18147_06771.png filter=lfs diff=lfs merge=lfs -text
 face_anon_simple-main/teaser.jpg filter=lfs diff=lfs merge=lfs -text
+my_dataset/test/00482.png filter=lfs diff=lfs merge=lfs -text
+my_dataset/test/14795.png filter=lfs diff=lfs merge=lfs -text
+my_dataset/test/friends.jpg filter=lfs diff=lfs merge=lfs -text
+my_dataset/train/celeb/fake/18147_06771-01758_01758.png filter=lfs diff=lfs merge=lfs -text
+my_dataset/train/celeb/real/01758_01758.png filter=lfs diff=lfs merge=lfs -text
+my_dataset/train/celeb/real/01758_09704.png filter=lfs diff=lfs merge=lfs -text
+my_dataset/train/celeb/real/18147_06771.png filter=lfs diff=lfs merge=lfs -text
+teaser.jpg filter=lfs diff=lfs merge=lfs -text

LICENSE

@@ -0,0 +1,661 @@
+                    GNU AFFERO GENERAL PUBLIC LICENSE
+                       Version 3, 19 November 2007
+
+ Copyright (C) 2007 Free Software Foundation, Inc. <https://fsf.org/>
+ Everyone is permitted to copy and distribute verbatim copies
+ of this license document, but changing it is not allowed.
+
+                            Preamble
+
+  The GNU Affero General Public License is a free, copyleft license for
+software and other kinds of works, specifically designed to ensure
+cooperation with the community in the case of network server software.
+
+  The licenses for most software and other practical works are designed
+to take away your freedom to share and change the works.  By contrast,
+our General Public Licenses are intended to guarantee your freedom to
+share and change all versions of a program--to make sure it remains free
+software for all its users.
+
+  When we speak of free software, we are referring to freedom, not
+price.  Our General Public Licenses are designed to make sure that you
+have the freedom to distribute copies of free software (and charge for
+them if you wish), that you receive source code or can get it if you
+want it, that you can change the software or use pieces of it in new
+free programs, and that you know you can do these things.
+
+  Developers that use our General Public Licenses protect your rights
+with two steps: (1) assert copyright on the software, and (2) offer
+you this License which gives you legal permission to copy, distribute
+and/or modify the software.
+
+  A secondary benefit of defending all users' freedom is that
+improvements made in alternate versions of the program, if they
+receive widespread use, become available for other developers to
+incorporate.  Many developers of free software are heartened and
+encouraged by the resulting cooperation.  However, in the case of
+software used on network servers, this result may fail to come about.
+The GNU General Public License permits making a modified version and
+letting the public access it on a server without ever releasing its
+source code to the public.
+
+  The GNU Affero General Public License is designed specifically to
+ensure that, in such cases, the modified source code becomes available
+to the community.  It requires the operator of a network server to
+provide the source code of the modified version running there to the
+users of that server.  Therefore, public use of a modified version, on
+a publicly accessible server, gives the public access to the source
+code of the modified version.
+
+  An older license, called the Affero General Public License and
+published by Affero, was designed to accomplish similar goals.  This is
+a different license, not a version of the Affero GPL, but Affero has
+released a new version of the Affero GPL which permits relicensing under
+this license.
+
+  The precise terms and conditions for copying, distribution and
+modification follow.
+
+                       TERMS AND CONDITIONS
+
+  0. Definitions.
+
+  "This License" refers to version 3 of the GNU Affero General Public License.
+
+  "Copyright" also means copyright-like laws that apply to other kinds of
+works, such as semiconductor masks.
+
+  "The Program" refers to any copyrightable work licensed under this
+License.  Each licensee is addressed as "you".  "Licensees" and
+"recipients" may be individuals or organizations.
+
+  To "modify" a work means to copy from or adapt all or part of the work
+in a fashion requiring copyright permission, other than the making of an
+exact copy.  The resulting work is called a "modified version" of the
+earlier work or a work "based on" the earlier work.
+
+  A "covered work" means either the unmodified Program or a work based
+on the Program.
+
+  To "propagate" a work means to do anything with it that, without
+permission, would make you directly or secondarily liable for
+infringement under applicable copyright law, except executing it on a
+computer or modifying a private copy.  Propagation includes copying,
+distribution (with or without modification), making available to the
+public, and in some countries other activities as well.
+
+  To "convey" a work means any kind of propagation that enables other
+parties to make or receive copies.  Mere interaction with a user through
+a computer network, with no transfer of a copy, is not conveying.
+
+  An interactive user interface displays "Appropriate Legal Notices"
+to the extent that it includes a convenient and prominently visible
+feature that (1) displays an appropriate copyright notice, and (2)
+tells the user that there is no warranty for the work (except to the
+extent that warranties are provided), that licensees may convey the
+work under this License, and how to view a copy of this License.  If
+the interface presents a list of user commands or options, such as a
+menu, a prominent item in the list meets this criterion.
+
+  1. Source Code.
+
+  The "source code" for a work means the preferred form of the work
+for making modifications to it.  "Object code" means any non-source
+form of a work.
+
+  A "Standard Interface" means an interface that either is an official
+standard defined by a recognized standards body, or, in the case of
+interfaces specified for a particular programming language, one that
+is widely used among developers working in that language.
+
+  The "System Libraries" of an executable work include anything, other
+than the work as a whole, that (a) is included in the normal form of
+packaging a Major Component, but which is not part of that Major
+Component, and (b) serves only to enable use of the work with that
+Major Component, or to implement a Standard Interface for which an
+implementation is available to the public in source code form.  A
+"Major Component", in this context, means a major essential component
+(kernel, window system, and so on) of the specific operating system
+(if any) on which the executable work runs, or a compiler used to
+produce the work, or an object code interpreter used to run it.
+
+  The "Corresponding Source" for a work in object code form means all
+the source code needed to generate, install, and (for an executable
+work) run the object code and to modify the work, including scripts to
+control those activities.  However, it does not include the work's
+System Libraries, or general-purpose tools or generally available free
+programs which are used unmodified in performing those activities but
+which are not part of the work.  For example, Corresponding Source
+includes interface definition files associated with source files for
+the work, and the source code for shared libraries and dynamically
+linked subprograms that the work is specifically designed to require,
+such as by intimate data communication or control flow between those
+subprograms and other parts of the work.
+
+  The Corresponding Source need not include anything that users
+can regenerate automatically from other parts of the Corresponding
+Source.
+
+  The Corresponding Source for a work in source code form is that
+same work.
+
+  2. Basic Permissions.
+
+  All rights granted under this License are granted for the term of
+copyright on the Program, and are irrevocable provided the stated
+conditions are met.  This License explicitly affirms your unlimited
+permission to run the unmodified Program.  The output from running a
+covered work is covered by this License only if the output, given its
+content, constitutes a covered work.  This License acknowledges your
+rights of fair use or other equivalent, as provided by copyright law.
+
+  You may make, run and propagate covered works that you do not
+convey, without conditions so long as your license otherwise remains
+in force.  You may convey covered works to others for the sole purpose
+of having them make modifications exclusively for you, or provide you
+with facilities for running those works, provided that you comply with
+the terms of this License in conveying all material for which you do
+not control copyright.  Those thus making or running the covered works
+for you must do so exclusively on your behalf, under your direction
+and control, on terms that prohibit them from making any copies of
+your copyrighted material outside their relationship with you.
+
+  Conveying under any other circumstances is permitted solely under
+the conditions stated below.  Sublicensing is not allowed; section 10
+makes it unnecessary.
+
+  3. Protecting Users' Legal Rights From Anti-Circumvention Law.
+
+  No covered work shall be deemed part of an effective technological
+measure under any applicable law fulfilling obligations under article
+11 of the WIPO copyright treaty adopted on 20 December 1996, or
+similar laws prohibiting or restricting circumvention of such
+measures.
+
+  When you convey a covered work, you waive any legal power to forbid
+circumvention of technological measures to the extent such circumvention
+is effected by exercising rights under this License with respect to
+the covered work, and you disclaim any intention to limit operation or
+modification of the work as a means of enforcing, against the work's
+users, your or third parties' legal rights to forbid circumvention of
+technological measures.
+
+  4. Conveying Verbatim Copies.
+
+  You may convey verbatim copies of the Program's source code as you
+receive it, in any medium, provided that you conspicuously and
+appropriately publish on each copy an appropriate copyright notice;
+keep intact all notices stating that this License and any
+non-permissive terms added in accord with section 7 apply to the code;
+keep intact all notices of the absence of any warranty; and give all
+recipients a copy of this License along with the Program.
+
+  You may charge any price or no price for each copy that you convey,
+and you may offer support or warranty protection for a fee.
+
+  5. Conveying Modified Source Versions.
+
+  You may convey a work based on the Program, or the modifications to
+produce it from the Program, in the form of source code under the
+terms of section 4, provided that you also meet all of these conditions:
+
+    a) The work must carry prominent notices stating that you modified
+    it, and giving a relevant date.
+
+    b) The work must carry prominent notices stating that it is
+    released under this License and any conditions added under section
+    7.  This requirement modifies the requirement in section 4 to
+    "keep intact all notices".
+
+    c) You must license the entire work, as a whole, under this
+    License to anyone who comes into possession of a copy.  This
+    License will therefore apply, along with any applicable section 7
+    additional terms, to the whole of the work, and all its parts,
+    regardless of how they are packaged.  This License gives no
+    permission to license the work in any other way, but it does not
+    invalidate such permission if you have separately received it.
+
+    d) If the work has interactive user interfaces, each must display
+    Appropriate Legal Notices; however, if the Program has interactive
+    interfaces that do not display Appropriate Legal Notices, your
+    work need not make them do so.
+
+  A compilation of a covered work with other separate and independent
+works, which are not by their nature extensions of the covered work,
+and which are not combined with it such as to form a larger program,
+in or on a volume of a storage or distribution medium, is called an
+"aggregate" if the compilation and its resulting copyright are not
+used to limit the access or legal rights of the compilation's users
+beyond what the individual works permit.  Inclusion of a covered work
+in an aggregate does not cause this License to apply to the other
+parts of the aggregate.
+
+  6. Conveying Non-Source Forms.
+
+  You may convey a covered work in object code form under the terms
+of sections 4 and 5, provided that you also convey the
+machine-readable Corresponding Source under the terms of this License,
+in one of these ways:
+
+    a) Convey the object code in, or embodied in, a physical product
+    (including a physical distribution medium), accompanied by the
+    Corresponding Source fixed on a durable physical medium
+    customarily used for software interchange.
+
+    b) Convey the object code in, or embodied in, a physical product
+    (including a physical distribution medium), accompanied by a
+    written offer, valid for at least three years and valid for as
+    long as you offer spare parts or customer support for that product
+    model, to give anyone who possesses the object code either (1) a
+    copy of the Corresponding Source for all the software in the
+    product that is covered by this License, on a durable physical
+    medium customarily used for software interchange, for a price no
+    more than your reasonable cost of physically performing this
+    conveying of source, or (2) access to copy the
+    Corresponding Source from a network server at no charge.
+
+    c) Convey individual copies of the object code with a copy of the
+    written offer to provide the Corresponding Source.  This
+    alternative is allowed only occasionally and noncommercially, and
+    only if you received the object code with such an offer, in accord
+    with subsection 6b.
+
+    d) Convey the object code by offering access from a designated
+    place (gratis or for a charge), and offer equivalent access to the
+    Corresponding Source in the same way through the same place at no
+    further charge.  You need not require recipients to copy the
+    Corresponding Source along with the object code.  If the place to
+    copy the object code is a network server, the Corresponding Source
+    may be on a different server (operated by you or a third party)
+    that supports equivalent copying facilities, provided you maintain
+    clear directions next to the object code saying where to find the
+    Corresponding Source.  Regardless of what server hosts the
+    Corresponding Source, you remain obligated to ensure that it is
+    available for as long as needed to satisfy these requirements.
+
+    e) Convey the object code using peer-to-peer transmission, provided
+    you inform other peers where the object code and Corresponding
+    Source of the work are being offered to the general public at no
+    charge under subsection 6d.
+
+  A separable portion of the object code, whose source code is excluded
+from the Corresponding Source as a System Library, need not be
+included in conveying the object code work.
+
+  A "User Product" is either (1) a "consumer product", which means any
+tangible personal property which is normally used for personal, family,
+or household purposes, or (2) anything designed or sold for incorporation
+into a dwelling.  In determining whether a product is a consumer product,
+doubtful cases shall be resolved in favor of coverage.  For a particular
+product received by a particular user, "normally used" refers to a
+typical or common use of that class of product, regardless of the status
+of the particular user or of the way in which the particular user
+actually uses, or expects or is expected to use, the product.  A product
+is a consumer product regardless of whether the product has substantial
+commercial, industrial or non-consumer uses, unless such uses represent
+the only significant mode of use of the product.
+
+  "Installation Information" for a User Product means any methods,
+procedures, authorization keys, or other information required to install
+and execute modified versions of a covered work in that User Product from
+a modified version of its Corresponding Source.  The information must
+suffice to ensure that the continued functioning of the modified object
+code is in no case prevented or interfered with solely because
+modification has been made.
+
+  If you convey an object code work under this section in, or with, or
+specifically for use in, a User Product, and the conveying occurs as
+part of a transaction in which the right of possession and use of the
+User Product is transferred to the recipient in perpetuity or for a
+fixed term (regardless of how the transaction is characterized), the
+Corresponding Source conveyed under this section must be accompanied
+by the Installation Information.  But this requirement does not apply
+if neither you nor any third party retains the ability to install
+modified object code on the User Product (for example, the work has
+been installed in ROM).
+
+  The requirement to provide Installation Information does not include a
+requirement to continue to provide support service, warranty, or updates
+for a work that has been modified or installed by the recipient, or for
+the User Product in which it has been modified or installed.  Access to a
+network may be denied when the modification itself materially and
+adversely affects the operation of the network or violates the rules and
+protocols for communication across the network.
+
+  Corresponding Source conveyed, and Installation Information provided,
+in accord with this section must be in a format that is publicly
+documented (and with an implementation available to the public in
+source code form), and must require no special password or key for
+unpacking, reading or copying.
+
+  7. Additional Terms.
+
+  "Additional permissions" are terms that supplement the terms of this
+License by making exceptions from one or more of its conditions.
+Additional permissions that are applicable to the entire Program shall
+be treated as though they were included in this License, to the extent
+that they are valid under applicable law.  If additional permissions
+apply only to part of the Program, that part may be used separately
+under those permissions, but the entire Program remains governed by
+this License without regard to the additional permissions.
+
+  When you convey a copy of a covered work, you may at your option
+remove any additional permissions from that copy, or from any part of
+it.  (Additional permissions may be written to require their own
+removal in certain cases when you modify the work.)  You may place
+additional permissions on material, added by you to a covered work,
+for which you have or can give appropriate copyright permission.
+
+  Notwithstanding any other provision of this License, for material you
+add to a covered work, you may (if authorized by the copyright holders of
+that material) supplement the terms of this License with terms:
+
+    a) Disclaiming warranty or limiting liability differently from the
+    terms of sections 15 and 16 of this License; or
+
+    b) Requiring preservation of specified reasonable legal notices or
+    author attributions in that material or in the Appropriate Legal
+    Notices displayed by works containing it; or
+
+    c) Prohibiting misrepresentation of the origin of that material, or
+    requiring that modified versions of such material be marked in
+    reasonable ways as different from the original version; or
+
+    d) Limiting the use for publicity purposes of names of licensors or
+    authors of the material; or
+
+    e) Declining to grant rights under trademark law for use of some
+    trade names, trademarks, or service marks; or
+
+    f) Requiring indemnification of licensors and authors of that
+    material by anyone who conveys the material (or modified versions of
+    it) with contractual assumptions of liability to the recipient, for
+    any liability that these contractual assumptions directly impose on
+    those licensors and authors.
+
+  All other non-permissive additional terms are considered "further
+restrictions" within the meaning of section 10.  If the Program as you
+received it, or any part of it, contains a notice stating that it is
+governed by this License along with a term that is a further
+restriction, you may remove that term.  If a license document contains
+a further restriction but permits relicensing or conveying under this
+License, you may add to a covered work material governed by the terms
+of that license document, provided that the further restriction does
+not survive such relicensing or conveying.
+
+  If you add terms to a covered work in accord with this section, you
+must place, in the relevant source files, a statement of the
+additional terms that apply to those files, or a notice indicating
+where to find the applicable terms.
+
+  Additional terms, permissive or non-permissive, may be stated in the
+form of a separately written license, or stated as exceptions;
+the above requirements apply either way.
+
+  8. Termination.
+
+  You may not propagate or modify a covered work except as expressly
+provided under this License.  Any attempt otherwise to propagate or
+modify it is void, and will automatically terminate your rights under
+this License (including any patent licenses granted under the third
+paragraph of section 11).
+
+  However, if you cease all violation of this License, then your
+license from a particular copyright holder is reinstated (a)
+provisionally, unless and until the copyright holder explicitly and
+finally terminates your license, and (b) permanently, if the copyright
+holder fails to notify you of the violation by some reasonable means
+prior to 60 days after the cessation.
+
+  Moreover, your license from a particular copyright holder is
+reinstated permanently if the copyright holder notifies you of the
+violation by some reasonable means, this is the first time you have
+received notice of violation of this License (for any work) from that
+copyright holder, and you cure the violation prior to 30 days after
+your receipt of the notice.
+
+  Termination of your rights under this section does not terminate the
+licenses of parties who have received copies or rights from you under
+this License.  If your rights have been terminated and not permanently
+reinstated, you do not qualify to receive new licenses for the same
+material under section 10.
+
+  9. Acceptance Not Required for Having Copies.
+
+  You are not required to accept this License in order to receive or
+run a copy of the Program.  Ancillary propagation of a covered work
+occurring solely as a consequence of using peer-to-peer transmission
+to receive a copy likewise does not require acceptance.  However,
+nothing other than this License grants you permission to propagate or
+modify any covered work.  These actions infringe copyright if you do
+not accept this License.  Therefore, by modifying or propagating a
+covered work, you indicate your acceptance of this License to do so.
+
+  10. Automatic Licensing of Downstream Recipients.
+
+  Each time you convey a covered work, the recipient automatically
+receives a license from the original licensors, to run, modify and
+propagate that work, subject to this License.  You are not responsible
+for enforcing compliance by third parties with this License.
+
+  An "entity transaction" is a transaction transferring control of an
+organization, or substantially all assets of one, or subdividing an
+organization, or merging organizations.  If propagation of a covered
+work results from an entity transaction, each party to that
+transaction who receives a copy of the work also receives whatever
+licenses to the work the party's predecessor in interest had or could
+give under the previous paragraph, plus a right to possession of the
+Corresponding Source of the work from the predecessor in interest, if
+the predecessor has it or can get it with reasonable efforts.
+
+  You may not impose any further restrictions on the exercise of the
+rights granted or affirmed under this License.  For example, you may
+not impose a license fee, royalty, or other charge for exercise of
+rights granted under this License, and you may not initiate litigation
+(including a cross-claim or counterclaim in a lawsuit) alleging that
+any patent claim is infringed by making, using, selling, offering for
+sale, or importing the Program or any portion of it.
+
+  11. Patents.
+
+  A "contributor" is a copyright holder who authorizes use under this
+License of the Program or a work on which the Program is based.  The
+work thus licensed is called the contributor's "contributor version".
+
+  A contributor's "essential patent claims" are all patent claims
+owned or controlled by the contributor, whether already acquired or
+hereafter acquired, that would be infringed by some manner, permitted
+by this License, of making, using, or selling its contributor version,
+but do not include claims that would be infringed only as a
+consequence of further modification of the contributor version.  For
+purposes of this definition, "control" includes the right to grant
+patent sublicenses in a manner consistent with the requirements of
+this License.
+
+  Each contributor grants you a non-exclusive, worldwide, royalty-free
+patent license under the contributor's essential patent claims, to
+make, use, sell, offer for sale, import and otherwise run, modify and
+propagate the contents of its contributor version.
+
+  In the following three paragraphs, a "patent license" is any express
+agreement or commitment, however denominated, not to enforce a patent
+(such as an express permission to practice a patent or covenant not to
+sue for patent infringement).  To "grant" such a patent license to a
+party means to make such an agreement or commitment not to enforce a
+patent against the party.
+
+  If you convey a covered work, knowingly relying on a patent license,
+and the Corresponding Source of the work is not available for anyone
+to copy, free of charge and under the terms of this License, through a
+publicly available network server or other readily accessible means,
+then you must either (1) cause the Corresponding Source to be so
+available, or (2) arrange to deprive yourself of the benefit of the
+patent license for this particular work, or (3) arrange, in a manner
+consistent with the requirements of this License, to extend the patent
+license to downstream recipients.  "Knowingly relying" means you have
+actual knowledge that, but for the patent license, your conveying the
+covered work in a country, or your recipient's use of the covered work
+in a country, would infringe one or more identifiable patents in that
+country that you have reason to believe are valid.
+
+  If, pursuant to or in connection with a single transaction or
+arrangement, you convey, or propagate by procuring conveyance of, a
+covered work, and grant a patent license to some of the parties
+receiving the covered work authorizing them to use, propagate, modify
+or convey a specific copy of the covered work, then the patent license
+you grant is automatically extended to all recipients of the covered
+work and works based on it.
+
+  A patent license is "discriminatory" if it does not include within
+the scope of its coverage, prohibits the exercise of, or is
+conditioned on the non-exercise of one or more of the rights that are
+specifically granted under this License.  You may not convey a covered
+work if you are a party to an arrangement with a third party that is
+in the business of distributing software, under which you make payment
+to the third party based on the extent of your activity of conveying
+the work, and under which the third party grants, to any of the
+parties who would receive the covered work from you, a discriminatory
+patent license (a) in connection with copies of the covered work
+conveyed by you (or copies made from those copies), or (b) primarily
+for and in connection with specific products or compilations that
+contain the covered work, unless you entered into that arrangement,
+or that patent license was granted, prior to 28 March 2007.
+
+  Nothing in this License shall be construed as excluding or limiting
+any implied license or other defenses to infringement that may
+otherwise be available to you under applicable patent law.
+
+  12. No Surrender of Others' Freedom.
+
+  If conditions are imposed on you (whether by court order, agreement or
+otherwise) that contradict the conditions of this License, they do not
+excuse you from the conditions of this License.  If you cannot convey a
+covered work so as to satisfy simultaneously your obligations under this
+License and any other pertinent obligations, then as a consequence you may
+not convey it at all.  For example, if you agree to terms that obligate you
+to collect a royalty for further conveying from those to whom you convey
+the Program, the only way you could satisfy both those terms and this
+License would be to refrain entirely from conveying the Program.
+
+  13. Remote Network Interaction; Use with the GNU General Public License.
+
+  Notwithstanding any other provision of this License, if you modify the
+Program, your modified version must prominently offer all users
+interacting with it remotely through a computer network (if your version
+supports such interaction) an opportunity to receive the Corresponding
+Source of your version by providing access to the Corresponding Source
+from a network server at no charge, through some standard or customary
+means of facilitating copying of software.  This Corresponding Source
+shall include the Corresponding Source for any work covered by version 3
+of the GNU General Public License that is incorporated pursuant to the
+following paragraph.
+
+  Notwithstanding any other provision of this License, you have
+permission to link or combine any covered work with a work licensed
+under version 3 of the GNU General Public License into a single
+combined work, and to convey the resulting work.  The terms of this
+License will continue to apply to the part which is the covered work,
+but the work with which it is combined will remain governed by version
+3 of the GNU General Public License.
+
+  14. Revised Versions of this License.
+
+  The Free Software Foundation may publish revised and/or new versions of
+the GNU Affero General Public License from time to time.  Such new versions
+will be similar in spirit to the present version, but may differ in detail to
+address new problems or concerns.
+
+  Each version is given a distinguishing version number.  If the
+Program specifies that a certain numbered version of the GNU Affero General
+Public License "or any later version" applies to it, you have the
+option of following the terms and conditions either of that numbered
+version or of any later version published by the Free Software
+Foundation.  If the Program does not specify a version number of the
+GNU Affero General Public License, you may choose any version ever published
+by the Free Software Foundation.
+
+  If the Program specifies that a proxy can decide which future
+versions of the GNU Affero General Public License can be used, that proxy's
+public statement of acceptance of a version permanently authorizes you
+to choose that version for the Program.
+
+  Later license versions may give you additional or different
+permissions.  However, no additional obligations are imposed on any
+author or copyright holder as a result of your choosing to follow a
+later version.
+
+  15. Disclaimer of Warranty.
+
+  THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY
+APPLICABLE LAW.  EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT
+HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY
+OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO,
+THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+PURPOSE.  THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM
+IS WITH YOU.  SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF
+ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
+
+  16. Limitation of Liability.
+
+  IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
+WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS
+THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY
+GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE
+USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF
+DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD
+PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS),
+EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF
+SUCH DAMAGES.
+
+  17. Interpretation of Sections 15 and 16.
+
+  If the disclaimer of warranty and limitation of liability provided
+above cannot be given local legal effect according to their terms,
+reviewing courts shall apply local law that most closely approximates
+an absolute waiver of all civil liability in connection with the
+Program, unless a warranty or assumption of liability accompanies a
+copy of the Program in return for a fee.
+
+                     END OF TERMS AND CONDITIONS
+
+            How to Apply These Terms to Your New Programs
+
+  If you develop a new program, and you want it to be of the greatest
+possible use to the public, the best way to achieve this is to make it
+free software which everyone can redistribute and change under these terms.
+
+  To do so, attach the following notices to the program.  It is safest
+to attach them to the start of each source file to most effectively
+state the exclusion of warranty; and each file should have at least
+the "copyright" line and a pointer to where the full notice is found.
+
+    <one line to give the program's name and a brief idea of what it does.>
+    Copyright (C) <year>  <name of author>
+
+    This program is free software: you can redistribute it and/or modify
+    it under the terms of the GNU Affero General Public License as published
+    by the Free Software Foundation, either version 3 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU Affero General Public License for more details.
+
+    You should have received a copy of the GNU Affero General Public License
+    along with this program.  If not, see <https://www.gnu.org/licenses/>.
+
+Also add information on how to contact you by electronic and paper mail.
+
+  If your software can interact with users remotely through a computer
+network, you should also make sure that it provides a way for users to
+get its source.  For example, if your program is a web application, its
+interface could display a "Source" link that leads users to an archive
+of the code.  There are many ways you could offer source, and different
+solutions will be better for different programs; see section 13 for the
+specific requirements.
+
+  You should also get your employer (if you work as a programmer) or school,
+if any, to sign a "copyright disclaimer" for the program, if necessary.
+For more information on this, and how to apply and follow the GNU AGPL, see
+<https://www.gnu.org/licenses/>.

README.md

@@ -1,12 +1,272 @@
----
-title: Anonymizeface
-emoji: 🐠
-colorFrom: green
-colorTo: pink
-sdk: gradio
-sdk_version: 5.49.1
-app_file: app.py
-pinned: false
----
-
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# Face Anonymization Made Simple (WACV 2025)
+
+[arXiv](http://arxiv.org/abs/2411.00762)
+
+![teaser](teaser.jpg)
+
+Our face anonymization technique preserves the original facial expressions, head positioning, eye direction, and background elements, effectively masking identity while retaining other crucial details. The anonymized face blends seamlessly into its original photograph, making it ideal for diverse real-world applications.
+
+## Setup
+
+1. Clone the repository.
+
+```bash
+git clone https://github.com/hanweikung/face_anon_simple.git
+```
+
+2. Create a Python environment from the `environment.yml` file.
+
+```bash
+conda env create -f environment.yml
+```
+
+## Usage
+1. Import the library.
+
+```python
+import torch
+from transformers import CLIPImageProcessor, CLIPVisionModel
+
+from diffusers import AutoencoderKL, DDPMScheduler
+from diffusers.utils import load_image
+from src.diffusers.models.referencenet.referencenet_unet_2d_condition import (
+    ReferenceNetModel,
+)
+from src.diffusers.models.referencenet.unet_2d_condition import UNet2DConditionModel
+from src.diffusers.pipelines.referencenet.pipeline_referencenet import (
+    StableDiffusionReferenceNetPipeline,
+)
+```
+
+2. Create & load models.
+
+```python
+face_model_id = "hkung/face-anon-simple"
+clip_model_id = "openai/clip-vit-large-patch14"
+sd_model_id = "stabilityai/stable-diffusion-2-1"
+
+unet = UNet2DConditionModel.from_pretrained(
+    face_model_id, subfolder="unet", use_safetensors=True
+)
+referencenet = ReferenceNetModel.from_pretrained(
+    face_model_id, subfolder="referencenet", use_safetensors=True
+)
+conditioning_referencenet = ReferenceNetModel.from_pretrained(
+    face_model_id, subfolder="conditioning_referencenet", use_safetensors=True
+)
+vae = AutoencoderKL.from_pretrained(sd_model_id, subfolder="vae", use_safetensors=True)
+scheduler = DDPMScheduler.from_pretrained(
+    sd_model_id, subfolder="scheduler", use_safetensors=True
+)
+feature_extractor = CLIPImageProcessor.from_pretrained(
+    clip_model_id, use_safetensors=True
+)
+image_encoder = CLIPVisionModel.from_pretrained(clip_model_id, use_safetensors=True)
+
+pipe = StableDiffusionReferenceNetPipeline(
+    unet=unet,
+    referencenet=referencenet,
+    conditioning_referencenet=conditioning_referencenet,
+    vae=vae,
+    feature_extractor=feature_extractor,
+    image_encoder=image_encoder,
+    scheduler=scheduler,
+)
+pipe = pipe.to("cuda")
+
+generator = torch.manual_seed(1)
+```
+
+### Anonymize images with a single aligned face
+
+Create an anonymized version of an image if the image contains a single face and that face has already been aligned similarly to those in the [FFHQ](https://github.com/NVlabs/ffhq-dataset) or [CelebA-HQ](https://github.com/tkarras/progressive_growing_of_gans) datasets.
+
+```python
+# get an input image for anonymization
+original_image = load_image("my_dataset/test/14795.png")
+
+# generate an image that anonymizes faces
+anon_image = pipe(
+    source_image=original_image,
+    conditioning_image=original_image,
+    num_inference_steps=200,
+    guidance_scale=4.0,
+    generator=generator,
+    anonymization_degree=1.25,
+    width=512,
+    height=512,
+).images[0]
+anon_image.save("anon.png")
+```
+
+### Anonymize images with one or multiple unaligned faces
+
+Create an anonymized version of an image if it contains one or more unaligned faces.
+
+```python
+import face_alignment
+from utils.anonymize_faces_in_image import anonymize_faces_in_image
+
+# get an input image for anonymization
+original_image = load_image("my_dataset/test/friends.jpg")
+
+# SFD (likely best results, but slower)
+fa = face_alignment.FaceAlignment(
+    face_alignment.LandmarksType.TWO_D, face_detector="sfd"
+)
+
+# generate an image that anonymizes faces
+anon_image = anonymize_faces_in_image(
+    image=original_image,
+    face_alignment=fa,
+    pipe=pipe,
+    generator=generator,
+    face_image_size=512,
+    num_inference_steps=25,
+    guidance_scale=4.0,
+    anonymization_degree=1.25,
+)
+anon_image.save("anon.png")
+```
+
+### Swap faces between two images
+
+Create an image that swap faces.
+
+```python
+# get source and conditioning (driving) images for face swap
+source_image = load_image("my_dataset/test/00482.png")
+conditioning_image = load_image("my_dataset/test/14795.png")
+
+# generate an image that swaps faces
+swap_image = pipe(
+    source_image=source_image,
+    conditioning_image=conditioning_image,
+    num_inference_steps=200,
+    guidance_scale=4.0,
+    generator=generator,
+    anonymization_degree=0.0,
+    width=512,
+    height=512,
+).images[0]
+swap_image.save("swap.png")
+```
+
+We also provide the [demo.ipynb](https://github.com/hanweikung/face_anon_simple/blob/main/demo.ipynb) notebook, which guides you through the steps mentioned above.
+
+### Note on image resolution
+
+Our model was trained on 512x512 images. To ensure correct results, always set `width=512` and `height=512` in the `pipe` function, and `face_image_size=512` in the `anonymize_faces_in_image` function. This ensures that input images are resized correctly for the diffusion pipeline. If you're using a model trained on different sizes, like 768x768, adjust these parameters accordingly.
+
+## Training
+
+Our model learns face swapping for anonymization. You can train it using your own face-swapped images.
+
+### Training data structure
+
+Sample training data is available in the `my_dataset/train` directory. Real images are stored in the `real` subdirectory, while face-swapped images are stored in the `fake` subdirectory.
+
+```bash
+my_dataset/
+├── train
+│   ├── celeb
+│   │   ├── fake
+│   │   │   └── 18147_06771-01758_01758.png
+│   │   └── real
+│   │       ├── 01758_01758.png
+│   │       ├── 01758_09704.png
+│   │       └── 18147_06771.png
+│   └── train.jsonl
+└── train_dataset_loading_script.py
+```
+
+### Data loading and configuration
+
+Training data is loaded using a JSON lines file (`my_dataset/train.jsonl`) and a dataset loading script (`my_dataset/train_dataset_loading_script.py`), both provided as examples.
+
+The JSON lines file includes two sample entries specifying the source image, conditioning (driving) image, and ground truth image, with file paths based on the sample training data. Adjust these paths to match your own data:
+
+```json
+{"source_image": "celeb/real/18147_06771.png", "conditioning_image": "celeb/real/01758_01758.png", "ground_truth": "celeb/fake/18147_06771-01758_01758.png"}
+{"source_image": "celeb/real/01758_09704.png", "conditioning_image": "celeb/fake/18147_06771-01758_01758.png", "ground_truth": "celeb/real/01758_01758.png"}
+```
+
+To simulate face-swapping behavior, the source and conditioning images should have different identities. The source and ground truth should share the same identity, while the conditioning and ground truth should share the same pose and expression. When no actual ground truth is available (e.g., the first entry), the face-swapped image serves as the ground truth. When a ground truth image is available (e.g., the second entry), the swapped version of the ground truth is used as the conditioning image.
+
+Our dataset loading script follows [Hugging Face's documentation](https://huggingface.co/docs/datasets/en/dataset_script). Please update the `metadata_path` and `images_dir` file paths in the script to match your dataset:
+
+```python
+_URLS = {
+    "metadata_path": "/path/to/face_anon_simple/my_dataset/train/train.jsonl",
+    "images_dir": "/path/to/face_anon_simple/my_dataset/train/",
+}
+```
+
+### Training script setup
+
+A bash script, `train_referencenet.sh`, with the training command is provided. Update the file paths and adjust parameters as needed:
+
+```bash
+export MODEL_DIR="/path/to/stable-diffusion-2-1/"
+export CLIP_MODEL_DIR="/path/to/clip-vit-large-patch14/"
+export OUTPUT_DIR="./runs/celeb/"
+export NCCL_P2P_DISABLE=1
+export DATASET_LOADING_SCRIPT_PATH="./my_dataset/train_dataset_loading_script.py"
+export TORCH_DISTRIBUTED_DEBUG="INFO"
+export WANDB__SERVICE_WAIT="300"
+
+accelerate launch --main_process_port=29500 --mixed_precision="fp16" --multi_gpu -m examples.referencenet.train_referencenet \
+	--pretrained_model_name_or_path=$MODEL_DIR \
+	--pretrained_clip_model_name_or_path=$CLIP_MODEL_DIR \
+	--output_dir=$OUTPUT_DIR \
+	--dataset_loading_script_path=$DATASET_LOADING_SCRIPT_PATH \
+	--resolution=512 \
+	--learning_rate=1e-5 \
+	--validation_source_image "./my_dataset/test/00482.png" \
+	--validation_conditioning_image "./my_dataset/test/14795.png" \
+	--train_batch_size=1 \
+	--tracker_project_name="celeb" \
+	--checkpointing_steps=10000 \
+	--num_validation_images=1 \
+	--validation_steps=1000 \
+	--mixed_precision="fp16" \
+	--gradient_checkpointing \
+	--use_8bit_adam \
+	--enable_xformers_memory_efficient_attention \
+	--gradient_accumulation_steps=8 \
+	--resume_from_checkpoint="latest" \
+	--set_grads_to_none \
+	--max_train_steps=60000 \
+	--conditioning_dropout_prob=0.1 \
+	--seed=0 \
+	--report_to="wandb" \
+	--random_flip \
+	--dataloader_num_workers=8
+```
+
+To train your model, run:
+
+```bash
+bash train_referencenet.sh
+```
+
+## Test images
+
+In our paper, we selected 1,000 images each from [CelebA-HQ](https://github.com/tkarras/progressive_growing_of_gans) and [FFHQ](https://github.com/NVlabs/ffhq-dataset) for quantitative analysis. The list of test images can be found in our [Hugging Face Hub repository](https://huggingface.co/datasets/hkung/face-anon-simple-dataset).
+
+## Citation
+
+```bibtex
+@InProceedings{Kung_2025_WACV,
+    author    = {Kung, Han-Wei and Varanka, Tuomas and Saha, Sanjay and Sim, Terence and Sebe, Nicu},
+    title     = {Face Anonymization Made Simple},
+    booktitle = {Proceedings of the Winter Conference on Applications of Computer Vision (WACV)},
+    month     = {February},
+    year      = {2025},
+    pages     = {1040-1050}
+}
+```
+
+## Acknowledgements
+
+This work is built upon the [Diffusers](https://github.com/huggingface/diffusers) project. The [face extractor](https://github.com/hanweikung/face_anon_simple/blob/main/utils/extractor.py) is adapted from [DeepFaceLab](https://github.com/iperov/DeepFaceLab/blob/master/mainscripts/Extractor.py).

environment.yml

@@ -0,0 +1,18 @@
+name: face-anon-simple
+dependencies:
+  - python=3.8.18
+  - pip
+  - pip:
+    - torch==2.1
+    - torchvision==0.16.0
+    - huggingface_hub<0.26.0
+    - diffusers==0.25.1
+    - transformers==4.46.1
+    - accelerate==1.0.1
+    - datasets==3.1.0
+    - xformers==0.0.22.post7
+    - bitsandbytes==0.44.1
+    - ipykernel==6.29.5
+    - face-alignment==1.4.1
+    - albumentations==1.4.1
+    - wandb==0.16.3

examples/referencenet/infer_referencenet.py

@@ -0,0 +1,277 @@
+import argparse
+import os
+from pathlib import Path
+
+import numpy as np
+import torch
+import torchvision.transforms.v2 as transforms_v2
+from PIL import Image
+from transformers import CLIPImageProcessor, CLIPVisionModel
+from tqdm import tqdm
+from accelerate import Accelerator
+
+from datasets import load_dataset
+from diffusers import AutoencoderKL, DDPMScheduler
+from src.diffusers.models.referencenet.unet_2d_condition import UNet2DConditionModel
+from src.diffusers.models.referencenet.referencenet_unet_2d_condition import ReferenceNetModel
+from src.diffusers.pipelines.referencenet.pipeline_referencenet import StableDiffusionReferenceNetPipeline
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(description="Inference")
+    parser.add_argument(
+        "--pretrained_model_name_or_path",
+        type=str,
+        default="stabilityai/stable-diffusion-2-1",
+        required=False,
+        help="Path to pretrained model or model identifier from huggingface.co/models.",
+    )
+    parser.add_argument(
+        "--pretrained_clip_model_name_or_path",
+        type=str,
+        default="openai/clip-vit-large-patch14",
+        required=False,
+        help="Path to pretrained CLIP model or model identifier from huggingface.co/models.",
+    )
+    parser.add_argument(
+        "--model_path",
+        type=str,
+        default=None,
+        required=True,
+        help="Path to the model trained by yourself",
+    )
+    parser.add_argument(
+        "--dataset_loading_script_path",
+        type=str,
+        default=None,
+        required=True,
+        help="Path to the dataset loading script file",
+    )
+    parser.add_argument(
+        "--output_dir",
+        type=str,
+        default="./test-infer/",
+        help="The output directory where predictions are saved",
+    )
+    parser.add_argument(
+        "--resolution",
+        type=int,
+        default=512,
+        help="The resolution for input images, all the images in the test dataset will be resized to this resolution",
+    )
+    parser.add_argument("--guidance_scale", type=float, default=2.5)
+    parser.add_argument("--num_inference_steps", type=int, default=50)
+    parser.add_argument("--seed", type=int, default=None, help="A seed for reproducible inference.")
+    parser.add_argument(
+        "--anonymization_degree_start",
+        type=float,
+        default=0.0,
+        help="Increasing the anonymization scale value encourages the model to produce images that diverge significantly from the conditioning image.",
+    )
+    parser.add_argument("--anonymization_degree_end", type=float, default=0.0)
+    parser.add_argument("--num_anonymization_degrees", type=int, default=1)
+    parser.add_argument(
+        "--center_crop",
+        default=False,
+        action="store_true",
+        help=(
+            "Whether to center crop the input images to the resolution. If not set, the images will be randomly"
+            " cropped. The images will be resized to the resolution first before cropping."
+        ),
+    )
+    parser.add_argument(
+        "--max_test_samples",
+        type=int,
+        default=None,
+        help="Truncate the number of test examples to this value if set.",
+    )
+    parser.add_argument(
+        "--vis_input",
+        action="store_true",
+        help="If set, save the input and generated images together as a single output image for easy visualization",
+    )
+    parser.add_argument(
+        "--test_batch_size",
+        type=int,
+        default=1,
+        help=(
+            "The batch size for the test dataloader per device should be set to 1."
+            "This setting does not affect performance, no matter how large the batch size is."
+        ),
+    )
+    parser.add_argument(
+        "--dataloader_num_workers",
+        type=int,
+        default=0,
+        help=(
+            "Number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process."
+        ),
+    )
+
+    args = parser.parse_args()
+    return args
+
+
+def combine_images(images):
+    # Get the total width and maximum height of all images
+    total_width = sum(img.width for img in images)
+    max_height = max(img.height for img in images)
+
+    # Create a new image with the combined width and maximum height
+    new_image = Image.new("RGB", (total_width, max_height))
+
+    # Paste each image onto the new image horizontally
+    x_offset = 0
+    for img in images:
+        new_image.paste(img, (x_offset, 0))
+        x_offset += img.width
+
+    return new_image
+
+
+def make_test_dataset(args):
+    ds = load_dataset(path=args.dataset_loading_script_path, split="test", trust_remote_code=True)
+
+    # Preprocessing the datasets.
+    image_transforms = transforms_v2.Compose(
+        [
+            transforms_v2.Resize(args.resolution, interpolation=transforms_v2.InterpolationMode.BILINEAR),
+            transforms_v2.CenterCrop(args.resolution)
+            if args.center_crop
+            else transforms_v2.RandomCrop(args.resolution),
+        ]
+    )
+
+    def preprocess_test(examples):
+        images = [image.convert("RGB") for image in examples["source_image"]]
+        images = [image_transforms(image) for image in images]
+
+        conditioning_images = [image.convert("RGB") for image in examples["conditioning_image"]]
+        conditioning_images = [image_transforms(image) for image in conditioning_images]
+
+        examples["source_image"] = images
+        examples["conditioning_image"] = conditioning_images
+
+        return examples
+
+    if args.max_test_samples is not None:
+        max_test_samples = min(args.max_test_samples, len(ds))
+        ds = ds.select(range(max_test_samples))
+
+    test_dataset = ds.with_transform(preprocess_test)
+    return test_dataset
+
+
+def collate_fn(examples):
+    source_images = [example["source_image"] for example in examples]
+    conditioning_images = [example["conditioning_image"] for example in examples]
+    source_image_paths = [example["source_image_path"] for example in examples]
+    conditioning_image_paths = [example["conditioning_image_path"] for example in examples]
+
+    return {
+        "source_images": source_images,
+        "conditioning_images": conditioning_images,
+        "source_image_paths": source_image_paths,
+        "conditioning_image_paths": conditioning_image_paths,
+    }
+
+
+if __name__ == "__main__":
+    args = parse_args()
+
+    accelerator = Accelerator()
+    device = accelerator.device
+
+    os.makedirs(args.output_dir, exist_ok=True)
+
+    if args.vis_input:
+        output_vis_dir = Path(args.output_dir, "vis")
+        output_vis_dir.mkdir(parents=True, exist_ok=True)
+
+    generator = None
+
+    # create & load model
+    unet = UNet2DConditionModel.from_pretrained(args.model_path, subfolder="unet")
+    referencenet = ReferenceNetModel.from_pretrained(args.model_path, subfolder="referencenet")
+    conditioning_referencenet = ReferenceNetModel.from_pretrained(
+        args.model_path, subfolder="conditioning_referencenet"
+    )
+    vae = AutoencoderKL.from_pretrained(args.pretrained_model_name_or_path, subfolder="vae")
+    feature_extractor = CLIPImageProcessor.from_pretrained(args.pretrained_clip_model_name_or_path)
+    image_encoder = CLIPVisionModel.from_pretrained(args.pretrained_clip_model_name_or_path)
+    scheduler = DDPMScheduler.from_pretrained(args.pretrained_model_name_or_path, subfolder="scheduler")
+
+    pipe = StableDiffusionReferenceNetPipeline(
+        unet=unet,
+        referencenet=referencenet,
+        conditioning_referencenet=conditioning_referencenet,
+        vae=vae,
+        feature_extractor=feature_extractor,
+        image_encoder=image_encoder,
+        scheduler=scheduler,
+    )
+    pipe = pipe.to(device)
+
+    test_dataset = make_test_dataset(args)
+    test_dataloader = torch.utils.data.DataLoader(
+        test_dataset,
+        shuffle=False,
+        collate_fn=collate_fn,
+        batch_size=args.test_batch_size,
+        num_workers=args.dataloader_num_workers,
+    )
+    test_dataloader = accelerator.prepare(test_dataloader)
+
+    # Generate the list of evenly spaced numbers
+    anonymization_degrees = np.linspace(
+        args.anonymization_degree_start, args.anonymization_degree_end, args.num_anonymization_degrees
+    )
+
+    for step, batch in enumerate(tqdm(test_dataloader)):
+        # Group corresponding items from each key together
+        grouped_items = list(
+            zip(
+                batch["source_images"],
+                batch["conditioning_images"],
+                batch["source_image_paths"],
+                batch["conditioning_image_paths"],
+            )
+        )
+
+        for source_image, conditioning_image, source_image_path, conditioning_image_path in grouped_items:
+            source_image_name = Path(source_image_path).stem
+            conditioning_image_name = Path(conditioning_image_path).stem
+
+            for index, anonymization_degree in enumerate(anonymization_degrees):
+                filename = f"{source_image_name}-{conditioning_image_name}_{index:03}.png"
+                save_to = Path(args.output_dir, filename)
+
+                if save_to.is_file():
+                    continue
+
+                if args.seed is not None:
+                    # create a generator for reproducibility; notice you don't place it on the GPU!
+                    generator = torch.manual_seed(args.seed)
+
+                image = pipe(
+                    source_image=source_image,
+                    conditioning_image=conditioning_image,
+                    height=args.resolution,
+                    width=args.resolution,
+                    num_inference_steps=args.num_inference_steps,
+                    guidance_scale=args.guidance_scale,
+                    generator=generator,
+                    anonymization_degree=anonymization_degree,
+                ).images[0]
+
+                image.save(save_to)
+                save_vis_to = Path(output_vis_dir, filename)
+
+                if args.vis_input and not save_vis_to.is_file():
+                    if anonymization_degree > 0.0:
+                        # face anonymization
+                        combined_image = combine_images([conditioning_image, image])
+                    else:
+                        # face swapping
+                        combined_image = combine_images([source_image, conditioning_image, image])
+                    combined_image.save(save_vis_to)

examples/referencenet/train_referencenet.py

@@ -0,0 +1,1304 @@
+#!/usr/bin/env python
+# coding=utf-8
+# Copyright 2023 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+
+import argparse
+import logging
+import math
+import os
+import shutil
+from pathlib import Path
+
+import accelerate
+import datasets
+import numpy as np
+import torch
+import torch.nn.functional as F
+import torch.utils.checkpoint
+import transformers
+from accelerate import Accelerator
+from accelerate.logging import get_logger
+from accelerate.state import AcceleratorState
+from accelerate.utils import ProjectConfiguration, set_seed
+from datasets import load_dataset
+from huggingface_hub import create_repo
+from packaging import version
+from torchvision import transforms
+from tqdm.auto import tqdm
+from transformers import CLIPImageProcessor, CLIPVisionModel
+from transformers.utils import ContextManagers
+
+import diffusers
+from diffusers import AutoencoderKL, DDPMScheduler
+from diffusers.optimization import get_scheduler
+from diffusers.training_utils import EMAModel, compute_snr
+from diffusers.utils import check_min_version, deprecate, is_wandb_available
+from diffusers.utils.import_utils import is_xformers_available
+
+from src.diffusers.models.referencenet.unet_2d_condition import UNet2DConditionModel
+from src.diffusers.models.referencenet.referencenet_unet_2d_condition import ReferenceNetModel
+from src.diffusers.pipelines.referencenet.pipeline_referencenet import (
+    StableDiffusionReferenceNetPipeline,
+    cat_referencenet_states,
+)
+from examples.referencenet.infer_referencenet import combine_images
+from PIL import Image
+from PIL import ImageFile
+
+import torchvision.transforms.v2 as transforms_v2
+
+import albumentations as A
+
+ImageFile.LOAD_TRUNCATED_IMAGES = True
+
+if is_wandb_available():
+    import wandb
+
+
+# Will error if the minimal version of diffusers is not installed. Remove at your own risks.
+check_min_version("0.25.0.dev0")
+
+logger = get_logger(__name__, log_level="INFO")
+
+
+def torch_dfs(model: torch.nn.Module):
+    result = [model]
+    for child in model.children():
+        result += torch_dfs(child)
+    return result
+
+
+def set_parts_of_model_for_gradient_computation(module):
+    # Include attention blocks in gradient computation
+    for attn_processor_name, attn_processor in module.attn_processors.items():
+        attn_module = module
+        for n in attn_processor_name.split(".")[:-1]:
+            attn_module = getattr(attn_module, n)
+        attn_module.requires_grad_(True)
+
+    # Include transformer blocks in gradient computation
+    tb_type = type(module.down_blocks[0].attentions[0].transformer_blocks[0])
+    attn_modules = [module for module in torch_dfs(module) if isinstance(module, tb_type)]
+    attn_modules = sorted(attn_modules, key=lambda x: -x.norm1.normalized_shape[0])
+    [attn_module.requires_grad_(True) for attn_module in attn_modules]
+
+    return module
+
+
+def recursive_multiply(element, tensor):
+    if isinstance(element, tuple) or isinstance(element, list):
+        for element in element:
+            recursive_multiply(element, tensor)
+    elif torch.is_tensor(element):
+        # In-place multiplication
+        element.mul_(tensor)
+    else:
+        raise ValueError("Invalid type encountered in the element")
+
+
+def log_validation(
+    vae, unet, feature_extractor, image_encoder, referencenet, conditioning_referencenet, args, accelerator, step
+):
+    logger.info("Running validation... ")
+    scheduler = DDPMScheduler.from_pretrained(args.pretrained_model_name_or_path, subfolder="scheduler")
+    pipeline = StableDiffusionReferenceNetPipeline(
+        unet=accelerator.unwrap_model(unet),
+        referencenet=accelerator.unwrap_model(referencenet),
+        conditioning_referencenet=accelerator.unwrap_model(conditioning_referencenet),
+        vae=vae,
+        feature_extractor=feature_extractor,
+        image_encoder=image_encoder,
+        scheduler=scheduler,
+    )
+
+    pipeline.set_progress_bar_config(disable=True)
+
+    if args.enable_xformers_memory_efficient_attention:
+        pipeline.enable_xformers_memory_efficient_attention()
+
+    if args.seed is None:
+        generator = None
+    else:
+        generator = torch.Generator(device=accelerator.device).manual_seed(args.seed)
+
+    if len(args.validation_conditioning_image) == len(args.validation_source_image):
+        validation_conditioning_images = args.validation_conditioning_image
+        validation_source_images = args.validation_source_image
+    elif len(args.validation_conditioning_image) == 1:
+        validation_conditioning_images = args.validation_conditioning_image * len(args.validation_source_image)
+        validation_source_images = args.validation_source_image
+    elif len(args.validation_source_image) == 1:
+        validation_conditioning_images = args.validation_conditioning_image
+        validation_source_images = args.validation_source_image * len(args.validation_conditioning_image)
+    else:
+        raise ValueError(
+            "number of `args.validation_conditioning_image` and `args.validation_source_image` should be checked in `parse_args`"
+        )
+    image_logs = []
+
+    image_path = os.path.join(args.output_dir, "outputs")
+    if not os.path.exists(image_path):
+        os.makedirs(image_path)
+
+    for i, (validation_source_image, validation_conditioning_image) in enumerate(
+        zip(validation_source_images, validation_conditioning_images)
+    ):
+        source_image_filename_without_ext = Path(validation_source_image).stem
+        conditioning_image_filename_without_ext = Path(validation_conditioning_image).stem
+
+        # Source images
+        validation_source_image = Image.open(validation_source_image).convert("RGB")
+        validation_source_image = transforms_v2.Resize(size=(args.resolution, args.resolution), antialias=True)(
+            validation_source_image
+        )
+
+        # Driving images
+        validation_conditioning_image = Image.open(validation_conditioning_image).convert("RGB")
+        validation_conditioning_image = transforms_v2.Resize(size=(args.resolution, args.resolution), antialias=True)(
+            validation_conditioning_image
+        )
+
+        images = []
+        for n in range(args.num_validation_images):
+            with torch.autocast("cuda"):
+                image = pipeline(
+                    source_image=validation_source_image,
+                    conditioning_image=validation_conditioning_image,
+                    height=args.resolution,
+                    width=args.resolution,
+                    num_inference_steps=200,
+                    guidance_scale=4.0,
+                    generator=generator,
+                ).images[0]
+
+            images.append(image)
+
+            combined_images = combine_images([validation_source_image, validation_conditioning_image, image])
+            save_to = Path(
+                image_path,
+                f"{i:03}_src_{source_image_filename_without_ext}_drv_{conditioning_image_filename_without_ext}_{n:03}.png",
+            )
+            combined_images.save(save_to, format="PNG")
+
+        image_logs.append(
+            {
+                "validation_source_image": validation_source_image,
+                "validation_conditioning_image": validation_conditioning_image,
+                "images": images,
+            }
+        )
+
+    for tracker in accelerator.trackers:
+        if tracker.name == "tensorboard":
+            for i, log in enumerate(image_logs):
+                images = log["images"]
+                validation_source_image = log["validation_source_image"]
+                validation_conditioning_image = log["validation_conditioning_image"]
+
+                formatted_images = []
+
+                formatted_images.append(np.asarray(validation_source_image))
+                formatted_images.append(np.asarray(validation_conditioning_image))
+
+                for image in images:
+                    formatted_images.append(np.asarray(image))
+
+                formatted_images = np.stack(formatted_images)
+
+                tracker.writer.add_images(f"{i:05}", formatted_images, step, dataformats="NHWC")
+        elif tracker.name == "wandb":
+            formatted_images = []
+
+            for log in image_logs:
+                images = log["images"]
+                validation_source_image = log["validation_source_image"]
+                validation_conditioning_image = log["validation_conditioning_image"]
+
+                formatted_images.append(wandb.Image(validation_source_image, caption="Source"))
+                formatted_images.append(wandb.Image(validation_conditioning_image, caption="Conditioning"))
+
+                for image in images:
+                    image = wandb.Image(image, caption="Generated")
+                    formatted_images.append(image)
+
+            tracker.log({"validation": formatted_images})
+        else:
+            logger.warn(f"image logging not implemented for {tracker.name}")
+
+        return image_logs
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(description="Simple example of a training script.")
+    parser.add_argument(
+        "--input_perturbation", type=float, default=0, help="The scale of input perturbation. Recommended 0.1."
+    )
+    parser.add_argument(
+        "--pretrained_model_name_or_path",
+        type=str,
+        default=None,
+        required=True,
+        help="Path to pretrained model or model identifier from huggingface.co/models.",
+    )
+    parser.add_argument(
+        "--pretrained_clip_model_name_or_path",
+        type=str,
+        default="openai/clip-vit-large-patch14",
+        required=False,
+        help="Path to pretrained CLIP model or model identifier from huggingface.co/models.",
+    )
+    parser.add_argument(
+        "--revision",
+        type=str,
+        default=None,
+        required=False,
+        help="Revision of pretrained model identifier from huggingface.co/models.",
+    )
+    parser.add_argument(
+        "--variant",
+        type=str,
+        default=None,
+        help="Variant of the model files of the pretrained model identifier from huggingface.co/models, 'e.g.' fp16",
+    )
+
+    parser.add_argument(
+        "--dataset_config_name",
+        type=str,
+        default=None,
+        help="The config of the Dataset, leave as None if there's only one config.",
+    )
+    parser.add_argument(
+        "--dataset_loading_script_path",
+        type=str,
+        default=None,
+        required=True,
+        help="Path to the dataset loading script file",
+    )
+    parser.add_argument(
+        "--source_image_column",
+        type=str,
+        default="source_image",
+        help="The column of the dataset containing the referencenet source image.",
+    )
+    parser.add_argument(
+        "--conditioning_image_column",
+        type=str,
+        default="conditioning_image",
+        help="The column of the dataset containing the referencenet conditioning image.",
+    )
+    parser.add_argument(
+        "--ground_truth_column",
+        type=str,
+        default="ground_truth",
+        help="The column of the dataset containing the ground truth image.",
+    )
+    parser.add_argument(
+        "--output_dir",
+        type=str,
+        default="sd-model-finetuned",
+        help="The output directory where the model predictions and checkpoints will be written.",
+    )
+    parser.add_argument(
+        "--cache_dir",
+        type=str,
+        default=None,
+        help="The directory where the downloaded models and datasets will be stored.",
+    )
+    parser.add_argument("--seed", type=int, default=None, help="A seed for reproducible training.")
+    parser.add_argument(
+        "--resolution",
+        type=int,
+        default=512,
+        help=(
+            "The resolution for input images, all the images in the train/validation dataset will be resized to this"
+            " resolution"
+        ),
+    )
+    parser.add_argument(
+        "--center_crop",
+        default=False,
+        action="store_true",
+        help=(
+            "Whether to center crop the input images to the resolution. If not set, the images will be randomly"
+            " cropped. The images will be resized to the resolution first before cropping."
+        ),
+    )
+    parser.add_argument(
+        "--random_flip",
+        action="store_true",
+        help="whether to randomly flip images horizontally",
+    )
+    parser.add_argument(
+        "--train_batch_size", type=int, default=16, help="Batch size (per device) for the training dataloader."
+    )
+    parser.add_argument("--num_train_epochs", type=int, default=100)
+    parser.add_argument(
+        "--max_train_steps",
+        type=int,
+        default=None,
+        help="Total number of training steps to perform.  If provided, overrides num_train_epochs.",
+    )
+    parser.add_argument(
+        "--max_train_samples",
+        type=int,
+        default=None,
+        help=(
+            "For debugging purposes or quicker training, truncate the number of training examples to this "
+            "value if set."
+        ),
+    )
+    parser.add_argument(
+        "--gradient_accumulation_steps",
+        type=int,
+        default=1,
+        help="Number of updates steps to accumulate before performing a backward/update pass.",
+    )
+    parser.add_argument(
+        "--gradient_checkpointing",
+        action="store_true",
+        help="Whether or not to use gradient checkpointing to save memory at the expense of slower backward pass.",
+    )
+    parser.add_argument(
+        "--learning_rate",
+        type=float,
+        default=1e-4,
+        help="Initial learning rate (after the potential warmup period) to use.",
+    )
+    parser.add_argument(
+        "--scale_lr",
+        action="store_true",
+        default=False,
+        help="Scale the learning rate by the number of GPUs, gradient accumulation steps, and batch size.",
+    )
+    parser.add_argument(
+        "--lr_scheduler",
+        type=str,
+        default="constant",
+        help=(
+            'The scheduler type to use. Choose between ["linear", "cosine", "cosine_with_restarts", "polynomial",'
+            ' "constant", "constant_with_warmup"]'
+        ),
+    )
+    parser.add_argument(
+        "--lr_warmup_steps", type=int, default=500, help="Number of steps for the warmup in the lr scheduler."
+    )
+    parser.add_argument(
+        "--snr_gamma",
+        type=float,
+        default=None,
+        help="SNR weighting gamma to be used if rebalancing the loss. Recommended value is 5.0. "
+        "More details here: https://arxiv.org/abs/2303.09556.",
+    )
+    parser.add_argument(
+        "--use_8bit_adam", action="store_true", help="Whether or not to use 8-bit Adam from bitsandbytes."
+    )
+    parser.add_argument(
+        "--allow_tf32",
+        action="store_true",
+        help=(
+            "Whether or not to allow TF32 on Ampere GPUs. Can be used to speed up training. For more information, see"
+            " https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices"
+        ),
+    )
+    parser.add_argument("--use_ema", action="store_true", help="Whether to use EMA model.")
+    parser.add_argument(
+        "--non_ema_revision",
+        type=str,
+        default=None,
+        required=False,
+        help=(
+            "Revision of pretrained non-ema model identifier. Must be a branch, tag or git identifier of the local or"
+            " remote repository specified with --pretrained_model_name_or_path."
+        ),
+    )
+    parser.add_argument(
+        "--dataloader_num_workers",
+        type=int,
+        default=0,
+        help=(
+            "Number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process."
+        ),
+    )
+    parser.add_argument("--adam_beta1", type=float, default=0.9, help="The beta1 parameter for the Adam optimizer.")
+    parser.add_argument("--adam_beta2", type=float, default=0.999, help="The beta2 parameter for the Adam optimizer.")
+    parser.add_argument("--adam_weight_decay", type=float, default=1e-2, help="Weight decay to use.")
+    parser.add_argument("--adam_epsilon", type=float, default=1e-08, help="Epsilon value for the Adam optimizer")
+    parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.")
+    parser.add_argument("--push_to_hub", action="store_true", help="Whether or not to push the model to the Hub.")
+    parser.add_argument("--hub_token", type=str, default=None, help="The token to use to push to the Model Hub.")
+    parser.add_argument(
+        "--prediction_type",
+        type=str,
+        default=None,
+        help="The prediction_type that shall be used for training. Choose between 'epsilon' or 'v_prediction' or leave `None`. If left to `None` the default prediction type of the scheduler: `noise_scheduler.config.prediciton_type` is chosen.",
+    )
+    parser.add_argument(
+        "--hub_model_id",
+        type=str,
+        default=None,
+        help="The name of the repository to keep in sync with the local `output_dir`.",
+    )
+    parser.add_argument(
+        "--logging_dir",
+        type=str,
+        default="logs",
+        help=(
+            "[TensorBoard](https://www.tensorflow.org/tensorboard) log directory. Will default to"
+            " *output_dir/runs/**CURRENT_DATETIME_HOSTNAME***."
+        ),
+    )
+    parser.add_argument(
+        "--mixed_precision",
+        type=str,
+        default=None,
+        choices=["no", "fp16", "bf16"],
+        help=(
+            "Whether to use mixed precision. Choose between fp16 and bf16 (bfloat16). Bf16 requires PyTorch >="
+            " 1.10.and an Nvidia Ampere GPU.  Default to the value of accelerate config of the current system or the"
+            " flag passed with the `accelerate.launch` command. Use this argument to override the accelerate config."
+        ),
+    )
+    parser.add_argument(
+        "--report_to",
+        type=str,
+        default="tensorboard",
+        help=(
+            'The integration to report the results and logs to. Supported platforms are `"tensorboard"`'
+            ' (default), `"wandb"` and `"comet_ml"`. Use `"all"` to report to all integrations.'
+        ),
+    )
+    parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank")
+    parser.add_argument(
+        "--checkpointing_steps",
+        type=int,
+        default=500,
+        help=(
+            "Save a checkpoint of the training state every X updates. These checkpoints are only suitable for resuming"
+            " training using `--resume_from_checkpoint`."
+        ),
+    )
+    parser.add_argument(
+        "--checkpoints_total_limit",
+        type=int,
+        default=None,
+        help=("Max number of checkpoints to store."),
+    )
+    parser.add_argument(
+        "--resume_from_checkpoint",
+        type=str,
+        default=None,
+        help=(
+            "Whether training should be resumed from a previous checkpoint. Use a path saved by"
+            ' `--checkpointing_steps`, or `"latest"` to automatically select the last available checkpoint.'
+        ),
+    )
+    parser.add_argument(
+        "--enable_xformers_memory_efficient_attention", action="store_true", help="Whether or not to use xformers."
+    )
+    parser.add_argument(
+        "--set_grads_to_none",
+        action="store_true",
+        help=(
+            "Save more memory by using setting grads to None instead of zero. Be aware, that this changes certain"
+            " behaviors, so disable this argument if it causes any problems. More info:"
+            " https://pytorch.org/docs/stable/generated/torch.optim.Optimizer.zero_grad.html"
+        ),
+    )
+    parser.add_argument("--noise_offset", type=float, default=0, help="The scale of noise offset.")
+    parser.add_argument(
+        "--validation_conditioning_image",
+        type=str,
+        default=None,
+        nargs="+",
+        help=(
+            "A set of paths to the referencenet conditioning image be evaluated every `--validation_steps`"
+            " and logged to `--report_to`. Provide either a matching number of `--validation_source_image`s, a"
+            " a single `--validation_source_image` to be used with all `--validation_conditioning_image`s, or a single"
+            " `--validation_conditioning_image` that will be used with all `--validation_source_image`s."
+        ),
+    )
+    parser.add_argument(
+        "--validation_source_image",
+        type=str,
+        default=None,
+        nargs="+",
+        help=(
+            "A set of source images evaluated every `--validation_steps` and logged to `--report_to`."
+            " Provide either a matching number of `--validation_conditioning_image`s, a single `--validation_conditioning_image`"
+            " to be used with all source images, or a single source image that will be used with all `--validation_conditioning_image`s."
+        ),
+    )
+    parser.add_argument(
+        "--num_validation_images",
+        type=int,
+        default=4,
+        help="Number of images to be generated for each `--validation_conditioning_image`, `--validation_source_image` pair",
+    )
+    parser.add_argument(
+        "--validation_epochs",
+        type=int,
+        default=5,
+        help="Run validation every X epochs.",
+    )
+    parser.add_argument(
+        "--validation_steps",
+        type=int,
+        default=100,
+        help=(
+            "Run validation every X steps. Validation consists of running the prompt"
+            " `args.validation_source_image` multiple times: `args.num_validation_images`"
+            " and logging the images."
+        ),
+    )
+    parser.add_argument(
+        "--tracker_project_name",
+        type=str,
+        default="text2image-fine-tune",
+        help=(
+            "The `project_name` argument passed to Accelerator.init_trackers for"
+            " more information see https://huggingface.co/docs/accelerate/v0.17.0/en/package_reference/accelerator#accelerate.Accelerator"
+        ),
+    )
+    parser.add_argument(
+        "--conditioning_dropout_prob",
+        type=float,
+        default=None,
+        help="Conditioning dropout probability. Drops out the conditioning image used during training",
+    )
+
+    args = parser.parse_args()
+    env_local_rank = int(os.environ.get("LOCAL_RANK", -1))
+    if env_local_rank != -1 and env_local_rank != args.local_rank:
+        args.local_rank = env_local_rank
+
+    # Sanity checks
+    if args.dataset_loading_script_path is None:
+        raise ValueError("Need a script to load training dataset.")
+
+    # default to using the same revision for the non-ema model if not specified
+    if args.non_ema_revision is None:
+        args.non_ema_revision = args.revision
+
+    return args
+
+
+def make_train_dataset(args, feature_extractor, accelerator):
+    # In distributed training, the load_dataset function guarantees that only one local process can concurrently
+    # download the dataset.
+    dataset = load_dataset(path=args.dataset_loading_script_path, split="train", trust_remote_code=True)
+
+    # Preprocessing the datasets.
+    # We need to tokenize inputs and targets.
+    column_names = dataset.column_names
+
+    # 6. Get the column names for input/target.
+    if args.source_image_column is None:
+        source_image_column = column_names[0]
+        logger.info(f"image column defaulting to {source_image_column}")
+    else:
+        source_image_column = args.source_image_column
+        if source_image_column not in column_names:
+            raise ValueError(
+                f"`--source_image_column` value '{args.source_image_column}' not found in dataset columns. Dataset columns are: {', '.join(column_names)}"
+            )
+
+    if args.conditioning_image_column is None:
+        conditioning_image_column = column_names[1]
+        logger.info(f"conditioning image column defaulting to {conditioning_image_column}")
+    else:
+        conditioning_image_column = args.conditioning_image_column
+        if conditioning_image_column not in column_names:
+            raise ValueError(
+                f"`--conditioning_image_column` value '{args.conditioning_image_column}' not found in dataset columns. Dataset columns are: {', '.join(column_names)}"
+            )
+
+    if args.ground_truth_column is None:
+        ground_truth_column = column_names[2]
+        logger.info(f"ground truth column defaulting to {ground_truth_column}")
+    else:
+        ground_truth_column = args.ground_truth_column
+        if ground_truth_column not in column_names:
+            raise ValueError(
+                f"`--ground_truth_column` value '{args.ground_truth_column}' not found in dataset columns. Dataset columns are: {', '.join(column_names)}"
+            )
+
+    def extract_features(images):
+        features = []
+        for image in images:
+            feature = feature_extractor(images=image, do_rescale=False, return_tensors="pt").pixel_values[0]
+            features.append(feature)
+        return features
+
+    # The pipeline expects two images as inputs named image and image0 and will output numpy arrays.
+    albumentations_transform = A.Compose(
+        [
+            A.HorizontalFlip(p=0.5 if args.random_flip else 0.0),
+        ],
+        additional_targets={"image0": "image", "image1": "image"},
+    )
+
+    torchvision_transforms = transforms_v2.Compose(
+        [
+            transforms_v2.ToImage(),
+            transforms_v2.Resize(args.resolution, interpolation=transforms.InterpolationMode.BILINEAR, antialias=True),
+            transforms_v2.CenterCrop(args.resolution),
+            transforms_v2.ToDtype(torch.float32, scale=True),
+        ]
+    )
+
+    # Create a Normalize transform
+    normalize_transforms = transforms_v2.Compose([transforms_v2.Normalize(mean=[0.5], std=[0.5])])
+
+    def preprocess_train(examples):
+        source_images = [image.convert("RGB") for image in examples[source_image_column]]
+        conditioning_images = [image.convert("RGB") for image in examples[conditioning_image_column]]
+        ground_truth = [image.convert("RGB") for image in examples[ground_truth_column]]
+
+        for i, (src, cond, gt) in enumerate(zip(source_images, conditioning_images, ground_truth)):
+            # Convert PIL image to numpy array
+            src = np.array(src)
+            cond = np.array(cond)
+            gt = np.array(gt)
+
+            # Apply the same augmentation with the same parameters to multiple images
+            augmented = albumentations_transform(image=src, image0=cond, image1=gt)
+
+            # Convert numpy array to PIL image
+            source_images[i] = Image.fromarray(augmented["image"])
+            conditioning_images[i] = Image.fromarray(augmented["image0"])
+            ground_truth[i] = Image.fromarray(augmented["image1"])
+
+        source_images = [torchvision_transforms(image) for image in source_images]
+        conditioning_images = [torchvision_transforms(image) for image in conditioning_images]
+        ground_truth = [torchvision_transforms(image) for image in ground_truth]
+
+        examples["source_images"] = [normalize_transforms(image) for image in source_images]
+        examples["conditioning_images"] = [normalize_transforms(image) for image in conditioning_images]
+        examples["ground_truth"] = [normalize_transforms(image) for image in ground_truth]
+
+        examples["clip_source_images"] = extract_features(source_images)
+        examples["clip_conditioning_images"] = extract_features(conditioning_images)
+        examples["clip_ground_truth"] = extract_features(ground_truth)
+
+        return examples
+
+    with accelerator.main_process_first():
+        if args.max_train_samples is not None:
+            dataset = dataset.shuffle(seed=args.seed).select(range(args.max_train_samples))
+        # Set the training transforms
+        train_dataset = dataset.with_transform(preprocess_train)
+
+    return train_dataset
+
+
+def collate_fn(examples):
+    ground_truth = [example["ground_truth"] for example in examples]
+    source_images = [example["source_images"] for example in examples]
+    conditioning_images = [example["conditioning_images"] for example in examples]
+    clip_ground_truth = [example["clip_ground_truth"] for example in examples]
+    clip_source_images = [example["clip_source_images"] for example in examples]
+    clip_conditioning_images = [example["clip_conditioning_images"] for example in examples]
+
+    ground_truth = torch.stack(ground_truth)
+    ground_truth = ground_truth.to(memory_format=torch.contiguous_format).float()
+
+    source_images = torch.stack(source_images)
+    source_images = source_images.to(memory_format=torch.contiguous_format).float()
+
+    conditioning_images = torch.stack(conditioning_images)
+    conditioning_images = conditioning_images.to(memory_format=torch.contiguous_format).float()
+
+    clip_ground_truth = torch.stack(clip_ground_truth)
+    clip_ground_truth = clip_ground_truth.to(memory_format=torch.contiguous_format).float()
+
+    clip_source_images = torch.stack(clip_source_images)
+    clip_source_images = clip_source_images.to(memory_format=torch.contiguous_format).float()
+
+    clip_conditioning_images = torch.stack(clip_conditioning_images)
+    clip_conditioning_images = clip_conditioning_images.to(memory_format=torch.contiguous_format).float()
+
+    return {
+        "ground_truth": ground_truth,
+        "source_images": source_images,
+        "conditioning_images": conditioning_images,
+        "clip_ground_truth": clip_ground_truth,
+        "clip_source_images": clip_source_images,
+        "clip_conditioning_images": clip_conditioning_images,
+    }
+
+
+def main():
+    args = parse_args()
+
+    if args.non_ema_revision is not None:
+        deprecate(
+            "non_ema_revision!=None",
+            "0.15.0",
+            message=(
+                "Downloading 'non_ema' weights from revision branches of the Hub is deprecated. Please make sure to"
+                " use `--variant=non_ema` instead."
+            ),
+        )
+    logging_dir = os.path.join(args.output_dir, args.logging_dir)
+
+    accelerator_project_config = ProjectConfiguration(project_dir=args.output_dir, logging_dir=logging_dir)
+    # ddp_kwargs = DistributedDataParallelKwargs(find_unused_parameters=True)
+    accelerator = Accelerator(
+        gradient_accumulation_steps=args.gradient_accumulation_steps,
+        mixed_precision=args.mixed_precision,
+        log_with=args.report_to,
+        project_config=accelerator_project_config,
+        # kwargs_handlers=[ddp_kwargs]
+    )
+
+    generator = None if args.seed is None else torch.Generator(device=accelerator.device).manual_seed(args.seed)
+
+    # Make one log on every process with the configuration for debugging.
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
+        datefmt="%m/%d/%Y %H:%M:%S",
+        level=logging.INFO,
+    )
+    logger.info(accelerator.state, main_process_only=False)
+    if accelerator.is_local_main_process:
+        datasets.utils.logging.set_verbosity_warning()
+        transformers.utils.logging.set_verbosity_warning()
+        diffusers.utils.logging.set_verbosity_info()
+    else:
+        datasets.utils.logging.set_verbosity_error()
+        transformers.utils.logging.set_verbosity_error()
+        diffusers.utils.logging.set_verbosity_error()
+
+    # If passed along, set the training seed now.
+    if args.seed is not None:
+        set_seed(args.seed)
+
+    # Handle the repository creation
+    if accelerator.is_main_process:
+        if args.output_dir is not None:
+            os.makedirs(args.output_dir, exist_ok=True)
+
+        if args.push_to_hub:
+            repo_id = create_repo(
+                repo_id=args.hub_model_id or Path(args.output_dir).name, exist_ok=True, token=args.hub_token
+            ).repo_id
+
+    # Load scheduler and models.
+    noise_scheduler = DDPMScheduler.from_pretrained(args.pretrained_model_name_or_path, subfolder="scheduler")
+    feature_extractor = CLIPImageProcessor.from_pretrained(args.pretrained_clip_model_name_or_path)
+
+    def deepspeed_zero_init_disabled_context_manager():
+        """
+        returns either a context list that includes one that will disable zero.Init or an empty context list
+        """
+        deepspeed_plugin = AcceleratorState().deepspeed_plugin if accelerate.state.is_initialized() else None
+        if deepspeed_plugin is None:
+            return []
+
+        return [deepspeed_plugin.zero3_init_context_manager(enable=False)]
+
+    # Currently Accelerate doesn't know how to handle multiple models under Deepspeed ZeRO stage 3.
+    # For this to work properly all models must be run through `accelerate.prepare`. But accelerate
+    # will try to assign the same optimizer with the same weights to all models during
+    # `deepspeed.initialize`, which of course doesn't work.
+    #
+    # For now the following workaround will partially support Deepspeed ZeRO-3, by excluding the 2
+    # frozen models from being partitioned during `zero.Init` which gets called during
+    # `from_pretrained` So CLIPTextModel and AutoencoderKL will not enjoy the parameter sharding
+    # across multiple gpus and only UNet2DConditionModel will get ZeRO sharded.
+    with ContextManagers(deepspeed_zero_init_disabled_context_manager()):
+        image_encoder = CLIPVisionModel.from_pretrained(args.pretrained_clip_model_name_or_path)
+        vae = AutoencoderKL.from_pretrained(
+            args.pretrained_model_name_or_path, subfolder="vae", revision=args.revision, variant=args.variant
+        )
+
+    unet = UNet2DConditionModel.from_pretrained(
+        args.pretrained_model_name_or_path, subfolder="unet", revision=args.non_ema_revision
+    )
+    referencenet = ReferenceNetModel.from_pretrained(
+        args.pretrained_model_name_or_path, subfolder="unet", revision=args.non_ema_revision
+    )
+    conditioning_referencenet = ReferenceNetModel.from_pretrained(
+        args.pretrained_model_name_or_path, subfolder="unet", revision=args.non_ema_revision
+    )
+
+    # Freeze vae and image_encoder
+    vae.requires_grad_(False)
+    image_encoder.requires_grad_(False)
+    unet.train()
+    # unet.requires_grad_(False)
+    # for p in unet.down_blocks[0].attentions[0].transformer_blocks[0].parameters():
+    #     print(p.requires_grad)  # Expect to be False
+    #     break
+    # unet = set_parts_of_model_for_gradient_computation(unet)
+    # Check if gradient will be calculated on the tensor
+    # for p in unet.down_blocks[0].attentions[0].transformer_blocks[0].parameters():
+    #     print(p.requires_grad)  # Expect to be True
+    #     break
+
+    # referencenet.train()
+    referencenet.requires_grad_(False)
+    referencenet = set_parts_of_model_for_gradient_computation(referencenet)
+
+    # conditioning_referencenet.train()
+    conditioning_referencenet.requires_grad_(False)
+    conditioning_referencenet = set_parts_of_model_for_gradient_computation(conditioning_referencenet)
+
+    # Create EMA for the unet.
+    if args.use_ema:
+        ema_unet = UNet2DConditionModel.from_pretrained(
+            args.pretrained_model_name_or_path, subfolder="unet", revision=args.revision, variant=args.variant
+        )
+        ema_unet = EMAModel(ema_unet.parameters(), model_cls=UNet2DConditionModel, model_config=ema_unet.config)
+
+    if args.enable_xformers_memory_efficient_attention:
+        if is_xformers_available():
+            import xformers
+
+            xformers_version = version.parse(xformers.__version__)
+            if xformers_version == version.parse("0.0.16"):
+                logger.warn(
+                    "xFormers 0.0.16 cannot be used for training in some GPUs. If you observe problems during training, please update xFormers to at least 0.0.17. See https://huggingface.co/docs/diffusers/main/en/optimization/xformers for more details."
+                )
+            unet.enable_xformers_memory_efficient_attention()
+        else:
+            raise ValueError("xformers is not available. Make sure it is installed correctly")
+
+    # `accelerate` 0.16.0 will have better support for customized saving
+    if False:
+        # create custom saving & loading hooks so that `accelerator.save_state(...)` serializes in a nice format
+        def save_model_hook(models, weights, output_dir):
+            if accelerator.is_main_process:
+                if args.use_ema:
+                    ema_unet.save_pretrained(os.path.join(output_dir, "unet_ema"))
+
+                for model in models:
+                    sub_dir = "unet" if isinstance(model, type(accelerator.unwrap_model(unet))) else "referencenet"
+                    model.save_pretrained(os.path.join(output_dir, sub_dir))
+
+                    # make sure to pop weight so that corresponding model is not saved again
+                    weights.pop()
+
+        def load_model_hook(models, input_dir):
+            if args.use_ema:
+                load_model = EMAModel.from_pretrained(os.path.join(input_dir, "unet_ema"), UNet2DConditionModel)
+                ema_unet.load_state_dict(load_model.state_dict())
+                ema_unet.to(accelerator.device)
+                del load_model
+
+            while len(models) > 0:
+                # pop models so that they are not loaded again
+                model = models.pop()
+
+                if isinstance(model, type(accelerator.unwrap_model(referencenet))):
+                    # load transformers style into model
+                    load_model = ReferenceNetModel.from_pretrained(input_dir, subfolder="referencenet")
+                else:
+                    # load diffusers style into model
+                    load_model = UNet2DConditionModel.from_pretrained(input_dir, subfolder="unet")
+                model.register_to_config(**load_model.config)
+
+                model.load_state_dict(load_model.state_dict())
+                del load_model
+
+        accelerator.register_save_state_pre_hook(save_model_hook)
+        accelerator.register_load_state_pre_hook(load_model_hook)
+
+    if args.gradient_checkpointing:
+        unet.enable_gradient_checkpointing()
+
+    # Enable TF32 for faster training on Ampere GPUs,
+    # cf https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices
+    if args.allow_tf32:
+        torch.backends.cuda.matmul.allow_tf32 = True
+
+    if args.scale_lr:
+        args.learning_rate = (
+            args.learning_rate * args.gradient_accumulation_steps * args.train_batch_size * accelerator.num_processes
+        )
+
+    # Initialize the optimizer
+    if args.use_8bit_adam:
+        try:
+            import bitsandbytes as bnb
+        except ImportError:
+            raise ImportError(
+                "Please install bitsandbytes to use 8-bit Adam. You can do so by running `pip install bitsandbytes`"
+            )
+
+        optimizer_cls = bnb.optim.AdamW8bit
+    else:
+        optimizer_cls = torch.optim.AdamW
+
+    trainable_params = list(filter(lambda p: p.requires_grad, unet.parameters()))
+    trainable_params += list(filter(lambda p: p.requires_grad, referencenet.parameters()))
+    trainable_params += list(filter(lambda p: p.requires_grad, conditioning_referencenet.parameters()))
+
+    optimizer = optimizer_cls(
+        trainable_params,
+        lr=args.learning_rate,
+        betas=(args.adam_beta1, args.adam_beta2),
+        weight_decay=args.adam_weight_decay,
+        eps=args.adam_epsilon,
+    )
+
+    # Dataset and DataLoaders creation:
+    train_dataset = make_train_dataset(args, feature_extractor, accelerator)
+
+    train_dataloader = torch.utils.data.DataLoader(
+        train_dataset,
+        shuffle=True,
+        collate_fn=collate_fn,
+        batch_size=args.train_batch_size,
+        num_workers=args.dataloader_num_workers,
+    )
+
+    # Scheduler and math around the number of training steps.
+    overrode_max_train_steps = False
+    num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
+    if args.max_train_steps is None:
+        args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
+        overrode_max_train_steps = True
+
+    lr_scheduler = get_scheduler(
+        args.lr_scheduler,
+        optimizer=optimizer,
+        num_warmup_steps=args.lr_warmup_steps * accelerator.num_processes,
+        num_training_steps=args.max_train_steps * accelerator.num_processes,
+    )
+
+    # Prepare everything with our `accelerator`.
+    unet, referencenet, conditioning_referencenet, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
+        unet, referencenet, conditioning_referencenet, optimizer, train_dataloader, lr_scheduler
+    )
+
+    if args.use_ema:
+        ema_unet.to(accelerator.device)
+
+    # For mixed precision training we cast all non-trainable weigths (vae, non-lora image_encoder and non-lora unet) to half-precision
+    # as these weights are only used for inference, keeping weights in full precision is not required.
+    weight_dtype = torch.float32
+    if accelerator.mixed_precision == "fp16":
+        weight_dtype = torch.float16
+        args.mixed_precision = accelerator.mixed_precision
+    elif accelerator.mixed_precision == "bf16":
+        weight_dtype = torch.bfloat16
+        args.mixed_precision = accelerator.mixed_precision
+
+    # Move image_encoder and vae to gpu and cast to weight_dtype
+    image_encoder.to(accelerator.device, dtype=weight_dtype)
+    vae.to(accelerator.device, dtype=weight_dtype)
+
+    # We need to recalculate our total training steps as the size of the training dataloader may have changed.
+    num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
+    if overrode_max_train_steps:
+        args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
+    # Afterwards we recalculate our number of training epochs
+    args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)
+
+    # We need to initialize the trackers we use, and also store our configuration.
+    # The trackers initializes automatically on the main process.
+    if accelerator.is_main_process:
+        tracker_config = dict(vars(args))
+        tracker_config.pop("validation_conditioning_image")
+        tracker_config.pop("validation_source_image")
+        accelerator.init_trackers(args.tracker_project_name, tracker_config)
+
+    # Train!
+    total_batch_size = args.train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps
+
+    logger.info("***** Running training *****")
+    logger.info(f"  Num examples = {len(train_dataset)}")
+    logger.info(f"  Num Epochs = {args.num_train_epochs}")
+    logger.info(f"  Instantaneous batch size per device = {args.train_batch_size}")
+    logger.info(f"  Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}")
+    logger.info(f"  Gradient Accumulation steps = {args.gradient_accumulation_steps}")
+    logger.info(f"  Total optimization steps = {args.max_train_steps}")
+    global_step = 0
+    first_epoch = 0
+
+    # Potentially load in the weights and states from a previous save
+    if args.resume_from_checkpoint:
+        if args.resume_from_checkpoint != "latest":
+            path = os.path.basename(args.resume_from_checkpoint)
+        else:
+            # Get the most recent checkpoint
+            dirs = os.listdir(args.output_dir)
+            dirs = [d for d in dirs if d.startswith("checkpoint")]
+            dirs = sorted(dirs, key=lambda x: int(x.split("-")[1]))
+            path = dirs[-1] if len(dirs) > 0 else None
+
+        if path is None:
+            accelerator.print(
+                f"Checkpoint '{args.resume_from_checkpoint}' does not exist. Starting a new training run."
+            )
+            args.resume_from_checkpoint = None
+            initial_global_step = 0
+        else:
+            accelerator.print(f"Resuming from checkpoint {path}")
+            accelerator.load_state(os.path.join(args.output_dir, path))
+            global_step = int(path.split("-")[1])
+
+            initial_global_step = global_step
+            first_epoch = global_step // num_update_steps_per_epoch
+
+    else:
+        initial_global_step = 0
+
+    progress_bar = tqdm(
+        range(0, args.max_train_steps),
+        initial=initial_global_step,
+        desc="Steps",
+        # Only show the progress bar once on each machine.
+        disable=not accelerator.is_local_main_process,
+    )
+
+    last_step = noise_scheduler.config.num_train_timesteps - 1
+    for epoch in range(first_epoch, args.num_train_epochs):
+        train_loss = 0.0
+        for step, batch in enumerate(train_dataloader):
+            with accelerator.accumulate(unet, referencenet, conditioning_referencenet):
+                # Ground truth
+                ground_truth = batch["ground_truth"].to(weight_dtype)
+                clip_ground_truth = batch["clip_ground_truth"].to(weight_dtype)
+
+                # Source images
+                source_images = batch["source_images"].to(weight_dtype)
+                clip_source_images = batch["clip_source_images"].to(weight_dtype)
+
+                # Driving images
+                conditioning_images = batch["conditioning_images"].to(weight_dtype)
+                clip_conditioning_images = batch["clip_conditioning_images"].to(weight_dtype)
+
+                # Convert images to latent space
+                latents = vae.encode(ground_truth).latent_dist.sample()
+                latents *= vae.config.scaling_factor
+                source_latents = vae.encode(source_images).latent_dist.sample()
+                source_latents *= vae.config.scaling_factor
+                conditioning_latents = vae.encode(conditioning_images).latent_dist.sample()
+                conditioning_latents *= vae.config.scaling_factor
+
+                # Sample noise that we'll add to the latents
+                noise = torch.randn_like(latents)
+                if args.noise_offset:
+                    # https://www.crosslabs.org//blog/diffusion-with-offset-noise
+                    noise += args.noise_offset * torch.randn(
+                        (latents.shape[0], latents.shape[1], 1, 1), device=latents.device
+                    )
+                if args.input_perturbation:
+                    new_noise = noise + args.input_perturbation * torch.randn_like(noise)
+                bsz = latents.shape[0]
+                # Sample a random timestep for each image
+                timesteps = torch.randint(0, noise_scheduler.config.num_train_timesteps, (bsz,), device=latents.device)
+                timesteps = timesteps.long()
+                ref_timesteps = torch.zeros_like(timesteps)
+
+                # Add noise to the latents according to the noise magnitude at each timestep
+                # (this is the forward diffusion process)
+                if args.input_perturbation:
+                    noisy_latents = noise_scheduler.add_noise(latents, new_noise, timesteps)
+                else:
+                    noisy_latents = noise_scheduler.add_noise(latents, noise, timesteps)
+
+                # Fix zero PSNR
+                noisy_latents[timesteps == last_step] = noise[timesteps == last_step]
+
+                # Get the source image embedding
+                source_image_embeds = image_encoder(clip_source_images).pooler_output.unsqueeze(1)
+
+                # Get the conditioning image embedding
+                conditioning_image_embeds = image_encoder(clip_conditioning_images).pooler_output.unsqueeze(1)
+
+                # Conditioning dropout to support classifier-free guidance during inference.
+                random_p = torch.rand(bsz, device=accelerator.device, generator=generator)
+                if args.conditioning_dropout_prob is not None:
+                    # Sample masks for the source images.
+                    image_mask = 1 - (random_p < args.conditioning_dropout_prob).to(source_images.dtype)
+                    # Final image conditioning.
+                    image_mask = image_mask.reshape(bsz, 1, 1, 1)
+                    source_latents = image_mask * source_latents
+
+                    image_mask = image_mask.reshape(bsz, 1, 1)
+                    source_image_embeds = image_mask * source_image_embeds
+
+                # Get the target for loss depending on the prediction type
+                if args.prediction_type is not None:
+                    # set prediction_type of scheduler if defined
+                    noise_scheduler.register_to_config(prediction_type=args.prediction_type)
+
+                if noise_scheduler.config.prediction_type == "epsilon":
+                    target = noise
+                elif noise_scheduler.config.prediction_type == "v_prediction":
+                    target = noise_scheduler.get_velocity(latents, noise, timesteps)
+                else:
+                    raise ValueError(f"Unknown prediction type {noise_scheduler.config.prediction_type}")
+
+                # Referencenet pass
+                referencenet_sample, referencenet_states = referencenet(
+                    sample=source_latents,
+                    timestep=ref_timesteps,
+                    encoder_hidden_states=source_image_embeds,
+                    return_dict=False,
+                )
+
+                if False:
+                    # Sample masks for the referencenet states.
+                    referencenet_states_mask = 1 - (random_p < args.conditioning_dropout_prob).to(referencenet.dtype)
+                    # Final referencenet states conditioning.
+                    referencenet_states_mask = referencenet_states_mask.reshape(bsz, 1, 1)
+                    recursive_multiply(referencenet_states, referencenet_states_mask)
+
+                conditioning_referencenet_sample, conditioning_referencenet_states = conditioning_referencenet(
+                    sample=conditioning_latents,
+                    timestep=ref_timesteps,
+                    encoder_hidden_states=conditioning_image_embeds,
+                    return_dict=False,
+                )
+
+                concatenated_embeds = torch.cat([source_image_embeds, conditioning_image_embeds], dim=1)
+                concatenated_referencenet_states = cat_referencenet_states(
+                    referencenet_states,
+                    conditioning_referencenet_states,
+                    dim=1,
+                )
+
+                # Predict the noise residual and compute loss
+                model_pred = unet(
+                    sample=noisy_latents,
+                    timestep=timesteps,
+                    encoder_hidden_states=concatenated_embeds,
+                    referencenet_states=concatenated_referencenet_states,
+                ).sample
+                # Fix error by adding with 0 weight
+                model_pred += 0 * (referencenet_sample + conditioning_referencenet_sample)
+
+                if args.snr_gamma is None:
+                    loss = F.mse_loss(model_pred.float(), target.float(), reduction="mean")
+                else:
+                    # Compute loss-weights as per Section 3.4 of https://arxiv.org/abs/2303.09556.
+                    # Since we predict the noise instead of x_0, the original formulation is slightly changed.
+                    # This is discussed in Section 4.2 of the same paper.
+                    snr = compute_snr(noise_scheduler, timesteps)
+                    if noise_scheduler.config.prediction_type == "v_prediction":
+                        # Velocity objective requires that we add one to SNR values before we divide by them.
+                        snr = snr + 1
+                    mse_loss_weights = (
+                        torch.stack([snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1).min(dim=1)[0] / snr
+                    )
+
+                    loss = F.mse_loss(model_pred.float(), target.float(), reduction="none")
+                    loss = loss.mean(dim=list(range(1, len(loss.shape)))) * mse_loss_weights
+                    loss = loss.mean()
+
+                # Gather the losses across all processes for logging (if we use distributed training).
+                avg_loss = accelerator.gather(loss.repeat(args.train_batch_size)).mean()
+                train_loss += avg_loss.item() / args.gradient_accumulation_steps
+
+                # Backpropagate
+                accelerator.backward(loss)
+                if accelerator.sync_gradients:
+                    accelerator.clip_grad_norm_(unet.parameters(), args.max_grad_norm)
+                optimizer.step()
+                lr_scheduler.step()
+                optimizer.zero_grad(set_to_none=args.set_grads_to_none)
+
+            # Checks if the accelerator has performed an optimization step behind the scenes
+            if accelerator.sync_gradients:
+                if args.use_ema:
+                    ema_unet.step(unet.parameters())
+                progress_bar.update(1)
+                global_step += 1
+                accelerator.log({"train_loss": train_loss}, step=global_step)
+                train_loss = 0.0
+
+                if accelerator.is_main_process:
+                    if global_step % args.checkpointing_steps == 0:
+                        # _before_ saving state, check if this save would set us over the `checkpoints_total_limit`
+                        if args.checkpoints_total_limit is not None:
+                            checkpoints = os.listdir(args.output_dir)
+                            checkpoints = [d for d in checkpoints if d.startswith("checkpoint")]
+                            checkpoints = sorted(checkpoints, key=lambda x: int(x.split("-")[1]))
+
+                            # before we save the new checkpoint, we need to have at _most_ `checkpoints_total_limit - 1` checkpoints
+                            if len(checkpoints) >= args.checkpoints_total_limit:
+                                num_to_remove = len(checkpoints) - args.checkpoints_total_limit + 1
+                                removing_checkpoints = checkpoints[0:num_to_remove]
+
+                                logger.info(
+                                    f"{len(checkpoints)} checkpoints already exist, removing {len(removing_checkpoints)} checkpoints"
+                                )
+                                logger.info(f"removing checkpoints: {', '.join(removing_checkpoints)}")
+
+                                for removing_checkpoint in removing_checkpoints:
+                                    removing_checkpoint = os.path.join(args.output_dir, removing_checkpoint)
+                                    shutil.rmtree(removing_checkpoint)
+
+                        save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}")
+                        accelerator.save_state(save_path)
+                        logger.info(f"Saved state to {save_path}")
+
+                    if args.validation_conditioning_image is not None and global_step % args.validation_steps == 0:
+                        if args.use_ema:
+                            # Store the UNet parameters temporarily and load the EMA parameters to perform inference.
+                            ema_unet.store(unet.parameters())
+                            ema_unet.copy_to(unet.parameters())
+                        log_validation(
+                            vae=vae,
+                            unet=unet,
+                            feature_extractor=feature_extractor,
+                            image_encoder=image_encoder,
+                            referencenet=referencenet,
+                            conditioning_referencenet=conditioning_referencenet,
+                            args=args,
+                            accelerator=accelerator,
+                            step=global_step,
+                        )
+                        if args.use_ema:
+                            # Switch back to the original UNet parameters.
+                            ema_unet.restore(unet.parameters())
+
+            logs = {"step_loss": loss.detach().item(), "lr": lr_scheduler.get_last_lr()[0]}
+            progress_bar.set_postfix(**logs)
+
+            if global_step >= args.max_train_steps:
+                break
+
+    # Create the pipeline using the trained modules and save it.
+    accelerator.wait_for_everyone()
+    if accelerator.is_main_process:
+        unet = accelerator.unwrap_model(unet)
+        unet.save_pretrained(Path(args.output_dir, "unet"))
+        referencenet = accelerator.unwrap_model(referencenet)
+        referencenet.save_pretrained(Path(args.output_dir, "referencenet"))
+        conditioning_referencenet = accelerator.unwrap_model(conditioning_referencenet)
+        conditioning_referencenet.save_pretrained(Path(args.output_dir, "conditioning_referencenet"))
+
+        # Run a final round of inference.
+        log_validation(
+            vae=vae,
+            unet=unet,
+            feature_extractor=feature_extractor,
+            image_encoder=image_encoder,
+            referencenet=referencenet,
+            conditioning_referencenet=conditioning_referencenet,
+            args=args,
+            accelerator=accelerator,
+            step=global_step,
+        )
+
+    accelerator.end_training()
+
+
+if __name__ == "__main__":
+    main()

my_dataset/train/train.jsonl

@@ -0,0 +1,2 @@
+{"source_image": "celeb/real/18147_06771.png", "conditioning_image": "celeb/real/01758_01758.png", "ground_truth": "celeb/fake/18147_06771-01758_01758.png"}
+{"source_image": "celeb/real/01758_09704.png", "conditioning_image": "celeb/fake/18147_06771-01758_01758.png", "ground_truth": "celeb/real/01758_01758.png"}

my_dataset/train_dataset_loading_script.py

@@ -0,0 +1,143 @@
+# Copyright 2020 The HuggingFace Datasets Authors and the current dataset script contributor.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# TODO: Address all TODOs and remove all explanatory comments
+"""TODO: Add a description here."""
+
+
+import os
+
+import datasets
+import pandas as pd
+
+
+# TODO: Add BibTeX citation
+# Find for instance the citation on arxiv or on the dataset repo/website
+_CITATION = """\
+@InProceedings{huggingface:dataset,
+title = {A great new dataset},
+author={huggingface, Inc.
+},
+year={2020}
+}
+"""
+
+# TODO: Add description of the dataset here
+# You can copy an official description
+_DESCRIPTION = """\
+This new dataset is designed to solve this great NLP task and is crafted with a lot of care.
+"""
+
+# TODO: Add a link to an official homepage for the dataset here
+_HOMEPAGE = ""
+
+# TODO: Add the licence for the dataset here if you can find it
+_LICENSE = ""
+
+# TODO: Add link to the official dataset URLs here
+_URLS = {
+    "metadata_path": "/path/to/face_anon_simple/my_dataset/train/train.jsonl",
+    "images_dir": "/path/to/face_anon_simple/my_dataset/train/",
+}
+
+
+# TODO: Name of the dataset usually matches the script name with CamelCase instead of snake_case
+class NewDataset(datasets.GeneratorBasedBuilder):
+    """TODO: Short description of my dataset."""
+
+    VERSION = datasets.Version("1.1.0")
+
+    def _info(self):
+        # TODO: This method specifies the datasets.DatasetInfo object which contains informations and typings for the dataset
+        features = datasets.Features(
+            {
+                "source_image": datasets.Image(),
+                "conditioning_image": datasets.Image(),
+                "ground_truth": datasets.Image(),
+                "source_image_path": datasets.Value("string"),
+                "conditioning_image_path": datasets.Value("string"),
+                "ground_truth_path": datasets.Value("string"),
+            }
+        )
+
+        return datasets.DatasetInfo(
+            # This is the description that will appear on the datasets page.
+            description=_DESCRIPTION,
+            # This defines the different columns of the dataset and their types
+            features=features,  # Here we define them above because they are different between the two configurations
+            # If there's a common (input, target) tuple from the features, uncomment supervised_keys line below and
+            # specify them. They'll be used if as_supervised=True in builder.as_dataset.
+            # supervised_keys=("sentence", "label"),
+            # Homepage of the dataset for documentation
+            homepage=_HOMEPAGE,
+            # License for the dataset if available
+            license=_LICENSE,
+            # Citation for the dataset
+            citation=_CITATION,
+        )
+
+    def _split_generators(self, dl_manager):
+        # TODO: This method is tasked with downloading/extracting the data and defining the splits depending on the configuration
+        # If several configurations are possible (listed in BUILDER_CONFIGS), the configuration selected by the user is in self.config.name
+        metadata_path = _URLS["metadata_path"]
+        images_dir = _URLS["images_dir"]
+        return [
+            datasets.SplitGenerator(
+                name=datasets.Split.TRAIN,
+                # These kwargs will be passed to _generate_examples
+                gen_kwargs={
+                    "metadata_path": metadata_path,
+                    "images_dir": images_dir,
+                },
+            ),
+        ]
+
+    # method parameters are unpacked from `gen_kwargs` as given in `_split_generators`
+    def _generate_examples(self, metadata_path, images_dir):
+        # TODO: This method handles input defined in _split_generators to yield (key, example) tuples from the dataset.
+        # The `key` is for legacy reasons (tfds) and is not important in itself, but must be unique for each example.
+        metadata = pd.read_json(metadata_path, lines=True)
+
+        for _, row in metadata.iterrows():
+            source_image_path = row["source_image"]
+            source_image_path = os.path.join(images_dir, source_image_path)
+            source_image = open(source_image_path, "rb").read()
+
+            conditioning_image_path = row["conditioning_image"]
+            conditioning_image_path = os.path.join(images_dir, conditioning_image_path)
+            conditioning_image = open(conditioning_image_path, "rb").read()
+
+            ground_truth_path = row["ground_truth"]
+            ground_truth_path = os.path.join(images_dir, ground_truth_path)
+            ground_truth = open(ground_truth_path, "rb").read()
+
+            yield (
+                "-".join([source_image_path, conditioning_image_path]),
+                {
+                    "source_image": {
+                        "path": source_image_path,
+                        "bytes": source_image,
+                    },
+                    "conditioning_image": {
+                        "path": conditioning_image_path,
+                        "bytes": conditioning_image,
+                    },
+                    "ground_truth": {
+                        "path": ground_truth_path,
+                        "bytes": ground_truth,
+                    },
+                    "source_image_path": source_image_path,
+                    "conditioning_image_path": conditioning_image_path,
+                    "ground_truth_path": ground_truth_path,
+                },
+            )

src/diffusers/__init__.py

@@ -0,0 +1,758 @@
+__version__ = "0.25.1"
+
+from typing import TYPE_CHECKING
+
+from .utils import (
+    DIFFUSERS_SLOW_IMPORT,
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    is_flax_available,
+    is_k_diffusion_available,
+    is_librosa_available,
+    is_note_seq_available,
+    is_onnx_available,
+    is_scipy_available,
+    is_torch_available,
+    is_torchsde_available,
+    is_transformers_available,
+)
+
+
+# Lazy Import based on
+# https://github.com/huggingface/transformers/blob/main/src/transformers/__init__.py
+
+# When adding a new object to this init, please add it to `_import_structure`. The `_import_structure` is a dictionary submodule to list of object names,
+# and is used to defer the actual importing for when the objects are requested.
+# This way `import diffusers` provides the names in the namespace without actually importing anything (and especially none of the backends).
+
+_import_structure = {
+    "configuration_utils": ["ConfigMixin"],
+    "models": [],
+    "pipelines": [],
+    "schedulers": [],
+    "utils": [
+        "OptionalDependencyNotAvailable",
+        "is_flax_available",
+        "is_inflect_available",
+        "is_invisible_watermark_available",
+        "is_k_diffusion_available",
+        "is_k_diffusion_version",
+        "is_librosa_available",
+        "is_note_seq_available",
+        "is_onnx_available",
+        "is_scipy_available",
+        "is_torch_available",
+        "is_torchsde_available",
+        "is_transformers_available",
+        "is_transformers_version",
+        "is_unidecode_available",
+        "logging",
+    ],
+}
+
+try:
+    if not is_onnx_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    from .utils import dummy_onnx_objects  # noqa F403
+
+    _import_structure["utils.dummy_onnx_objects"] = [
+        name for name in dir(dummy_onnx_objects) if not name.startswith("_")
+    ]
+
+else:
+    _import_structure["pipelines"].extend(["OnnxRuntimeModel"])
+
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    from .utils import dummy_pt_objects  # noqa F403
+
+    _import_structure["utils.dummy_pt_objects"] = [name for name in dir(dummy_pt_objects) if not name.startswith("_")]
+
+else:
+    _import_structure["models"].extend(
+        [
+            "AsymmetricAutoencoderKL",
+            "AutoencoderKL",
+            "AutoencoderKLTemporalDecoder",
+            "AutoencoderTiny",
+            "ConsistencyDecoderVAE",
+            "ControlNetModel",
+            "Kandinsky3UNet",
+            "ModelMixin",
+            "MotionAdapter",
+            "MultiAdapter",
+            "PriorTransformer",
+            "T2IAdapter",
+            "T5FilmDecoder",
+            "Transformer2DModel",
+            "UNet1DModel",
+            "UNet2DConditionModel",
+            "UNet2DModel",
+            "UNet3DConditionModel",
+            "UNetMotionModel",
+            "UNetSpatioTemporalConditionModel",
+            "UVit2DModel",
+            "VQModel",
+        ]
+    )
+
+    _import_structure["optimization"] = [
+        "get_constant_schedule",
+        "get_constant_schedule_with_warmup",
+        "get_cosine_schedule_with_warmup",
+        "get_cosine_with_hard_restarts_schedule_with_warmup",
+        "get_linear_schedule_with_warmup",
+        "get_polynomial_decay_schedule_with_warmup",
+        "get_scheduler",
+    ]
+    _import_structure["pipelines"].extend(
+        [
+            "AudioPipelineOutput",
+            "AutoPipelineForImage2Image",
+            "AutoPipelineForInpainting",
+            "AutoPipelineForText2Image",
+            "ConsistencyModelPipeline",
+            "DanceDiffusionPipeline",
+            "DDIMPipeline",
+            "DDPMPipeline",
+            "DiffusionPipeline",
+            "DiTPipeline",
+            "ImagePipelineOutput",
+            "KarrasVePipeline",
+            "LDMPipeline",
+            "LDMSuperResolutionPipeline",
+            "PNDMPipeline",
+            "RePaintPipeline",
+            "ScoreSdeVePipeline",
+        ]
+    )
+    _import_structure["schedulers"].extend(
+        [
+            "AmusedScheduler",
+            "CMStochasticIterativeScheduler",
+            "DDIMInverseScheduler",
+            "DDIMParallelScheduler",
+            "DDIMScheduler",
+            "DDPMParallelScheduler",
+            "DDPMScheduler",
+            "DDPMWuerstchenScheduler",
+            "DEISMultistepScheduler",
+            "DPMSolverMultistepInverseScheduler",
+            "DPMSolverMultistepScheduler",
+            "DPMSolverSinglestepScheduler",
+            "EulerAncestralDiscreteScheduler",
+            "EulerDiscreteScheduler",
+            "HeunDiscreteScheduler",
+            "IPNDMScheduler",
+            "KarrasVeScheduler",
+            "KDPM2AncestralDiscreteScheduler",
+            "KDPM2DiscreteScheduler",
+            "LCMScheduler",
+            "PNDMScheduler",
+            "RePaintScheduler",
+            "SchedulerMixin",
+            "ScoreSdeVeScheduler",
+            "UnCLIPScheduler",
+            "UniPCMultistepScheduler",
+            "VQDiffusionScheduler",
+        ]
+    )
+    _import_structure["training_utils"] = ["EMAModel"]
+
+try:
+    if not (is_torch_available() and is_scipy_available()):
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    from .utils import dummy_torch_and_scipy_objects  # noqa F403
+
+    _import_structure["utils.dummy_torch_and_scipy_objects"] = [
+        name for name in dir(dummy_torch_and_scipy_objects) if not name.startswith("_")
+    ]
+
+else:
+    _import_structure["schedulers"].extend(["LMSDiscreteScheduler"])
+
+try:
+    if not (is_torch_available() and is_torchsde_available()):
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    from .utils import dummy_torch_and_torchsde_objects  # noqa F403
+
+    _import_structure["utils.dummy_torch_and_torchsde_objects"] = [
+        name for name in dir(dummy_torch_and_torchsde_objects) if not name.startswith("_")
+    ]
+
+else:
+    _import_structure["schedulers"].extend(["DPMSolverSDEScheduler"])
+
+try:
+    if not (is_torch_available() and is_transformers_available()):
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    from .utils import dummy_torch_and_transformers_objects  # noqa F403
+
+    _import_structure["utils.dummy_torch_and_transformers_objects"] = [
+        name for name in dir(dummy_torch_and_transformers_objects) if not name.startswith("_")
+    ]
+
+else:
+    _import_structure["pipelines"].extend(
+        [
+            "AltDiffusionImg2ImgPipeline",
+            "AltDiffusionPipeline",
+            "AmusedImg2ImgPipeline",
+            "AmusedInpaintPipeline",
+            "AmusedPipeline",
+            "AnimateDiffPipeline",
+            "AudioLDM2Pipeline",
+            "AudioLDM2ProjectionModel",
+            "AudioLDM2UNet2DConditionModel",
+            "AudioLDMPipeline",
+            "BlipDiffusionControlNetPipeline",
+            "BlipDiffusionPipeline",
+            "CLIPImageProjection",
+            "CycleDiffusionPipeline",
+            "IFImg2ImgPipeline",
+            "IFImg2ImgSuperResolutionPipeline",
+            "IFInpaintingPipeline",
+            "IFInpaintingSuperResolutionPipeline",
+            "IFPipeline",
+            "IFSuperResolutionPipeline",
+            "ImageTextPipelineOutput",
+            "Kandinsky3Img2ImgPipeline",
+            "Kandinsky3Pipeline",
+            "KandinskyCombinedPipeline",
+            "KandinskyImg2ImgCombinedPipeline",
+            "KandinskyImg2ImgPipeline",
+            "KandinskyInpaintCombinedPipeline",
+            "KandinskyInpaintPipeline",
+            "KandinskyPipeline",
+            "KandinskyPriorPipeline",
+            "KandinskyV22CombinedPipeline",
+            "KandinskyV22ControlnetImg2ImgPipeline",
+            "KandinskyV22ControlnetPipeline",
+            "KandinskyV22Img2ImgCombinedPipeline",
+            "KandinskyV22Img2ImgPipeline",
+            "KandinskyV22InpaintCombinedPipeline",
+            "KandinskyV22InpaintPipeline",
+            "KandinskyV22Pipeline",
+            "KandinskyV22PriorEmb2EmbPipeline",
+            "KandinskyV22PriorPipeline",
+            "LatentConsistencyModelImg2ImgPipeline",
+            "LatentConsistencyModelPipeline",
+            "LDMTextToImagePipeline",
+            "MusicLDMPipeline",
+            "PaintByExamplePipeline",
+            "PixArtAlphaPipeline",
+            "SemanticStableDiffusionPipeline",
+            "ShapEImg2ImgPipeline",
+            "ShapEPipeline",
+            "StableDiffusionAdapterPipeline",
+            "StableDiffusionAttendAndExcitePipeline",
+            "StableDiffusionControlNetImg2ImgPipeline",
+            "StableDiffusionControlNetInpaintPipeline",
+            "StableDiffusionControlNetPipeline",
+            "StableDiffusionDepth2ImgPipeline",
+            "StableDiffusionDiffEditPipeline",
+            "StableDiffusionGLIGENPipeline",
+            "StableDiffusionGLIGENTextImagePipeline",
+            "StableDiffusionImageVariationPipeline",
+            "StableDiffusionImg2ImgPipeline",
+            "StableDiffusionInpaintPipeline",
+            "StableDiffusionInpaintPipelineLegacy",
+            "StableDiffusionInstructPix2PixPipeline",
+            "StableDiffusionLatentUpscalePipeline",
+            "StableDiffusionLDM3DPipeline",
+            "StableDiffusionModelEditingPipeline",
+            "StableDiffusionPanoramaPipeline",
+            "StableDiffusionParadigmsPipeline",
+            "StableDiffusionPipeline",
+            "StableDiffusionPipelineSafe",
+            "StableDiffusionPix2PixZeroPipeline",
+            "StableDiffusionSAGPipeline",
+            "StableDiffusionUpscalePipeline",
+            "StableDiffusionXLAdapterPipeline",
+            "StableDiffusionXLControlNetImg2ImgPipeline",
+            "StableDiffusionXLControlNetInpaintPipeline",
+            "StableDiffusionXLControlNetPipeline",
+            "StableDiffusionXLImg2ImgPipeline",
+            "StableDiffusionXLInpaintPipeline",
+            "StableDiffusionXLInstructPix2PixPipeline",
+            "StableDiffusionXLPipeline",
+            "StableUnCLIPImg2ImgPipeline",
+            "StableUnCLIPPipeline",
+            "StableVideoDiffusionPipeline",
+            "TextToVideoSDPipeline",
+            "TextToVideoZeroPipeline",
+            "TextToVideoZeroSDXLPipeline",
+            "UnCLIPImageVariationPipeline",
+            "UnCLIPPipeline",
+            "UniDiffuserModel",
+            "UniDiffuserPipeline",
+            "UniDiffuserTextDecoder",
+            "VersatileDiffusionDualGuidedPipeline",
+            "VersatileDiffusionImageVariationPipeline",
+            "VersatileDiffusionPipeline",
+            "VersatileDiffusionTextToImagePipeline",
+            "VideoToVideoSDPipeline",
+            "VQDiffusionPipeline",
+            "WuerstchenCombinedPipeline",
+            "WuerstchenDecoderPipeline",
+            "WuerstchenPriorPipeline",
+        ]
+    )
+
+try:
+    if not (is_torch_available() and is_transformers_available() and is_k_diffusion_available()):
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    from .utils import dummy_torch_and_transformers_and_k_diffusion_objects  # noqa F403
+
+    _import_structure["utils.dummy_torch_and_transformers_and_k_diffusion_objects"] = [
+        name for name in dir(dummy_torch_and_transformers_and_k_diffusion_objects) if not name.startswith("_")
+    ]
+
+else:
+    _import_structure["pipelines"].extend(["StableDiffusionKDiffusionPipeline"])
+
+try:
+    if not (is_torch_available() and is_transformers_available() and is_onnx_available()):
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    from .utils import dummy_torch_and_transformers_and_onnx_objects  # noqa F403
+
+    _import_structure["utils.dummy_torch_and_transformers_and_onnx_objects"] = [
+        name for name in dir(dummy_torch_and_transformers_and_onnx_objects) if not name.startswith("_")
+    ]
+
+else:
+    _import_structure["pipelines"].extend(
+        [
+            "OnnxStableDiffusionImg2ImgPipeline",
+            "OnnxStableDiffusionInpaintPipeline",
+            "OnnxStableDiffusionInpaintPipelineLegacy",
+            "OnnxStableDiffusionPipeline",
+            "OnnxStableDiffusionUpscalePipeline",
+            "StableDiffusionOnnxPipeline",
+        ]
+    )
+
+try:
+    if not (is_torch_available() and is_librosa_available()):
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    from .utils import dummy_torch_and_librosa_objects  # noqa F403
+
+    _import_structure["utils.dummy_torch_and_librosa_objects"] = [
+        name for name in dir(dummy_torch_and_librosa_objects) if not name.startswith("_")
+    ]
+
+else:
+    _import_structure["pipelines"].extend(["AudioDiffusionPipeline", "Mel"])
+
+try:
+    if not (is_transformers_available() and is_torch_available() and is_note_seq_available()):
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    from .utils import dummy_transformers_and_torch_and_note_seq_objects  # noqa F403
+
+    _import_structure["utils.dummy_transformers_and_torch_and_note_seq_objects"] = [
+        name for name in dir(dummy_transformers_and_torch_and_note_seq_objects) if not name.startswith("_")
+    ]
+
+
+else:
+    _import_structure["pipelines"].extend(["SpectrogramDiffusionPipeline"])
+
+try:
+    if not is_flax_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    from .utils import dummy_flax_objects  # noqa F403
+
+    _import_structure["utils.dummy_flax_objects"] = [
+        name for name in dir(dummy_flax_objects) if not name.startswith("_")
+    ]
+
+
+else:
+    _import_structure["models.controlnet_flax"] = ["FlaxControlNetModel"]
+    _import_structure["models.modeling_flax_utils"] = ["FlaxModelMixin"]
+    _import_structure["models.unet_2d_condition_flax"] = ["FlaxUNet2DConditionModel"]
+    _import_structure["models.vae_flax"] = ["FlaxAutoencoderKL"]
+    _import_structure["pipelines"].extend(["FlaxDiffusionPipeline"])
+    _import_structure["schedulers"].extend(
+        [
+            "FlaxDDIMScheduler",
+            "FlaxDDPMScheduler",
+            "FlaxDPMSolverMultistepScheduler",
+            "FlaxEulerDiscreteScheduler",
+            "FlaxKarrasVeScheduler",
+            "FlaxLMSDiscreteScheduler",
+            "FlaxPNDMScheduler",
+            "FlaxSchedulerMixin",
+            "FlaxScoreSdeVeScheduler",
+        ]
+    )
+
+
+try:
+    if not (is_flax_available() and is_transformers_available()):
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    from .utils import dummy_flax_and_transformers_objects  # noqa F403
+
+    _import_structure["utils.dummy_flax_and_transformers_objects"] = [
+        name for name in dir(dummy_flax_and_transformers_objects) if not name.startswith("_")
+    ]
+
+
+else:
+    _import_structure["pipelines"].extend(
+        [
+            "FlaxStableDiffusionControlNetPipeline",
+            "FlaxStableDiffusionImg2ImgPipeline",
+            "FlaxStableDiffusionInpaintPipeline",
+            "FlaxStableDiffusionPipeline",
+            "FlaxStableDiffusionXLPipeline",
+        ]
+    )
+
+try:
+    if not (is_note_seq_available()):
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    from .utils import dummy_note_seq_objects  # noqa F403
+
+    _import_structure["utils.dummy_note_seq_objects"] = [
+        name for name in dir(dummy_note_seq_objects) if not name.startswith("_")
+    ]
+
+
+else:
+    _import_structure["pipelines"].extend(["MidiProcessor"])
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+    from .configuration_utils import ConfigMixin
+
+    try:
+        if not is_onnx_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        from .utils.dummy_onnx_objects import *  # noqa F403
+    else:
+        from .pipelines import OnnxRuntimeModel
+
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        from .utils.dummy_pt_objects import *  # noqa F403
+    else:
+        from .models import (
+            AsymmetricAutoencoderKL,
+            AutoencoderKL,
+            AutoencoderKLTemporalDecoder,
+            AutoencoderTiny,
+            ConsistencyDecoderVAE,
+            ControlNetModel,
+            Kandinsky3UNet,
+            ModelMixin,
+            MotionAdapter,
+            MultiAdapter,
+            PriorTransformer,
+            T2IAdapter,
+            T5FilmDecoder,
+            Transformer2DModel,
+            UNet1DModel,
+            UNet2DConditionModel,
+            UNet2DModel,
+            UNet3DConditionModel,
+            UNetMotionModel,
+            UNetSpatioTemporalConditionModel,
+            UVit2DModel,
+            VQModel,
+        )
+        from .optimization import (
+            get_constant_schedule,
+            get_constant_schedule_with_warmup,
+            get_cosine_schedule_with_warmup,
+            get_cosine_with_hard_restarts_schedule_with_warmup,
+            get_linear_schedule_with_warmup,
+            get_polynomial_decay_schedule_with_warmup,
+            get_scheduler,
+        )
+        from .pipelines import (
+            AudioPipelineOutput,
+            AutoPipelineForImage2Image,
+            AutoPipelineForInpainting,
+            AutoPipelineForText2Image,
+            BlipDiffusionControlNetPipeline,
+            BlipDiffusionPipeline,
+            CLIPImageProjection,
+            ConsistencyModelPipeline,
+            DanceDiffusionPipeline,
+            DDIMPipeline,
+            DDPMPipeline,
+            DiffusionPipeline,
+            DiTPipeline,
+            ImagePipelineOutput,
+            KarrasVePipeline,
+            LDMPipeline,
+            LDMSuperResolutionPipeline,
+            PNDMPipeline,
+            RePaintPipeline,
+            ScoreSdeVePipeline,
+        )
+        from .schedulers import (
+            AmusedScheduler,
+            CMStochasticIterativeScheduler,
+            DDIMInverseScheduler,
+            DDIMParallelScheduler,
+            DDIMScheduler,
+            DDPMParallelScheduler,
+            DDPMScheduler,
+            DDPMWuerstchenScheduler,
+            DEISMultistepScheduler,
+            DPMSolverMultistepInverseScheduler,
+            DPMSolverMultistepScheduler,
+            DPMSolverSinglestepScheduler,
+            EulerAncestralDiscreteScheduler,
+            EulerDiscreteScheduler,
+            HeunDiscreteScheduler,
+            IPNDMScheduler,
+            KarrasVeScheduler,
+            KDPM2AncestralDiscreteScheduler,
+            KDPM2DiscreteScheduler,
+            LCMScheduler,
+            PNDMScheduler,
+            RePaintScheduler,
+            SchedulerMixin,
+            ScoreSdeVeScheduler,
+            UnCLIPScheduler,
+            UniPCMultistepScheduler,
+            VQDiffusionScheduler,
+        )
+        from .training_utils import EMAModel
+
+    try:
+        if not (is_torch_available() and is_scipy_available()):
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        from .utils.dummy_torch_and_scipy_objects import *  # noqa F403
+    else:
+        from .schedulers import LMSDiscreteScheduler
+
+    try:
+        if not (is_torch_available() and is_torchsde_available()):
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        from .utils.dummy_torch_and_torchsde_objects import *  # noqa F403
+    else:
+        from .schedulers import DPMSolverSDEScheduler
+
+    try:
+        if not (is_torch_available() and is_transformers_available()):
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        from .utils.dummy_torch_and_transformers_objects import *  # noqa F403
+    else:
+        from .pipelines import (
+            AltDiffusionImg2ImgPipeline,
+            AltDiffusionPipeline,
+            AmusedImg2ImgPipeline,
+            AmusedInpaintPipeline,
+            AmusedPipeline,
+            AnimateDiffPipeline,
+            AudioLDM2Pipeline,
+            AudioLDM2ProjectionModel,
+            AudioLDM2UNet2DConditionModel,
+            AudioLDMPipeline,
+            CLIPImageProjection,
+            CycleDiffusionPipeline,
+            IFImg2ImgPipeline,
+            IFImg2ImgSuperResolutionPipeline,
+            IFInpaintingPipeline,
+            IFInpaintingSuperResolutionPipeline,
+            IFPipeline,
+            IFSuperResolutionPipeline,
+            ImageTextPipelineOutput,
+            Kandinsky3Img2ImgPipeline,
+            Kandinsky3Pipeline,
+            KandinskyCombinedPipeline,
+            KandinskyImg2ImgCombinedPipeline,
+            KandinskyImg2ImgPipeline,
+            KandinskyInpaintCombinedPipeline,
+            KandinskyInpaintPipeline,
+            KandinskyPipeline,
+            KandinskyPriorPipeline,
+            KandinskyV22CombinedPipeline,
+            KandinskyV22ControlnetImg2ImgPipeline,
+            KandinskyV22ControlnetPipeline,
+            KandinskyV22Img2ImgCombinedPipeline,
+            KandinskyV22Img2ImgPipeline,
+            KandinskyV22InpaintCombinedPipeline,
+            KandinskyV22InpaintPipeline,
+            KandinskyV22Pipeline,
+            KandinskyV22PriorEmb2EmbPipeline,
+            KandinskyV22PriorPipeline,
+            LatentConsistencyModelImg2ImgPipeline,
+            LatentConsistencyModelPipeline,
+            LDMTextToImagePipeline,
+            MusicLDMPipeline,
+            PaintByExamplePipeline,
+            PixArtAlphaPipeline,
+            SemanticStableDiffusionPipeline,
+            ShapEImg2ImgPipeline,
+            ShapEPipeline,
+            StableDiffusionAdapterPipeline,
+            StableDiffusionAttendAndExcitePipeline,
+            StableDiffusionControlNetImg2ImgPipeline,
+            StableDiffusionControlNetInpaintPipeline,
+            StableDiffusionControlNetPipeline,
+            StableDiffusionDepth2ImgPipeline,
+            StableDiffusionDiffEditPipeline,
+            StableDiffusionGLIGENPipeline,
+            StableDiffusionGLIGENTextImagePipeline,
+            StableDiffusionImageVariationPipeline,
+            StableDiffusionImg2ImgPipeline,
+            StableDiffusionInpaintPipeline,
+            StableDiffusionInpaintPipelineLegacy,
+            StableDiffusionInstructPix2PixPipeline,
+            StableDiffusionLatentUpscalePipeline,
+            StableDiffusionLDM3DPipeline,
+            StableDiffusionModelEditingPipeline,
+            StableDiffusionPanoramaPipeline,
+            StableDiffusionParadigmsPipeline,
+            StableDiffusionPipeline,
+            StableDiffusionPipelineSafe,
+            StableDiffusionPix2PixZeroPipeline,
+            StableDiffusionSAGPipeline,
+            StableDiffusionUpscalePipeline,
+            StableDiffusionXLAdapterPipeline,
+            StableDiffusionXLControlNetImg2ImgPipeline,
+            StableDiffusionXLControlNetInpaintPipeline,
+            StableDiffusionXLControlNetPipeline,
+            StableDiffusionXLImg2ImgPipeline,
+            StableDiffusionXLInpaintPipeline,
+            StableDiffusionXLInstructPix2PixPipeline,
+            StableDiffusionXLPipeline,
+            StableUnCLIPImg2ImgPipeline,
+            StableUnCLIPPipeline,
+            StableVideoDiffusionPipeline,
+            TextToVideoSDPipeline,
+            TextToVideoZeroPipeline,
+            TextToVideoZeroSDXLPipeline,
+            UnCLIPImageVariationPipeline,
+            UnCLIPPipeline,
+            UniDiffuserModel,
+            UniDiffuserPipeline,
+            UniDiffuserTextDecoder,
+            VersatileDiffusionDualGuidedPipeline,
+            VersatileDiffusionImageVariationPipeline,
+            VersatileDiffusionPipeline,
+            VersatileDiffusionTextToImagePipeline,
+            VideoToVideoSDPipeline,
+            VQDiffusionPipeline,
+            WuerstchenCombinedPipeline,
+            WuerstchenDecoderPipeline,
+            WuerstchenPriorPipeline,
+        )
+
+    try:
+        if not (is_torch_available() and is_transformers_available() and is_k_diffusion_available()):
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        from .utils.dummy_torch_and_transformers_and_k_diffusion_objects import *  # noqa F403
+    else:
+        from .pipelines import StableDiffusionKDiffusionPipeline
+
+    try:
+        if not (is_torch_available() and is_transformers_available() and is_onnx_available()):
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        from .utils.dummy_torch_and_transformers_and_onnx_objects import *  # noqa F403
+    else:
+        from .pipelines import (
+            OnnxStableDiffusionImg2ImgPipeline,
+            OnnxStableDiffusionInpaintPipeline,
+            OnnxStableDiffusionInpaintPipelineLegacy,
+            OnnxStableDiffusionPipeline,
+            OnnxStableDiffusionUpscalePipeline,
+            StableDiffusionOnnxPipeline,
+        )
+
+    try:
+        if not (is_torch_available() and is_librosa_available()):
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        from .utils.dummy_torch_and_librosa_objects import *  # noqa F403
+    else:
+        from .pipelines import AudioDiffusionPipeline, Mel
+
+    try:
+        if not (is_transformers_available() and is_torch_available() and is_note_seq_available()):
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        from .utils.dummy_transformers_and_torch_and_note_seq_objects import *  # noqa F403
+    else:
+        from .pipelines import SpectrogramDiffusionPipeline
+
+    try:
+        if not is_flax_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        from .utils.dummy_flax_objects import *  # noqa F403
+    else:
+        from .models.controlnet_flax import FlaxControlNetModel
+        from .models.modeling_flax_utils import FlaxModelMixin
+        from .models.unet_2d_condition_flax import FlaxUNet2DConditionModel
+        from .models.vae_flax import FlaxAutoencoderKL
+        from .pipelines import FlaxDiffusionPipeline
+        from .schedulers import (
+            FlaxDDIMScheduler,
+            FlaxDDPMScheduler,
+            FlaxDPMSolverMultistepScheduler,
+            FlaxEulerDiscreteScheduler,
+            FlaxKarrasVeScheduler,
+            FlaxLMSDiscreteScheduler,
+            FlaxPNDMScheduler,
+            FlaxSchedulerMixin,
+            FlaxScoreSdeVeScheduler,
+        )
+
+    try:
+        if not (is_flax_available() and is_transformers_available()):
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        from .utils.dummy_flax_and_transformers_objects import *  # noqa F403
+    else:
+        from .pipelines import (
+            FlaxStableDiffusionControlNetPipeline,
+            FlaxStableDiffusionImg2ImgPipeline,
+            FlaxStableDiffusionInpaintPipeline,
+            FlaxStableDiffusionPipeline,
+            FlaxStableDiffusionXLPipeline,
+        )
+
+    try:
+        if not (is_note_seq_available()):
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        from .utils.dummy_note_seq_objects import *  # noqa F403
+    else:
+        from .pipelines import MidiProcessor
+
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(
+        __name__,
+        globals()["__file__"],
+        _import_structure,
+        module_spec=__spec__,
+        extra_objects={"__version__": __version__},
+    )

src/diffusers/commands/__init__.py

@@ -0,0 +1,27 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from abc import ABC, abstractmethod
+from argparse import ArgumentParser
+
+
+class BaseDiffusersCLICommand(ABC):
+    @staticmethod
+    @abstractmethod
+    def register_subcommand(parser: ArgumentParser):
+        raise NotImplementedError()
+
+    @abstractmethod
+    def run(self):
+        raise NotImplementedError()

src/diffusers/commands/diffusers_cli.py

@@ -0,0 +1,43 @@
+#!/usr/bin/env python
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from argparse import ArgumentParser
+
+from .env import EnvironmentCommand
+from .fp16_safetensors import FP16SafetensorsCommand
+
+
+def main():
+    parser = ArgumentParser("Diffusers CLI tool", usage="diffusers-cli <command> [<args>]")
+    commands_parser = parser.add_subparsers(help="diffusers-cli command helpers")
+
+    # Register commands
+    EnvironmentCommand.register_subcommand(commands_parser)
+    FP16SafetensorsCommand.register_subcommand(commands_parser)
+
+    # Let's go
+    args = parser.parse_args()
+
+    if not hasattr(args, "func"):
+        parser.print_help()
+        exit(1)
+
+    # Run
+    service = args.func(args)
+    service.run()
+
+
+if __name__ == "__main__":
+    main()

src/diffusers/commands/env.py

@@ -0,0 +1,84 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import platform
+from argparse import ArgumentParser
+
+import huggingface_hub
+
+from .. import __version__ as version
+from ..utils import is_accelerate_available, is_torch_available, is_transformers_available, is_xformers_available
+from . import BaseDiffusersCLICommand
+
+
+def info_command_factory(_):
+    return EnvironmentCommand()
+
+
+class EnvironmentCommand(BaseDiffusersCLICommand):
+    @staticmethod
+    def register_subcommand(parser: ArgumentParser):
+        download_parser = parser.add_parser("env")
+        download_parser.set_defaults(func=info_command_factory)
+
+    def run(self):
+        hub_version = huggingface_hub.__version__
+
+        pt_version = "not installed"
+        pt_cuda_available = "NA"
+        if is_torch_available():
+            import torch
+
+            pt_version = torch.__version__
+            pt_cuda_available = torch.cuda.is_available()
+
+        transformers_version = "not installed"
+        if is_transformers_available():
+            import transformers
+
+            transformers_version = transformers.__version__
+
+        accelerate_version = "not installed"
+        if is_accelerate_available():
+            import accelerate
+
+            accelerate_version = accelerate.__version__
+
+        xformers_version = "not installed"
+        if is_xformers_available():
+            import xformers
+
+            xformers_version = xformers.__version__
+
+        info = {
+            "`diffusers` version": version,
+            "Platform": platform.platform(),
+            "Python version": platform.python_version(),
+            "PyTorch version (GPU?)": f"{pt_version} ({pt_cuda_available})",
+            "Huggingface_hub version": hub_version,
+            "Transformers version": transformers_version,
+            "Accelerate version": accelerate_version,
+            "xFormers version": xformers_version,
+            "Using GPU in script?": "<fill in>",
+            "Using distributed or parallel set-up in script?": "<fill in>",
+        }
+
+        print("\nCopy-and-paste the text below in your GitHub issue and FILL OUT the two last points.\n")
+        print(self.format_dict(info))
+
+        return info
+
+    @staticmethod
+    def format_dict(d):
+        return "\n".join([f"- {prop}: {val}" for prop, val in d.items()]) + "\n"

src/diffusers/commands/fp16_safetensors.py

@@ -0,0 +1,132 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Usage example:
+    diffusers-cli fp16_safetensors --ckpt_id=openai/shap-e --fp16 --use_safetensors
+"""
+
+import glob
+import json
+import warnings
+from argparse import ArgumentParser, Namespace
+from importlib import import_module
+
+import huggingface_hub
+import torch
+from huggingface_hub import hf_hub_download
+from packaging import version
+
+from ..utils import logging
+from . import BaseDiffusersCLICommand
+
+
+def conversion_command_factory(args: Namespace):
+    if args.use_auth_token:
+        warnings.warn(
+            "The `--use_auth_token` flag is deprecated and will be removed in a future version. Authentication is now"
+            " handled automatically if user is logged in."
+        )
+    return FP16SafetensorsCommand(args.ckpt_id, args.fp16, args.use_safetensors)
+
+
+class FP16SafetensorsCommand(BaseDiffusersCLICommand):
+    @staticmethod
+    def register_subcommand(parser: ArgumentParser):
+        conversion_parser = parser.add_parser("fp16_safetensors")
+        conversion_parser.add_argument(
+            "--ckpt_id",
+            type=str,
+            help="Repo id of the checkpoints on which to run the conversion. Example: 'openai/shap-e'.",
+        )
+        conversion_parser.add_argument(
+            "--fp16", action="store_true", help="If serializing the variables in FP16 precision."
+        )
+        conversion_parser.add_argument(
+            "--use_safetensors", action="store_true", help="If serializing in the safetensors format."
+        )
+        conversion_parser.add_argument(
+            "--use_auth_token",
+            action="store_true",
+            help="When working with checkpoints having private visibility. When used `huggingface-cli login` needs to be run beforehand.",
+        )
+        conversion_parser.set_defaults(func=conversion_command_factory)
+
+    def __init__(self, ckpt_id: str, fp16: bool, use_safetensors: bool):
+        self.logger = logging.get_logger("diffusers-cli/fp16_safetensors")
+        self.ckpt_id = ckpt_id
+        self.local_ckpt_dir = f"/tmp/{ckpt_id}"
+        self.fp16 = fp16
+
+        self.use_safetensors = use_safetensors
+
+        if not self.use_safetensors and not self.fp16:
+            raise NotImplementedError(
+                "When `use_safetensors` and `fp16` both are False, then this command is of no use."
+            )
+
+    def run(self):
+        if version.parse(huggingface_hub.__version__) < version.parse("0.9.0"):
+            raise ImportError(
+                "The huggingface_hub version must be >= 0.9.0 to use this command. Please update your huggingface_hub"
+                " installation."
+            )
+        else:
+            from huggingface_hub import create_commit
+            from huggingface_hub._commit_api import CommitOperationAdd
+
+        model_index = hf_hub_download(repo_id=self.ckpt_id, filename="model_index.json")
+        with open(model_index, "r") as f:
+            pipeline_class_name = json.load(f)["_class_name"]
+        pipeline_class = getattr(import_module("diffusers"), pipeline_class_name)
+        self.logger.info(f"Pipeline class imported: {pipeline_class_name}.")
+
+        # Load the appropriate pipeline. We could have use `DiffusionPipeline`
+        # here, but just to avoid any rough edge cases.
+        pipeline = pipeline_class.from_pretrained(
+            self.ckpt_id, torch_dtype=torch.float16 if self.fp16 else torch.float32
+        )
+        pipeline.save_pretrained(
+            self.local_ckpt_dir,
+            safe_serialization=True if self.use_safetensors else False,
+            variant="fp16" if self.fp16 else None,
+        )
+        self.logger.info(f"Pipeline locally saved to {self.local_ckpt_dir}.")
+
+        # Fetch all the paths.
+        if self.fp16:
+            modified_paths = glob.glob(f"{self.local_ckpt_dir}/*/*.fp16.*")
+        elif self.use_safetensors:
+            modified_paths = glob.glob(f"{self.local_ckpt_dir}/*/*.safetensors")
+
+        # Prepare for the PR.
+        commit_message = f"Serialize variables with FP16: {self.fp16} and safetensors: {self.use_safetensors}."
+        operations = []
+        for path in modified_paths:
+            operations.append(CommitOperationAdd(path_in_repo="/".join(path.split("/")[4:]), path_or_fileobj=path))
+
+        # Open the PR.
+        commit_description = (
+            "Variables converted by the [`diffusers`' `fp16_safetensors`"
+            " CLI](https://github.com/huggingface/diffusers/blob/main/src/diffusers/commands/fp16_safetensors.py)."
+        )
+        hub_pr_url = create_commit(
+            repo_id=self.ckpt_id,
+            operations=operations,
+            commit_message=commit_message,
+            commit_description=commit_description,
+            repo_type="model",
+            create_pr=True,
+        ).pr_url
+        self.logger.info(f"PR created here: {hub_pr_url}.")

src/diffusers/configuration_utils.py

@@ -0,0 +1,699 @@
+# coding=utf-8
+# Copyright 2023 The HuggingFace Inc. team.
+# Copyright (c) 2022, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" ConfigMixin base class and utilities."""
+import dataclasses
+import functools
+import importlib
+import inspect
+import json
+import os
+import re
+from collections import OrderedDict
+from pathlib import PosixPath
+from typing import Any, Dict, Tuple, Union
+
+import numpy as np
+from huggingface_hub import create_repo, hf_hub_download
+from huggingface_hub.utils import (
+    EntryNotFoundError,
+    RepositoryNotFoundError,
+    RevisionNotFoundError,
+    validate_hf_hub_args,
+)
+from requests import HTTPError
+
+from . import __version__
+from .utils import (
+    HUGGINGFACE_CO_RESOLVE_ENDPOINT,
+    DummyObject,
+    deprecate,
+    extract_commit_hash,
+    http_user_agent,
+    logging,
+)
+
+
+logger = logging.get_logger(__name__)
+
+_re_configuration_file = re.compile(r"config\.(.*)\.json")
+
+
+class FrozenDict(OrderedDict):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        for key, value in self.items():
+            setattr(self, key, value)
+
+        self.__frozen = True
+
+    def __delitem__(self, *args, **kwargs):
+        raise Exception(f"You cannot use ``__delitem__`` on a {self.__class__.__name__} instance.")
+
+    def setdefault(self, *args, **kwargs):
+        raise Exception(f"You cannot use ``setdefault`` on a {self.__class__.__name__} instance.")
+
+    def pop(self, *args, **kwargs):
+        raise Exception(f"You cannot use ``pop`` on a {self.__class__.__name__} instance.")
+
+    def update(self, *args, **kwargs):
+        raise Exception(f"You cannot use ``update`` on a {self.__class__.__name__} instance.")
+
+    def __setattr__(self, name, value):
+        if hasattr(self, "__frozen") and self.__frozen:
+            raise Exception(f"You cannot use ``__setattr__`` on a {self.__class__.__name__} instance.")
+        super().__setattr__(name, value)
+
+    def __setitem__(self, name, value):
+        if hasattr(self, "__frozen") and self.__frozen:
+            raise Exception(f"You cannot use ``__setattr__`` on a {self.__class__.__name__} instance.")
+        super().__setitem__(name, value)
+
+
+class ConfigMixin:
+    r"""
+    Base class for all configuration classes. All configuration parameters are stored under `self.config`. Also
+    provides the [`~ConfigMixin.from_config`] and [`~ConfigMixin.save_config`] methods for loading, downloading, and
+    saving classes that inherit from [`ConfigMixin`].
+
+    Class attributes:
+        - **config_name** (`str`) -- A filename under which the config should stored when calling
+          [`~ConfigMixin.save_config`] (should be overridden by parent class).
+        - **ignore_for_config** (`List[str]`) -- A list of attributes that should not be saved in the config (should be
+          overridden by subclass).
+        - **has_compatibles** (`bool`) -- Whether the class has compatible classes (should be overridden by subclass).
+        - **_deprecated_kwargs** (`List[str]`) -- Keyword arguments that are deprecated. Note that the `init` function
+          should only have a `kwargs` argument if at least one argument is deprecated (should be overridden by
+          subclass).
+    """
+
+    config_name = None
+    ignore_for_config = []
+    has_compatibles = False
+
+    _deprecated_kwargs = []
+
+    def register_to_config(self, **kwargs):
+        if self.config_name is None:
+            raise NotImplementedError(f"Make sure that {self.__class__} has defined a class name `config_name`")
+        # Special case for `kwargs` used in deprecation warning added to schedulers
+        # TODO: remove this when we remove the deprecation warning, and the `kwargs` argument,
+        # or solve in a more general way.
+        kwargs.pop("kwargs", None)
+
+        if not hasattr(self, "_internal_dict"):
+            internal_dict = kwargs
+        else:
+            previous_dict = dict(self._internal_dict)
+            internal_dict = {**self._internal_dict, **kwargs}
+            logger.debug(f"Updating config from {previous_dict} to {internal_dict}")
+
+        self._internal_dict = FrozenDict(internal_dict)
+
+    def __getattr__(self, name: str) -> Any:
+        """The only reason we overwrite `getattr` here is to gracefully deprecate accessing
+        config attributes directly. See https://github.com/huggingface/diffusers/pull/3129
+
+        Tihs funtion is mostly copied from PyTorch's __getattr__ overwrite:
+        https://pytorch.org/docs/stable/_modules/torch/nn/modules/module.html#Module
+        """
+
+        is_in_config = "_internal_dict" in self.__dict__ and hasattr(self.__dict__["_internal_dict"], name)
+        is_attribute = name in self.__dict__
+
+        if is_in_config and not is_attribute:
+            deprecation_message = f"Accessing config attribute `{name}` directly via '{type(self).__name__}' object attribute is deprecated. Please access '{name}' over '{type(self).__name__}'s config object instead, e.g. 'scheduler.config.{name}'."
+            deprecate("direct config name access", "1.0.0", deprecation_message, standard_warn=False)
+            return self._internal_dict[name]
+
+        raise AttributeError(f"'{type(self).__name__}' object has no attribute '{name}'")
+
+    def save_config(self, save_directory: Union[str, os.PathLike], push_to_hub: bool = False, **kwargs):
+        """
+        Save a configuration object to the directory specified in `save_directory` so that it can be reloaded using the
+        [`~ConfigMixin.from_config`] class method.
+
+        Args:
+            save_directory (`str` or `os.PathLike`):
+                Directory where the configuration JSON file is saved (will be created if it does not exist).
+            push_to_hub (`bool`, *optional*, defaults to `False`):
+                Whether or not to push your model to the Hugging Face Hub after saving it. You can specify the
+                repository you want to push to with `repo_id` (will default to the name of `save_directory` in your
+                namespace).
+            kwargs (`Dict[str, Any]`, *optional*):
+                Additional keyword arguments passed along to the [`~utils.PushToHubMixin.push_to_hub`] method.
+        """
+        if os.path.isfile(save_directory):
+            raise AssertionError(f"Provided path ({save_directory}) should be a directory, not a file")
+
+        os.makedirs(save_directory, exist_ok=True)
+
+        # If we save using the predefined names, we can load using `from_config`
+        output_config_file = os.path.join(save_directory, self.config_name)
+
+        self.to_json_file(output_config_file)
+        logger.info(f"Configuration saved in {output_config_file}")
+
+        if push_to_hub:
+            commit_message = kwargs.pop("commit_message", None)
+            private = kwargs.pop("private", False)
+            create_pr = kwargs.pop("create_pr", False)
+            token = kwargs.pop("token", None)
+            repo_id = kwargs.pop("repo_id", save_directory.split(os.path.sep)[-1])
+            repo_id = create_repo(repo_id, exist_ok=True, private=private, token=token).repo_id
+
+            self._upload_folder(
+                save_directory,
+                repo_id,
+                token=token,
+                commit_message=commit_message,
+                create_pr=create_pr,
+            )
+
+    @classmethod
+    def from_config(cls, config: Union[FrozenDict, Dict[str, Any]] = None, return_unused_kwargs=False, **kwargs):
+        r"""
+        Instantiate a Python class from a config dictionary.
+
+        Parameters:
+            config (`Dict[str, Any]`):
+                A config dictionary from which the Python class is instantiated. Make sure to only load configuration
+                files of compatible classes.
+            return_unused_kwargs (`bool`, *optional*, defaults to `False`):
+                Whether kwargs that are not consumed by the Python class should be returned or not.
+            kwargs (remaining dictionary of keyword arguments, *optional*):
+                Can be used to update the configuration object (after it is loaded) and initiate the Python class.
+                `**kwargs` are passed directly to the underlying scheduler/model's `__init__` method and eventually
+                overwrite the same named arguments in `config`.
+
+        Returns:
+            [`ModelMixin`] or [`SchedulerMixin`]:
+                A model or scheduler object instantiated from a config dictionary.
+
+        Examples:
+
+        ```python
+        >>> from diffusers import DDPMScheduler, DDIMScheduler, PNDMScheduler
+
+        >>> # Download scheduler from huggingface.co and cache.
+        >>> scheduler = DDPMScheduler.from_pretrained("google/ddpm-cifar10-32")
+
+        >>> # Instantiate DDIM scheduler class with same config as DDPM
+        >>> scheduler = DDIMScheduler.from_config(scheduler.config)
+
+        >>> # Instantiate PNDM scheduler class with same config as DDPM
+        >>> scheduler = PNDMScheduler.from_config(scheduler.config)
+        ```
+        """
+        # <===== TO BE REMOVED WITH DEPRECATION
+        # TODO(Patrick) - make sure to remove the following lines when config=="model_path" is deprecated
+        if "pretrained_model_name_or_path" in kwargs:
+            config = kwargs.pop("pretrained_model_name_or_path")
+
+        if config is None:
+            raise ValueError("Please make sure to provide a config as the first positional argument.")
+        # ======>
+
+        if not isinstance(config, dict):
+            deprecation_message = "It is deprecated to pass a pretrained model name or path to `from_config`."
+            if "Scheduler" in cls.__name__:
+                deprecation_message += (
+                    f"If you were trying to load a scheduler, please use {cls}.from_pretrained(...) instead."
+                    " Otherwise, please make sure to pass a configuration dictionary instead. This functionality will"
+                    " be removed in v1.0.0."
+                )
+            elif "Model" in cls.__name__:
+                deprecation_message += (
+                    f"If you were trying to load a model, please use {cls}.load_config(...) followed by"
+                    f" {cls}.from_config(...) instead. Otherwise, please make sure to pass a configuration dictionary"
+                    " instead. This functionality will be removed in v1.0.0."
+                )
+            deprecate("config-passed-as-path", "1.0.0", deprecation_message, standard_warn=False)
+            config, kwargs = cls.load_config(pretrained_model_name_or_path=config, return_unused_kwargs=True, **kwargs)
+
+        init_dict, unused_kwargs, hidden_dict = cls.extract_init_dict(config, **kwargs)
+
+        # Allow dtype to be specified on initialization
+        if "dtype" in unused_kwargs:
+            init_dict["dtype"] = unused_kwargs.pop("dtype")
+
+        # add possible deprecated kwargs
+        for deprecated_kwarg in cls._deprecated_kwargs:
+            if deprecated_kwarg in unused_kwargs:
+                init_dict[deprecated_kwarg] = unused_kwargs.pop(deprecated_kwarg)
+
+        # Return model and optionally state and/or unused_kwargs
+        model = cls(**init_dict)
+
+        # make sure to also save config parameters that might be used for compatible classes
+        model.register_to_config(**hidden_dict)
+
+        # add hidden kwargs of compatible classes to unused_kwargs
+        unused_kwargs = {**unused_kwargs, **hidden_dict}
+
+        if return_unused_kwargs:
+            return (model, unused_kwargs)
+        else:
+            return model
+
+    @classmethod
+    def get_config_dict(cls, *args, **kwargs):
+        deprecation_message = (
+            f" The function get_config_dict is deprecated. Please use {cls}.load_config instead. This function will be"
+            " removed in version v1.0.0"
+        )
+        deprecate("get_config_dict", "1.0.0", deprecation_message, standard_warn=False)
+        return cls.load_config(*args, **kwargs)
+
+    @classmethod
+    @validate_hf_hub_args
+    def load_config(
+        cls,
+        pretrained_model_name_or_path: Union[str, os.PathLike],
+        return_unused_kwargs=False,
+        return_commit_hash=False,
+        **kwargs,
+    ) -> Tuple[Dict[str, Any], Dict[str, Any]]:
+        r"""
+        Load a model or scheduler configuration.
+
+        Parameters:
+            pretrained_model_name_or_path (`str` or `os.PathLike`, *optional*):
+                Can be either:
+
+                    - A string, the *model id* (for example `google/ddpm-celebahq-256`) of a pretrained model hosted on
+                      the Hub.
+                    - A path to a *directory* (for example `./my_model_directory`) containing model weights saved with
+                      [`~ConfigMixin.save_config`].
+
+            cache_dir (`Union[str, os.PathLike]`, *optional*):
+                Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
+                is not used.
+            force_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to force the (re-)download of the model weights and configuration files, overriding the
+                cached versions if they exist.
+            resume_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to resume downloading the model weights and configuration files. If set to `False`, any
+                incompletely downloaded files are deleted.
+            proxies (`Dict[str, str]`, *optional*):
+                A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
+                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
+            output_loading_info(`bool`, *optional*, defaults to `False`):
+                Whether or not to also return a dictionary containing missing keys, unexpected keys and error messages.
+            local_files_only (`bool`, *optional*, defaults to `False`):
+                Whether to only load local model weights and configuration files or not. If set to `True`, the model
+                won't be downloaded from the Hub.
+            token (`str` or *bool*, *optional*):
+                The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
+                `diffusers-cli login` (stored in `~/.huggingface`) is used.
+            revision (`str`, *optional*, defaults to `"main"`):
+                The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
+                allowed by Git.
+            subfolder (`str`, *optional*, defaults to `""`):
+                The subfolder location of a model file within a larger model repository on the Hub or locally.
+            return_unused_kwargs (`bool`, *optional*, defaults to `False):
+                Whether unused keyword arguments of the config are returned.
+            return_commit_hash (`bool`, *optional*, defaults to `False):
+                Whether the `commit_hash` of the loaded configuration are returned.
+
+        Returns:
+            `dict`:
+                A dictionary of all the parameters stored in a JSON configuration file.
+
+        """
+        cache_dir = kwargs.pop("cache_dir", None)
+        force_download = kwargs.pop("force_download", False)
+        resume_download = kwargs.pop("resume_download", False)
+        proxies = kwargs.pop("proxies", None)
+        token = kwargs.pop("token", None)
+        local_files_only = kwargs.pop("local_files_only", False)
+        revision = kwargs.pop("revision", None)
+        _ = kwargs.pop("mirror", None)
+        subfolder = kwargs.pop("subfolder", None)
+        user_agent = kwargs.pop("user_agent", {})
+
+        user_agent = {**user_agent, "file_type": "config"}
+        user_agent = http_user_agent(user_agent)
+
+        pretrained_model_name_or_path = str(pretrained_model_name_or_path)
+
+        if cls.config_name is None:
+            raise ValueError(
+                "`self.config_name` is not defined. Note that one should not load a config from "
+                "`ConfigMixin`. Please make sure to define `config_name` in a class inheriting from `ConfigMixin`"
+            )
+
+        if os.path.isfile(pretrained_model_name_or_path):
+            config_file = pretrained_model_name_or_path
+        elif os.path.isdir(pretrained_model_name_or_path):
+            if os.path.isfile(os.path.join(pretrained_model_name_or_path, cls.config_name)):
+                # Load from a PyTorch checkpoint
+                config_file = os.path.join(pretrained_model_name_or_path, cls.config_name)
+            elif subfolder is not None and os.path.isfile(
+                os.path.join(pretrained_model_name_or_path, subfolder, cls.config_name)
+            ):
+                config_file = os.path.join(pretrained_model_name_or_path, subfolder, cls.config_name)
+            else:
+                raise EnvironmentError(
+                    f"Error no file named {cls.config_name} found in directory {pretrained_model_name_or_path}."
+                )
+        else:
+            try:
+                # Load from URL or cache if already cached
+                config_file = hf_hub_download(
+                    pretrained_model_name_or_path,
+                    filename=cls.config_name,
+                    cache_dir=cache_dir,
+                    force_download=force_download,
+                    proxies=proxies,
+                    resume_download=resume_download,
+                    local_files_only=local_files_only,
+                    token=token,
+                    user_agent=user_agent,
+                    subfolder=subfolder,
+                    revision=revision,
+                )
+            except RepositoryNotFoundError:
+                raise EnvironmentError(
+                    f"{pretrained_model_name_or_path} is not a local folder and is not a valid model identifier"
+                    " listed on 'https://huggingface.co/models'\nIf this is a private repository, make sure to pass a"
+                    " token having permission to this repo with `token` or log in with `huggingface-cli login`."
+                )
+            except RevisionNotFoundError:
+                raise EnvironmentError(
+                    f"{revision} is not a valid git identifier (branch name, tag name or commit id) that exists for"
+                    " this model name. Check the model page at"
+                    f" 'https://huggingface.co/{pretrained_model_name_or_path}' for available revisions."
+                )
+            except EntryNotFoundError:
+                raise EnvironmentError(
+                    f"{pretrained_model_name_or_path} does not appear to have a file named {cls.config_name}."
+                )
+            except HTTPError as err:
+                raise EnvironmentError(
+                    "There was a specific connection error when trying to load"
+                    f" {pretrained_model_name_or_path}:\n{err}"
+                )
+            except ValueError:
+                raise EnvironmentError(
+                    f"We couldn't connect to '{HUGGINGFACE_CO_RESOLVE_ENDPOINT}' to load this model, couldn't find it"
+                    f" in the cached files and it looks like {pretrained_model_name_or_path} is not the path to a"
+                    f" directory containing a {cls.config_name} file.\nCheckout your internet connection or see how to"
+                    " run the library in offline mode at"
+                    " 'https://huggingface.co/docs/diffusers/installation#offline-mode'."
+                )
+            except EnvironmentError:
+                raise EnvironmentError(
+                    f"Can't load config for '{pretrained_model_name_or_path}'. If you were trying to load it from "
+                    "'https://huggingface.co/models', make sure you don't have a local directory with the same name. "
+                    f"Otherwise, make sure '{pretrained_model_name_or_path}' is the correct path to a directory "
+                    f"containing a {cls.config_name} file"
+                )
+
+        try:
+            # Load config dict
+            config_dict = cls._dict_from_json_file(config_file)
+
+            commit_hash = extract_commit_hash(config_file)
+        except (json.JSONDecodeError, UnicodeDecodeError):
+            raise EnvironmentError(f"It looks like the config file at '{config_file}' is not a valid JSON file.")
+
+        if not (return_unused_kwargs or return_commit_hash):
+            return config_dict
+
+        outputs = (config_dict,)
+
+        if return_unused_kwargs:
+            outputs += (kwargs,)
+
+        if return_commit_hash:
+            outputs += (commit_hash,)
+
+        return outputs
+
+    @staticmethod
+    def _get_init_keys(cls):
+        return set(dict(inspect.signature(cls.__init__).parameters).keys())
+
+    @classmethod
+    def extract_init_dict(cls, config_dict, **kwargs):
+        # Skip keys that were not present in the original config, so default __init__ values were used
+        used_defaults = config_dict.get("_use_default_values", [])
+        config_dict = {k: v for k, v in config_dict.items() if k not in used_defaults and k != "_use_default_values"}
+
+        # 0. Copy origin config dict
+        original_dict = dict(config_dict.items())
+
+        # 1. Retrieve expected config attributes from __init__ signature
+        expected_keys = cls._get_init_keys(cls)
+        expected_keys.remove("self")
+        # remove general kwargs if present in dict
+        if "kwargs" in expected_keys:
+            expected_keys.remove("kwargs")
+        # remove flax internal keys
+        if hasattr(cls, "_flax_internal_args"):
+            for arg in cls._flax_internal_args:
+                expected_keys.remove(arg)
+
+        # 2. Remove attributes that cannot be expected from expected config attributes
+        # remove keys to be ignored
+        if len(cls.ignore_for_config) > 0:
+            expected_keys = expected_keys - set(cls.ignore_for_config)
+
+        # load diffusers library to import compatible and original scheduler
+        diffusers_library = importlib.import_module(__name__.split(".")[0])
+
+        if cls.has_compatibles:
+            compatible_classes = [c for c in cls._get_compatibles() if not isinstance(c, DummyObject)]
+        else:
+            compatible_classes = []
+
+        expected_keys_comp_cls = set()
+        for c in compatible_classes:
+            expected_keys_c = cls._get_init_keys(c)
+            expected_keys_comp_cls = expected_keys_comp_cls.union(expected_keys_c)
+        expected_keys_comp_cls = expected_keys_comp_cls - cls._get_init_keys(cls)
+        config_dict = {k: v for k, v in config_dict.items() if k not in expected_keys_comp_cls}
+
+        # remove attributes from orig class that cannot be expected
+        orig_cls_name = config_dict.pop("_class_name", cls.__name__)
+        if (
+            isinstance(orig_cls_name, str)
+            and orig_cls_name != cls.__name__
+            and hasattr(diffusers_library, orig_cls_name)
+        ):
+            orig_cls = getattr(diffusers_library, orig_cls_name)
+            unexpected_keys_from_orig = cls._get_init_keys(orig_cls) - expected_keys
+            config_dict = {k: v for k, v in config_dict.items() if k not in unexpected_keys_from_orig}
+        elif not isinstance(orig_cls_name, str) and not isinstance(orig_cls_name, (list, tuple)):
+            raise ValueError(
+                "Make sure that the `_class_name` is of type string or list of string (for custom pipelines)."
+            )
+
+        # remove private attributes
+        config_dict = {k: v for k, v in config_dict.items() if not k.startswith("_")}
+
+        # 3. Create keyword arguments that will be passed to __init__ from expected keyword arguments
+        init_dict = {}
+        for key in expected_keys:
+            # if config param is passed to kwarg and is present in config dict
+            # it should overwrite existing config dict key
+            if key in kwargs and key in config_dict:
+                config_dict[key] = kwargs.pop(key)
+
+            if key in kwargs:
+                # overwrite key
+                init_dict[key] = kwargs.pop(key)
+            elif key in config_dict:
+                # use value from config dict
+                init_dict[key] = config_dict.pop(key)
+
+        # 4. Give nice warning if unexpected values have been passed
+        if len(config_dict) > 0:
+            logger.warning(
+                f"The config attributes {config_dict} were passed to {cls.__name__}, "
+                "but are not expected and will be ignored. Please verify your "
+                f"{cls.config_name} configuration file."
+            )
+
+        # 5. Give nice info if config attributes are initiliazed to default because they have not been passed
+        passed_keys = set(init_dict.keys())
+        if len(expected_keys - passed_keys) > 0:
+            logger.info(
+                f"{expected_keys - passed_keys} was not found in config. Values will be initialized to default values."
+            )
+
+        # 6. Define unused keyword arguments
+        unused_kwargs = {**config_dict, **kwargs}
+
+        # 7. Define "hidden" config parameters that were saved for compatible classes
+        hidden_config_dict = {k: v for k, v in original_dict.items() if k not in init_dict}
+
+        return init_dict, unused_kwargs, hidden_config_dict
+
+    @classmethod
+    def _dict_from_json_file(cls, json_file: Union[str, os.PathLike]):
+        with open(json_file, "r", encoding="utf-8") as reader:
+            text = reader.read()
+        return json.loads(text)
+
+    def __repr__(self):
+        return f"{self.__class__.__name__} {self.to_json_string()}"
+
+    @property
+    def config(self) -> Dict[str, Any]:
+        """
+        Returns the config of the class as a frozen dictionary
+
+        Returns:
+            `Dict[str, Any]`: Config of the class.
+        """
+        return self._internal_dict
+
+    def to_json_string(self) -> str:
+        """
+        Serializes the configuration instance to a JSON string.
+
+        Returns:
+            `str`:
+                String containing all the attributes that make up the configuration instance in JSON format.
+        """
+        config_dict = self._internal_dict if hasattr(self, "_internal_dict") else {}
+        config_dict["_class_name"] = self.__class__.__name__
+        config_dict["_diffusers_version"] = __version__
+
+        def to_json_saveable(value):
+            if isinstance(value, np.ndarray):
+                value = value.tolist()
+            elif isinstance(value, PosixPath):
+                value = str(value)
+            return value
+
+        config_dict = {k: to_json_saveable(v) for k, v in config_dict.items()}
+        # Don't save "_ignore_files" or "_use_default_values"
+        config_dict.pop("_ignore_files", None)
+        config_dict.pop("_use_default_values", None)
+
+        return json.dumps(config_dict, indent=2, sort_keys=True) + "\n"
+
+    def to_json_file(self, json_file_path: Union[str, os.PathLike]):
+        """
+        Save the configuration instance's parameters to a JSON file.
+
+        Args:
+            json_file_path (`str` or `os.PathLike`):
+                Path to the JSON file to save a configuration instance's parameters.
+        """
+        with open(json_file_path, "w", encoding="utf-8") as writer:
+            writer.write(self.to_json_string())
+
+
+def register_to_config(init):
+    r"""
+    Decorator to apply on the init of classes inheriting from [`ConfigMixin`] so that all the arguments are
+    automatically sent to `self.register_for_config`. To ignore a specific argument accepted by the init but that
+    shouldn't be registered in the config, use the `ignore_for_config` class variable
+
+    Warning: Once decorated, all private arguments (beginning with an underscore) are trashed and not sent to the init!
+    """
+
+    @functools.wraps(init)
+    def inner_init(self, *args, **kwargs):
+        # Ignore private kwargs in the init.
+        init_kwargs = {k: v for k, v in kwargs.items() if not k.startswith("_")}
+        config_init_kwargs = {k: v for k, v in kwargs.items() if k.startswith("_")}
+        if not isinstance(self, ConfigMixin):
+            raise RuntimeError(
+                f"`@register_for_config` was applied to {self.__class__.__name__} init method, but this class does "
+                "not inherit from `ConfigMixin`."
+            )
+
+        ignore = getattr(self, "ignore_for_config", [])
+        # Get positional arguments aligned with kwargs
+        new_kwargs = {}
+        signature = inspect.signature(init)
+        parameters = {
+            name: p.default for i, (name, p) in enumerate(signature.parameters.items()) if i > 0 and name not in ignore
+        }
+        for arg, name in zip(args, parameters.keys()):
+            new_kwargs[name] = arg
+
+        # Then add all kwargs
+        new_kwargs.update(
+            {
+                k: init_kwargs.get(k, default)
+                for k, default in parameters.items()
+                if k not in ignore and k not in new_kwargs
+            }
+        )
+
+        # Take note of the parameters that were not present in the loaded config
+        if len(set(new_kwargs.keys()) - set(init_kwargs)) > 0:
+            new_kwargs["_use_default_values"] = list(set(new_kwargs.keys()) - set(init_kwargs))
+
+        new_kwargs = {**config_init_kwargs, **new_kwargs}
+        getattr(self, "register_to_config")(**new_kwargs)
+        init(self, *args, **init_kwargs)
+
+    return inner_init
+
+
+def flax_register_to_config(cls):
+    original_init = cls.__init__
+
+    @functools.wraps(original_init)
+    def init(self, *args, **kwargs):
+        if not isinstance(self, ConfigMixin):
+            raise RuntimeError(
+                f"`@register_for_config` was applied to {self.__class__.__name__} init method, but this class does "
+                "not inherit from `ConfigMixin`."
+            )
+
+        # Ignore private kwargs in the init. Retrieve all passed attributes
+        init_kwargs = dict(kwargs.items())
+
+        # Retrieve default values
+        fields = dataclasses.fields(self)
+        default_kwargs = {}
+        for field in fields:
+            # ignore flax specific attributes
+            if field.name in self._flax_internal_args:
+                continue
+            if type(field.default) == dataclasses._MISSING_TYPE:
+                default_kwargs[field.name] = None
+            else:
+                default_kwargs[field.name] = getattr(self, field.name)
+
+        # Make sure init_kwargs override default kwargs
+        new_kwargs = {**default_kwargs, **init_kwargs}
+        # dtype should be part of `init_kwargs`, but not `new_kwargs`
+        if "dtype" in new_kwargs:
+            new_kwargs.pop("dtype")
+
+        # Get positional arguments aligned with kwargs
+        for i, arg in enumerate(args):
+            name = fields[i].name
+            new_kwargs[name] = arg
+
+        # Take note of the parameters that were not present in the loaded config
+        if len(set(new_kwargs.keys()) - set(init_kwargs)) > 0:
+            new_kwargs["_use_default_values"] = list(set(new_kwargs.keys()) - set(init_kwargs))
+
+        getattr(self, "register_to_config")(**new_kwargs)
+        original_init(self, *args, **kwargs)
+
+    cls.__init__ = init
+    return cls

src/diffusers/dependency_versions_check.py

@@ -0,0 +1,34 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from .dependency_versions_table import deps
+from .utils.versions import require_version, require_version_core
+
+
+# define which module versions we always want to check at run time
+# (usually the ones defined in `install_requires` in setup.py)
+#
+# order specific notes:
+# - tqdm must be checked before tokenizers
+
+pkgs_to_check_at_runtime = "python requests filelock numpy".split()
+for pkg in pkgs_to_check_at_runtime:
+    if pkg in deps:
+        require_version_core(deps[pkg])
+    else:
+        raise ValueError(f"can't find {pkg} in {deps.keys()}, check dependency_versions_table.py")
+
+
+def dep_version_check(pkg, hint=None):
+    require_version(deps[pkg], hint)

src/diffusers/dependency_versions_table.py

@@ -0,0 +1,46 @@
+# THIS FILE HAS BEEN AUTOGENERATED. To update:
+# 1. modify the `_deps` dict in setup.py
+# 2. run `make deps_table_update`
+deps = {
+    "Pillow": "Pillow",
+    "accelerate": "accelerate>=0.11.0",
+    "compel": "compel==0.1.8",
+    "datasets": "datasets",
+    "filelock": "filelock",
+    "flax": "flax>=0.4.1",
+    "hf-doc-builder": "hf-doc-builder>=0.3.0",
+    "huggingface-hub": "huggingface-hub>=0.20.2",
+    "requests-mock": "requests-mock==1.10.0",
+    "importlib_metadata": "importlib_metadata",
+    "invisible-watermark": "invisible-watermark>=0.2.0",
+    "isort": "isort>=5.5.4",
+    "jax": "jax>=0.4.1",
+    "jaxlib": "jaxlib>=0.4.1",
+    "Jinja2": "Jinja2",
+    "k-diffusion": "k-diffusion>=0.0.12",
+    "torchsde": "torchsde",
+    "note_seq": "note_seq",
+    "librosa": "librosa",
+    "numpy": "numpy",
+    "omegaconf": "omegaconf",
+    "parameterized": "parameterized",
+    "peft": "peft>=0.6.0",
+    "protobuf": "protobuf>=3.20.3,<4",
+    "pytest": "pytest",
+    "pytest-timeout": "pytest-timeout",
+    "pytest-xdist": "pytest-xdist",
+    "python": "python>=3.8.0",
+    "ruff": "ruff==0.1.5",
+    "safetensors": "safetensors>=0.3.1",
+    "sentencepiece": "sentencepiece>=0.1.91,!=0.1.92",
+    "GitPython": "GitPython<3.1.19",
+    "scipy": "scipy",
+    "onnx": "onnx",
+    "regex": "regex!=2019.12.17",
+    "requests": "requests",
+    "tensorboard": "tensorboard",
+    "torch": "torch>=1.4",
+    "torchvision": "torchvision",
+    "transformers": "transformers>=4.25.1",
+    "urllib3": "urllib3<=2.0.0",
+}

src/diffusers/experimental/README.md

@@ -0,0 +1,5 @@
+# 🧨 Diffusers Experimental
+
+We are adding experimental code to support novel applications and usages of the Diffusers library.
+Currently, the following experiments are supported:
+* Reinforcement learning via an implementation of the [Diffuser](https://arxiv.org/abs/2205.09991) model.

src/diffusers/experimental/__init__.py

@@ -0,0 +1 @@
+from .rl import ValueGuidedRLPipeline

src/diffusers/experimental/rl/__init__.py

@@ -0,0 +1 @@
+from .value_guided_sampling import ValueGuidedRLPipeline

src/diffusers/experimental/rl/value_guided_sampling.py

@@ -0,0 +1,154 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import numpy as np
+import torch
+import tqdm
+
+from ...models.unet_1d import UNet1DModel
+from ...pipelines import DiffusionPipeline
+from ...utils.dummy_pt_objects import DDPMScheduler
+from ...utils.torch_utils import randn_tensor
+
+
+class ValueGuidedRLPipeline(DiffusionPipeline):
+    r"""
+    Pipeline for value-guided sampling from a diffusion model trained to predict sequences of states.
+
+    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+    implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+    Parameters:
+        value_function ([`UNet1DModel`]):
+            A specialized UNet for fine-tuning trajectories base on reward.
+        unet ([`UNet1DModel`]):
+            UNet architecture to denoise the encoded trajectories.
+        scheduler ([`SchedulerMixin`]):
+            A scheduler to be used in combination with `unet` to denoise the encoded trajectories. Default for this
+            application is [`DDPMScheduler`].
+        env ():
+            An environment following the OpenAI gym API to act in. For now only Hopper has pretrained models.
+    """
+
+    def __init__(
+        self,
+        value_function: UNet1DModel,
+        unet: UNet1DModel,
+        scheduler: DDPMScheduler,
+        env,
+    ):
+        super().__init__()
+        self.value_function = value_function
+        self.unet = unet
+        self.scheduler = scheduler
+        self.env = env
+        self.data = env.get_dataset()
+        self.means = {}
+        for key in self.data.keys():
+            try:
+                self.means[key] = self.data[key].mean()
+            except:  # noqa: E722
+                pass
+        self.stds = {}
+        for key in self.data.keys():
+            try:
+                self.stds[key] = self.data[key].std()
+            except:  # noqa: E722
+                pass
+        self.state_dim = env.observation_space.shape[0]
+        self.action_dim = env.action_space.shape[0]
+
+    def normalize(self, x_in, key):
+        return (x_in - self.means[key]) / self.stds[key]
+
+    def de_normalize(self, x_in, key):
+        return x_in * self.stds[key] + self.means[key]
+
+    def to_torch(self, x_in):
+        if isinstance(x_in, dict):
+            return {k: self.to_torch(v) for k, v in x_in.items()}
+        elif torch.is_tensor(x_in):
+            return x_in.to(self.unet.device)
+        return torch.tensor(x_in, device=self.unet.device)
+
+    def reset_x0(self, x_in, cond, act_dim):
+        for key, val in cond.items():
+            x_in[:, key, act_dim:] = val.clone()
+        return x_in
+
+    def run_diffusion(self, x, conditions, n_guide_steps, scale):
+        batch_size = x.shape[0]
+        y = None
+        for i in tqdm.tqdm(self.scheduler.timesteps):
+            # create batch of timesteps to pass into model
+            timesteps = torch.full((batch_size,), i, device=self.unet.device, dtype=torch.long)
+            for _ in range(n_guide_steps):
+                with torch.enable_grad():
+                    x.requires_grad_()
+
+                    # permute to match dimension for pre-trained models
+                    y = self.value_function(x.permute(0, 2, 1), timesteps).sample
+                    grad = torch.autograd.grad([y.sum()], [x])[0]
+
+                    posterior_variance = self.scheduler._get_variance(i)
+                    model_std = torch.exp(0.5 * posterior_variance)
+                    grad = model_std * grad
+
+                grad[timesteps < 2] = 0
+                x = x.detach()
+                x = x + scale * grad
+                x = self.reset_x0(x, conditions, self.action_dim)
+
+            prev_x = self.unet(x.permute(0, 2, 1), timesteps).sample.permute(0, 2, 1)
+
+            # TODO: verify deprecation of this kwarg
+            x = self.scheduler.step(prev_x, i, x)["prev_sample"]
+
+            # apply conditions to the trajectory (set the initial state)
+            x = self.reset_x0(x, conditions, self.action_dim)
+            x = self.to_torch(x)
+        return x, y
+
+    def __call__(self, obs, batch_size=64, planning_horizon=32, n_guide_steps=2, scale=0.1):
+        # normalize the observations and create  batch dimension
+        obs = self.normalize(obs, "observations")
+        obs = obs[None].repeat(batch_size, axis=0)
+
+        conditions = {0: self.to_torch(obs)}
+        shape = (batch_size, planning_horizon, self.state_dim + self.action_dim)
+
+        # generate initial noise and apply our conditions (to make the trajectories start at current state)
+        x1 = randn_tensor(shape, device=self.unet.device)
+        x = self.reset_x0(x1, conditions, self.action_dim)
+        x = self.to_torch(x)
+
+        # run the diffusion process
+        x, y = self.run_diffusion(x, conditions, n_guide_steps, scale)
+
+        # sort output trajectories by value
+        sorted_idx = y.argsort(0, descending=True).squeeze()
+        sorted_values = x[sorted_idx]
+        actions = sorted_values[:, :, : self.action_dim]
+        actions = actions.detach().cpu().numpy()
+        denorm_actions = self.de_normalize(actions, key="actions")
+
+        # select the action with the highest value
+        if y is not None:
+            selected_index = 0
+        else:
+            # if we didn't run value guiding, select a random action
+            selected_index = np.random.randint(0, batch_size)
+
+        denorm_actions = denorm_actions[selected_index, 0]
+        return denorm_actions

src/diffusers/image_processor.py

@@ -0,0 +1,888 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import warnings
+from typing import List, Optional, Tuple, Union
+
+import numpy as np
+import PIL.Image
+import torch
+from PIL import Image, ImageFilter, ImageOps
+
+from .configuration_utils import ConfigMixin, register_to_config
+from .utils import CONFIG_NAME, PIL_INTERPOLATION, deprecate
+
+
+PipelineImageInput = Union[
+    PIL.Image.Image,
+    np.ndarray,
+    torch.FloatTensor,
+    List[PIL.Image.Image],
+    List[np.ndarray],
+    List[torch.FloatTensor],
+]
+
+PipelineDepthInput = Union[
+    PIL.Image.Image,
+    np.ndarray,
+    torch.FloatTensor,
+    List[PIL.Image.Image],
+    List[np.ndarray],
+    List[torch.FloatTensor],
+]
+
+
+class VaeImageProcessor(ConfigMixin):
+    """
+    Image processor for VAE.
+
+    Args:
+        do_resize (`bool`, *optional*, defaults to `True`):
+            Whether to downscale the image's (height, width) dimensions to multiples of `vae_scale_factor`. Can accept
+            `height` and `width` arguments from [`image_processor.VaeImageProcessor.preprocess`] method.
+        vae_scale_factor (`int`, *optional*, defaults to `8`):
+            VAE scale factor. If `do_resize` is `True`, the image is automatically resized to multiples of this factor.
+        resample (`str`, *optional*, defaults to `lanczos`):
+            Resampling filter to use when resizing the image.
+        do_normalize (`bool`, *optional*, defaults to `True`):
+            Whether to normalize the image to [-1,1].
+        do_binarize (`bool`, *optional*, defaults to `False`):
+            Whether to binarize the image to 0/1.
+        do_convert_rgb (`bool`, *optional*, defaults to be `False`):
+            Whether to convert the images to RGB format.
+        do_convert_grayscale (`bool`, *optional*, defaults to be `False`):
+            Whether to convert the images to grayscale format.
+    """
+
+    config_name = CONFIG_NAME
+
+    @register_to_config
+    def __init__(
+        self,
+        do_resize: bool = True,
+        vae_scale_factor: int = 8,
+        resample: str = "lanczos",
+        do_normalize: bool = True,
+        do_binarize: bool = False,
+        do_convert_rgb: bool = False,
+        do_convert_grayscale: bool = False,
+    ):
+        super().__init__()
+        if do_convert_rgb and do_convert_grayscale:
+            raise ValueError(
+                "`do_convert_rgb` and `do_convert_grayscale` can not both be set to `True`,"
+                " if you intended to convert the image into RGB format, please set `do_convert_grayscale = False`.",
+                " if you intended to convert the image into grayscale format, please set `do_convert_rgb = False`",
+            )
+            self.config.do_convert_rgb = False
+
+    @staticmethod
+    def numpy_to_pil(images: np.ndarray) -> List[PIL.Image.Image]:
+        """
+        Convert a numpy image or a batch of images to a PIL image.
+        """
+        if images.ndim == 3:
+            images = images[None, ...]
+        images = (images * 255).round().astype("uint8")
+        if images.shape[-1] == 1:
+            # special case for grayscale (single channel) images
+            pil_images = [Image.fromarray(image.squeeze(), mode="L") for image in images]
+        else:
+            pil_images = [Image.fromarray(image) for image in images]
+
+        return pil_images
+
+    @staticmethod
+    def pil_to_numpy(images: Union[List[PIL.Image.Image], PIL.Image.Image]) -> np.ndarray:
+        """
+        Convert a PIL image or a list of PIL images to NumPy arrays.
+        """
+        if not isinstance(images, list):
+            images = [images]
+        images = [np.array(image).astype(np.float32) / 255.0 for image in images]
+        images = np.stack(images, axis=0)
+
+        return images
+
+    @staticmethod
+    def numpy_to_pt(images: np.ndarray) -> torch.FloatTensor:
+        """
+        Convert a NumPy image to a PyTorch tensor.
+        """
+        if images.ndim == 3:
+            images = images[..., None]
+
+        images = torch.from_numpy(images.transpose(0, 3, 1, 2))
+        return images
+
+    @staticmethod
+    def pt_to_numpy(images: torch.FloatTensor) -> np.ndarray:
+        """
+        Convert a PyTorch tensor to a NumPy image.
+        """
+        images = images.cpu().permute(0, 2, 3, 1).float().numpy()
+        return images
+
+    @staticmethod
+    def normalize(images: Union[np.ndarray, torch.Tensor]) -> Union[np.ndarray, torch.Tensor]:
+        """
+        Normalize an image array to [-1,1].
+        """
+        return 2.0 * images - 1.0
+
+    @staticmethod
+    def denormalize(images: Union[np.ndarray, torch.Tensor]) -> Union[np.ndarray, torch.Tensor]:
+        """
+        Denormalize an image array to [0,1].
+        """
+        return (images / 2 + 0.5).clamp(0, 1)
+
+    @staticmethod
+    def convert_to_rgb(image: PIL.Image.Image) -> PIL.Image.Image:
+        """
+        Converts a PIL image to RGB format.
+        """
+        image = image.convert("RGB")
+
+        return image
+
+    @staticmethod
+    def convert_to_grayscale(image: PIL.Image.Image) -> PIL.Image.Image:
+        """
+        Converts a PIL image to grayscale format.
+        """
+        image = image.convert("L")
+
+        return image
+
+    @staticmethod
+    def blur(image: PIL.Image.Image, blur_factor: int = 4) -> PIL.Image.Image:
+        """
+        Blurs an image.
+        """
+        image = image.filter(ImageFilter.GaussianBlur(blur_factor))
+
+        return image
+
+    @staticmethod
+    def get_crop_region(mask_image: PIL.Image.Image, width: int, height: int, pad=0):
+        """
+        Finds a rectangular region that contains all masked ares in an image, and expands region to match the aspect ratio of the original image;
+        for example, if user drew mask in a 128x32 region, and the dimensions for processing are 512x512, the region will be expanded to 128x128.
+
+        Args:
+            mask_image (PIL.Image.Image): Mask image.
+            width (int): Width of the image to be processed.
+            height (int): Height of the image to be processed.
+            pad (int, optional): Padding to be added to the crop region. Defaults to 0.
+
+        Returns:
+            tuple: (x1, y1, x2, y2) represent a rectangular region that contains all masked ares in an image and matches the original aspect ratio.
+        """
+
+        mask_image = mask_image.convert("L")
+        mask = np.array(mask_image)
+
+        # 1. find a rectangular region that contains all masked ares in an image
+        h, w = mask.shape
+        crop_left = 0
+        for i in range(w):
+            if not (mask[:, i] == 0).all():
+                break
+            crop_left += 1
+
+        crop_right = 0
+        for i in reversed(range(w)):
+            if not (mask[:, i] == 0).all():
+                break
+            crop_right += 1
+
+        crop_top = 0
+        for i in range(h):
+            if not (mask[i] == 0).all():
+                break
+            crop_top += 1
+
+        crop_bottom = 0
+        for i in reversed(range(h)):
+            if not (mask[i] == 0).all():
+                break
+            crop_bottom += 1
+
+        # 2. add padding to the crop region
+        x1, y1, x2, y2 = (
+            int(max(crop_left - pad, 0)),
+            int(max(crop_top - pad, 0)),
+            int(min(w - crop_right + pad, w)),
+            int(min(h - crop_bottom + pad, h)),
+        )
+
+        # 3. expands crop region to match the aspect ratio of the image to be processed
+        ratio_crop_region = (x2 - x1) / (y2 - y1)
+        ratio_processing = width / height
+
+        if ratio_crop_region > ratio_processing:
+            desired_height = (x2 - x1) / ratio_processing
+            desired_height_diff = int(desired_height - (y2 - y1))
+            y1 -= desired_height_diff // 2
+            y2 += desired_height_diff - desired_height_diff // 2
+            if y2 >= mask_image.height:
+                diff = y2 - mask_image.height
+                y2 -= diff
+                y1 -= diff
+            if y1 < 0:
+                y2 -= y1
+                y1 -= y1
+            if y2 >= mask_image.height:
+                y2 = mask_image.height
+        else:
+            desired_width = (y2 - y1) * ratio_processing
+            desired_width_diff = int(desired_width - (x2 - x1))
+            x1 -= desired_width_diff // 2
+            x2 += desired_width_diff - desired_width_diff // 2
+            if x2 >= mask_image.width:
+                diff = x2 - mask_image.width
+                x2 -= diff
+                x1 -= diff
+            if x1 < 0:
+                x2 -= x1
+                x1 -= x1
+            if x2 >= mask_image.width:
+                x2 = mask_image.width
+
+        return x1, y1, x2, y2
+
+    def _resize_and_fill(
+        self,
+        image: PIL.Image.Image,
+        width: int,
+        height: int,
+    ) -> PIL.Image.Image:
+        """
+        Resize the image to fit within the specified width and height, maintaining the aspect ratio, and then center the image within the dimensions, filling empty with data from image.
+
+        Args:
+            image: The image to resize.
+            width: The width to resize the image to.
+            height: The height to resize the image to.
+        """
+
+        ratio = width / height
+        src_ratio = image.width / image.height
+
+        src_w = width if ratio < src_ratio else image.width * height // image.height
+        src_h = height if ratio >= src_ratio else image.height * width // image.width
+
+        resized = image.resize((src_w, src_h), resample=PIL_INTERPOLATION["lanczos"])
+        res = Image.new("RGB", (width, height))
+        res.paste(resized, box=(width // 2 - src_w // 2, height // 2 - src_h // 2))
+
+        if ratio < src_ratio:
+            fill_height = height // 2 - src_h // 2
+            if fill_height > 0:
+                res.paste(resized.resize((width, fill_height), box=(0, 0, width, 0)), box=(0, 0))
+                res.paste(
+                    resized.resize((width, fill_height), box=(0, resized.height, width, resized.height)),
+                    box=(0, fill_height + src_h),
+                )
+        elif ratio > src_ratio:
+            fill_width = width // 2 - src_w // 2
+            if fill_width > 0:
+                res.paste(resized.resize((fill_width, height), box=(0, 0, 0, height)), box=(0, 0))
+                res.paste(
+                    resized.resize((fill_width, height), box=(resized.width, 0, resized.width, height)),
+                    box=(fill_width + src_w, 0),
+                )
+
+        return res
+
+    def _resize_and_crop(
+        self,
+        image: PIL.Image.Image,
+        width: int,
+        height: int,
+    ) -> PIL.Image.Image:
+        """
+        Resize the image to fit within the specified width and height, maintaining the aspect ratio, and then center the image within the dimensions, cropping the excess.
+
+        Args:
+            image: The image to resize.
+            width: The width to resize the image to.
+            height: The height to resize the image to.
+        """
+        ratio = width / height
+        src_ratio = image.width / image.height
+
+        src_w = width if ratio > src_ratio else image.width * height // image.height
+        src_h = height if ratio <= src_ratio else image.height * width // image.width
+
+        resized = image.resize((src_w, src_h), resample=PIL_INTERPOLATION["lanczos"])
+        res = Image.new("RGB", (width, height))
+        res.paste(resized, box=(width // 2 - src_w // 2, height // 2 - src_h // 2))
+        return res
+
+    def resize(
+        self,
+        image: Union[PIL.Image.Image, np.ndarray, torch.Tensor],
+        height: int,
+        width: int,
+        resize_mode: str = "default",  # "defalt", "fill", "crop"
+    ) -> Union[PIL.Image.Image, np.ndarray, torch.Tensor]:
+        """
+        Resize image.
+
+        Args:
+            image (`PIL.Image.Image`, `np.ndarray` or `torch.Tensor`):
+                The image input, can be a PIL image, numpy array or pytorch tensor.
+            height (`int`):
+                The height to resize to.
+            width (`int`):
+                The width to resize to.
+            resize_mode (`str`, *optional*, defaults to `default`):
+                The resize mode to use, can be one of `default` or `fill`. If `default`, will resize the image to fit
+                within the specified width and height, and it may not maintaining the original aspect ratio.
+                If `fill`, will resize the image to fit within the specified width and height, maintaining the aspect ratio, and then center the image
+                within the dimensions, filling empty with data from image.
+                If `crop`, will resize the image to fit within the specified width and height, maintaining the aspect ratio, and then center the image
+                within the dimensions, cropping the excess.
+                Note that resize_mode `fill` and `crop` are only supported for PIL image input.
+
+        Returns:
+            `PIL.Image.Image`, `np.ndarray` or `torch.Tensor`:
+                The resized image.
+        """
+        if resize_mode != "default" and not isinstance(image, PIL.Image.Image):
+            raise ValueError(f"Only PIL image input is supported for resize_mode {resize_mode}")
+        if isinstance(image, PIL.Image.Image):
+            if resize_mode == "default":
+                image = image.resize((width, height), resample=PIL_INTERPOLATION[self.config.resample])
+            elif resize_mode == "fill":
+                image = self._resize_and_fill(image, width, height)
+            elif resize_mode == "crop":
+                image = self._resize_and_crop(image, width, height)
+            else:
+                raise ValueError(f"resize_mode {resize_mode} is not supported")
+
+        elif isinstance(image, torch.Tensor):
+            image = torch.nn.functional.interpolate(
+                image,
+                size=(height, width),
+            )
+        elif isinstance(image, np.ndarray):
+            image = self.numpy_to_pt(image)
+            image = torch.nn.functional.interpolate(
+                image,
+                size=(height, width),
+            )
+            image = self.pt_to_numpy(image)
+        return image
+
+    def binarize(self, image: PIL.Image.Image) -> PIL.Image.Image:
+        """
+        Create a mask.
+
+        Args:
+            image (`PIL.Image.Image`):
+                The image input, should be a PIL image.
+
+        Returns:
+            `PIL.Image.Image`:
+                The binarized image. Values less than 0.5 are set to 0, values greater than 0.5 are set to 1.
+        """
+        image[image < 0.5] = 0
+        image[image >= 0.5] = 1
+        return image
+
+    def get_default_height_width(
+        self,
+        image: Union[PIL.Image.Image, np.ndarray, torch.Tensor],
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+    ) -> Tuple[int, int]:
+        """
+        This function return the height and width that are downscaled to the next integer multiple of
+        `vae_scale_factor`.
+
+        Args:
+            image(`PIL.Image.Image`, `np.ndarray` or `torch.Tensor`):
+                The image input, can be a PIL image, numpy array or pytorch tensor. if it is a numpy array, should have
+                shape `[batch, height, width]` or `[batch, height, width, channel]` if it is a pytorch tensor, should
+                have shape `[batch, channel, height, width]`.
+            height (`int`, *optional*, defaults to `None`):
+                The height in preprocessed image. If `None`, will use the height of `image` input.
+            width (`int`, *optional*`, defaults to `None`):
+                The width in preprocessed. If `None`, will use the width of the `image` input.
+        """
+
+        if height is None:
+            if isinstance(image, PIL.Image.Image):
+                height = image.height
+            elif isinstance(image, torch.Tensor):
+                height = image.shape[2]
+            else:
+                height = image.shape[1]
+
+        if width is None:
+            if isinstance(image, PIL.Image.Image):
+                width = image.width
+            elif isinstance(image, torch.Tensor):
+                width = image.shape[3]
+            else:
+                width = image.shape[2]
+
+        width, height = (
+            x - x % self.config.vae_scale_factor for x in (width, height)
+        )  # resize to integer multiple of vae_scale_factor
+
+        return height, width
+
+    def preprocess(
+        self,
+        image: PipelineImageInput,
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        resize_mode: str = "default",  # "defalt", "fill", "crop"
+        crops_coords: Optional[Tuple[int, int, int, int]] = None,
+    ) -> torch.Tensor:
+        """
+        Preprocess the image input.
+
+        Args:
+            image (`pipeline_image_input`):
+                The image input, accepted formats are PIL images, NumPy arrays, PyTorch tensors; Also accept list of supported formats.
+            height (`int`, *optional*, defaults to `None`):
+                The height in preprocessed image. If `None`, will use the `get_default_height_width()` to get default height.
+            width (`int`, *optional*`, defaults to `None`):
+                The width in preprocessed. If `None`, will use  get_default_height_width()` to get the default width.
+            resize_mode (`str`, *optional*, defaults to `default`):
+                The resize mode, can be one of `default` or `fill`. If `default`, will resize the image to fit
+                within the specified width and height, and it may not maintaining the original aspect ratio.
+                If `fill`, will resize the image to fit within the specified width and height, maintaining the aspect ratio, and then center the image
+                within the dimensions, filling empty with data from image.
+                If `crop`, will resize the image to fit within the specified width and height, maintaining the aspect ratio, and then center the image
+                within the dimensions, cropping the excess.
+                Note that resize_mode `fill` and `crop` are only supported for PIL image input.
+            crops_coords (`List[Tuple[int, int, int, int]]`, *optional*, defaults to `None`):
+                The crop coordinates for each image in the batch. If `None`, will not crop the image.
+        """
+        supported_formats = (PIL.Image.Image, np.ndarray, torch.Tensor)
+
+        # Expand the missing dimension for 3-dimensional pytorch tensor or numpy array that represents grayscale image
+        if self.config.do_convert_grayscale and isinstance(image, (torch.Tensor, np.ndarray)) and image.ndim == 3:
+            if isinstance(image, torch.Tensor):
+                # if image is a pytorch tensor could have 2 possible shapes:
+                #    1. batch x height x width: we should insert the channel dimension at position 1
+                #    2. channnel x height x width: we should insert batch dimension at position 0,
+                #       however, since both channel and batch dimension has same size 1, it is same to insert at position 1
+                #    for simplicity, we insert a dimension of size 1 at position 1 for both cases
+                image = image.unsqueeze(1)
+            else:
+                # if it is a numpy array, it could have 2 possible shapes:
+                #   1. batch x height x width: insert channel dimension on last position
+                #   2. height x width x channel: insert batch dimension on first position
+                if image.shape[-1] == 1:
+                    image = np.expand_dims(image, axis=0)
+                else:
+                    image = np.expand_dims(image, axis=-1)
+
+        if isinstance(image, supported_formats):
+            image = [image]
+        elif not (isinstance(image, list) and all(isinstance(i, supported_formats) for i in image)):
+            raise ValueError(
+                f"Input is in incorrect format: {[type(i) for i in image]}. Currently, we only support {', '.join(supported_formats)}"
+            )
+
+        if isinstance(image[0], PIL.Image.Image):
+            if crops_coords is not None:
+                image = [i.crop(crops_coords) for i in image]
+            if self.config.do_resize:
+                height, width = self.get_default_height_width(image[0], height, width)
+                image = [self.resize(i, height, width, resize_mode=resize_mode) for i in image]
+            if self.config.do_convert_rgb:
+                image = [self.convert_to_rgb(i) for i in image]
+            elif self.config.do_convert_grayscale:
+                image = [self.convert_to_grayscale(i) for i in image]
+            image = self.pil_to_numpy(image)  # to np
+            image = self.numpy_to_pt(image)  # to pt
+
+        elif isinstance(image[0], np.ndarray):
+            image = np.concatenate(image, axis=0) if image[0].ndim == 4 else np.stack(image, axis=0)
+
+            image = self.numpy_to_pt(image)
+
+            height, width = self.get_default_height_width(image, height, width)
+            if self.config.do_resize:
+                image = self.resize(image, height, width)
+
+        elif isinstance(image[0], torch.Tensor):
+            image = torch.cat(image, axis=0) if image[0].ndim == 4 else torch.stack(image, axis=0)
+
+            if self.config.do_convert_grayscale and image.ndim == 3:
+                image = image.unsqueeze(1)
+
+            channel = image.shape[1]
+            # don't need any preprocess if the image is latents
+            if channel == 4:
+                return image
+
+            height, width = self.get_default_height_width(image, height, width)
+            if self.config.do_resize:
+                image = self.resize(image, height, width)
+
+        # expected range [0,1], normalize to [-1,1]
+        do_normalize = self.config.do_normalize
+        if do_normalize and image.min() < 0:
+            warnings.warn(
+                "Passing `image` as torch tensor with value range in [-1,1] is deprecated. The expected value range for image tensor is [0,1] "
+                f"when passing as pytorch tensor or numpy Array. You passed `image` with value range [{image.min()},{image.max()}]",
+                FutureWarning,
+            )
+            do_normalize = False
+
+        if do_normalize:
+            image = self.normalize(image)
+
+        if self.config.do_binarize:
+            image = self.binarize(image)
+
+        return image
+
+    def postprocess(
+        self,
+        image: torch.FloatTensor,
+        output_type: str = "pil",
+        do_denormalize: Optional[List[bool]] = None,
+    ) -> Union[PIL.Image.Image, np.ndarray, torch.FloatTensor]:
+        """
+        Postprocess the image output from tensor to `output_type`.
+
+        Args:
+            image (`torch.FloatTensor`):
+                The image input, should be a pytorch tensor with shape `B x C x H x W`.
+            output_type (`str`, *optional*, defaults to `pil`):
+                The output type of the image, can be one of `pil`, `np`, `pt`, `latent`.
+            do_denormalize (`List[bool]`, *optional*, defaults to `None`):
+                Whether to denormalize the image to [0,1]. If `None`, will use the value of `do_normalize` in the
+                `VaeImageProcessor` config.
+
+        Returns:
+            `PIL.Image.Image`, `np.ndarray` or `torch.FloatTensor`:
+                The postprocessed image.
+        """
+        if not isinstance(image, torch.Tensor):
+            raise ValueError(
+                f"Input for postprocessing is in incorrect format: {type(image)}. We only support pytorch tensor"
+            )
+        if output_type not in ["latent", "pt", "np", "pil"]:
+            deprecation_message = (
+                f"the output_type {output_type} is outdated and has been set to `np`. Please make sure to set it to one of these instead: "
+                "`pil`, `np`, `pt`, `latent`"
+            )
+            deprecate("Unsupported output_type", "1.0.0", deprecation_message, standard_warn=False)
+            output_type = "np"
+
+        if output_type == "latent":
+            return image
+
+        if do_denormalize is None:
+            do_denormalize = [self.config.do_normalize] * image.shape[0]
+
+        image = torch.stack(
+            [self.denormalize(image[i]) if do_denormalize[i] else image[i] for i in range(image.shape[0])]
+        )
+
+        if output_type == "pt":
+            return image
+
+        image = self.pt_to_numpy(image)
+
+        if output_type == "np":
+            return image
+
+        if output_type == "pil":
+            return self.numpy_to_pil(image)
+
+    def apply_overlay(
+        self,
+        mask: PIL.Image.Image,
+        init_image: PIL.Image.Image,
+        image: PIL.Image.Image,
+        crop_coords: Optional[Tuple[int, int, int, int]] = None,
+    ) -> PIL.Image.Image:
+        """
+        overlay the inpaint output to the original image
+        """
+
+        width, height = image.width, image.height
+
+        init_image = self.resize(init_image, width=width, height=height)
+        mask = self.resize(mask, width=width, height=height)
+
+        init_image_masked = PIL.Image.new("RGBa", (width, height))
+        init_image_masked.paste(init_image.convert("RGBA").convert("RGBa"), mask=ImageOps.invert(mask.convert("L")))
+        init_image_masked = init_image_masked.convert("RGBA")
+
+        if crop_coords is not None:
+            x, y, w, h = crop_coords
+            base_image = PIL.Image.new("RGBA", (width, height))
+            image = self.resize(image, height=h, width=w, resize_mode="crop")
+            base_image.paste(image, (x, y))
+            image = base_image.convert("RGB")
+
+        image = image.convert("RGBA")
+        image.alpha_composite(init_image_masked)
+        image = image.convert("RGB")
+
+        return image
+
+
+class VaeImageProcessorLDM3D(VaeImageProcessor):
+    """
+    Image processor for VAE LDM3D.
+
+    Args:
+        do_resize (`bool`, *optional*, defaults to `True`):
+            Whether to downscale the image's (height, width) dimensions to multiples of `vae_scale_factor`.
+        vae_scale_factor (`int`, *optional*, defaults to `8`):
+            VAE scale factor. If `do_resize` is `True`, the image is automatically resized to multiples of this factor.
+        resample (`str`, *optional*, defaults to `lanczos`):
+            Resampling filter to use when resizing the image.
+        do_normalize (`bool`, *optional*, defaults to `True`):
+            Whether to normalize the image to [-1,1].
+    """
+
+    config_name = CONFIG_NAME
+
+    @register_to_config
+    def __init__(
+        self,
+        do_resize: bool = True,
+        vae_scale_factor: int = 8,
+        resample: str = "lanczos",
+        do_normalize: bool = True,
+    ):
+        super().__init__()
+
+    @staticmethod
+    def numpy_to_pil(images: np.ndarray) -> List[PIL.Image.Image]:
+        """
+        Convert a NumPy image or a batch of images to a PIL image.
+        """
+        if images.ndim == 3:
+            images = images[None, ...]
+        images = (images * 255).round().astype("uint8")
+        if images.shape[-1] == 1:
+            # special case for grayscale (single channel) images
+            pil_images = [Image.fromarray(image.squeeze(), mode="L") for image in images]
+        else:
+            pil_images = [Image.fromarray(image[:, :, :3]) for image in images]
+
+        return pil_images
+
+    @staticmethod
+    def depth_pil_to_numpy(images: Union[List[PIL.Image.Image], PIL.Image.Image]) -> np.ndarray:
+        """
+        Convert a PIL image or a list of PIL images to NumPy arrays.
+        """
+        if not isinstance(images, list):
+            images = [images]
+
+        images = [np.array(image).astype(np.float32) / (2**16 - 1) for image in images]
+        images = np.stack(images, axis=0)
+        return images
+
+    @staticmethod
+    def rgblike_to_depthmap(image: Union[np.ndarray, torch.Tensor]) -> Union[np.ndarray, torch.Tensor]:
+        """
+        Args:
+            image: RGB-like depth image
+
+        Returns: depth map
+
+        """
+        return image[:, :, 1] * 2**8 + image[:, :, 2]
+
+    def numpy_to_depth(self, images: np.ndarray) -> List[PIL.Image.Image]:
+        """
+        Convert a NumPy depth image or a batch of images to a PIL image.
+        """
+        if images.ndim == 3:
+            images = images[None, ...]
+        images_depth = images[:, :, :, 3:]
+        if images.shape[-1] == 6:
+            images_depth = (images_depth * 255).round().astype("uint8")
+            pil_images = [
+                Image.fromarray(self.rgblike_to_depthmap(image_depth), mode="I;16") for image_depth in images_depth
+            ]
+        elif images.shape[-1] == 4:
+            images_depth = (images_depth * 65535.0).astype(np.uint16)
+            pil_images = [Image.fromarray(image_depth, mode="I;16") for image_depth in images_depth]
+        else:
+            raise Exception("Not supported")
+
+        return pil_images
+
+    def postprocess(
+        self,
+        image: torch.FloatTensor,
+        output_type: str = "pil",
+        do_denormalize: Optional[List[bool]] = None,
+    ) -> Union[PIL.Image.Image, np.ndarray, torch.FloatTensor]:
+        """
+        Postprocess the image output from tensor to `output_type`.
+
+        Args:
+            image (`torch.FloatTensor`):
+                The image input, should be a pytorch tensor with shape `B x C x H x W`.
+            output_type (`str`, *optional*, defaults to `pil`):
+                The output type of the image, can be one of `pil`, `np`, `pt`, `latent`.
+            do_denormalize (`List[bool]`, *optional*, defaults to `None`):
+                Whether to denormalize the image to [0,1]. If `None`, will use the value of `do_normalize` in the
+                `VaeImageProcessor` config.
+
+        Returns:
+            `PIL.Image.Image`, `np.ndarray` or `torch.FloatTensor`:
+                The postprocessed image.
+        """
+        if not isinstance(image, torch.Tensor):
+            raise ValueError(
+                f"Input for postprocessing is in incorrect format: {type(image)}. We only support pytorch tensor"
+            )
+        if output_type not in ["latent", "pt", "np", "pil"]:
+            deprecation_message = (
+                f"the output_type {output_type} is outdated and has been set to `np`. Please make sure to set it to one of these instead: "
+                "`pil`, `np`, `pt`, `latent`"
+            )
+            deprecate("Unsupported output_type", "1.0.0", deprecation_message, standard_warn=False)
+            output_type = "np"
+
+        if do_denormalize is None:
+            do_denormalize = [self.config.do_normalize] * image.shape[0]
+
+        image = torch.stack(
+            [self.denormalize(image[i]) if do_denormalize[i] else image[i] for i in range(image.shape[0])]
+        )
+
+        image = self.pt_to_numpy(image)
+
+        if output_type == "np":
+            if image.shape[-1] == 6:
+                image_depth = np.stack([self.rgblike_to_depthmap(im[:, :, 3:]) for im in image], axis=0)
+            else:
+                image_depth = image[:, :, :, 3:]
+            return image[:, :, :, :3], image_depth
+
+        if output_type == "pil":
+            return self.numpy_to_pil(image), self.numpy_to_depth(image)
+        else:
+            raise Exception(f"This type {output_type} is not supported")
+
+    def preprocess(
+        self,
+        rgb: Union[torch.FloatTensor, PIL.Image.Image, np.ndarray],
+        depth: Union[torch.FloatTensor, PIL.Image.Image, np.ndarray],
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        target_res: Optional[int] = None,
+    ) -> torch.Tensor:
+        """
+        Preprocess the image input. Accepted formats are PIL images, NumPy arrays or PyTorch tensors.
+        """
+        supported_formats = (PIL.Image.Image, np.ndarray, torch.Tensor)
+
+        # Expand the missing dimension for 3-dimensional pytorch tensor or numpy array that represents grayscale image
+        if self.config.do_convert_grayscale and isinstance(rgb, (torch.Tensor, np.ndarray)) and rgb.ndim == 3:
+            raise Exception("This is not yet supported")
+
+        if isinstance(rgb, supported_formats):
+            rgb = [rgb]
+            depth = [depth]
+        elif not (isinstance(rgb, list) and all(isinstance(i, supported_formats) for i in rgb)):
+            raise ValueError(
+                f"Input is in incorrect format: {[type(i) for i in rgb]}. Currently, we only support {', '.join(supported_formats)}"
+            )
+
+        if isinstance(rgb[0], PIL.Image.Image):
+            if self.config.do_convert_rgb:
+                raise Exception("This is not yet supported")
+                # rgb = [self.convert_to_rgb(i) for i in rgb]
+                # depth = [self.convert_to_depth(i) for i in depth]  #TODO define convert_to_depth
+            if self.config.do_resize or target_res:
+                height, width = self.get_default_height_width(rgb[0], height, width) if not target_res else target_res
+                rgb = [self.resize(i, height, width) for i in rgb]
+                depth = [self.resize(i, height, width) for i in depth]
+            rgb = self.pil_to_numpy(rgb)  # to np
+            rgb = self.numpy_to_pt(rgb)  # to pt
+
+            depth = self.depth_pil_to_numpy(depth)  # to np
+            depth = self.numpy_to_pt(depth)  # to pt
+
+        elif isinstance(rgb[0], np.ndarray):
+            rgb = np.concatenate(rgb, axis=0) if rgb[0].ndim == 4 else np.stack(rgb, axis=0)
+            rgb = self.numpy_to_pt(rgb)
+            height, width = self.get_default_height_width(rgb, height, width)
+            if self.config.do_resize:
+                rgb = self.resize(rgb, height, width)
+
+            depth = np.concatenate(depth, axis=0) if rgb[0].ndim == 4 else np.stack(depth, axis=0)
+            depth = self.numpy_to_pt(depth)
+            height, width = self.get_default_height_width(depth, height, width)
+            if self.config.do_resize:
+                depth = self.resize(depth, height, width)
+
+        elif isinstance(rgb[0], torch.Tensor):
+            raise Exception("This is not yet supported")
+            # rgb = torch.cat(rgb, axis=0) if rgb[0].ndim == 4 else torch.stack(rgb, axis=0)
+
+            # if self.config.do_convert_grayscale and rgb.ndim == 3:
+            #     rgb = rgb.unsqueeze(1)
+
+            # channel = rgb.shape[1]
+
+            # height, width = self.get_default_height_width(rgb, height, width)
+            # if self.config.do_resize:
+            #     rgb = self.resize(rgb, height, width)
+
+            # depth = torch.cat(depth, axis=0) if depth[0].ndim == 4 else torch.stack(depth, axis=0)
+
+            # if self.config.do_convert_grayscale and depth.ndim == 3:
+            #     depth = depth.unsqueeze(1)
+
+            # channel = depth.shape[1]
+            # # don't need any preprocess if the image is latents
+            # if depth == 4:
+            #     return rgb, depth
+
+            # height, width = self.get_default_height_width(depth, height, width)
+            # if self.config.do_resize:
+            #     depth = self.resize(depth, height, width)
+        # expected range [0,1], normalize to [-1,1]
+        do_normalize = self.config.do_normalize
+        if rgb.min() < 0 and do_normalize:
+            warnings.warn(
+                "Passing `image` as torch tensor with value range in [-1,1] is deprecated. The expected value range for image tensor is [0,1] "
+                f"when passing as pytorch tensor or numpy Array. You passed `image` with value range [{rgb.min()},{rgb.max()}]",
+                FutureWarning,
+            )
+            do_normalize = False
+
+        if do_normalize:
+            rgb = self.normalize(rgb)
+            depth = self.normalize(depth)
+
+        if self.config.do_binarize:
+            rgb = self.binarize(rgb)
+            depth = self.binarize(depth)
+
+        return rgb, depth

src/diffusers/loaders/__init__.py

@@ -0,0 +1,82 @@
+from typing import TYPE_CHECKING
+
+from ..utils import DIFFUSERS_SLOW_IMPORT, _LazyModule, deprecate
+from ..utils.import_utils import is_torch_available, is_transformers_available
+
+
+def text_encoder_lora_state_dict(text_encoder):
+    deprecate(
+        "text_encoder_load_state_dict in `models`",
+        "0.27.0",
+        "`text_encoder_lora_state_dict` is deprecated and will be removed in 0.27.0. Make sure to retrieve the weights using `get_peft_model`. See https://huggingface.co/docs/peft/v0.6.2/en/quicktour#peftmodel for more information.",
+    )
+    state_dict = {}
+
+    for name, module in text_encoder_attn_modules(text_encoder):
+        for k, v in module.q_proj.lora_linear_layer.state_dict().items():
+            state_dict[f"{name}.q_proj.lora_linear_layer.{k}"] = v
+
+        for k, v in module.k_proj.lora_linear_layer.state_dict().items():
+            state_dict[f"{name}.k_proj.lora_linear_layer.{k}"] = v
+
+        for k, v in module.v_proj.lora_linear_layer.state_dict().items():
+            state_dict[f"{name}.v_proj.lora_linear_layer.{k}"] = v
+
+        for k, v in module.out_proj.lora_linear_layer.state_dict().items():
+            state_dict[f"{name}.out_proj.lora_linear_layer.{k}"] = v
+
+    return state_dict
+
+
+if is_transformers_available():
+
+    def text_encoder_attn_modules(text_encoder):
+        deprecate(
+            "text_encoder_attn_modules in `models`",
+            "0.27.0",
+            "`text_encoder_lora_state_dict` is deprecated and will be removed in 0.27.0. Make sure to retrieve the weights using `get_peft_model`. See https://huggingface.co/docs/peft/v0.6.2/en/quicktour#peftmodel for more information.",
+        )
+        from transformers import CLIPTextModel, CLIPTextModelWithProjection
+
+        attn_modules = []
+
+        if isinstance(text_encoder, (CLIPTextModel, CLIPTextModelWithProjection)):
+            for i, layer in enumerate(text_encoder.text_model.encoder.layers):
+                name = f"text_model.encoder.layers.{i}.self_attn"
+                mod = layer.self_attn
+                attn_modules.append((name, mod))
+        else:
+            raise ValueError(f"do not know how to get attention modules for: {text_encoder.__class__.__name__}")
+
+        return attn_modules
+
+
+_import_structure = {}
+
+if is_torch_available():
+    _import_structure["single_file"] = ["FromOriginalControlnetMixin", "FromOriginalVAEMixin"]
+    _import_structure["unet"] = ["UNet2DConditionLoadersMixin"]
+    _import_structure["utils"] = ["AttnProcsLayers"]
+
+    if is_transformers_available():
+        _import_structure["single_file"].extend(["FromSingleFileMixin"])
+        _import_structure["lora"] = ["LoraLoaderMixin", "StableDiffusionXLLoraLoaderMixin"]
+        _import_structure["textual_inversion"] = ["TextualInversionLoaderMixin"]
+        _import_structure["ip_adapter"] = ["IPAdapterMixin"]
+
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+    if is_torch_available():
+        from .single_file import FromOriginalControlnetMixin, FromOriginalVAEMixin
+        from .unet import UNet2DConditionLoadersMixin
+        from .utils import AttnProcsLayers
+
+        if is_transformers_available():
+            from .ip_adapter import IPAdapterMixin
+            from .lora import LoraLoaderMixin, StableDiffusionXLLoraLoaderMixin
+            from .single_file import FromSingleFileMixin
+            from .textual_inversion import TextualInversionLoaderMixin
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)

src/diffusers/loaders/ip_adapter.py

@@ -0,0 +1,159 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import os
+from typing import Dict, Union
+
+import torch
+from huggingface_hub.utils import validate_hf_hub_args
+from safetensors import safe_open
+
+from ..utils import (
+    _get_model_file,
+    is_transformers_available,
+    logging,
+)
+
+
+if is_transformers_available():
+    from transformers import (
+        CLIPImageProcessor,
+        CLIPVisionModelWithProjection,
+    )
+
+    from ..models.attention_processor import (
+        IPAdapterAttnProcessor,
+        IPAdapterAttnProcessor2_0,
+    )
+
+logger = logging.get_logger(__name__)
+
+
+class IPAdapterMixin:
+    """Mixin for handling IP Adapters."""
+
+    @validate_hf_hub_args
+    def load_ip_adapter(
+        self,
+        pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]],
+        subfolder: str,
+        weight_name: str,
+        **kwargs,
+    ):
+        """
+        Parameters:
+            pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
+                Can be either:
+
+                    - A string, the *model id* (for example `google/ddpm-celebahq-256`) of a pretrained model hosted on
+                      the Hub.
+                    - A path to a *directory* (for example `./my_model_directory`) containing the model weights saved
+                      with [`ModelMixin.save_pretrained`].
+                    - A [torch state
+                      dict](https://pytorch.org/tutorials/beginner/saving_loading_models.html#what-is-a-state-dict).
+
+            cache_dir (`Union[str, os.PathLike]`, *optional*):
+                Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
+                is not used.
+            force_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to force the (re-)download of the model weights and configuration files, overriding the
+                cached versions if they exist.
+            resume_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to resume downloading the model weights and configuration files. If set to `False`, any
+                incompletely downloaded files are deleted.
+            proxies (`Dict[str, str]`, *optional*):
+                A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
+                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
+            local_files_only (`bool`, *optional*, defaults to `False`):
+                Whether to only load local model weights and configuration files or not. If set to `True`, the model
+                won't be downloaded from the Hub.
+            token (`str` or *bool*, *optional*):
+                The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
+                `diffusers-cli login` (stored in `~/.huggingface`) is used.
+            revision (`str`, *optional*, defaults to `"main"`):
+                The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
+                allowed by Git.
+            subfolder (`str`, *optional*, defaults to `""`):
+                The subfolder location of a model file within a larger model repository on the Hub or locally.
+        """
+
+        # Load the main state dict first.
+        cache_dir = kwargs.pop("cache_dir", None)
+        force_download = kwargs.pop("force_download", False)
+        resume_download = kwargs.pop("resume_download", False)
+        proxies = kwargs.pop("proxies", None)
+        local_files_only = kwargs.pop("local_files_only", None)
+        token = kwargs.pop("token", None)
+        revision = kwargs.pop("revision", None)
+
+        user_agent = {
+            "file_type": "attn_procs_weights",
+            "framework": "pytorch",
+        }
+
+        if not isinstance(pretrained_model_name_or_path_or_dict, dict):
+            model_file = _get_model_file(
+                pretrained_model_name_or_path_or_dict,
+                weights_name=weight_name,
+                cache_dir=cache_dir,
+                force_download=force_download,
+                resume_download=resume_download,
+                proxies=proxies,
+                local_files_only=local_files_only,
+                token=token,
+                revision=revision,
+                subfolder=subfolder,
+                user_agent=user_agent,
+            )
+            if weight_name.endswith(".safetensors"):
+                state_dict = {"image_proj": {}, "ip_adapter": {}}
+                with safe_open(model_file, framework="pt", device="cpu") as f:
+                    for key in f.keys():
+                        if key.startswith("image_proj."):
+                            state_dict["image_proj"][key.replace("image_proj.", "")] = f.get_tensor(key)
+                        elif key.startswith("ip_adapter."):
+                            state_dict["ip_adapter"][key.replace("ip_adapter.", "")] = f.get_tensor(key)
+            else:
+                state_dict = torch.load(model_file, map_location="cpu")
+        else:
+            state_dict = pretrained_model_name_or_path_or_dict
+
+        keys = list(state_dict.keys())
+        if keys != ["image_proj", "ip_adapter"]:
+            raise ValueError("Required keys are (`image_proj` and `ip_adapter`) missing from the state dict.")
+
+        # load CLIP image encoer here if it has not been registered to the pipeline yet
+        if hasattr(self, "image_encoder") and getattr(self, "image_encoder", None) is None:
+            if not isinstance(pretrained_model_name_or_path_or_dict, dict):
+                logger.info(f"loading image_encoder from {pretrained_model_name_or_path_or_dict}")
+                image_encoder = CLIPVisionModelWithProjection.from_pretrained(
+                    pretrained_model_name_or_path_or_dict,
+                    subfolder=os.path.join(subfolder, "image_encoder"),
+                ).to(self.device, dtype=self.dtype)
+                self.image_encoder = image_encoder
+            else:
+                raise ValueError("`image_encoder` cannot be None when using IP Adapters.")
+
+        # create feature extractor if it has not been registered to the pipeline yet
+        if hasattr(self, "feature_extractor") and getattr(self, "feature_extractor", None) is None:
+            self.feature_extractor = CLIPImageProcessor()
+
+        # load ip-adapter into unet
+        unet = getattr(self, self.unet_name) if not hasattr(self, "unet") else self.unet
+        unet._load_ip_adapter_weights(state_dict)
+
+    def set_ip_adapter_scale(self, scale):
+        unet = getattr(self, self.unet_name) if not hasattr(self, "unet") else self.unet
+        for attn_processor in unet.attn_processors.values():
+            if isinstance(attn_processor, (IPAdapterAttnProcessor, IPAdapterAttnProcessor2_0)):
+                attn_processor.scale = scale

src/diffusers/loaders/lora.py

@@ -0,0 +1,1553 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import inspect
+import os
+from contextlib import nullcontext
+from typing import Callable, Dict, List, Optional, Union
+
+import safetensors
+import torch
+from huggingface_hub import model_info
+from huggingface_hub.constants import HF_HUB_OFFLINE
+from huggingface_hub.utils import validate_hf_hub_args
+from packaging import version
+from torch import nn
+
+from .. import __version__
+from ..models.modeling_utils import _LOW_CPU_MEM_USAGE_DEFAULT, load_model_dict_into_meta
+from ..utils import (
+    USE_PEFT_BACKEND,
+    _get_model_file,
+    convert_state_dict_to_diffusers,
+    convert_state_dict_to_peft,
+    convert_unet_state_dict_to_peft,
+    delete_adapter_layers,
+    deprecate,
+    get_adapter_name,
+    get_peft_kwargs,
+    is_accelerate_available,
+    is_transformers_available,
+    logging,
+    recurse_remove_peft_layers,
+    scale_lora_layers,
+    set_adapter_layers,
+    set_weights_and_activate_adapters,
+)
+from .lora_conversion_utils import _convert_kohya_lora_to_diffusers, _maybe_map_sgm_blocks_to_diffusers
+
+
+if is_transformers_available():
+    from transformers import PreTrainedModel
+
+    from ..models.lora import PatchedLoraProjection, text_encoder_attn_modules, text_encoder_mlp_modules
+
+if is_accelerate_available():
+    from accelerate import init_empty_weights
+    from accelerate.hooks import AlignDevicesHook, CpuOffload, remove_hook_from_module
+
+logger = logging.get_logger(__name__)
+
+TEXT_ENCODER_NAME = "text_encoder"
+UNET_NAME = "unet"
+TRANSFORMER_NAME = "transformer"
+
+LORA_WEIGHT_NAME = "pytorch_lora_weights.bin"
+LORA_WEIGHT_NAME_SAFE = "pytorch_lora_weights.safetensors"
+
+LORA_DEPRECATION_MESSAGE = "You are using an old version of LoRA backend. This will be deprecated in the next releases in favor of PEFT make sure to install the latest PEFT and transformers packages in the future."
+
+
+class LoraLoaderMixin:
+    r"""
+    Load LoRA layers into [`UNet2DConditionModel`] and
+    [`CLIPTextModel`](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel).
+    """
+
+    text_encoder_name = TEXT_ENCODER_NAME
+    unet_name = UNET_NAME
+    transformer_name = TRANSFORMER_NAME
+    num_fused_loras = 0
+
+    def load_lora_weights(
+        self, pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]], adapter_name=None, **kwargs
+    ):
+        """
+        Load LoRA weights specified in `pretrained_model_name_or_path_or_dict` into `self.unet` and
+        `self.text_encoder`.
+
+        All kwargs are forwarded to `self.lora_state_dict`.
+
+        See [`~loaders.LoraLoaderMixin.lora_state_dict`] for more details on how the state dict is loaded.
+
+        See [`~loaders.LoraLoaderMixin.load_lora_into_unet`] for more details on how the state dict is loaded into
+        `self.unet`.
+
+        See [`~loaders.LoraLoaderMixin.load_lora_into_text_encoder`] for more details on how the state dict is loaded
+        into `self.text_encoder`.
+
+        Parameters:
+            pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
+                See [`~loaders.LoraLoaderMixin.lora_state_dict`].
+            kwargs (`dict`, *optional*):
+                See [`~loaders.LoraLoaderMixin.lora_state_dict`].
+            adapter_name (`str`, *optional*):
+                Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
+                `default_{i}` where i is the total number of adapters being loaded.
+        """
+        # First, ensure that the checkpoint is a compatible one and can be successfully loaded.
+        state_dict, network_alphas = self.lora_state_dict(pretrained_model_name_or_path_or_dict, **kwargs)
+
+        is_correct_format = all("lora" in key for key in state_dict.keys())
+        if not is_correct_format:
+            raise ValueError("Invalid LoRA checkpoint.")
+
+        low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", _LOW_CPU_MEM_USAGE_DEFAULT)
+
+        self.load_lora_into_unet(
+            state_dict,
+            network_alphas=network_alphas,
+            unet=getattr(self, self.unet_name) if not hasattr(self, "unet") else self.unet,
+            low_cpu_mem_usage=low_cpu_mem_usage,
+            adapter_name=adapter_name,
+            _pipeline=self,
+        )
+        self.load_lora_into_text_encoder(
+            state_dict,
+            network_alphas=network_alphas,
+            text_encoder=getattr(self, self.text_encoder_name)
+            if not hasattr(self, "text_encoder")
+            else self.text_encoder,
+            lora_scale=self.lora_scale,
+            low_cpu_mem_usage=low_cpu_mem_usage,
+            adapter_name=adapter_name,
+            _pipeline=self,
+        )
+
+    @classmethod
+    @validate_hf_hub_args
+    def lora_state_dict(
+        cls,
+        pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]],
+        **kwargs,
+    ):
+        r"""
+        Return state dict for lora weights and the network alphas.
+
+        <Tip warning={true}>
+
+        We support loading A1111 formatted LoRA checkpoints in a limited capacity.
+
+        This function is experimental and might change in the future.
+
+        </Tip>
+
+        Parameters:
+            pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
+                Can be either:
+
+                    - A string, the *model id* (for example `google/ddpm-celebahq-256`) of a pretrained model hosted on
+                      the Hub.
+                    - A path to a *directory* (for example `./my_model_directory`) containing the model weights saved
+                      with [`ModelMixin.save_pretrained`].
+                    - A [torch state
+                      dict](https://pytorch.org/tutorials/beginner/saving_loading_models.html#what-is-a-state-dict).
+
+            cache_dir (`Union[str, os.PathLike]`, *optional*):
+                Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
+                is not used.
+            force_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to force the (re-)download of the model weights and configuration files, overriding the
+                cached versions if they exist.
+            resume_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to resume downloading the model weights and configuration files. If set to `False`, any
+                incompletely downloaded files are deleted.
+            proxies (`Dict[str, str]`, *optional*):
+                A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
+                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
+            local_files_only (`bool`, *optional*, defaults to `False`):
+                Whether to only load local model weights and configuration files or not. If set to `True`, the model
+                won't be downloaded from the Hub.
+            token (`str` or *bool*, *optional*):
+                The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
+                `diffusers-cli login` (stored in `~/.huggingface`) is used.
+            revision (`str`, *optional*, defaults to `"main"`):
+                The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
+                allowed by Git.
+            subfolder (`str`, *optional*, defaults to `""`):
+                The subfolder location of a model file within a larger model repository on the Hub or locally.
+            low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 else `False`):
+                Speed up model loading only loading the pretrained weights and not initializing the weights. This also
+                tries to not use more than 1x model size in CPU memory (including peak memory) while loading the model.
+                Only supported for PyTorch >= 1.9.0. If you are using an older version of PyTorch, setting this
+                argument to `True` will raise an error.
+            mirror (`str`, *optional*):
+                Mirror source to resolve accessibility issues if you're downloading a model in China. We do not
+                guarantee the timeliness or safety of the source, and you should refer to the mirror site for more
+                information.
+
+        """
+        # Load the main state dict first which has the LoRA layers for either of
+        # UNet and text encoder or both.
+        cache_dir = kwargs.pop("cache_dir", None)
+        force_download = kwargs.pop("force_download", False)
+        resume_download = kwargs.pop("resume_download", False)
+        proxies = kwargs.pop("proxies", None)
+        local_files_only = kwargs.pop("local_files_only", None)
+        token = kwargs.pop("token", None)
+        revision = kwargs.pop("revision", None)
+        subfolder = kwargs.pop("subfolder", None)
+        weight_name = kwargs.pop("weight_name", None)
+        unet_config = kwargs.pop("unet_config", None)
+        use_safetensors = kwargs.pop("use_safetensors", None)
+
+        allow_pickle = False
+        if use_safetensors is None:
+            use_safetensors = True
+            allow_pickle = True
+
+        user_agent = {
+            "file_type": "attn_procs_weights",
+            "framework": "pytorch",
+        }
+
+        model_file = None
+        if not isinstance(pretrained_model_name_or_path_or_dict, dict):
+            # Let's first try to load .safetensors weights
+            if (use_safetensors and weight_name is None) or (
+                weight_name is not None and weight_name.endswith(".safetensors")
+            ):
+                try:
+                    # Here we're relaxing the loading check to enable more Inference API
+                    # friendliness where sometimes, it's not at all possible to automatically
+                    # determine `weight_name`.
+                    if weight_name is None:
+                        weight_name = cls._best_guess_weight_name(
+                            pretrained_model_name_or_path_or_dict,
+                            file_extension=".safetensors",
+                            local_files_only=local_files_only,
+                        )
+                    model_file = _get_model_file(
+                        pretrained_model_name_or_path_or_dict,
+                        weights_name=weight_name or LORA_WEIGHT_NAME_SAFE,
+                        cache_dir=cache_dir,
+                        force_download=force_download,
+                        resume_download=resume_download,
+                        proxies=proxies,
+                        local_files_only=local_files_only,
+                        token=token,
+                        revision=revision,
+                        subfolder=subfolder,
+                        user_agent=user_agent,
+                    )
+                    state_dict = safetensors.torch.load_file(model_file, device="cpu")
+                except (IOError, safetensors.SafetensorError) as e:
+                    if not allow_pickle:
+                        raise e
+                    # try loading non-safetensors weights
+                    model_file = None
+                    pass
+
+            if model_file is None:
+                if weight_name is None:
+                    weight_name = cls._best_guess_weight_name(
+                        pretrained_model_name_or_path_or_dict, file_extension=".bin", local_files_only=local_files_only
+                    )
+                model_file = _get_model_file(
+                    pretrained_model_name_or_path_or_dict,
+                    weights_name=weight_name or LORA_WEIGHT_NAME,
+                    cache_dir=cache_dir,
+                    force_download=force_download,
+                    resume_download=resume_download,
+                    proxies=proxies,
+                    local_files_only=local_files_only,
+                    token=token,
+                    revision=revision,
+                    subfolder=subfolder,
+                    user_agent=user_agent,
+                )
+                state_dict = torch.load(model_file, map_location="cpu")
+        else:
+            state_dict = pretrained_model_name_or_path_or_dict
+
+        network_alphas = None
+        # TODO: replace it with a method from `state_dict_utils`
+        if all(
+            (
+                k.startswith("lora_te_")
+                or k.startswith("lora_unet_")
+                or k.startswith("lora_te1_")
+                or k.startswith("lora_te2_")
+            )
+            for k in state_dict.keys()
+        ):
+            # Map SDXL blocks correctly.
+            if unet_config is not None:
+                # use unet config to remap block numbers
+                state_dict = _maybe_map_sgm_blocks_to_diffusers(state_dict, unet_config)
+            state_dict, network_alphas = _convert_kohya_lora_to_diffusers(state_dict)
+
+        return state_dict, network_alphas
+
+    @classmethod
+    def _best_guess_weight_name(
+        cls, pretrained_model_name_or_path_or_dict, file_extension=".safetensors", local_files_only=False
+    ):
+        if local_files_only or HF_HUB_OFFLINE:
+            raise ValueError("When using the offline mode, you must specify a `weight_name`.")
+
+        targeted_files = []
+
+        if os.path.isfile(pretrained_model_name_or_path_or_dict):
+            return
+        elif os.path.isdir(pretrained_model_name_or_path_or_dict):
+            targeted_files = [
+                f for f in os.listdir(pretrained_model_name_or_path_or_dict) if f.endswith(file_extension)
+            ]
+        else:
+            files_in_repo = model_info(pretrained_model_name_or_path_or_dict).siblings
+            targeted_files = [f.rfilename for f in files_in_repo if f.rfilename.endswith(file_extension)]
+        if len(targeted_files) == 0:
+            return
+
+        # "scheduler" does not correspond to a LoRA checkpoint.
+        # "optimizer" does not correspond to a LoRA checkpoint
+        # only top-level checkpoints are considered and not the other ones, hence "checkpoint".
+        unallowed_substrings = {"scheduler", "optimizer", "checkpoint"}
+        targeted_files = list(
+            filter(lambda x: all(substring not in x for substring in unallowed_substrings), targeted_files)
+        )
+
+        if any(f.endswith(LORA_WEIGHT_NAME) for f in targeted_files):
+            targeted_files = list(filter(lambda x: x.endswith(LORA_WEIGHT_NAME), targeted_files))
+        elif any(f.endswith(LORA_WEIGHT_NAME_SAFE) for f in targeted_files):
+            targeted_files = list(filter(lambda x: x.endswith(LORA_WEIGHT_NAME_SAFE), targeted_files))
+
+        if len(targeted_files) > 1:
+            raise ValueError(
+                f"Provided path contains more than one weights file in the {file_extension} format. Either specify `weight_name` in `load_lora_weights` or make sure there's only one  `.safetensors` or `.bin` file in  {pretrained_model_name_or_path_or_dict}."
+            )
+        weight_name = targeted_files[0]
+        return weight_name
+
+    @classmethod
+    def _optionally_disable_offloading(cls, _pipeline):
+        """
+        Optionally removes offloading in case the pipeline has been already sequentially offloaded to CPU.
+
+        Args:
+            _pipeline (`DiffusionPipeline`):
+                The pipeline to disable offloading for.
+
+        Returns:
+            tuple:
+                A tuple indicating if `is_model_cpu_offload` or `is_sequential_cpu_offload` is True.
+        """
+        is_model_cpu_offload = False
+        is_sequential_cpu_offload = False
+
+        if _pipeline is not None:
+            for _, component in _pipeline.components.items():
+                if isinstance(component, nn.Module) and hasattr(component, "_hf_hook"):
+                    if not is_model_cpu_offload:
+                        is_model_cpu_offload = isinstance(component._hf_hook, CpuOffload)
+                    if not is_sequential_cpu_offload:
+                        is_sequential_cpu_offload = isinstance(component._hf_hook, AlignDevicesHook)
+
+                    logger.info(
+                        "Accelerate hooks detected. Since you have called `load_lora_weights()`, the previous hooks will be first removed. Then the LoRA parameters will be loaded and the hooks will be applied again."
+                    )
+                    remove_hook_from_module(component, recurse=is_sequential_cpu_offload)
+
+        return (is_model_cpu_offload, is_sequential_cpu_offload)
+
+    @classmethod
+    def load_lora_into_unet(
+        cls, state_dict, network_alphas, unet, low_cpu_mem_usage=None, adapter_name=None, _pipeline=None
+    ):
+        """
+        This will load the LoRA layers specified in `state_dict` into `unet`.
+
+        Parameters:
+            state_dict (`dict`):
+                A standard state dict containing the lora layer parameters. The keys can either be indexed directly
+                into the unet or prefixed with an additional `unet` which can be used to distinguish between text
+                encoder lora layers.
+            network_alphas (`Dict[str, float]`):
+                See `LoRALinearLayer` for more details.
+            unet (`UNet2DConditionModel`):
+                The UNet model to load the LoRA layers into.
+            low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 else `False`):
+                Speed up model loading only loading the pretrained weights and not initializing the weights. This also
+                tries to not use more than 1x model size in CPU memory (including peak memory) while loading the model.
+                Only supported for PyTorch >= 1.9.0. If you are using an older version of PyTorch, setting this
+                argument to `True` will raise an error.
+            adapter_name (`str`, *optional*):
+                Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
+                `default_{i}` where i is the total number of adapters being loaded.
+        """
+        low_cpu_mem_usage = low_cpu_mem_usage if low_cpu_mem_usage is not None else _LOW_CPU_MEM_USAGE_DEFAULT
+        # If the serialization format is new (introduced in https://github.com/huggingface/diffusers/pull/2918),
+        # then the `state_dict` keys should have `cls.unet_name` and/or `cls.text_encoder_name` as
+        # their prefixes.
+        keys = list(state_dict.keys())
+
+        if all(key.startswith("unet.unet") for key in keys):
+            deprecation_message = "Keys starting with 'unet.unet' are deprecated."
+            deprecate("unet.unet keys", "0.27", deprecation_message)
+
+        if all(key.startswith(cls.unet_name) or key.startswith(cls.text_encoder_name) for key in keys):
+            # Load the layers corresponding to UNet.
+            logger.info(f"Loading {cls.unet_name}.")
+
+            unet_keys = [k for k in keys if k.startswith(cls.unet_name)]
+            state_dict = {k.replace(f"{cls.unet_name}.", ""): v for k, v in state_dict.items() if k in unet_keys}
+
+            if network_alphas is not None:
+                alpha_keys = [k for k in network_alphas.keys() if k.startswith(cls.unet_name)]
+                network_alphas = {
+                    k.replace(f"{cls.unet_name}.", ""): v for k, v in network_alphas.items() if k in alpha_keys
+                }
+
+        else:
+            # Otherwise, we're dealing with the old format. This means the `state_dict` should only
+            # contain the module names of the `unet` as its keys WITHOUT any prefix.
+            if not USE_PEFT_BACKEND:
+                warn_message = "You have saved the LoRA weights using the old format. To convert the old LoRA weights to the new format, you can first load them in a dictionary and then create a new dictionary like the following: `new_state_dict = {f'unet.{module_name}': params for module_name, params in old_state_dict.items()}`."
+                logger.warn(warn_message)
+
+        if USE_PEFT_BACKEND and len(state_dict.keys()) > 0:
+            from peft import LoraConfig, inject_adapter_in_model, set_peft_model_state_dict
+
+            if adapter_name in getattr(unet, "peft_config", {}):
+                raise ValueError(
+                    f"Adapter name {adapter_name} already in use in the Unet - please select a new adapter name."
+                )
+
+            state_dict = convert_unet_state_dict_to_peft(state_dict)
+
+            if network_alphas is not None:
+                # The alphas state dict have the same structure as Unet, thus we convert it to peft format using
+                # `convert_unet_state_dict_to_peft` method.
+                network_alphas = convert_unet_state_dict_to_peft(network_alphas)
+
+            rank = {}
+            for key, val in state_dict.items():
+                if "lora_B" in key:
+                    rank[key] = val.shape[1]
+
+            lora_config_kwargs = get_peft_kwargs(rank, network_alphas, state_dict, is_unet=True)
+            lora_config = LoraConfig(**lora_config_kwargs)
+
+            # adapter_name
+            if adapter_name is None:
+                adapter_name = get_adapter_name(unet)
+
+            # In case the pipeline has been already offloaded to CPU - temporarily remove the hooks
+            # otherwise loading LoRA weights will lead to an error
+            is_model_cpu_offload, is_sequential_cpu_offload = cls._optionally_disable_offloading(_pipeline)
+
+            inject_adapter_in_model(lora_config, unet, adapter_name=adapter_name)
+            incompatible_keys = set_peft_model_state_dict(unet, state_dict, adapter_name)
+
+            if incompatible_keys is not None:
+                # check only for unexpected keys
+                unexpected_keys = getattr(incompatible_keys, "unexpected_keys", None)
+                if unexpected_keys:
+                    logger.warning(
+                        f"Loading adapter weights from state_dict led to unexpected keys not found in the model: "
+                        f" {unexpected_keys}. "
+                    )
+
+            # Offload back.
+            if is_model_cpu_offload:
+                _pipeline.enable_model_cpu_offload()
+            elif is_sequential_cpu_offload:
+                _pipeline.enable_sequential_cpu_offload()
+            # Unsafe code />
+
+        unet.load_attn_procs(
+            state_dict, network_alphas=network_alphas, low_cpu_mem_usage=low_cpu_mem_usage, _pipeline=_pipeline
+        )
+
+    @classmethod
+    def load_lora_into_text_encoder(
+        cls,
+        state_dict,
+        network_alphas,
+        text_encoder,
+        prefix=None,
+        lora_scale=1.0,
+        low_cpu_mem_usage=None,
+        adapter_name=None,
+        _pipeline=None,
+    ):
+        """
+        This will load the LoRA layers specified in `state_dict` into `text_encoder`
+
+        Parameters:
+            state_dict (`dict`):
+                A standard state dict containing the lora layer parameters. The key should be prefixed with an
+                additional `text_encoder` to distinguish between unet lora layers.
+            network_alphas (`Dict[str, float]`):
+                See `LoRALinearLayer` for more details.
+            text_encoder (`CLIPTextModel`):
+                The text encoder model to load the LoRA layers into.
+            prefix (`str`):
+                Expected prefix of the `text_encoder` in the `state_dict`.
+            lora_scale (`float`):
+                How much to scale the output of the lora linear layer before it is added with the output of the regular
+                lora layer.
+            low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 else `False`):
+                Speed up model loading only loading the pretrained weights and not initializing the weights. This also
+                tries to not use more than 1x model size in CPU memory (including peak memory) while loading the model.
+                Only supported for PyTorch >= 1.9.0. If you are using an older version of PyTorch, setting this
+                argument to `True` will raise an error.
+            adapter_name (`str`, *optional*):
+                Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
+                `default_{i}` where i is the total number of adapters being loaded.
+        """
+        low_cpu_mem_usage = low_cpu_mem_usage if low_cpu_mem_usage is not None else _LOW_CPU_MEM_USAGE_DEFAULT
+
+        # If the serialization format is new (introduced in https://github.com/huggingface/diffusers/pull/2918),
+        # then the `state_dict` keys should have `self.unet_name` and/or `self.text_encoder_name` as
+        # their prefixes.
+        keys = list(state_dict.keys())
+        prefix = cls.text_encoder_name if prefix is None else prefix
+
+        # Safe prefix to check with.
+        if any(cls.text_encoder_name in key for key in keys):
+            # Load the layers corresponding to text encoder and make necessary adjustments.
+            text_encoder_keys = [k for k in keys if k.startswith(prefix) and k.split(".")[0] == prefix]
+            text_encoder_lora_state_dict = {
+                k.replace(f"{prefix}.", ""): v for k, v in state_dict.items() if k in text_encoder_keys
+            }
+
+            if len(text_encoder_lora_state_dict) > 0:
+                logger.info(f"Loading {prefix}.")
+                rank = {}
+                text_encoder_lora_state_dict = convert_state_dict_to_diffusers(text_encoder_lora_state_dict)
+
+                if USE_PEFT_BACKEND:
+                    # convert state dict
+                    text_encoder_lora_state_dict = convert_state_dict_to_peft(text_encoder_lora_state_dict)
+
+                    for name, _ in text_encoder_attn_modules(text_encoder):
+                        rank_key = f"{name}.out_proj.lora_B.weight"
+                        rank[rank_key] = text_encoder_lora_state_dict[rank_key].shape[1]
+
+                    patch_mlp = any(".mlp." in key for key in text_encoder_lora_state_dict.keys())
+                    if patch_mlp:
+                        for name, _ in text_encoder_mlp_modules(text_encoder):
+                            rank_key_fc1 = f"{name}.fc1.lora_B.weight"
+                            rank_key_fc2 = f"{name}.fc2.lora_B.weight"
+
+                            rank[rank_key_fc1] = text_encoder_lora_state_dict[rank_key_fc1].shape[1]
+                            rank[rank_key_fc2] = text_encoder_lora_state_dict[rank_key_fc2].shape[1]
+                else:
+                    for name, _ in text_encoder_attn_modules(text_encoder):
+                        rank_key = f"{name}.out_proj.lora_linear_layer.up.weight"
+                        rank.update({rank_key: text_encoder_lora_state_dict[rank_key].shape[1]})
+
+                    patch_mlp = any(".mlp." in key for key in text_encoder_lora_state_dict.keys())
+                    if patch_mlp:
+                        for name, _ in text_encoder_mlp_modules(text_encoder):
+                            rank_key_fc1 = f"{name}.fc1.lora_linear_layer.up.weight"
+                            rank_key_fc2 = f"{name}.fc2.lora_linear_layer.up.weight"
+                            rank[rank_key_fc1] = text_encoder_lora_state_dict[rank_key_fc1].shape[1]
+                            rank[rank_key_fc2] = text_encoder_lora_state_dict[rank_key_fc2].shape[1]
+
+                if network_alphas is not None:
+                    alpha_keys = [
+                        k for k in network_alphas.keys() if k.startswith(prefix) and k.split(".")[0] == prefix
+                    ]
+                    network_alphas = {
+                        k.replace(f"{prefix}.", ""): v for k, v in network_alphas.items() if k in alpha_keys
+                    }
+
+                if USE_PEFT_BACKEND:
+                    from peft import LoraConfig
+
+                    lora_config_kwargs = get_peft_kwargs(
+                        rank, network_alphas, text_encoder_lora_state_dict, is_unet=False
+                    )
+
+                    lora_config = LoraConfig(**lora_config_kwargs)
+
+                    # adapter_name
+                    if adapter_name is None:
+                        adapter_name = get_adapter_name(text_encoder)
+
+                    is_model_cpu_offload, is_sequential_cpu_offload = cls._optionally_disable_offloading(_pipeline)
+
+                    # inject LoRA layers and load the state dict
+                    # in transformers we automatically check whether the adapter name is already in use or not
+                    text_encoder.load_adapter(
+                        adapter_name=adapter_name,
+                        adapter_state_dict=text_encoder_lora_state_dict,
+                        peft_config=lora_config,
+                    )
+
+                    # scale LoRA layers with `lora_scale`
+                    scale_lora_layers(text_encoder, weight=lora_scale)
+                else:
+                    cls._modify_text_encoder(
+                        text_encoder,
+                        lora_scale,
+                        network_alphas,
+                        rank=rank,
+                        patch_mlp=patch_mlp,
+                        low_cpu_mem_usage=low_cpu_mem_usage,
+                    )
+
+                    is_pipeline_offloaded = _pipeline is not None and any(
+                        isinstance(c, torch.nn.Module) and hasattr(c, "_hf_hook")
+                        for c in _pipeline.components.values()
+                    )
+                    if is_pipeline_offloaded and low_cpu_mem_usage:
+                        low_cpu_mem_usage = True
+                        logger.info(
+                            f"Pipeline {_pipeline.__class__} is offloaded. Therefore low cpu mem usage loading is forced."
+                        )
+
+                    if low_cpu_mem_usage:
+                        device = next(iter(text_encoder_lora_state_dict.values())).device
+                        dtype = next(iter(text_encoder_lora_state_dict.values())).dtype
+                        unexpected_keys = load_model_dict_into_meta(
+                            text_encoder, text_encoder_lora_state_dict, device=device, dtype=dtype
+                        )
+                    else:
+                        load_state_dict_results = text_encoder.load_state_dict(
+                            text_encoder_lora_state_dict, strict=False
+                        )
+                        unexpected_keys = load_state_dict_results.unexpected_keys
+
+                    if len(unexpected_keys) != 0:
+                        raise ValueError(
+                            f"failed to load text encoder state dict, unexpected keys: {load_state_dict_results.unexpected_keys}"
+                        )
+
+                    # <Unsafe code
+                    # We can be sure that the following works as all we do is change the dtype and device of the text encoder
+                    # Now we remove any existing hooks to
+                    is_model_cpu_offload = False
+                    is_sequential_cpu_offload = False
+                    if _pipeline is not None:
+                        for _, component in _pipeline.components.items():
+                            if isinstance(component, torch.nn.Module):
+                                if hasattr(component, "_hf_hook"):
+                                    is_model_cpu_offload = isinstance(getattr(component, "_hf_hook"), CpuOffload)
+                                    is_sequential_cpu_offload = isinstance(
+                                        getattr(component, "_hf_hook"), AlignDevicesHook
+                                    )
+                                    logger.info(
+                                        "Accelerate hooks detected. Since you have called `load_lora_weights()`, the previous hooks will be first removed. Then the LoRA parameters will be loaded and the hooks will be applied again."
+                                    )
+                                    remove_hook_from_module(component, recurse=is_sequential_cpu_offload)
+
+                text_encoder.to(device=text_encoder.device, dtype=text_encoder.dtype)
+
+                # Offload back.
+                if is_model_cpu_offload:
+                    _pipeline.enable_model_cpu_offload()
+                elif is_sequential_cpu_offload:
+                    _pipeline.enable_sequential_cpu_offload()
+                # Unsafe code />
+
+    @classmethod
+    def load_lora_into_transformer(
+        cls, state_dict, network_alphas, transformer, low_cpu_mem_usage=None, adapter_name=None, _pipeline=None
+    ):
+        """
+        This will load the LoRA layers specified in `state_dict` into `transformer`.
+
+        Parameters:
+            state_dict (`dict`):
+                A standard state dict containing the lora layer parameters. The keys can either be indexed directly
+                into the unet or prefixed with an additional `unet` which can be used to distinguish between text
+                encoder lora layers.
+            network_alphas (`Dict[str, float]`):
+                See `LoRALinearLayer` for more details.
+            unet (`UNet2DConditionModel`):
+                The UNet model to load the LoRA layers into.
+            low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 else `False`):
+                Speed up model loading only loading the pretrained weights and not initializing the weights. This also
+                tries to not use more than 1x model size in CPU memory (including peak memory) while loading the model.
+                Only supported for PyTorch >= 1.9.0. If you are using an older version of PyTorch, setting this
+                argument to `True` will raise an error.
+            adapter_name (`str`, *optional*):
+                Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
+                `default_{i}` where i is the total number of adapters being loaded.
+        """
+        low_cpu_mem_usage = low_cpu_mem_usage if low_cpu_mem_usage is not None else _LOW_CPU_MEM_USAGE_DEFAULT
+
+        keys = list(state_dict.keys())
+
+        transformer_keys = [k for k in keys if k.startswith(cls.transformer_name)]
+        state_dict = {
+            k.replace(f"{cls.transformer_name}.", ""): v for k, v in state_dict.items() if k in transformer_keys
+        }
+
+        if network_alphas is not None:
+            alpha_keys = [k for k in network_alphas.keys() if k.startswith(cls.transformer_name)]
+            network_alphas = {
+                k.replace(f"{cls.transformer_name}.", ""): v for k, v in network_alphas.items() if k in alpha_keys
+            }
+
+        if len(state_dict.keys()) > 0:
+            from peft import LoraConfig, inject_adapter_in_model, set_peft_model_state_dict
+
+            if adapter_name in getattr(transformer, "peft_config", {}):
+                raise ValueError(
+                    f"Adapter name {adapter_name} already in use in the transformer - please select a new adapter name."
+                )
+
+            rank = {}
+            for key, val in state_dict.items():
+                if "lora_B" in key:
+                    rank[key] = val.shape[1]
+
+            lora_config_kwargs = get_peft_kwargs(rank, network_alphas, state_dict)
+            lora_config = LoraConfig(**lora_config_kwargs)
+
+            # adapter_name
+            if adapter_name is None:
+                adapter_name = get_adapter_name(transformer)
+
+            # In case the pipeline has been already offloaded to CPU - temporarily remove the hooks
+            # otherwise loading LoRA weights will lead to an error
+            is_model_cpu_offload, is_sequential_cpu_offload = cls._optionally_disable_offloading(_pipeline)
+
+            inject_adapter_in_model(lora_config, transformer, adapter_name=adapter_name)
+            incompatible_keys = set_peft_model_state_dict(transformer, state_dict, adapter_name)
+
+            if incompatible_keys is not None:
+                # check only for unexpected keys
+                unexpected_keys = getattr(incompatible_keys, "unexpected_keys", None)
+                if unexpected_keys:
+                    logger.warning(
+                        f"Loading adapter weights from state_dict led to unexpected keys not found in the model: "
+                        f" {unexpected_keys}. "
+                    )
+
+            # Offload back.
+            if is_model_cpu_offload:
+                _pipeline.enable_model_cpu_offload()
+            elif is_sequential_cpu_offload:
+                _pipeline.enable_sequential_cpu_offload()
+            # Unsafe code />
+
+    @property
+    def lora_scale(self) -> float:
+        # property function that returns the lora scale which can be set at run time by the pipeline.
+        # if _lora_scale has not been set, return 1
+        return self._lora_scale if hasattr(self, "_lora_scale") else 1.0
+
+    def _remove_text_encoder_monkey_patch(self):
+        if USE_PEFT_BACKEND:
+            remove_method = recurse_remove_peft_layers
+        else:
+            remove_method = self._remove_text_encoder_monkey_patch_classmethod
+
+        if hasattr(self, "text_encoder"):
+            remove_method(self.text_encoder)
+
+            # In case text encoder have no Lora attached
+            if USE_PEFT_BACKEND and getattr(self.text_encoder, "peft_config", None) is not None:
+                del self.text_encoder.peft_config
+                self.text_encoder._hf_peft_config_loaded = None
+        if hasattr(self, "text_encoder_2"):
+            remove_method(self.text_encoder_2)
+            if USE_PEFT_BACKEND:
+                del self.text_encoder_2.peft_config
+                self.text_encoder_2._hf_peft_config_loaded = None
+
+    @classmethod
+    def _remove_text_encoder_monkey_patch_classmethod(cls, text_encoder):
+        deprecate("_remove_text_encoder_monkey_patch_classmethod", "0.27", LORA_DEPRECATION_MESSAGE)
+
+        for _, attn_module in text_encoder_attn_modules(text_encoder):
+            if isinstance(attn_module.q_proj, PatchedLoraProjection):
+                attn_module.q_proj.lora_linear_layer = None
+                attn_module.k_proj.lora_linear_layer = None
+                attn_module.v_proj.lora_linear_layer = None
+                attn_module.out_proj.lora_linear_layer = None
+
+        for _, mlp_module in text_encoder_mlp_modules(text_encoder):
+            if isinstance(mlp_module.fc1, PatchedLoraProjection):
+                mlp_module.fc1.lora_linear_layer = None
+                mlp_module.fc2.lora_linear_layer = None
+
+    @classmethod
+    def _modify_text_encoder(
+        cls,
+        text_encoder,
+        lora_scale=1,
+        network_alphas=None,
+        rank: Union[Dict[str, int], int] = 4,
+        dtype=None,
+        patch_mlp=False,
+        low_cpu_mem_usage=False,
+    ):
+        r"""
+        Monkey-patches the forward passes of attention modules of the text encoder.
+        """
+        deprecate("_modify_text_encoder", "0.27", LORA_DEPRECATION_MESSAGE)
+
+        def create_patched_linear_lora(model, network_alpha, rank, dtype, lora_parameters):
+            linear_layer = model.regular_linear_layer if isinstance(model, PatchedLoraProjection) else model
+            ctx = init_empty_weights if low_cpu_mem_usage else nullcontext
+            with ctx():
+                model = PatchedLoraProjection(linear_layer, lora_scale, network_alpha, rank, dtype=dtype)
+
+            lora_parameters.extend(model.lora_linear_layer.parameters())
+            return model
+
+        # First, remove any monkey-patch that might have been applied before
+        cls._remove_text_encoder_monkey_patch_classmethod(text_encoder)
+
+        lora_parameters = []
+        network_alphas = {} if network_alphas is None else network_alphas
+        is_network_alphas_populated = len(network_alphas) > 0
+
+        for name, attn_module in text_encoder_attn_modules(text_encoder):
+            query_alpha = network_alphas.pop(name + ".to_q_lora.down.weight.alpha", None)
+            key_alpha = network_alphas.pop(name + ".to_k_lora.down.weight.alpha", None)
+            value_alpha = network_alphas.pop(name + ".to_v_lora.down.weight.alpha", None)
+            out_alpha = network_alphas.pop(name + ".to_out_lora.down.weight.alpha", None)
+
+            if isinstance(rank, dict):
+                current_rank = rank.pop(f"{name}.out_proj.lora_linear_layer.up.weight")
+            else:
+                current_rank = rank
+
+            attn_module.q_proj = create_patched_linear_lora(
+                attn_module.q_proj, query_alpha, current_rank, dtype, lora_parameters
+            )
+            attn_module.k_proj = create_patched_linear_lora(
+                attn_module.k_proj, key_alpha, current_rank, dtype, lora_parameters
+            )
+            attn_module.v_proj = create_patched_linear_lora(
+                attn_module.v_proj, value_alpha, current_rank, dtype, lora_parameters
+            )
+            attn_module.out_proj = create_patched_linear_lora(
+                attn_module.out_proj, out_alpha, current_rank, dtype, lora_parameters
+            )
+
+        if patch_mlp:
+            for name, mlp_module in text_encoder_mlp_modules(text_encoder):
+                fc1_alpha = network_alphas.pop(name + ".fc1.lora_linear_layer.down.weight.alpha", None)
+                fc2_alpha = network_alphas.pop(name + ".fc2.lora_linear_layer.down.weight.alpha", None)
+
+                current_rank_fc1 = rank.pop(f"{name}.fc1.lora_linear_layer.up.weight")
+                current_rank_fc2 = rank.pop(f"{name}.fc2.lora_linear_layer.up.weight")
+
+                mlp_module.fc1 = create_patched_linear_lora(
+                    mlp_module.fc1, fc1_alpha, current_rank_fc1, dtype, lora_parameters
+                )
+                mlp_module.fc2 = create_patched_linear_lora(
+                    mlp_module.fc2, fc2_alpha, current_rank_fc2, dtype, lora_parameters
+                )
+
+        if is_network_alphas_populated and len(network_alphas) > 0:
+            raise ValueError(
+                f"The `network_alphas` has to be empty at this point but has the following keys \n\n {', '.join(network_alphas.keys())}"
+            )
+
+        return lora_parameters
+
+    @classmethod
+    def save_lora_weights(
+        cls,
+        save_directory: Union[str, os.PathLike],
+        unet_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None,
+        text_encoder_lora_layers: Dict[str, torch.nn.Module] = None,
+        transformer_lora_layers: Dict[str, torch.nn.Module] = None,
+        is_main_process: bool = True,
+        weight_name: str = None,
+        save_function: Callable = None,
+        safe_serialization: bool = True,
+    ):
+        r"""
+        Save the LoRA parameters corresponding to the UNet and text encoder.
+
+        Arguments:
+            save_directory (`str` or `os.PathLike`):
+                Directory to save LoRA parameters to. Will be created if it doesn't exist.
+            unet_lora_layers (`Dict[str, torch.nn.Module]` or `Dict[str, torch.Tensor]`):
+                State dict of the LoRA layers corresponding to the `unet`.
+            text_encoder_lora_layers (`Dict[str, torch.nn.Module]` or `Dict[str, torch.Tensor]`):
+                State dict of the LoRA layers corresponding to the `text_encoder`. Must explicitly pass the text
+                encoder LoRA state dict because it comes from 🤗 Transformers.
+            is_main_process (`bool`, *optional*, defaults to `True`):
+                Whether the process calling this is the main process or not. Useful during distributed training and you
+                need to call this function on all processes. In this case, set `is_main_process=True` only on the main
+                process to avoid race conditions.
+            save_function (`Callable`):
+                The function to use to save the state dictionary. Useful during distributed training when you need to
+                replace `torch.save` with another method. Can be configured with the environment variable
+                `DIFFUSERS_SAVE_MODE`.
+            safe_serialization (`bool`, *optional*, defaults to `True`):
+                Whether to save the model using `safetensors` or the traditional PyTorch way with `pickle`.
+        """
+        state_dict = {}
+
+        def pack_weights(layers, prefix):
+            layers_weights = layers.state_dict() if isinstance(layers, torch.nn.Module) else layers
+            layers_state_dict = {f"{prefix}.{module_name}": param for module_name, param in layers_weights.items()}
+            return layers_state_dict
+
+        if not (unet_lora_layers or text_encoder_lora_layers or transformer_lora_layers):
+            raise ValueError(
+                "You must pass at least one of `unet_lora_layers`, `text_encoder_lora_layers`, or `transformer_lora_layers`."
+            )
+
+        if unet_lora_layers:
+            state_dict.update(pack_weights(unet_lora_layers, cls.unet_name))
+
+        if text_encoder_lora_layers:
+            state_dict.update(pack_weights(text_encoder_lora_layers, cls.text_encoder_name))
+
+        if transformer_lora_layers:
+            state_dict.update(pack_weights(transformer_lora_layers, "transformer"))
+
+        # Save the model
+        cls.write_lora_layers(
+            state_dict=state_dict,
+            save_directory=save_directory,
+            is_main_process=is_main_process,
+            weight_name=weight_name,
+            save_function=save_function,
+            safe_serialization=safe_serialization,
+        )
+
+    @staticmethod
+    def write_lora_layers(
+        state_dict: Dict[str, torch.Tensor],
+        save_directory: str,
+        is_main_process: bool,
+        weight_name: str,
+        save_function: Callable,
+        safe_serialization: bool,
+    ):
+        if os.path.isfile(save_directory):
+            logger.error(f"Provided path ({save_directory}) should be a directory, not a file")
+            return
+
+        if save_function is None:
+            if safe_serialization:
+
+                def save_function(weights, filename):
+                    return safetensors.torch.save_file(weights, filename, metadata={"format": "pt"})
+
+            else:
+                save_function = torch.save
+
+        os.makedirs(save_directory, exist_ok=True)
+
+        if weight_name is None:
+            if safe_serialization:
+                weight_name = LORA_WEIGHT_NAME_SAFE
+            else:
+                weight_name = LORA_WEIGHT_NAME
+
+        save_function(state_dict, os.path.join(save_directory, weight_name))
+        logger.info(f"Model weights saved in {os.path.join(save_directory, weight_name)}")
+
+    def unload_lora_weights(self):
+        """
+        Unloads the LoRA parameters.
+
+        Examples:
+
+        ```python
+        >>> # Assuming `pipeline` is already loaded with the LoRA parameters.
+        >>> pipeline.unload_lora_weights()
+        >>> ...
+        ```
+        """
+        unet = getattr(self, self.unet_name) if not hasattr(self, "unet") else self.unet
+
+        if not USE_PEFT_BACKEND:
+            if version.parse(__version__) > version.parse("0.23"):
+                logger.warn(
+                    "You are using `unload_lora_weights` to disable and unload lora weights. If you want to iteratively enable and disable adapter weights,"
+                    "you can use `pipe.enable_lora()` or `pipe.disable_lora()`. After installing the latest version of PEFT."
+                )
+
+            for _, module in unet.named_modules():
+                if hasattr(module, "set_lora_layer"):
+                    module.set_lora_layer(None)
+        else:
+            recurse_remove_peft_layers(unet)
+            if hasattr(unet, "peft_config"):
+                del unet.peft_config
+
+        # Safe to call the following regardless of LoRA.
+        self._remove_text_encoder_monkey_patch()
+
+    def fuse_lora(
+        self,
+        fuse_unet: bool = True,
+        fuse_text_encoder: bool = True,
+        lora_scale: float = 1.0,
+        safe_fusing: bool = False,
+        adapter_names: Optional[List[str]] = None,
+    ):
+        r"""
+        Fuses the LoRA parameters into the original parameters of the corresponding blocks.
+
+        <Tip warning={true}>
+
+        This is an experimental API.
+
+        </Tip>
+
+        Args:
+            fuse_unet (`bool`, defaults to `True`): Whether to fuse the UNet LoRA parameters.
+            fuse_text_encoder (`bool`, defaults to `True`):
+                Whether to fuse the text encoder LoRA parameters. If the text encoder wasn't monkey-patched with the
+                LoRA parameters then it won't have any effect.
+            lora_scale (`float`, defaults to 1.0):
+                Controls how much to influence the outputs with the LoRA parameters.
+            safe_fusing (`bool`, defaults to `False`):
+                Whether to check fused weights for NaN values before fusing and if values are NaN not fusing them.
+            adapter_names (`List[str]`, *optional*):
+                Adapter names to be used for fusing. If nothing is passed, all active adapters will be fused.
+
+        Example:
+
+        ```py
+        from diffusers import DiffusionPipeline
+        import torch
+
+        pipeline = DiffusionPipeline.from_pretrained(
+            "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16
+        ).to("cuda")
+        pipeline.load_lora_weights("nerijs/pixel-art-xl", weight_name="pixel-art-xl.safetensors", adapter_name="pixel")
+        pipeline.fuse_lora(lora_scale=0.7)
+        ```
+        """
+        if fuse_unet or fuse_text_encoder:
+            self.num_fused_loras += 1
+            if self.num_fused_loras > 1:
+                logger.warn(
+                    "The current API is supported for operating with a single LoRA file. You are trying to load and fuse more than one LoRA which is not well-supported.",
+                )
+
+        if fuse_unet:
+            unet = getattr(self, self.unet_name) if not hasattr(self, "unet") else self.unet
+            unet.fuse_lora(lora_scale, safe_fusing=safe_fusing, adapter_names=adapter_names)
+
+        if USE_PEFT_BACKEND:
+            from peft.tuners.tuners_utils import BaseTunerLayer
+
+            def fuse_text_encoder_lora(text_encoder, lora_scale=1.0, safe_fusing=False, adapter_names=None):
+                merge_kwargs = {"safe_merge": safe_fusing}
+
+                for module in text_encoder.modules():
+                    if isinstance(module, BaseTunerLayer):
+                        if lora_scale != 1.0:
+                            module.scale_layer(lora_scale)
+
+                        # For BC with previous PEFT versions, we need to check the signature
+                        # of the `merge` method to see if it supports the `adapter_names` argument.
+                        supported_merge_kwargs = list(inspect.signature(module.merge).parameters)
+                        if "adapter_names" in supported_merge_kwargs:
+                            merge_kwargs["adapter_names"] = adapter_names
+                        elif "adapter_names" not in supported_merge_kwargs and adapter_names is not None:
+                            raise ValueError(
+                                "The `adapter_names` argument is not supported with your PEFT version. "
+                                "Please upgrade to the latest version of PEFT. `pip install -U peft`"
+                            )
+
+                        module.merge(**merge_kwargs)
+
+        else:
+            deprecate("fuse_text_encoder_lora", "0.27", LORA_DEPRECATION_MESSAGE)
+
+            def fuse_text_encoder_lora(text_encoder, lora_scale=1.0, safe_fusing=False, **kwargs):
+                if "adapter_names" in kwargs and kwargs["adapter_names"] is not None:
+                    raise ValueError(
+                        "The `adapter_names` argument is not supported in your environment. Please switch to PEFT "
+                        "backend to use this argument by installing latest PEFT and transformers."
+                        " `pip install -U peft transformers`"
+                    )
+
+                for _, attn_module in text_encoder_attn_modules(text_encoder):
+                    if isinstance(attn_module.q_proj, PatchedLoraProjection):
+                        attn_module.q_proj._fuse_lora(lora_scale, safe_fusing)
+                        attn_module.k_proj._fuse_lora(lora_scale, safe_fusing)
+                        attn_module.v_proj._fuse_lora(lora_scale, safe_fusing)
+                        attn_module.out_proj._fuse_lora(lora_scale, safe_fusing)
+
+                for _, mlp_module in text_encoder_mlp_modules(text_encoder):
+                    if isinstance(mlp_module.fc1, PatchedLoraProjection):
+                        mlp_module.fc1._fuse_lora(lora_scale, safe_fusing)
+                        mlp_module.fc2._fuse_lora(lora_scale, safe_fusing)
+
+        if fuse_text_encoder:
+            if hasattr(self, "text_encoder"):
+                fuse_text_encoder_lora(self.text_encoder, lora_scale, safe_fusing, adapter_names=adapter_names)
+            if hasattr(self, "text_encoder_2"):
+                fuse_text_encoder_lora(self.text_encoder_2, lora_scale, safe_fusing, adapter_names=adapter_names)
+
+    def unfuse_lora(self, unfuse_unet: bool = True, unfuse_text_encoder: bool = True):
+        r"""
+        Reverses the effect of
+        [`pipe.fuse_lora()`](https://huggingface.co/docs/diffusers/main/en/api/loaders#diffusers.loaders.LoraLoaderMixin.fuse_lora).
+
+        <Tip warning={true}>
+
+        This is an experimental API.
+
+        </Tip>
+
+        Args:
+            unfuse_unet (`bool`, defaults to `True`): Whether to unfuse the UNet LoRA parameters.
+            unfuse_text_encoder (`bool`, defaults to `True`):
+                Whether to unfuse the text encoder LoRA parameters. If the text encoder wasn't monkey-patched with the
+                LoRA parameters then it won't have any effect.
+        """
+        unet = getattr(self, self.unet_name) if not hasattr(self, "unet") else self.unet
+        if unfuse_unet:
+            if not USE_PEFT_BACKEND:
+                unet.unfuse_lora()
+            else:
+                from peft.tuners.tuners_utils import BaseTunerLayer
+
+                for module in unet.modules():
+                    if isinstance(module, BaseTunerLayer):
+                        module.unmerge()
+
+        if USE_PEFT_BACKEND:
+            from peft.tuners.tuners_utils import BaseTunerLayer
+
+            def unfuse_text_encoder_lora(text_encoder):
+                for module in text_encoder.modules():
+                    if isinstance(module, BaseTunerLayer):
+                        module.unmerge()
+
+        else:
+            deprecate("unfuse_text_encoder_lora", "0.27", LORA_DEPRECATION_MESSAGE)
+
+            def unfuse_text_encoder_lora(text_encoder):
+                for _, attn_module in text_encoder_attn_modules(text_encoder):
+                    if isinstance(attn_module.q_proj, PatchedLoraProjection):
+                        attn_module.q_proj._unfuse_lora()
+                        attn_module.k_proj._unfuse_lora()
+                        attn_module.v_proj._unfuse_lora()
+                        attn_module.out_proj._unfuse_lora()
+
+                for _, mlp_module in text_encoder_mlp_modules(text_encoder):
+                    if isinstance(mlp_module.fc1, PatchedLoraProjection):
+                        mlp_module.fc1._unfuse_lora()
+                        mlp_module.fc2._unfuse_lora()
+
+        if unfuse_text_encoder:
+            if hasattr(self, "text_encoder"):
+                unfuse_text_encoder_lora(self.text_encoder)
+            if hasattr(self, "text_encoder_2"):
+                unfuse_text_encoder_lora(self.text_encoder_2)
+
+        self.num_fused_loras -= 1
+
+    def set_adapters_for_text_encoder(
+        self,
+        adapter_names: Union[List[str], str],
+        text_encoder: Optional["PreTrainedModel"] = None,  # noqa: F821
+        text_encoder_weights: List[float] = None,
+    ):
+        """
+        Sets the adapter layers for the text encoder.
+
+        Args:
+            adapter_names (`List[str]` or `str`):
+                The names of the adapters to use.
+            text_encoder (`torch.nn.Module`, *optional*):
+                The text encoder module to set the adapter layers for. If `None`, it will try to get the `text_encoder`
+                attribute.
+            text_encoder_weights (`List[float]`, *optional*):
+                The weights to use for the text encoder. If `None`, the weights are set to `1.0` for all the adapters.
+        """
+        if not USE_PEFT_BACKEND:
+            raise ValueError("PEFT backend is required for this method.")
+
+        def process_weights(adapter_names, weights):
+            if weights is None:
+                weights = [1.0] * len(adapter_names)
+            elif isinstance(weights, float):
+                weights = [weights]
+
+            if len(adapter_names) != len(weights):
+                raise ValueError(
+                    f"Length of adapter names {len(adapter_names)} is not equal to the length of the weights {len(weights)}"
+                )
+            return weights
+
+        adapter_names = [adapter_names] if isinstance(adapter_names, str) else adapter_names
+        text_encoder_weights = process_weights(adapter_names, text_encoder_weights)
+        text_encoder = text_encoder or getattr(self, "text_encoder", None)
+        if text_encoder is None:
+            raise ValueError(
+                "The pipeline does not have a default `pipe.text_encoder` class. Please make sure to pass a `text_encoder` instead."
+            )
+        set_weights_and_activate_adapters(text_encoder, adapter_names, text_encoder_weights)
+
+    def disable_lora_for_text_encoder(self, text_encoder: Optional["PreTrainedModel"] = None):
+        """
+        Disables the LoRA layers for the text encoder.
+
+        Args:
+            text_encoder (`torch.nn.Module`, *optional*):
+                The text encoder module to disable the LoRA layers for. If `None`, it will try to get the
+                `text_encoder` attribute.
+        """
+        if not USE_PEFT_BACKEND:
+            raise ValueError("PEFT backend is required for this method.")
+
+        text_encoder = text_encoder or getattr(self, "text_encoder", None)
+        if text_encoder is None:
+            raise ValueError("Text Encoder not found.")
+        set_adapter_layers(text_encoder, enabled=False)
+
+    def enable_lora_for_text_encoder(self, text_encoder: Optional["PreTrainedModel"] = None):
+        """
+        Enables the LoRA layers for the text encoder.
+
+        Args:
+            text_encoder (`torch.nn.Module`, *optional*):
+                The text encoder module to enable the LoRA layers for. If `None`, it will try to get the `text_encoder`
+                attribute.
+        """
+        if not USE_PEFT_BACKEND:
+            raise ValueError("PEFT backend is required for this method.")
+        text_encoder = text_encoder or getattr(self, "text_encoder", None)
+        if text_encoder is None:
+            raise ValueError("Text Encoder not found.")
+        set_adapter_layers(self.text_encoder, enabled=True)
+
+    def set_adapters(
+        self,
+        adapter_names: Union[List[str], str],
+        adapter_weights: Optional[List[float]] = None,
+    ):
+        unet = getattr(self, self.unet_name) if not hasattr(self, "unet") else self.unet
+        # Handle the UNET
+        unet.set_adapters(adapter_names, adapter_weights)
+
+        # Handle the Text Encoder
+        if hasattr(self, "text_encoder"):
+            self.set_adapters_for_text_encoder(adapter_names, self.text_encoder, adapter_weights)
+        if hasattr(self, "text_encoder_2"):
+            self.set_adapters_for_text_encoder(adapter_names, self.text_encoder_2, adapter_weights)
+
+    def disable_lora(self):
+        if not USE_PEFT_BACKEND:
+            raise ValueError("PEFT backend is required for this method.")
+
+        # Disable unet adapters
+        unet = getattr(self, self.unet_name) if not hasattr(self, "unet") else self.unet
+        unet.disable_lora()
+
+        # Disable text encoder adapters
+        if hasattr(self, "text_encoder"):
+            self.disable_lora_for_text_encoder(self.text_encoder)
+        if hasattr(self, "text_encoder_2"):
+            self.disable_lora_for_text_encoder(self.text_encoder_2)
+
+    def enable_lora(self):
+        if not USE_PEFT_BACKEND:
+            raise ValueError("PEFT backend is required for this method.")
+
+        # Enable unet adapters
+        unet = getattr(self, self.unet_name) if not hasattr(self, "unet") else self.unet
+        unet.enable_lora()
+
+        # Enable text encoder adapters
+        if hasattr(self, "text_encoder"):
+            self.enable_lora_for_text_encoder(self.text_encoder)
+        if hasattr(self, "text_encoder_2"):
+            self.enable_lora_for_text_encoder(self.text_encoder_2)
+
+    def delete_adapters(self, adapter_names: Union[List[str], str]):
+        """
+        Args:
+        Deletes the LoRA layers of `adapter_name` for the unet and text-encoder(s).
+            adapter_names (`Union[List[str], str]`):
+                The names of the adapter to delete. Can be a single string or a list of strings
+        """
+        if not USE_PEFT_BACKEND:
+            raise ValueError("PEFT backend is required for this method.")
+
+        if isinstance(adapter_names, str):
+            adapter_names = [adapter_names]
+
+        # Delete unet adapters
+        unet = getattr(self, self.unet_name) if not hasattr(self, "unet") else self.unet
+        unet.delete_adapters(adapter_names)
+
+        for adapter_name in adapter_names:
+            # Delete text encoder adapters
+            if hasattr(self, "text_encoder"):
+                delete_adapter_layers(self.text_encoder, adapter_name)
+            if hasattr(self, "text_encoder_2"):
+                delete_adapter_layers(self.text_encoder_2, adapter_name)
+
+    def get_active_adapters(self) -> List[str]:
+        """
+        Gets the list of the current active adapters.
+
+        Example:
+
+        ```python
+        from diffusers import DiffusionPipeline
+
+        pipeline = DiffusionPipeline.from_pretrained(
+            "stabilityai/stable-diffusion-xl-base-1.0",
+        ).to("cuda")
+        pipeline.load_lora_weights("CiroN2022/toy-face", weight_name="toy_face_sdxl.safetensors", adapter_name="toy")
+        pipeline.get_active_adapters()
+        ```
+        """
+        if not USE_PEFT_BACKEND:
+            raise ValueError(
+                "PEFT backend is required for this method. Please install the latest version of PEFT `pip install -U peft`"
+            )
+
+        from peft.tuners.tuners_utils import BaseTunerLayer
+
+        active_adapters = []
+        unet = getattr(self, self.unet_name) if not hasattr(self, "unet") else self.unet
+        for module in unet.modules():
+            if isinstance(module, BaseTunerLayer):
+                active_adapters = module.active_adapters
+                break
+
+        return active_adapters
+
+    def get_list_adapters(self) -> Dict[str, List[str]]:
+        """
+        Gets the current list of all available adapters in the pipeline.
+        """
+        if not USE_PEFT_BACKEND:
+            raise ValueError(
+                "PEFT backend is required for this method. Please install the latest version of PEFT `pip install -U peft`"
+            )
+
+        set_adapters = {}
+
+        if hasattr(self, "text_encoder") and hasattr(self.text_encoder, "peft_config"):
+            set_adapters["text_encoder"] = list(self.text_encoder.peft_config.keys())
+
+        if hasattr(self, "text_encoder_2") and hasattr(self.text_encoder_2, "peft_config"):
+            set_adapters["text_encoder_2"] = list(self.text_encoder_2.peft_config.keys())
+
+        unet = getattr(self, self.unet_name) if not hasattr(self, "unet") else self.unet
+        if hasattr(self, self.unet_name) and hasattr(unet, "peft_config"):
+            set_adapters[self.unet_name] = list(self.unet.peft_config.keys())
+
+        return set_adapters
+
+    def set_lora_device(self, adapter_names: List[str], device: Union[torch.device, str, int]) -> None:
+        """
+        Moves the LoRAs listed in `adapter_names` to a target device. Useful for offloading the LoRA to the CPU in case
+        you want to load multiple adapters and free some GPU memory.
+
+        Args:
+            adapter_names (`List[str]`):
+                List of adapters to send device to.
+            device (`Union[torch.device, str, int]`):
+                Device to send the adapters to. Can be either a torch device, a str or an integer.
+        """
+        if not USE_PEFT_BACKEND:
+            raise ValueError("PEFT backend is required for this method.")
+
+        from peft.tuners.tuners_utils import BaseTunerLayer
+
+        # Handle the UNET
+        unet = getattr(self, self.unet_name) if not hasattr(self, "unet") else self.unet
+        for unet_module in unet.modules():
+            if isinstance(unet_module, BaseTunerLayer):
+                for adapter_name in adapter_names:
+                    unet_module.lora_A[adapter_name].to(device)
+                    unet_module.lora_B[adapter_name].to(device)
+
+        # Handle the text encoder
+        modules_to_process = []
+        if hasattr(self, "text_encoder"):
+            modules_to_process.append(self.text_encoder)
+
+        if hasattr(self, "text_encoder_2"):
+            modules_to_process.append(self.text_encoder_2)
+
+        for text_encoder in modules_to_process:
+            # loop over submodules
+            for text_encoder_module in text_encoder.modules():
+                if isinstance(text_encoder_module, BaseTunerLayer):
+                    for adapter_name in adapter_names:
+                        text_encoder_module.lora_A[adapter_name].to(device)
+                        text_encoder_module.lora_B[adapter_name].to(device)
+
+
+class StableDiffusionXLLoraLoaderMixin(LoraLoaderMixin):
+    """This class overrides `LoraLoaderMixin` with LoRA loading/saving code that's specific to SDXL"""
+
+    # Overrride to properly handle the loading and unloading of the additional text encoder.
+    def load_lora_weights(
+        self,
+        pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]],
+        adapter_name: Optional[str] = None,
+        **kwargs,
+    ):
+        """
+        Load LoRA weights specified in `pretrained_model_name_or_path_or_dict` into `self.unet` and
+        `self.text_encoder`.
+
+        All kwargs are forwarded to `self.lora_state_dict`.
+
+        See [`~loaders.LoraLoaderMixin.lora_state_dict`] for more details on how the state dict is loaded.
+
+        See [`~loaders.LoraLoaderMixin.load_lora_into_unet`] for more details on how the state dict is loaded into
+        `self.unet`.
+
+        See [`~loaders.LoraLoaderMixin.load_lora_into_text_encoder`] for more details on how the state dict is loaded
+        into `self.text_encoder`.
+
+        Parameters:
+            pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
+                See [`~loaders.LoraLoaderMixin.lora_state_dict`].
+            adapter_name (`str`, *optional*):
+                Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
+                `default_{i}` where i is the total number of adapters being loaded.
+            kwargs (`dict`, *optional*):
+                See [`~loaders.LoraLoaderMixin.lora_state_dict`].
+        """
+        # We could have accessed the unet config from `lora_state_dict()` too. We pass
+        # it here explicitly to be able to tell that it's coming from an SDXL
+        # pipeline.
+
+        # First, ensure that the checkpoint is a compatible one and can be successfully loaded.
+        state_dict, network_alphas = self.lora_state_dict(
+            pretrained_model_name_or_path_or_dict,
+            unet_config=self.unet.config,
+            **kwargs,
+        )
+        is_correct_format = all("lora" in key for key in state_dict.keys())
+        if not is_correct_format:
+            raise ValueError("Invalid LoRA checkpoint.")
+
+        self.load_lora_into_unet(
+            state_dict, network_alphas=network_alphas, unet=self.unet, adapter_name=adapter_name, _pipeline=self
+        )
+        text_encoder_state_dict = {k: v for k, v in state_dict.items() if "text_encoder." in k}
+        if len(text_encoder_state_dict) > 0:
+            self.load_lora_into_text_encoder(
+                text_encoder_state_dict,
+                network_alphas=network_alphas,
+                text_encoder=self.text_encoder,
+                prefix="text_encoder",
+                lora_scale=self.lora_scale,
+                adapter_name=adapter_name,
+                _pipeline=self,
+            )
+
+        text_encoder_2_state_dict = {k: v for k, v in state_dict.items() if "text_encoder_2." in k}
+        if len(text_encoder_2_state_dict) > 0:
+            self.load_lora_into_text_encoder(
+                text_encoder_2_state_dict,
+                network_alphas=network_alphas,
+                text_encoder=self.text_encoder_2,
+                prefix="text_encoder_2",
+                lora_scale=self.lora_scale,
+                adapter_name=adapter_name,
+                _pipeline=self,
+            )
+
+    @classmethod
+    def save_lora_weights(
+        cls,
+        save_directory: Union[str, os.PathLike],
+        unet_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None,
+        text_encoder_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None,
+        text_encoder_2_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None,
+        is_main_process: bool = True,
+        weight_name: str = None,
+        save_function: Callable = None,
+        safe_serialization: bool = True,
+    ):
+        r"""
+        Save the LoRA parameters corresponding to the UNet and text encoder.
+
+        Arguments:
+            save_directory (`str` or `os.PathLike`):
+                Directory to save LoRA parameters to. Will be created if it doesn't exist.
+            unet_lora_layers (`Dict[str, torch.nn.Module]` or `Dict[str, torch.Tensor]`):
+                State dict of the LoRA layers corresponding to the `unet`.
+            text_encoder_lora_layers (`Dict[str, torch.nn.Module]` or `Dict[str, torch.Tensor]`):
+                State dict of the LoRA layers corresponding to the `text_encoder`. Must explicitly pass the text
+                encoder LoRA state dict because it comes from 🤗 Transformers.
+            is_main_process (`bool`, *optional*, defaults to `True`):
+                Whether the process calling this is the main process or not. Useful during distributed training and you
+                need to call this function on all processes. In this case, set `is_main_process=True` only on the main
+                process to avoid race conditions.
+            save_function (`Callable`):
+                The function to use to save the state dictionary. Useful during distributed training when you need to
+                replace `torch.save` with another method. Can be configured with the environment variable
+                `DIFFUSERS_SAVE_MODE`.
+            safe_serialization (`bool`, *optional*, defaults to `True`):
+                Whether to save the model using `safetensors` or the traditional PyTorch way with `pickle`.
+        """
+        state_dict = {}
+
+        def pack_weights(layers, prefix):
+            layers_weights = layers.state_dict() if isinstance(layers, torch.nn.Module) else layers
+            layers_state_dict = {f"{prefix}.{module_name}": param for module_name, param in layers_weights.items()}
+            return layers_state_dict
+
+        if not (unet_lora_layers or text_encoder_lora_layers or text_encoder_2_lora_layers):
+            raise ValueError(
+                "You must pass at least one of `unet_lora_layers`, `text_encoder_lora_layers` or `text_encoder_2_lora_layers`."
+            )
+
+        if unet_lora_layers:
+            state_dict.update(pack_weights(unet_lora_layers, "unet"))
+
+        if text_encoder_lora_layers and text_encoder_2_lora_layers:
+            state_dict.update(pack_weights(text_encoder_lora_layers, "text_encoder"))
+            state_dict.update(pack_weights(text_encoder_2_lora_layers, "text_encoder_2"))
+
+        cls.write_lora_layers(
+            state_dict=state_dict,
+            save_directory=save_directory,
+            is_main_process=is_main_process,
+            weight_name=weight_name,
+            save_function=save_function,
+            safe_serialization=safe_serialization,
+        )
+
+    def _remove_text_encoder_monkey_patch(self):
+        if USE_PEFT_BACKEND:
+            recurse_remove_peft_layers(self.text_encoder)
+            # TODO: @younesbelkada handle this in transformers side
+            if getattr(self.text_encoder, "peft_config", None) is not None:
+                del self.text_encoder.peft_config
+                self.text_encoder._hf_peft_config_loaded = None
+
+            recurse_remove_peft_layers(self.text_encoder_2)
+            if getattr(self.text_encoder_2, "peft_config", None) is not None:
+                del self.text_encoder_2.peft_config
+                self.text_encoder_2._hf_peft_config_loaded = None
+        else:
+            self._remove_text_encoder_monkey_patch_classmethod(self.text_encoder)
+            self._remove_text_encoder_monkey_patch_classmethod(self.text_encoder_2)

src/diffusers/loaders/lora_conversion_utils.py

@@ -0,0 +1,284 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import re
+
+from ..utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+
+def _maybe_map_sgm_blocks_to_diffusers(state_dict, unet_config, delimiter="_", block_slice_pos=5):
+    # 1. get all state_dict_keys
+    all_keys = list(state_dict.keys())
+    sgm_patterns = ["input_blocks", "middle_block", "output_blocks"]
+
+    # 2. check if needs remapping, if not return original dict
+    is_in_sgm_format = False
+    for key in all_keys:
+        if any(p in key for p in sgm_patterns):
+            is_in_sgm_format = True
+            break
+
+    if not is_in_sgm_format:
+        return state_dict
+
+    # 3. Else remap from SGM patterns
+    new_state_dict = {}
+    inner_block_map = ["resnets", "attentions", "upsamplers"]
+
+    # Retrieves # of down, mid and up blocks
+    input_block_ids, middle_block_ids, output_block_ids = set(), set(), set()
+
+    for layer in all_keys:
+        if "text" in layer:
+            new_state_dict[layer] = state_dict.pop(layer)
+        else:
+            layer_id = int(layer.split(delimiter)[:block_slice_pos][-1])
+            if sgm_patterns[0] in layer:
+                input_block_ids.add(layer_id)
+            elif sgm_patterns[1] in layer:
+                middle_block_ids.add(layer_id)
+            elif sgm_patterns[2] in layer:
+                output_block_ids.add(layer_id)
+            else:
+                raise ValueError(f"Checkpoint not supported because layer {layer} not supported.")
+
+    input_blocks = {
+        layer_id: [key for key in state_dict if f"input_blocks{delimiter}{layer_id}" in key]
+        for layer_id in input_block_ids
+    }
+    middle_blocks = {
+        layer_id: [key for key in state_dict if f"middle_block{delimiter}{layer_id}" in key]
+        for layer_id in middle_block_ids
+    }
+    output_blocks = {
+        layer_id: [key for key in state_dict if f"output_blocks{delimiter}{layer_id}" in key]
+        for layer_id in output_block_ids
+    }
+
+    # Rename keys accordingly
+    for i in input_block_ids:
+        block_id = (i - 1) // (unet_config.layers_per_block + 1)
+        layer_in_block_id = (i - 1) % (unet_config.layers_per_block + 1)
+
+        for key in input_blocks[i]:
+            inner_block_id = int(key.split(delimiter)[block_slice_pos])
+            inner_block_key = inner_block_map[inner_block_id] if "op" not in key else "downsamplers"
+            inner_layers_in_block = str(layer_in_block_id) if "op" not in key else "0"
+            new_key = delimiter.join(
+                key.split(delimiter)[: block_slice_pos - 1]
+                + [str(block_id), inner_block_key, inner_layers_in_block]
+                + key.split(delimiter)[block_slice_pos + 1 :]
+            )
+            new_state_dict[new_key] = state_dict.pop(key)
+
+    for i in middle_block_ids:
+        key_part = None
+        if i == 0:
+            key_part = [inner_block_map[0], "0"]
+        elif i == 1:
+            key_part = [inner_block_map[1], "0"]
+        elif i == 2:
+            key_part = [inner_block_map[0], "1"]
+        else:
+            raise ValueError(f"Invalid middle block id {i}.")
+
+        for key in middle_blocks[i]:
+            new_key = delimiter.join(
+                key.split(delimiter)[: block_slice_pos - 1] + key_part + key.split(delimiter)[block_slice_pos:]
+            )
+            new_state_dict[new_key] = state_dict.pop(key)
+
+    for i in output_block_ids:
+        block_id = i // (unet_config.layers_per_block + 1)
+        layer_in_block_id = i % (unet_config.layers_per_block + 1)
+
+        for key in output_blocks[i]:
+            inner_block_id = int(key.split(delimiter)[block_slice_pos])
+            inner_block_key = inner_block_map[inner_block_id]
+            inner_layers_in_block = str(layer_in_block_id) if inner_block_id < 2 else "0"
+            new_key = delimiter.join(
+                key.split(delimiter)[: block_slice_pos - 1]
+                + [str(block_id), inner_block_key, inner_layers_in_block]
+                + key.split(delimiter)[block_slice_pos + 1 :]
+            )
+            new_state_dict[new_key] = state_dict.pop(key)
+
+    if len(state_dict) > 0:
+        raise ValueError("At this point all state dict entries have to be converted.")
+
+    return new_state_dict
+
+
+def _convert_kohya_lora_to_diffusers(state_dict, unet_name="unet", text_encoder_name="text_encoder"):
+    unet_state_dict = {}
+    te_state_dict = {}
+    te2_state_dict = {}
+    network_alphas = {}
+
+    # every down weight has a corresponding up weight and potentially an alpha weight
+    lora_keys = [k for k in state_dict.keys() if k.endswith("lora_down.weight")]
+    for key in lora_keys:
+        lora_name = key.split(".")[0]
+        lora_name_up = lora_name + ".lora_up.weight"
+        lora_name_alpha = lora_name + ".alpha"
+
+        if lora_name.startswith("lora_unet_"):
+            diffusers_name = key.replace("lora_unet_", "").replace("_", ".")
+
+            if "input.blocks" in diffusers_name:
+                diffusers_name = diffusers_name.replace("input.blocks", "down_blocks")
+            else:
+                diffusers_name = diffusers_name.replace("down.blocks", "down_blocks")
+
+            if "middle.block" in diffusers_name:
+                diffusers_name = diffusers_name.replace("middle.block", "mid_block")
+            else:
+                diffusers_name = diffusers_name.replace("mid.block", "mid_block")
+            if "output.blocks" in diffusers_name:
+                diffusers_name = diffusers_name.replace("output.blocks", "up_blocks")
+            else:
+                diffusers_name = diffusers_name.replace("up.blocks", "up_blocks")
+
+            diffusers_name = diffusers_name.replace("transformer.blocks", "transformer_blocks")
+            diffusers_name = diffusers_name.replace("to.q.lora", "to_q_lora")
+            diffusers_name = diffusers_name.replace("to.k.lora", "to_k_lora")
+            diffusers_name = diffusers_name.replace("to.v.lora", "to_v_lora")
+            diffusers_name = diffusers_name.replace("to.out.0.lora", "to_out_lora")
+            diffusers_name = diffusers_name.replace("proj.in", "proj_in")
+            diffusers_name = diffusers_name.replace("proj.out", "proj_out")
+            diffusers_name = diffusers_name.replace("emb.layers", "time_emb_proj")
+
+            # SDXL specificity.
+            if "emb" in diffusers_name and "time.emb.proj" not in diffusers_name:
+                pattern = r"\.\d+(?=\D*$)"
+                diffusers_name = re.sub(pattern, "", diffusers_name, count=1)
+            if ".in." in diffusers_name:
+                diffusers_name = diffusers_name.replace("in.layers.2", "conv1")
+            if ".out." in diffusers_name:
+                diffusers_name = diffusers_name.replace("out.layers.3", "conv2")
+            if "downsamplers" in diffusers_name or "upsamplers" in diffusers_name:
+                diffusers_name = diffusers_name.replace("op", "conv")
+            if "skip" in diffusers_name:
+                diffusers_name = diffusers_name.replace("skip.connection", "conv_shortcut")
+
+            # LyCORIS specificity.
+            if "time.emb.proj" in diffusers_name:
+                diffusers_name = diffusers_name.replace("time.emb.proj", "time_emb_proj")
+            if "conv.shortcut" in diffusers_name:
+                diffusers_name = diffusers_name.replace("conv.shortcut", "conv_shortcut")
+
+            # General coverage.
+            if "transformer_blocks" in diffusers_name:
+                if "attn1" in diffusers_name or "attn2" in diffusers_name:
+                    diffusers_name = diffusers_name.replace("attn1", "attn1.processor")
+                    diffusers_name = diffusers_name.replace("attn2", "attn2.processor")
+                    unet_state_dict[diffusers_name] = state_dict.pop(key)
+                    unet_state_dict[diffusers_name.replace(".down.", ".up.")] = state_dict.pop(lora_name_up)
+                elif "ff" in diffusers_name:
+                    unet_state_dict[diffusers_name] = state_dict.pop(key)
+                    unet_state_dict[diffusers_name.replace(".down.", ".up.")] = state_dict.pop(lora_name_up)
+            elif any(key in diffusers_name for key in ("proj_in", "proj_out")):
+                unet_state_dict[diffusers_name] = state_dict.pop(key)
+                unet_state_dict[diffusers_name.replace(".down.", ".up.")] = state_dict.pop(lora_name_up)
+            else:
+                unet_state_dict[diffusers_name] = state_dict.pop(key)
+                unet_state_dict[diffusers_name.replace(".down.", ".up.")] = state_dict.pop(lora_name_up)
+
+        elif lora_name.startswith("lora_te_"):
+            diffusers_name = key.replace("lora_te_", "").replace("_", ".")
+            diffusers_name = diffusers_name.replace("text.model", "text_model")
+            diffusers_name = diffusers_name.replace("self.attn", "self_attn")
+            diffusers_name = diffusers_name.replace("q.proj.lora", "to_q_lora")
+            diffusers_name = diffusers_name.replace("k.proj.lora", "to_k_lora")
+            diffusers_name = diffusers_name.replace("v.proj.lora", "to_v_lora")
+            diffusers_name = diffusers_name.replace("out.proj.lora", "to_out_lora")
+            if "self_attn" in diffusers_name:
+                te_state_dict[diffusers_name] = state_dict.pop(key)
+                te_state_dict[diffusers_name.replace(".down.", ".up.")] = state_dict.pop(lora_name_up)
+            elif "mlp" in diffusers_name:
+                # Be aware that this is the new diffusers convention and the rest of the code might
+                # not utilize it yet.
+                diffusers_name = diffusers_name.replace(".lora.", ".lora_linear_layer.")
+                te_state_dict[diffusers_name] = state_dict.pop(key)
+                te_state_dict[diffusers_name.replace(".down.", ".up.")] = state_dict.pop(lora_name_up)
+
+        # (sayakpaul): Duplicate code. Needs to be cleaned.
+        elif lora_name.startswith("lora_te1_"):
+            diffusers_name = key.replace("lora_te1_", "").replace("_", ".")
+            diffusers_name = diffusers_name.replace("text.model", "text_model")
+            diffusers_name = diffusers_name.replace("self.attn", "self_attn")
+            diffusers_name = diffusers_name.replace("q.proj.lora", "to_q_lora")
+            diffusers_name = diffusers_name.replace("k.proj.lora", "to_k_lora")
+            diffusers_name = diffusers_name.replace("v.proj.lora", "to_v_lora")
+            diffusers_name = diffusers_name.replace("out.proj.lora", "to_out_lora")
+            if "self_attn" in diffusers_name:
+                te_state_dict[diffusers_name] = state_dict.pop(key)
+                te_state_dict[diffusers_name.replace(".down.", ".up.")] = state_dict.pop(lora_name_up)
+            elif "mlp" in diffusers_name:
+                # Be aware that this is the new diffusers convention and the rest of the code might
+                # not utilize it yet.
+                diffusers_name = diffusers_name.replace(".lora.", ".lora_linear_layer.")
+                te_state_dict[diffusers_name] = state_dict.pop(key)
+                te_state_dict[diffusers_name.replace(".down.", ".up.")] = state_dict.pop(lora_name_up)
+
+        # (sayakpaul): Duplicate code. Needs to be cleaned.
+        elif lora_name.startswith("lora_te2_"):
+            diffusers_name = key.replace("lora_te2_", "").replace("_", ".")
+            diffusers_name = diffusers_name.replace("text.model", "text_model")
+            diffusers_name = diffusers_name.replace("self.attn", "self_attn")
+            diffusers_name = diffusers_name.replace("q.proj.lora", "to_q_lora")
+            diffusers_name = diffusers_name.replace("k.proj.lora", "to_k_lora")
+            diffusers_name = diffusers_name.replace("v.proj.lora", "to_v_lora")
+            diffusers_name = diffusers_name.replace("out.proj.lora", "to_out_lora")
+            if "self_attn" in diffusers_name:
+                te2_state_dict[diffusers_name] = state_dict.pop(key)
+                te2_state_dict[diffusers_name.replace(".down.", ".up.")] = state_dict.pop(lora_name_up)
+            elif "mlp" in diffusers_name:
+                # Be aware that this is the new diffusers convention and the rest of the code might
+                # not utilize it yet.
+                diffusers_name = diffusers_name.replace(".lora.", ".lora_linear_layer.")
+                te2_state_dict[diffusers_name] = state_dict.pop(key)
+                te2_state_dict[diffusers_name.replace(".down.", ".up.")] = state_dict.pop(lora_name_up)
+
+        # Rename the alphas so that they can be mapped appropriately.
+        if lora_name_alpha in state_dict:
+            alpha = state_dict.pop(lora_name_alpha).item()
+            if lora_name_alpha.startswith("lora_unet_"):
+                prefix = "unet."
+            elif lora_name_alpha.startswith(("lora_te_", "lora_te1_")):
+                prefix = "text_encoder."
+            else:
+                prefix = "text_encoder_2."
+            new_name = prefix + diffusers_name.split(".lora.")[0] + ".alpha"
+            network_alphas.update({new_name: alpha})
+
+    if len(state_dict) > 0:
+        raise ValueError(f"The following keys have not been correctly be renamed: \n\n {', '.join(state_dict.keys())}")
+
+    logger.info("Kohya-style checkpoint detected.")
+    unet_state_dict = {f"{unet_name}.{module_name}": params for module_name, params in unet_state_dict.items()}
+    te_state_dict = {f"{text_encoder_name}.{module_name}": params for module_name, params in te_state_dict.items()}
+    te2_state_dict = (
+        {f"text_encoder_2.{module_name}": params for module_name, params in te2_state_dict.items()}
+        if len(te2_state_dict) > 0
+        else None
+    )
+    if te2_state_dict is not None:
+        te_state_dict.update(te2_state_dict)
+
+    new_state_dict = {**unet_state_dict, **te_state_dict}
+    return new_state_dict, network_alphas

src/diffusers/loaders/single_file.py

@@ -0,0 +1,637 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from contextlib import nullcontext
+from io import BytesIO
+from pathlib import Path
+
+import requests
+import torch
+from huggingface_hub import hf_hub_download
+from huggingface_hub.utils import validate_hf_hub_args
+
+from ..utils import (
+    deprecate,
+    is_accelerate_available,
+    is_omegaconf_available,
+    is_transformers_available,
+    logging,
+)
+from ..utils.import_utils import BACKENDS_MAPPING
+
+
+if is_transformers_available():
+    pass
+
+if is_accelerate_available():
+    from accelerate import init_empty_weights
+
+logger = logging.get_logger(__name__)
+
+
+class FromSingleFileMixin:
+    """
+    Load model weights saved in the `.ckpt` format into a [`DiffusionPipeline`].
+    """
+
+    @classmethod
+    def from_ckpt(cls, *args, **kwargs):
+        deprecation_message = "The function `from_ckpt` is deprecated in favor of `from_single_file` and will be removed in diffusers v.0.21. Please make sure to use `StableDiffusionPipeline.from_single_file(...)` instead."
+        deprecate("from_ckpt", "0.21.0", deprecation_message, standard_warn=False)
+        return cls.from_single_file(*args, **kwargs)
+
+    @classmethod
+    @validate_hf_hub_args
+    def from_single_file(cls, pretrained_model_link_or_path, **kwargs):
+        r"""
+        Instantiate a [`DiffusionPipeline`] from pretrained pipeline weights saved in the `.ckpt` or `.safetensors`
+        format. The pipeline is set in evaluation mode (`model.eval()`) by default.
+
+        Parameters:
+            pretrained_model_link_or_path (`str` or `os.PathLike`, *optional*):
+                Can be either:
+                    - A link to the `.ckpt` file (for example
+                      `"https://huggingface.co/<repo_id>/blob/main/<path_to_file>.ckpt"`) on the Hub.
+                    - A path to a *file* containing all pipeline weights.
+            torch_dtype (`str` or `torch.dtype`, *optional*):
+                Override the default `torch.dtype` and load the model with another dtype. If `"auto"` is passed, the
+                dtype is automatically derived from the model's weights.
+            force_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to force the (re-)download of the model weights and configuration files, overriding the
+                cached versions if they exist.
+            cache_dir (`Union[str, os.PathLike]`, *optional*):
+                Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
+                is not used.
+            resume_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to resume downloading the model weights and configuration files. If set to `False`, any
+                incompletely downloaded files are deleted.
+            proxies (`Dict[str, str]`, *optional*):
+                A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
+                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
+            local_files_only (`bool`, *optional*, defaults to `False`):
+                Whether to only load local model weights and configuration files or not. If set to `True`, the model
+                won't be downloaded from the Hub.
+            token (`str` or *bool*, *optional*):
+                The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
+                `diffusers-cli login` (stored in `~/.huggingface`) is used.
+            revision (`str`, *optional*, defaults to `"main"`):
+                The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
+                allowed by Git.
+            use_safetensors (`bool`, *optional*, defaults to `None`):
+                If set to `None`, the safetensors weights are downloaded if they're available **and** if the
+                safetensors library is installed. If set to `True`, the model is forcibly loaded from safetensors
+                weights. If set to `False`, safetensors weights are not loaded.
+            extract_ema (`bool`, *optional*, defaults to `False`):
+                Whether to extract the EMA weights or not. Pass `True` to extract the EMA weights which usually yield
+                higher quality images for inference. Non-EMA weights are usually better for continuing finetuning.
+            upcast_attention (`bool`, *optional*, defaults to `None`):
+                Whether the attention computation should always be upcasted.
+            image_size (`int`, *optional*, defaults to 512):
+                The image size the model was trained on. Use 512 for all Stable Diffusion v1 models and the Stable
+                Diffusion v2 base model. Use 768 for Stable Diffusion v2.
+            prediction_type (`str`, *optional*):
+                The prediction type the model was trained on. Use `'epsilon'` for all Stable Diffusion v1 models and
+                the Stable Diffusion v2 base model. Use `'v_prediction'` for Stable Diffusion v2.
+            num_in_channels (`int`, *optional*, defaults to `None`):
+                The number of input channels. If `None`, it is automatically inferred.
+            scheduler_type (`str`, *optional*, defaults to `"pndm"`):
+                Type of scheduler to use. Should be one of `["pndm", "lms", "heun", "euler", "euler-ancestral", "dpm",
+                "ddim"]`.
+            load_safety_checker (`bool`, *optional*, defaults to `True`):
+                Whether to load the safety checker or not.
+            text_encoder ([`~transformers.CLIPTextModel`], *optional*, defaults to `None`):
+                An instance of `CLIPTextModel` to use, specifically the
+                [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. If this
+                parameter is `None`, the function loads a new instance of `CLIPTextModel` by itself if needed.
+            vae (`AutoencoderKL`, *optional*, defaults to `None`):
+                Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. If
+                this parameter is `None`, the function will load a new instance of [CLIP] by itself, if needed.
+            tokenizer ([`~transformers.CLIPTokenizer`], *optional*, defaults to `None`):
+                An instance of `CLIPTokenizer` to use. If this parameter is `None`, the function loads a new instance
+                of `CLIPTokenizer` by itself if needed.
+            original_config_file (`str`):
+                Path to `.yaml` config file corresponding to the original architecture. If `None`, will be
+                automatically inferred by looking for a key that only exists in SD2.0 models.
+            kwargs (remaining dictionary of keyword arguments, *optional*):
+                Can be used to overwrite load and saveable variables (for example the pipeline components of the
+                specific pipeline class). The overwritten components are directly passed to the pipelines `__init__`
+                method. See example below for more information.
+
+        Examples:
+
+        ```py
+        >>> from diffusers import StableDiffusionPipeline
+
+        >>> # Download pipeline from huggingface.co and cache.
+        >>> pipeline = StableDiffusionPipeline.from_single_file(
+        ...     "https://huggingface.co/WarriorMama777/OrangeMixs/blob/main/Models/AbyssOrangeMix/AbyssOrangeMix.safetensors"
+        ... )
+
+        >>> # Download pipeline from local file
+        >>> # file is downloaded under ./v1-5-pruned-emaonly.ckpt
+        >>> pipeline = StableDiffusionPipeline.from_single_file("./v1-5-pruned-emaonly")
+
+        >>> # Enable float16 and move to GPU
+        >>> pipeline = StableDiffusionPipeline.from_single_file(
+        ...     "https://huggingface.co/runwayml/stable-diffusion-v1-5/blob/main/v1-5-pruned-emaonly.ckpt",
+        ...     torch_dtype=torch.float16,
+        ... )
+        >>> pipeline.to("cuda")
+        ```
+        """
+        # import here to avoid circular dependency
+        from ..pipelines.stable_diffusion.convert_from_ckpt import download_from_original_stable_diffusion_ckpt
+
+        original_config_file = kwargs.pop("original_config_file", None)
+        config_files = kwargs.pop("config_files", None)
+        cache_dir = kwargs.pop("cache_dir", None)
+        resume_download = kwargs.pop("resume_download", False)
+        force_download = kwargs.pop("force_download", False)
+        proxies = kwargs.pop("proxies", None)
+        local_files_only = kwargs.pop("local_files_only", None)
+        token = kwargs.pop("token", None)
+        revision = kwargs.pop("revision", None)
+        extract_ema = kwargs.pop("extract_ema", False)
+        image_size = kwargs.pop("image_size", None)
+        scheduler_type = kwargs.pop("scheduler_type", "pndm")
+        num_in_channels = kwargs.pop("num_in_channels", None)
+        upcast_attention = kwargs.pop("upcast_attention", None)
+        load_safety_checker = kwargs.pop("load_safety_checker", True)
+        prediction_type = kwargs.pop("prediction_type", None)
+        text_encoder = kwargs.pop("text_encoder", None)
+        text_encoder_2 = kwargs.pop("text_encoder_2", None)
+        vae = kwargs.pop("vae", None)
+        controlnet = kwargs.pop("controlnet", None)
+        adapter = kwargs.pop("adapter", None)
+        tokenizer = kwargs.pop("tokenizer", None)
+        tokenizer_2 = kwargs.pop("tokenizer_2", None)
+
+        torch_dtype = kwargs.pop("torch_dtype", None)
+
+        use_safetensors = kwargs.pop("use_safetensors", None)
+
+        pipeline_name = cls.__name__
+        file_extension = pretrained_model_link_or_path.rsplit(".", 1)[-1]
+        from_safetensors = file_extension == "safetensors"
+
+        if from_safetensors and use_safetensors is False:
+            raise ValueError("Make sure to install `safetensors` with `pip install safetensors`.")
+
+        # TODO: For now we only support stable diffusion
+        stable_unclip = None
+        model_type = None
+
+        if pipeline_name in [
+            "StableDiffusionControlNetPipeline",
+            "StableDiffusionControlNetImg2ImgPipeline",
+            "StableDiffusionControlNetInpaintPipeline",
+        ]:
+            from ..models.controlnet import ControlNetModel
+            from ..pipelines.controlnet.multicontrolnet import MultiControlNetModel
+
+            #  list/tuple or a single instance of ControlNetModel or MultiControlNetModel
+            if not (
+                isinstance(controlnet, (ControlNetModel, MultiControlNetModel))
+                or isinstance(controlnet, (list, tuple))
+                and isinstance(controlnet[0], ControlNetModel)
+            ):
+                raise ValueError("ControlNet needs to be passed if loading from ControlNet pipeline.")
+        elif "StableDiffusion" in pipeline_name:
+            # Model type will be inferred from the checkpoint.
+            pass
+        elif pipeline_name == "StableUnCLIPPipeline":
+            model_type = "FrozenOpenCLIPEmbedder"
+            stable_unclip = "txt2img"
+        elif pipeline_name == "StableUnCLIPImg2ImgPipeline":
+            model_type = "FrozenOpenCLIPEmbedder"
+            stable_unclip = "img2img"
+        elif pipeline_name == "PaintByExamplePipeline":
+            model_type = "PaintByExample"
+        elif pipeline_name == "LDMTextToImagePipeline":
+            model_type = "LDMTextToImage"
+        else:
+            raise ValueError(f"Unhandled pipeline class: {pipeline_name}")
+
+        # remove huggingface url
+        has_valid_url_prefix = False
+        valid_url_prefixes = ["https://huggingface.co/", "huggingface.co/", "hf.co/", "https://hf.co/"]
+        for prefix in valid_url_prefixes:
+            if pretrained_model_link_or_path.startswith(prefix):
+                pretrained_model_link_or_path = pretrained_model_link_or_path[len(prefix) :]
+                has_valid_url_prefix = True
+
+        # Code based on diffusers.pipelines.pipeline_utils.DiffusionPipeline.from_pretrained
+        ckpt_path = Path(pretrained_model_link_or_path)
+        if not ckpt_path.is_file():
+            if not has_valid_url_prefix:
+                raise ValueError(
+                    f"The provided path is either not a file or a valid huggingface URL was not provided. Valid URLs begin with {', '.join(valid_url_prefixes)}"
+                )
+
+            # get repo_id and (potentially nested) file path of ckpt in repo
+            repo_id = "/".join(ckpt_path.parts[:2])
+            file_path = "/".join(ckpt_path.parts[2:])
+
+            if file_path.startswith("blob/"):
+                file_path = file_path[len("blob/") :]
+
+            if file_path.startswith("main/"):
+                file_path = file_path[len("main/") :]
+
+            pretrained_model_link_or_path = hf_hub_download(
+                repo_id,
+                filename=file_path,
+                cache_dir=cache_dir,
+                resume_download=resume_download,
+                proxies=proxies,
+                local_files_only=local_files_only,
+                token=token,
+                revision=revision,
+                force_download=force_download,
+            )
+
+        pipe = download_from_original_stable_diffusion_ckpt(
+            pretrained_model_link_or_path,
+            pipeline_class=cls,
+            model_type=model_type,
+            stable_unclip=stable_unclip,
+            controlnet=controlnet,
+            adapter=adapter,
+            from_safetensors=from_safetensors,
+            extract_ema=extract_ema,
+            image_size=image_size,
+            scheduler_type=scheduler_type,
+            num_in_channels=num_in_channels,
+            upcast_attention=upcast_attention,
+            load_safety_checker=load_safety_checker,
+            prediction_type=prediction_type,
+            text_encoder=text_encoder,
+            text_encoder_2=text_encoder_2,
+            vae=vae,
+            tokenizer=tokenizer,
+            tokenizer_2=tokenizer_2,
+            original_config_file=original_config_file,
+            config_files=config_files,
+            local_files_only=local_files_only,
+        )
+
+        if torch_dtype is not None:
+            pipe.to(dtype=torch_dtype)
+
+        return pipe
+
+
+class FromOriginalVAEMixin:
+    """
+    Load pretrained ControlNet weights saved in the `.ckpt` or `.safetensors` format into an [`AutoencoderKL`].
+    """
+
+    @classmethod
+    @validate_hf_hub_args
+    def from_single_file(cls, pretrained_model_link_or_path, **kwargs):
+        r"""
+        Instantiate a [`AutoencoderKL`] from pretrained ControlNet weights saved in the original `.ckpt` or
+        `.safetensors` format. The pipeline is set in evaluation mode (`model.eval()`) by default.
+
+        Parameters:
+            pretrained_model_link_or_path (`str` or `os.PathLike`, *optional*):
+                Can be either:
+                    - A link to the `.ckpt` file (for example
+                      `"https://huggingface.co/<repo_id>/blob/main/<path_to_file>.ckpt"`) on the Hub.
+                    - A path to a *file* containing all pipeline weights.
+            torch_dtype (`str` or `torch.dtype`, *optional*):
+                Override the default `torch.dtype` and load the model with another dtype. If `"auto"` is passed, the
+                dtype is automatically derived from the model's weights.
+            force_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to force the (re-)download of the model weights and configuration files, overriding the
+                cached versions if they exist.
+            cache_dir (`Union[str, os.PathLike]`, *optional*):
+                Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
+                is not used.
+            resume_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to resume downloading the model weights and configuration files. If set to `False`, any
+                incompletely downloaded files are deleted.
+            proxies (`Dict[str, str]`, *optional*):
+                A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
+                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
+            local_files_only (`bool`, *optional*, defaults to `False`):
+                Whether to only load local model weights and configuration files or not. If set to True, the model
+                won't be downloaded from the Hub.
+            token (`str` or *bool*, *optional*):
+                The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
+                `diffusers-cli login` (stored in `~/.huggingface`) is used.
+            revision (`str`, *optional*, defaults to `"main"`):
+                The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
+                allowed by Git.
+            image_size (`int`, *optional*, defaults to 512):
+                The image size the model was trained on. Use 512 for all Stable Diffusion v1 models and the Stable
+                Diffusion v2 base model. Use 768 for Stable Diffusion v2.
+            use_safetensors (`bool`, *optional*, defaults to `None`):
+                If set to `None`, the safetensors weights are downloaded if they're available **and** if the
+                safetensors library is installed. If set to `True`, the model is forcibly loaded from safetensors
+                weights. If set to `False`, safetensors weights are not loaded.
+            upcast_attention (`bool`, *optional*, defaults to `None`):
+                Whether the attention computation should always be upcasted.
+            scaling_factor (`float`, *optional*, defaults to 0.18215):
+                The component-wise standard deviation of the trained latent space computed using the first batch of the
+                training set. This is used to scale the latent space to have unit variance when training the diffusion
+                model. The latents are scaled with the formula `z = z * scaling_factor` before being passed to the
+                diffusion model. When decoding, the latents are scaled back to the original scale with the formula: `z
+                = 1 / scaling_factor * z`. For more details, refer to sections 4.3.2 and D.1 of the [High-Resolution
+                Image Synthesis with Latent Diffusion Models](https://arxiv.org/abs/2112.10752) paper.
+            kwargs (remaining dictionary of keyword arguments, *optional*):
+                Can be used to overwrite load and saveable variables (for example the pipeline components of the
+                specific pipeline class). The overwritten components are directly passed to the pipelines `__init__`
+                method. See example below for more information.
+
+        <Tip warning={true}>
+
+            Make sure to pass both `image_size` and `scaling_factor` to `from_single_file()` if you're loading
+            a VAE from SDXL or a Stable Diffusion v2 model or higher.
+
+        </Tip>
+
+        Examples:
+
+        ```py
+        from diffusers import AutoencoderKL
+
+        url = "https://huggingface.co/stabilityai/sd-vae-ft-mse-original/blob/main/vae-ft-mse-840000-ema-pruned.safetensors"  # can also be local file
+        model = AutoencoderKL.from_single_file(url)
+        ```
+        """
+        if not is_omegaconf_available():
+            raise ValueError(BACKENDS_MAPPING["omegaconf"][1])
+
+        from omegaconf import OmegaConf
+
+        from ..models import AutoencoderKL
+
+        # import here to avoid circular dependency
+        from ..pipelines.stable_diffusion.convert_from_ckpt import (
+            convert_ldm_vae_checkpoint,
+            create_vae_diffusers_config,
+        )
+
+        config_file = kwargs.pop("config_file", None)
+        cache_dir = kwargs.pop("cache_dir", None)
+        resume_download = kwargs.pop("resume_download", False)
+        force_download = kwargs.pop("force_download", False)
+        proxies = kwargs.pop("proxies", None)
+        local_files_only = kwargs.pop("local_files_only", None)
+        token = kwargs.pop("token", None)
+        revision = kwargs.pop("revision", None)
+        image_size = kwargs.pop("image_size", None)
+        scaling_factor = kwargs.pop("scaling_factor", None)
+        kwargs.pop("upcast_attention", None)
+
+        torch_dtype = kwargs.pop("torch_dtype", None)
+
+        use_safetensors = kwargs.pop("use_safetensors", None)
+
+        file_extension = pretrained_model_link_or_path.rsplit(".", 1)[-1]
+        from_safetensors = file_extension == "safetensors"
+
+        if from_safetensors and use_safetensors is False:
+            raise ValueError("Make sure to install `safetensors` with `pip install safetensors`.")
+
+        # remove huggingface url
+        for prefix in ["https://huggingface.co/", "huggingface.co/", "hf.co/", "https://hf.co/"]:
+            if pretrained_model_link_or_path.startswith(prefix):
+                pretrained_model_link_or_path = pretrained_model_link_or_path[len(prefix) :]
+
+        # Code based on diffusers.pipelines.pipeline_utils.DiffusionPipeline.from_pretrained
+        ckpt_path = Path(pretrained_model_link_or_path)
+        if not ckpt_path.is_file():
+            # get repo_id and (potentially nested) file path of ckpt in repo
+            repo_id = "/".join(ckpt_path.parts[:2])
+            file_path = "/".join(ckpt_path.parts[2:])
+
+            if file_path.startswith("blob/"):
+                file_path = file_path[len("blob/") :]
+
+            if file_path.startswith("main/"):
+                file_path = file_path[len("main/") :]
+
+            pretrained_model_link_or_path = hf_hub_download(
+                repo_id,
+                filename=file_path,
+                cache_dir=cache_dir,
+                resume_download=resume_download,
+                proxies=proxies,
+                local_files_only=local_files_only,
+                token=token,
+                revision=revision,
+                force_download=force_download,
+            )
+
+        if from_safetensors:
+            from safetensors import safe_open
+
+            checkpoint = {}
+            with safe_open(pretrained_model_link_or_path, framework="pt", device="cpu") as f:
+                for key in f.keys():
+                    checkpoint[key] = f.get_tensor(key)
+        else:
+            checkpoint = torch.load(pretrained_model_link_or_path, map_location="cpu")
+
+        if "state_dict" in checkpoint:
+            checkpoint = checkpoint["state_dict"]
+
+        if config_file is None:
+            config_url = "https://raw.githubusercontent.com/CompVis/stable-diffusion/main/configs/stable-diffusion/v1-inference.yaml"
+            config_file = BytesIO(requests.get(config_url).content)
+
+        original_config = OmegaConf.load(config_file)
+
+        # default to sd-v1-5
+        image_size = image_size or 512
+
+        vae_config = create_vae_diffusers_config(original_config, image_size=image_size)
+        converted_vae_checkpoint = convert_ldm_vae_checkpoint(checkpoint, vae_config)
+
+        if scaling_factor is None:
+            if (
+                "model" in original_config
+                and "params" in original_config.model
+                and "scale_factor" in original_config.model.params
+            ):
+                vae_scaling_factor = original_config.model.params.scale_factor
+            else:
+                vae_scaling_factor = 0.18215  # default SD scaling factor
+
+        vae_config["scaling_factor"] = vae_scaling_factor
+
+        ctx = init_empty_weights if is_accelerate_available() else nullcontext
+        with ctx():
+            vae = AutoencoderKL(**vae_config)
+
+        if is_accelerate_available():
+            from ..models.modeling_utils import load_model_dict_into_meta
+
+            load_model_dict_into_meta(vae, converted_vae_checkpoint, device="cpu")
+        else:
+            vae.load_state_dict(converted_vae_checkpoint)
+
+        if torch_dtype is not None:
+            vae.to(dtype=torch_dtype)
+
+        return vae
+
+
+class FromOriginalControlnetMixin:
+    """
+    Load pretrained ControlNet weights saved in the `.ckpt` or `.safetensors` format into a [`ControlNetModel`].
+    """
+
+    @classmethod
+    @validate_hf_hub_args
+    def from_single_file(cls, pretrained_model_link_or_path, **kwargs):
+        r"""
+        Instantiate a [`ControlNetModel`] from pretrained ControlNet weights saved in the original `.ckpt` or
+        `.safetensors` format. The pipeline is set in evaluation mode (`model.eval()`) by default.
+
+        Parameters:
+            pretrained_model_link_or_path (`str` or `os.PathLike`, *optional*):
+                Can be either:
+                    - A link to the `.ckpt` file (for example
+                      `"https://huggingface.co/<repo_id>/blob/main/<path_to_file>.ckpt"`) on the Hub.
+                    - A path to a *file* containing all pipeline weights.
+            torch_dtype (`str` or `torch.dtype`, *optional*):
+                Override the default `torch.dtype` and load the model with another dtype. If `"auto"` is passed, the
+                dtype is automatically derived from the model's weights.
+            force_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to force the (re-)download of the model weights and configuration files, overriding the
+                cached versions if they exist.
+            cache_dir (`Union[str, os.PathLike]`, *optional*):
+                Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
+                is not used.
+            resume_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to resume downloading the model weights and configuration files. If set to `False`, any
+                incompletely downloaded files are deleted.
+            proxies (`Dict[str, str]`, *optional*):
+                A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
+                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
+            local_files_only (`bool`, *optional*, defaults to `False`):
+                Whether to only load local model weights and configuration files or not. If set to True, the model
+                won't be downloaded from the Hub.
+            token (`str` or *bool*, *optional*):
+                The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
+                `diffusers-cli login` (stored in `~/.huggingface`) is used.
+            revision (`str`, *optional*, defaults to `"main"`):
+                The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
+                allowed by Git.
+            use_safetensors (`bool`, *optional*, defaults to `None`):
+                If set to `None`, the safetensors weights are downloaded if they're available **and** if the
+                safetensors library is installed. If set to `True`, the model is forcibly loaded from safetensors
+                weights. If set to `False`, safetensors weights are not loaded.
+            image_size (`int`, *optional*, defaults to 512):
+                The image size the model was trained on. Use 512 for all Stable Diffusion v1 models and the Stable
+                Diffusion v2 base model. Use 768 for Stable Diffusion v2.
+            upcast_attention (`bool`, *optional*, defaults to `None`):
+                Whether the attention computation should always be upcasted.
+            kwargs (remaining dictionary of keyword arguments, *optional*):
+                Can be used to overwrite load and saveable variables (for example the pipeline components of the
+                specific pipeline class). The overwritten components are directly passed to the pipelines `__init__`
+                method. See example below for more information.
+
+        Examples:
+
+        ```py
+        from diffusers import StableDiffusionControlNetPipeline, ControlNetModel
+
+        url = "https://huggingface.co/lllyasviel/ControlNet-v1-1/blob/main/control_v11p_sd15_canny.pth"  # can also be a local path
+        model = ControlNetModel.from_single_file(url)
+
+        url = "https://huggingface.co/runwayml/stable-diffusion-v1-5/blob/main/v1-5-pruned.safetensors"  # can also be a local path
+        pipe = StableDiffusionControlNetPipeline.from_single_file(url, controlnet=controlnet)
+        ```
+        """
+        # import here to avoid circular dependency
+        from ..pipelines.stable_diffusion.convert_from_ckpt import download_controlnet_from_original_ckpt
+
+        config_file = kwargs.pop("config_file", None)
+        cache_dir = kwargs.pop("cache_dir", None)
+        resume_download = kwargs.pop("resume_download", False)
+        force_download = kwargs.pop("force_download", False)
+        proxies = kwargs.pop("proxies", None)
+        local_files_only = kwargs.pop("local_files_only", None)
+        token = kwargs.pop("token", None)
+        num_in_channels = kwargs.pop("num_in_channels", None)
+        use_linear_projection = kwargs.pop("use_linear_projection", None)
+        revision = kwargs.pop("revision", None)
+        extract_ema = kwargs.pop("extract_ema", False)
+        image_size = kwargs.pop("image_size", None)
+        upcast_attention = kwargs.pop("upcast_attention", None)
+
+        torch_dtype = kwargs.pop("torch_dtype", None)
+
+        use_safetensors = kwargs.pop("use_safetensors", None)
+
+        file_extension = pretrained_model_link_or_path.rsplit(".", 1)[-1]
+        from_safetensors = file_extension == "safetensors"
+
+        if from_safetensors and use_safetensors is False:
+            raise ValueError("Make sure to install `safetensors` with `pip install safetensors`.")
+
+        # remove huggingface url
+        for prefix in ["https://huggingface.co/", "huggingface.co/", "hf.co/", "https://hf.co/"]:
+            if pretrained_model_link_or_path.startswith(prefix):
+                pretrained_model_link_or_path = pretrained_model_link_or_path[len(prefix) :]
+
+        # Code based on diffusers.pipelines.pipeline_utils.DiffusionPipeline.from_pretrained
+        ckpt_path = Path(pretrained_model_link_or_path)
+        if not ckpt_path.is_file():
+            # get repo_id and (potentially nested) file path of ckpt in repo
+            repo_id = "/".join(ckpt_path.parts[:2])
+            file_path = "/".join(ckpt_path.parts[2:])
+
+            if file_path.startswith("blob/"):
+                file_path = file_path[len("blob/") :]
+
+            if file_path.startswith("main/"):
+                file_path = file_path[len("main/") :]
+
+            pretrained_model_link_or_path = hf_hub_download(
+                repo_id,
+                filename=file_path,
+                cache_dir=cache_dir,
+                resume_download=resume_download,
+                proxies=proxies,
+                local_files_only=local_files_only,
+                token=token,
+                revision=revision,
+                force_download=force_download,
+            )
+
+        if config_file is None:
+            config_url = "https://raw.githubusercontent.com/lllyasviel/ControlNet/main/models/cldm_v15.yaml"
+            config_file = BytesIO(requests.get(config_url).content)
+
+        image_size = image_size or 512
+
+        controlnet = download_controlnet_from_original_ckpt(
+            pretrained_model_link_or_path,
+            original_config_file=config_file,
+            image_size=image_size,
+            extract_ema=extract_ema,
+            num_in_channels=num_in_channels,
+            upcast_attention=upcast_attention,
+            from_safetensors=from_safetensors,
+            use_linear_projection=use_linear_projection,
+        )
+
+        if torch_dtype is not None:
+            controlnet.to(dtype=torch_dtype)
+
+        return controlnet

src/diffusers/loaders/textual_inversion.py

@@ -0,0 +1,455 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import Dict, List, Optional, Union
+
+import safetensors
+import torch
+from huggingface_hub.utils import validate_hf_hub_args
+from torch import nn
+
+from ..utils import _get_model_file, is_accelerate_available, is_transformers_available, logging
+
+
+if is_transformers_available():
+    from transformers import PreTrainedModel, PreTrainedTokenizer
+
+if is_accelerate_available():
+    from accelerate.hooks import AlignDevicesHook, CpuOffload, remove_hook_from_module
+
+logger = logging.get_logger(__name__)
+
+TEXT_INVERSION_NAME = "learned_embeds.bin"
+TEXT_INVERSION_NAME_SAFE = "learned_embeds.safetensors"
+
+
+@validate_hf_hub_args
+def load_textual_inversion_state_dicts(pretrained_model_name_or_paths, **kwargs):
+    cache_dir = kwargs.pop("cache_dir", None)
+    force_download = kwargs.pop("force_download", False)
+    resume_download = kwargs.pop("resume_download", False)
+    proxies = kwargs.pop("proxies", None)
+    local_files_only = kwargs.pop("local_files_only", None)
+    token = kwargs.pop("token", None)
+    revision = kwargs.pop("revision", None)
+    subfolder = kwargs.pop("subfolder", None)
+    weight_name = kwargs.pop("weight_name", None)
+    use_safetensors = kwargs.pop("use_safetensors", None)
+
+    allow_pickle = False
+    if use_safetensors is None:
+        use_safetensors = True
+        allow_pickle = True
+
+    user_agent = {
+        "file_type": "text_inversion",
+        "framework": "pytorch",
+    }
+    state_dicts = []
+    for pretrained_model_name_or_path in pretrained_model_name_or_paths:
+        if not isinstance(pretrained_model_name_or_path, (dict, torch.Tensor)):
+            # 3.1. Load textual inversion file
+            model_file = None
+
+            # Let's first try to load .safetensors weights
+            if (use_safetensors and weight_name is None) or (
+                weight_name is not None and weight_name.endswith(".safetensors")
+            ):
+                try:
+                    model_file = _get_model_file(
+                        pretrained_model_name_or_path,
+                        weights_name=weight_name or TEXT_INVERSION_NAME_SAFE,
+                        cache_dir=cache_dir,
+                        force_download=force_download,
+                        resume_download=resume_download,
+                        proxies=proxies,
+                        local_files_only=local_files_only,
+                        token=token,
+                        revision=revision,
+                        subfolder=subfolder,
+                        user_agent=user_agent,
+                    )
+                    state_dict = safetensors.torch.load_file(model_file, device="cpu")
+                except Exception as e:
+                    if not allow_pickle:
+                        raise e
+
+                    model_file = None
+
+            if model_file is None:
+                model_file = _get_model_file(
+                    pretrained_model_name_or_path,
+                    weights_name=weight_name or TEXT_INVERSION_NAME,
+                    cache_dir=cache_dir,
+                    force_download=force_download,
+                    resume_download=resume_download,
+                    proxies=proxies,
+                    local_files_only=local_files_only,
+                    token=token,
+                    revision=revision,
+                    subfolder=subfolder,
+                    user_agent=user_agent,
+                )
+                state_dict = torch.load(model_file, map_location="cpu")
+        else:
+            state_dict = pretrained_model_name_or_path
+
+        state_dicts.append(state_dict)
+
+    return state_dicts
+
+
+class TextualInversionLoaderMixin:
+    r"""
+    Load Textual Inversion tokens and embeddings to the tokenizer and text encoder.
+    """
+
+    def maybe_convert_prompt(self, prompt: Union[str, List[str]], tokenizer: "PreTrainedTokenizer"):  # noqa: F821
+        r"""
+        Processes prompts that include a special token corresponding to a multi-vector textual inversion embedding to
+        be replaced with multiple special tokens each corresponding to one of the vectors. If the prompt has no textual
+        inversion token or if the textual inversion token is a single vector, the input prompt is returned.
+
+        Parameters:
+            prompt (`str` or list of `str`):
+                The prompt or prompts to guide the image generation.
+            tokenizer (`PreTrainedTokenizer`):
+                The tokenizer responsible for encoding the prompt into input tokens.
+
+        Returns:
+            `str` or list of `str`: The converted prompt
+        """
+        if not isinstance(prompt, List):
+            prompts = [prompt]
+        else:
+            prompts = prompt
+
+        prompts = [self._maybe_convert_prompt(p, tokenizer) for p in prompts]
+
+        if not isinstance(prompt, List):
+            return prompts[0]
+
+        return prompts
+
+    def _maybe_convert_prompt(self, prompt: str, tokenizer: "PreTrainedTokenizer"):  # noqa: F821
+        r"""
+        Maybe convert a prompt into a "multi vector"-compatible prompt. If the prompt includes a token that corresponds
+        to a multi-vector textual inversion embedding, this function will process the prompt so that the special token
+        is replaced with multiple special tokens each corresponding to one of the vectors. If the prompt has no textual
+        inversion token or a textual inversion token that is a single vector, the input prompt is simply returned.
+
+        Parameters:
+            prompt (`str`):
+                The prompt to guide the image generation.
+            tokenizer (`PreTrainedTokenizer`):
+                The tokenizer responsible for encoding the prompt into input tokens.
+
+        Returns:
+            `str`: The converted prompt
+        """
+        tokens = tokenizer.tokenize(prompt)
+        unique_tokens = set(tokens)
+        for token in unique_tokens:
+            if token in tokenizer.added_tokens_encoder:
+                replacement = token
+                i = 1
+                while f"{token}_{i}" in tokenizer.added_tokens_encoder:
+                    replacement += f" {token}_{i}"
+                    i += 1
+
+                prompt = prompt.replace(token, replacement)
+
+        return prompt
+
+    def _check_text_inv_inputs(self, tokenizer, text_encoder, pretrained_model_name_or_paths, tokens):
+        if tokenizer is None:
+            raise ValueError(
+                f"{self.__class__.__name__} requires `self.tokenizer` or passing a `tokenizer` of type `PreTrainedTokenizer` for calling"
+                f" `{self.load_textual_inversion.__name__}`"
+            )
+
+        if text_encoder is None:
+            raise ValueError(
+                f"{self.__class__.__name__} requires `self.text_encoder` or passing a `text_encoder` of type `PreTrainedModel` for calling"
+                f" `{self.load_textual_inversion.__name__}`"
+            )
+
+        if len(pretrained_model_name_or_paths) > 1 and len(pretrained_model_name_or_paths) != len(tokens):
+            raise ValueError(
+                f"You have passed a list of models of length {len(pretrained_model_name_or_paths)}, and list of tokens of length {len(tokens)} "
+                f"Make sure both lists have the same length."
+            )
+
+        valid_tokens = [t for t in tokens if t is not None]
+        if len(set(valid_tokens)) < len(valid_tokens):
+            raise ValueError(f"You have passed a list of tokens that contains duplicates: {tokens}")
+
+    @staticmethod
+    def _retrieve_tokens_and_embeddings(tokens, state_dicts, tokenizer):
+        all_tokens = []
+        all_embeddings = []
+        for state_dict, token in zip(state_dicts, tokens):
+            if isinstance(state_dict, torch.Tensor):
+                if token is None:
+                    raise ValueError(
+                        "You are trying to load a textual inversion embedding that has been saved as a PyTorch tensor. Make sure to pass the name of the corresponding token in this case: `token=...`."
+                    )
+                loaded_token = token
+                embedding = state_dict
+            elif len(state_dict) == 1:
+                # diffusers
+                loaded_token, embedding = next(iter(state_dict.items()))
+            elif "string_to_param" in state_dict:
+                # A1111
+                loaded_token = state_dict["name"]
+                embedding = state_dict["string_to_param"]["*"]
+            else:
+                raise ValueError(
+                    f"Loaded state dictonary is incorrect: {state_dict}. \n\n"
+                    "Please verify that the loaded state dictionary of the textual embedding either only has a single key or includes the `string_to_param`"
+                    " input key."
+                )
+
+            if token is not None and loaded_token != token:
+                logger.info(f"The loaded token: {loaded_token} is overwritten by the passed token {token}.")
+            else:
+                token = loaded_token
+
+            if token in tokenizer.get_vocab():
+                raise ValueError(
+                    f"Token {token} already in tokenizer vocabulary. Please choose a different token name or remove {token} and embedding from the tokenizer and text encoder."
+                )
+
+            all_tokens.append(token)
+            all_embeddings.append(embedding)
+
+        return all_tokens, all_embeddings
+
+    @staticmethod
+    def _extend_tokens_and_embeddings(tokens, embeddings, tokenizer):
+        all_tokens = []
+        all_embeddings = []
+
+        for embedding, token in zip(embeddings, tokens):
+            if f"{token}_1" in tokenizer.get_vocab():
+                multi_vector_tokens = [token]
+                i = 1
+                while f"{token}_{i}" in tokenizer.added_tokens_encoder:
+                    multi_vector_tokens.append(f"{token}_{i}")
+                    i += 1
+
+                raise ValueError(
+                    f"Multi-vector Token {multi_vector_tokens} already in tokenizer vocabulary. Please choose a different token name or remove the {multi_vector_tokens} and embedding from the tokenizer and text encoder."
+                )
+
+            is_multi_vector = len(embedding.shape) > 1 and embedding.shape[0] > 1
+            if is_multi_vector:
+                all_tokens += [token] + [f"{token}_{i}" for i in range(1, embedding.shape[0])]
+                all_embeddings += [e for e in embedding]  # noqa: C416
+            else:
+                all_tokens += [token]
+                all_embeddings += [embedding[0]] if len(embedding.shape) > 1 else [embedding]
+
+        return all_tokens, all_embeddings
+
+    @validate_hf_hub_args
+    def load_textual_inversion(
+        self,
+        pretrained_model_name_or_path: Union[str, List[str], Dict[str, torch.Tensor], List[Dict[str, torch.Tensor]]],
+        token: Optional[Union[str, List[str]]] = None,
+        tokenizer: Optional["PreTrainedTokenizer"] = None,  # noqa: F821
+        text_encoder: Optional["PreTrainedModel"] = None,  # noqa: F821
+        **kwargs,
+    ):
+        r"""
+        Load Textual Inversion embeddings into the text encoder of [`StableDiffusionPipeline`] (both 🤗 Diffusers and
+        Automatic1111 formats are supported).
+
+        Parameters:
+            pretrained_model_name_or_path (`str` or `os.PathLike` or `List[str or os.PathLike]` or `Dict` or `List[Dict]`):
+                Can be either one of the following or a list of them:
+
+                    - A string, the *model id* (for example `sd-concepts-library/low-poly-hd-logos-icons`) of a
+                      pretrained model hosted on the Hub.
+                    - A path to a *directory* (for example `./my_text_inversion_directory/`) containing the textual
+                      inversion weights.
+                    - A path to a *file* (for example `./my_text_inversions.pt`) containing textual inversion weights.
+                    - A [torch state
+                      dict](https://pytorch.org/tutorials/beginner/saving_loading_models.html#what-is-a-state-dict).
+
+            token (`str` or `List[str]`, *optional*):
+                Override the token to use for the textual inversion weights. If `pretrained_model_name_or_path` is a
+                list, then `token` must also be a list of equal length.
+            text_encoder ([`~transformers.CLIPTextModel`], *optional*):
+                Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
+                If not specified, function will take self.tokenizer.
+            tokenizer ([`~transformers.CLIPTokenizer`], *optional*):
+                A `CLIPTokenizer` to tokenize text. If not specified, function will take self.tokenizer.
+            weight_name (`str`, *optional*):
+                Name of a custom weight file. This should be used when:
+
+                    - The saved textual inversion file is in 🤗 Diffusers format, but was saved under a specific weight
+                      name such as `text_inv.bin`.
+                    - The saved textual inversion file is in the Automatic1111 format.
+            cache_dir (`Union[str, os.PathLike]`, *optional*):
+                Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
+                is not used.
+            force_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to force the (re-)download of the model weights and configuration files, overriding the
+                cached versions if they exist.
+            resume_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to resume downloading the model weights and configuration files. If set to `False`, any
+                incompletely downloaded files are deleted.
+            proxies (`Dict[str, str]`, *optional*):
+                A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
+                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
+            local_files_only (`bool`, *optional*, defaults to `False`):
+                Whether to only load local model weights and configuration files or not. If set to `True`, the model
+                won't be downloaded from the Hub.
+            token (`str` or *bool*, *optional*):
+                The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
+                `diffusers-cli login` (stored in `~/.huggingface`) is used.
+            revision (`str`, *optional*, defaults to `"main"`):
+                The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
+                allowed by Git.
+            subfolder (`str`, *optional*, defaults to `""`):
+                The subfolder location of a model file within a larger model repository on the Hub or locally.
+            mirror (`str`, *optional*):
+                Mirror source to resolve accessibility issues if you're downloading a model in China. We do not
+                guarantee the timeliness or safety of the source, and you should refer to the mirror site for more
+                information.
+
+        Example:
+
+        To load a Textual Inversion embedding vector in 🤗 Diffusers format:
+
+        ```py
+        from diffusers import StableDiffusionPipeline
+        import torch
+
+        model_id = "runwayml/stable-diffusion-v1-5"
+        pipe = StableDiffusionPipeline.from_pretrained(model_id, torch_dtype=torch.float16).to("cuda")
+
+        pipe.load_textual_inversion("sd-concepts-library/cat-toy")
+
+        prompt = "A <cat-toy> backpack"
+
+        image = pipe(prompt, num_inference_steps=50).images[0]
+        image.save("cat-backpack.png")
+        ```
+
+        To load a Textual Inversion embedding vector in Automatic1111 format, make sure to download the vector first
+        (for example from [civitAI](https://civitai.com/models/3036?modelVersionId=9857)) and then load the vector
+        locally:
+
+        ```py
+        from diffusers import StableDiffusionPipeline
+        import torch
+
+        model_id = "runwayml/stable-diffusion-v1-5"
+        pipe = StableDiffusionPipeline.from_pretrained(model_id, torch_dtype=torch.float16).to("cuda")
+
+        pipe.load_textual_inversion("./charturnerv2.pt", token="charturnerv2")
+
+        prompt = "charturnerv2, multiple views of the same character in the same outfit, a character turnaround of a woman wearing a black jacket and red shirt, best quality, intricate details."
+
+        image = pipe(prompt, num_inference_steps=50).images[0]
+        image.save("character.png")
+        ```
+
+        """
+        # 1. Set correct tokenizer and text encoder
+        tokenizer = tokenizer or getattr(self, "tokenizer", None)
+        text_encoder = text_encoder or getattr(self, "text_encoder", None)
+
+        # 2. Normalize inputs
+        pretrained_model_name_or_paths = (
+            [pretrained_model_name_or_path]
+            if not isinstance(pretrained_model_name_or_path, list)
+            else pretrained_model_name_or_path
+        )
+        tokens = [token] if not isinstance(token, list) else token
+        if tokens[0] is None:
+            tokens = tokens * len(pretrained_model_name_or_paths)
+
+        # 3. Check inputs
+        self._check_text_inv_inputs(tokenizer, text_encoder, pretrained_model_name_or_paths, tokens)
+
+        # 4. Load state dicts of textual embeddings
+        state_dicts = load_textual_inversion_state_dicts(pretrained_model_name_or_paths, **kwargs)
+
+        # 4.1 Handle the special case when state_dict is a tensor that contains n embeddings for n tokens
+        if len(tokens) > 1 and len(state_dicts) == 1:
+            if isinstance(state_dicts[0], torch.Tensor):
+                state_dicts = list(state_dicts[0])
+                if len(tokens) != len(state_dicts):
+                    raise ValueError(
+                        f"You have passed a state_dict contains {len(state_dicts)} embeddings, and list of tokens of length {len(tokens)} "
+                        f"Make sure both have the same length."
+                    )
+
+        # 4. Retrieve tokens and embeddings
+        tokens, embeddings = self._retrieve_tokens_and_embeddings(tokens, state_dicts, tokenizer)
+
+        # 5. Extend tokens and embeddings for multi vector
+        tokens, embeddings = self._extend_tokens_and_embeddings(tokens, embeddings, tokenizer)
+
+        # 6. Make sure all embeddings have the correct size
+        expected_emb_dim = text_encoder.get_input_embeddings().weight.shape[-1]
+        if any(expected_emb_dim != emb.shape[-1] for emb in embeddings):
+            raise ValueError(
+                "Loaded embeddings are of incorrect shape. Expected each textual inversion embedding "
+                "to be of shape {input_embeddings.shape[-1]}, but are {embeddings.shape[-1]} "
+            )
+
+        # 7. Now we can be sure that loading the embedding matrix works
+        # < Unsafe code:
+
+        # 7.1 Offload all hooks in case the pipeline was cpu offloaded before make sure, we offload and onload again
+        is_model_cpu_offload = False
+        is_sequential_cpu_offload = False
+        for _, component in self.components.items():
+            if isinstance(component, nn.Module):
+                if hasattr(component, "_hf_hook"):
+                    is_model_cpu_offload = isinstance(getattr(component, "_hf_hook"), CpuOffload)
+                    is_sequential_cpu_offload = isinstance(getattr(component, "_hf_hook"), AlignDevicesHook)
+                    logger.info(
+                        "Accelerate hooks detected. Since you have called `load_textual_inversion()`, the previous hooks will be first removed. Then the textual inversion parameters will be loaded and the hooks will be applied again."
+                    )
+                    remove_hook_from_module(component, recurse=is_sequential_cpu_offload)
+
+        # 7.2 save expected device and dtype
+        device = text_encoder.device
+        dtype = text_encoder.dtype
+
+        # 7.3 Increase token embedding matrix
+        text_encoder.resize_token_embeddings(len(tokenizer) + len(tokens))
+        input_embeddings = text_encoder.get_input_embeddings().weight
+
+        # 7.4 Load token and embedding
+        for token, embedding in zip(tokens, embeddings):
+            # add tokens and get ids
+            tokenizer.add_tokens(token)
+            token_id = tokenizer.convert_tokens_to_ids(token)
+            input_embeddings.data[token_id] = embedding
+            logger.info(f"Loaded textual inversion embedding for {token}.")
+
+        input_embeddings.to(dtype=dtype, device=device)
+
+        # 7.5 Offload the model again
+        if is_model_cpu_offload:
+            self.enable_model_cpu_offload()
+        elif is_sequential_cpu_offload:
+            self.enable_sequential_cpu_offload()
+
+        # / Unsafe Code >

src/diffusers/loaders/unet.py

@@ -0,0 +1,828 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import inspect
+import os
+from collections import defaultdict
+from contextlib import nullcontext
+from functools import partial
+from typing import Callable, Dict, List, Optional, Union
+
+import safetensors
+import torch
+import torch.nn.functional as F
+from huggingface_hub.utils import validate_hf_hub_args
+from torch import nn
+
+from ..models.embeddings import ImageProjection, MLPProjection, Resampler
+from ..models.modeling_utils import _LOW_CPU_MEM_USAGE_DEFAULT, load_model_dict_into_meta
+from ..utils import (
+    USE_PEFT_BACKEND,
+    _get_model_file,
+    delete_adapter_layers,
+    is_accelerate_available,
+    logging,
+    set_adapter_layers,
+    set_weights_and_activate_adapters,
+)
+from .utils import AttnProcsLayers
+
+
+if is_accelerate_available():
+    from accelerate import init_empty_weights
+    from accelerate.hooks import AlignDevicesHook, CpuOffload, remove_hook_from_module
+
+logger = logging.get_logger(__name__)
+
+
+TEXT_ENCODER_NAME = "text_encoder"
+UNET_NAME = "unet"
+
+LORA_WEIGHT_NAME = "pytorch_lora_weights.bin"
+LORA_WEIGHT_NAME_SAFE = "pytorch_lora_weights.safetensors"
+
+CUSTOM_DIFFUSION_WEIGHT_NAME = "pytorch_custom_diffusion_weights.bin"
+CUSTOM_DIFFUSION_WEIGHT_NAME_SAFE = "pytorch_custom_diffusion_weights.safetensors"
+
+
+class UNet2DConditionLoadersMixin:
+    """
+    Load LoRA layers into a [`UNet2DCondtionModel`].
+    """
+
+    text_encoder_name = TEXT_ENCODER_NAME
+    unet_name = UNET_NAME
+
+    @validate_hf_hub_args
+    def load_attn_procs(self, pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]], **kwargs):
+        r"""
+        Load pretrained attention processor layers into [`UNet2DConditionModel`]. Attention processor layers have to be
+        defined in
+        [`attention_processor.py`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py)
+        and be a `torch.nn.Module` class.
+
+        Parameters:
+            pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
+                Can be either:
+
+                    - A string, the model id (for example `google/ddpm-celebahq-256`) of a pretrained model hosted on
+                      the Hub.
+                    - A path to a directory (for example `./my_model_directory`) containing the model weights saved
+                      with [`ModelMixin.save_pretrained`].
+                    - A [torch state
+                      dict](https://pytorch.org/tutorials/beginner/saving_loading_models.html#what-is-a-state-dict).
+
+            cache_dir (`Union[str, os.PathLike]`, *optional*):
+                Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
+                is not used.
+            force_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to force the (re-)download of the model weights and configuration files, overriding the
+                cached versions if they exist.
+            resume_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to resume downloading the model weights and configuration files. If set to `False`, any
+                incompletely downloaded files are deleted.
+            proxies (`Dict[str, str]`, *optional*):
+                A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
+                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
+            local_files_only (`bool`, *optional*, defaults to `False`):
+                Whether to only load local model weights and configuration files or not. If set to `True`, the model
+                won't be downloaded from the Hub.
+            token (`str` or *bool*, *optional*):
+                The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
+                `diffusers-cli login` (stored in `~/.huggingface`) is used.
+            low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 else `False`):
+                Speed up model loading only loading the pretrained weights and not initializing the weights. This also
+                tries to not use more than 1x model size in CPU memory (including peak memory) while loading the model.
+                Only supported for PyTorch >= 1.9.0. If you are using an older version of PyTorch, setting this
+                argument to `True` will raise an error.
+            revision (`str`, *optional*, defaults to `"main"`):
+                The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
+                allowed by Git.
+            subfolder (`str`, *optional*, defaults to `""`):
+                The subfolder location of a model file within a larger model repository on the Hub or locally.
+            mirror (`str`, *optional*):
+                Mirror source to resolve accessibility issues if you’re downloading a model in China. We do not
+                guarantee the timeliness or safety of the source, and you should refer to the mirror site for more
+                information.
+
+        Example:
+
+        ```py
+        from diffusers import AutoPipelineForText2Image
+        import torch
+
+        pipeline = AutoPipelineForText2Image.from_pretrained(
+            "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16
+        ).to("cuda")
+        pipeline.unet.load_attn_procs(
+            "jbilcke-hf/sdxl-cinematic-1", weight_name="pytorch_lora_weights.safetensors", adapter_name="cinematic"
+        )
+        ```
+        """
+        from ..models.attention_processor import CustomDiffusionAttnProcessor
+        from ..models.lora import LoRACompatibleConv, LoRACompatibleLinear, LoRAConv2dLayer, LoRALinearLayer
+
+        cache_dir = kwargs.pop("cache_dir", None)
+        force_download = kwargs.pop("force_download", False)
+        resume_download = kwargs.pop("resume_download", False)
+        proxies = kwargs.pop("proxies", None)
+        local_files_only = kwargs.pop("local_files_only", None)
+        token = kwargs.pop("token", None)
+        revision = kwargs.pop("revision", None)
+        subfolder = kwargs.pop("subfolder", None)
+        weight_name = kwargs.pop("weight_name", None)
+        use_safetensors = kwargs.pop("use_safetensors", None)
+        low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", _LOW_CPU_MEM_USAGE_DEFAULT)
+        # This value has the same meaning as the `--network_alpha` option in the kohya-ss trainer script.
+        # See https://github.com/darkstorm2150/sd-scripts/blob/main/docs/train_network_README-en.md#execute-learning
+        network_alphas = kwargs.pop("network_alphas", None)
+
+        _pipeline = kwargs.pop("_pipeline", None)
+
+        is_network_alphas_none = network_alphas is None
+
+        allow_pickle = False
+
+        if use_safetensors is None:
+            use_safetensors = True
+            allow_pickle = True
+
+        user_agent = {
+            "file_type": "attn_procs_weights",
+            "framework": "pytorch",
+        }
+
+        if low_cpu_mem_usage and not is_accelerate_available():
+            low_cpu_mem_usage = False
+            logger.warning(
+                "Cannot initialize model with low cpu memory usage because `accelerate` was not found in the"
+                " environment. Defaulting to `low_cpu_mem_usage=False`. It is strongly recommended to install"
+                " `accelerate` for faster and less memory-intense model loading. You can do so with: \n```\npip"
+                " install accelerate\n```\n."
+            )
+
+        model_file = None
+        if not isinstance(pretrained_model_name_or_path_or_dict, dict):
+            # Let's first try to load .safetensors weights
+            if (use_safetensors and weight_name is None) or (
+                weight_name is not None and weight_name.endswith(".safetensors")
+            ):
+                try:
+                    model_file = _get_model_file(
+                        pretrained_model_name_or_path_or_dict,
+                        weights_name=weight_name or LORA_WEIGHT_NAME_SAFE,
+                        cache_dir=cache_dir,
+                        force_download=force_download,
+                        resume_download=resume_download,
+                        proxies=proxies,
+                        local_files_only=local_files_only,
+                        token=token,
+                        revision=revision,
+                        subfolder=subfolder,
+                        user_agent=user_agent,
+                    )
+                    state_dict = safetensors.torch.load_file(model_file, device="cpu")
+                except IOError as e:
+                    if not allow_pickle:
+                        raise e
+                    # try loading non-safetensors weights
+                    pass
+            if model_file is None:
+                model_file = _get_model_file(
+                    pretrained_model_name_or_path_or_dict,
+                    weights_name=weight_name or LORA_WEIGHT_NAME,
+                    cache_dir=cache_dir,
+                    force_download=force_download,
+                    resume_download=resume_download,
+                    proxies=proxies,
+                    local_files_only=local_files_only,
+                    token=token,
+                    revision=revision,
+                    subfolder=subfolder,
+                    user_agent=user_agent,
+                )
+                state_dict = torch.load(model_file, map_location="cpu")
+        else:
+            state_dict = pretrained_model_name_or_path_or_dict
+
+        # fill attn processors
+        lora_layers_list = []
+
+        is_lora = all(("lora" in k or k.endswith(".alpha")) for k in state_dict.keys()) and not USE_PEFT_BACKEND
+        is_custom_diffusion = any("custom_diffusion" in k for k in state_dict.keys())
+
+        if is_lora:
+            # correct keys
+            state_dict, network_alphas = self.convert_state_dict_legacy_attn_format(state_dict, network_alphas)
+
+            if network_alphas is not None:
+                network_alphas_keys = list(network_alphas.keys())
+                used_network_alphas_keys = set()
+
+            lora_grouped_dict = defaultdict(dict)
+            mapped_network_alphas = {}
+
+            all_keys = list(state_dict.keys())
+            for key in all_keys:
+                value = state_dict.pop(key)
+                attn_processor_key, sub_key = ".".join(key.split(".")[:-3]), ".".join(key.split(".")[-3:])
+                lora_grouped_dict[attn_processor_key][sub_key] = value
+
+                # Create another `mapped_network_alphas` dictionary so that we can properly map them.
+                if network_alphas is not None:
+                    for k in network_alphas_keys:
+                        if k.replace(".alpha", "") in key:
+                            mapped_network_alphas.update({attn_processor_key: network_alphas.get(k)})
+                            used_network_alphas_keys.add(k)
+
+            if not is_network_alphas_none:
+                if len(set(network_alphas_keys) - used_network_alphas_keys) > 0:
+                    raise ValueError(
+                        f"The `network_alphas` has to be empty at this point but has the following keys \n\n {', '.join(network_alphas.keys())}"
+                    )
+
+            if len(state_dict) > 0:
+                raise ValueError(
+                    f"The `state_dict` has to be empty at this point but has the following keys \n\n {', '.join(state_dict.keys())}"
+                )
+
+            for key, value_dict in lora_grouped_dict.items():
+                attn_processor = self
+                for sub_key in key.split("."):
+                    attn_processor = getattr(attn_processor, sub_key)
+
+                # Process non-attention layers, which don't have to_{k,v,q,out_proj}_lora layers
+                # or add_{k,v,q,out_proj}_proj_lora layers.
+                rank = value_dict["lora.down.weight"].shape[0]
+
+                if isinstance(attn_processor, LoRACompatibleConv):
+                    in_features = attn_processor.in_channels
+                    out_features = attn_processor.out_channels
+                    kernel_size = attn_processor.kernel_size
+
+                    ctx = init_empty_weights if low_cpu_mem_usage else nullcontext
+                    with ctx():
+                        lora = LoRAConv2dLayer(
+                            in_features=in_features,
+                            out_features=out_features,
+                            rank=rank,
+                            kernel_size=kernel_size,
+                            stride=attn_processor.stride,
+                            padding=attn_processor.padding,
+                            network_alpha=mapped_network_alphas.get(key),
+                        )
+                elif isinstance(attn_processor, LoRACompatibleLinear):
+                    ctx = init_empty_weights if low_cpu_mem_usage else nullcontext
+                    with ctx():
+                        lora = LoRALinearLayer(
+                            attn_processor.in_features,
+                            attn_processor.out_features,
+                            rank,
+                            mapped_network_alphas.get(key),
+                        )
+                else:
+                    raise ValueError(f"Module {key} is not a LoRACompatibleConv or LoRACompatibleLinear module.")
+
+                value_dict = {k.replace("lora.", ""): v for k, v in value_dict.items()}
+                lora_layers_list.append((attn_processor, lora))
+
+                if low_cpu_mem_usage:
+                    device = next(iter(value_dict.values())).device
+                    dtype = next(iter(value_dict.values())).dtype
+                    load_model_dict_into_meta(lora, value_dict, device=device, dtype=dtype)
+                else:
+                    lora.load_state_dict(value_dict)
+
+        elif is_custom_diffusion:
+            attn_processors = {}
+            custom_diffusion_grouped_dict = defaultdict(dict)
+            for key, value in state_dict.items():
+                if len(value) == 0:
+                    custom_diffusion_grouped_dict[key] = {}
+                else:
+                    if "to_out" in key:
+                        attn_processor_key, sub_key = ".".join(key.split(".")[:-3]), ".".join(key.split(".")[-3:])
+                    else:
+                        attn_processor_key, sub_key = ".".join(key.split(".")[:-2]), ".".join(key.split(".")[-2:])
+                    custom_diffusion_grouped_dict[attn_processor_key][sub_key] = value
+
+            for key, value_dict in custom_diffusion_grouped_dict.items():
+                if len(value_dict) == 0:
+                    attn_processors[key] = CustomDiffusionAttnProcessor(
+                        train_kv=False, train_q_out=False, hidden_size=None, cross_attention_dim=None
+                    )
+                else:
+                    cross_attention_dim = value_dict["to_k_custom_diffusion.weight"].shape[1]
+                    hidden_size = value_dict["to_k_custom_diffusion.weight"].shape[0]
+                    train_q_out = True if "to_q_custom_diffusion.weight" in value_dict else False
+                    attn_processors[key] = CustomDiffusionAttnProcessor(
+                        train_kv=True,
+                        train_q_out=train_q_out,
+                        hidden_size=hidden_size,
+                        cross_attention_dim=cross_attention_dim,
+                    )
+                    attn_processors[key].load_state_dict(value_dict)
+        elif USE_PEFT_BACKEND:
+            # In that case we have nothing to do as loading the adapter weights is already handled above by `set_peft_model_state_dict`
+            # on the Unet
+            pass
+        else:
+            raise ValueError(
+                f"{model_file} does not seem to be in the correct format expected by LoRA or Custom Diffusion training."
+            )
+
+        # <Unsafe code
+        # We can be sure that the following works as it just sets attention processors, lora layers and puts all in the same dtype
+        # Now we remove any existing hooks to
+        is_model_cpu_offload = False
+        is_sequential_cpu_offload = False
+
+        # For PEFT backend the Unet is already offloaded at this stage as it is handled inside `lora_lora_weights_into_unet`
+        if not USE_PEFT_BACKEND:
+            if _pipeline is not None:
+                for _, component in _pipeline.components.items():
+                    if isinstance(component, nn.Module) and hasattr(component, "_hf_hook"):
+                        is_model_cpu_offload = isinstance(getattr(component, "_hf_hook"), CpuOffload)
+                        is_sequential_cpu_offload = isinstance(getattr(component, "_hf_hook"), AlignDevicesHook)
+
+                        logger.info(
+                            "Accelerate hooks detected. Since you have called `load_lora_weights()`, the previous hooks will be first removed. Then the LoRA parameters will be loaded and the hooks will be applied again."
+                        )
+                        remove_hook_from_module(component, recurse=is_sequential_cpu_offload)
+
+            # only custom diffusion needs to set attn processors
+            if is_custom_diffusion:
+                self.set_attn_processor(attn_processors)
+
+            # set lora layers
+            for target_module, lora_layer in lora_layers_list:
+                target_module.set_lora_layer(lora_layer)
+
+            self.to(dtype=self.dtype, device=self.device)
+
+            # Offload back.
+            if is_model_cpu_offload:
+                _pipeline.enable_model_cpu_offload()
+            elif is_sequential_cpu_offload:
+                _pipeline.enable_sequential_cpu_offload()
+            # Unsafe code />
+
+    def convert_state_dict_legacy_attn_format(self, state_dict, network_alphas):
+        is_new_lora_format = all(
+            key.startswith(self.unet_name) or key.startswith(self.text_encoder_name) for key in state_dict.keys()
+        )
+        if is_new_lora_format:
+            # Strip the `"unet"` prefix.
+            is_text_encoder_present = any(key.startswith(self.text_encoder_name) for key in state_dict.keys())
+            if is_text_encoder_present:
+                warn_message = "The state_dict contains LoRA params corresponding to the text encoder which are not being used here. To use both UNet and text encoder related LoRA params, use [`pipe.load_lora_weights()`](https://huggingface.co/docs/diffusers/main/en/api/loaders#diffusers.loaders.LoraLoaderMixin.load_lora_weights)."
+                logger.warn(warn_message)
+            unet_keys = [k for k in state_dict.keys() if k.startswith(self.unet_name)]
+            state_dict = {k.replace(f"{self.unet_name}.", ""): v for k, v in state_dict.items() if k in unet_keys}
+
+        # change processor format to 'pure' LoRACompatibleLinear format
+        if any("processor" in k.split(".") for k in state_dict.keys()):
+
+            def format_to_lora_compatible(key):
+                if "processor" not in key.split("."):
+                    return key
+                return key.replace(".processor", "").replace("to_out_lora", "to_out.0.lora").replace("_lora", ".lora")
+
+            state_dict = {format_to_lora_compatible(k): v for k, v in state_dict.items()}
+
+            if network_alphas is not None:
+                network_alphas = {format_to_lora_compatible(k): v for k, v in network_alphas.items()}
+        return state_dict, network_alphas
+
+    def save_attn_procs(
+        self,
+        save_directory: Union[str, os.PathLike],
+        is_main_process: bool = True,
+        weight_name: str = None,
+        save_function: Callable = None,
+        safe_serialization: bool = True,
+        **kwargs,
+    ):
+        r"""
+        Save attention processor layers to a directory so that it can be reloaded with the
+        [`~loaders.UNet2DConditionLoadersMixin.load_attn_procs`] method.
+
+        Arguments:
+            save_directory (`str` or `os.PathLike`):
+                Directory to save an attention processor to (will be created if it doesn't exist).
+            is_main_process (`bool`, *optional*, defaults to `True`):
+                Whether the process calling this is the main process or not. Useful during distributed training and you
+                need to call this function on all processes. In this case, set `is_main_process=True` only on the main
+                process to avoid race conditions.
+            save_function (`Callable`):
+                The function to use to save the state dictionary. Useful during distributed training when you need to
+                replace `torch.save` with another method. Can be configured with the environment variable
+                `DIFFUSERS_SAVE_MODE`.
+            safe_serialization (`bool`, *optional*, defaults to `True`):
+                Whether to save the model using `safetensors` or with `pickle`.
+
+        Example:
+
+        ```py
+        import torch
+        from diffusers import DiffusionPipeline
+
+        pipeline = DiffusionPipeline.from_pretrained(
+            "CompVis/stable-diffusion-v1-4",
+            torch_dtype=torch.float16,
+        ).to("cuda")
+        pipeline.unet.load_attn_procs("path-to-save-model", weight_name="pytorch_custom_diffusion_weights.bin")
+        pipeline.unet.save_attn_procs("path-to-save-model", weight_name="pytorch_custom_diffusion_weights.bin")
+        ```
+        """
+        from ..models.attention_processor import (
+            CustomDiffusionAttnProcessor,
+            CustomDiffusionAttnProcessor2_0,
+            CustomDiffusionXFormersAttnProcessor,
+        )
+
+        if os.path.isfile(save_directory):
+            logger.error(f"Provided path ({save_directory}) should be a directory, not a file")
+            return
+
+        if save_function is None:
+            if safe_serialization:
+
+                def save_function(weights, filename):
+                    return safetensors.torch.save_file(weights, filename, metadata={"format": "pt"})
+
+            else:
+                save_function = torch.save
+
+        os.makedirs(save_directory, exist_ok=True)
+
+        is_custom_diffusion = any(
+            isinstance(
+                x,
+                (CustomDiffusionAttnProcessor, CustomDiffusionAttnProcessor2_0, CustomDiffusionXFormersAttnProcessor),
+            )
+            for (_, x) in self.attn_processors.items()
+        )
+        if is_custom_diffusion:
+            model_to_save = AttnProcsLayers(
+                {
+                    y: x
+                    for (y, x) in self.attn_processors.items()
+                    if isinstance(
+                        x,
+                        (
+                            CustomDiffusionAttnProcessor,
+                            CustomDiffusionAttnProcessor2_0,
+                            CustomDiffusionXFormersAttnProcessor,
+                        ),
+                    )
+                }
+            )
+            state_dict = model_to_save.state_dict()
+            for name, attn in self.attn_processors.items():
+                if len(attn.state_dict()) == 0:
+                    state_dict[name] = {}
+        else:
+            model_to_save = AttnProcsLayers(self.attn_processors)
+            state_dict = model_to_save.state_dict()
+
+        if weight_name is None:
+            if safe_serialization:
+                weight_name = CUSTOM_DIFFUSION_WEIGHT_NAME_SAFE if is_custom_diffusion else LORA_WEIGHT_NAME_SAFE
+            else:
+                weight_name = CUSTOM_DIFFUSION_WEIGHT_NAME if is_custom_diffusion else LORA_WEIGHT_NAME
+
+        # Save the model
+        save_function(state_dict, os.path.join(save_directory, weight_name))
+        logger.info(f"Model weights saved in {os.path.join(save_directory, weight_name)}")
+
+    def fuse_lora(self, lora_scale=1.0, safe_fusing=False, adapter_names=None):
+        self.lora_scale = lora_scale
+        self._safe_fusing = safe_fusing
+        self.apply(partial(self._fuse_lora_apply, adapter_names=adapter_names))
+
+    def _fuse_lora_apply(self, module, adapter_names=None):
+        if not USE_PEFT_BACKEND:
+            if hasattr(module, "_fuse_lora"):
+                module._fuse_lora(self.lora_scale, self._safe_fusing)
+
+            if adapter_names is not None:
+                raise ValueError(
+                    "The `adapter_names` argument is not supported in your environment. Please switch"
+                    " to PEFT backend to use this argument by installing latest PEFT and transformers."
+                    " `pip install -U peft transformers`"
+                )
+        else:
+            from peft.tuners.tuners_utils import BaseTunerLayer
+
+            merge_kwargs = {"safe_merge": self._safe_fusing}
+
+            if isinstance(module, BaseTunerLayer):
+                if self.lora_scale != 1.0:
+                    module.scale_layer(self.lora_scale)
+
+                # For BC with prevous PEFT versions, we need to check the signature
+                # of the `merge` method to see if it supports the `adapter_names` argument.
+                supported_merge_kwargs = list(inspect.signature(module.merge).parameters)
+                if "adapter_names" in supported_merge_kwargs:
+                    merge_kwargs["adapter_names"] = adapter_names
+                elif "adapter_names" not in supported_merge_kwargs and adapter_names is not None:
+                    raise ValueError(
+                        "The `adapter_names` argument is not supported with your PEFT version. Please upgrade"
+                        " to the latest version of PEFT. `pip install -U peft`"
+                    )
+
+                module.merge(**merge_kwargs)
+
+    def unfuse_lora(self):
+        self.apply(self._unfuse_lora_apply)
+
+    def _unfuse_lora_apply(self, module):
+        if not USE_PEFT_BACKEND:
+            if hasattr(module, "_unfuse_lora"):
+                module._unfuse_lora()
+        else:
+            from peft.tuners.tuners_utils import BaseTunerLayer
+
+            if isinstance(module, BaseTunerLayer):
+                module.unmerge()
+
+    def set_adapters(
+        self,
+        adapter_names: Union[List[str], str],
+        weights: Optional[Union[List[float], float]] = None,
+    ):
+        """
+        Set the currently active adapters for use in the UNet.
+
+        Args:
+            adapter_names (`List[str]` or `str`):
+                The names of the adapters to use.
+            adapter_weights (`Union[List[float], float]`, *optional*):
+                The adapter(s) weights to use with the UNet. If `None`, the weights are set to `1.0` for all the
+                adapters.
+
+        Example:
+
+        ```py
+        from diffusers import AutoPipelineForText2Image
+        import torch
+
+        pipeline = AutoPipelineForText2Image.from_pretrained(
+            "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16
+        ).to("cuda")
+        pipeline.load_lora_weights(
+            "jbilcke-hf/sdxl-cinematic-1", weight_name="pytorch_lora_weights.safetensors", adapter_name="cinematic"
+        )
+        pipeline.load_lora_weights("nerijs/pixel-art-xl", weight_name="pixel-art-xl.safetensors", adapter_name="pixel")
+        pipeline.set_adapters(["cinematic", "pixel"], adapter_weights=[0.5, 0.5])
+        ```
+        """
+        if not USE_PEFT_BACKEND:
+            raise ValueError("PEFT backend is required for `set_adapters()`.")
+
+        adapter_names = [adapter_names] if isinstance(adapter_names, str) else adapter_names
+
+        if weights is None:
+            weights = [1.0] * len(adapter_names)
+        elif isinstance(weights, float):
+            weights = [weights] * len(adapter_names)
+
+        if len(adapter_names) != len(weights):
+            raise ValueError(
+                f"Length of adapter names {len(adapter_names)} is not equal to the length of their weights {len(weights)}."
+            )
+
+        set_weights_and_activate_adapters(self, adapter_names, weights)
+
+    def disable_lora(self):
+        """
+        Disable the UNet's active LoRA layers.
+
+        Example:
+
+        ```py
+        from diffusers import AutoPipelineForText2Image
+        import torch
+
+        pipeline = AutoPipelineForText2Image.from_pretrained(
+            "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16
+        ).to("cuda")
+        pipeline.load_lora_weights(
+            "jbilcke-hf/sdxl-cinematic-1", weight_name="pytorch_lora_weights.safetensors", adapter_name="cinematic"
+        )
+        pipeline.disable_lora()
+        ```
+        """
+        if not USE_PEFT_BACKEND:
+            raise ValueError("PEFT backend is required for this method.")
+        set_adapter_layers(self, enabled=False)
+
+    def enable_lora(self):
+        """
+        Enable the UNet's active LoRA layers.
+
+        Example:
+
+        ```py
+        from diffusers import AutoPipelineForText2Image
+        import torch
+
+        pipeline = AutoPipelineForText2Image.from_pretrained(
+            "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16
+        ).to("cuda")
+        pipeline.load_lora_weights(
+            "jbilcke-hf/sdxl-cinematic-1", weight_name="pytorch_lora_weights.safetensors", adapter_name="cinematic"
+        )
+        pipeline.enable_lora()
+        ```
+        """
+        if not USE_PEFT_BACKEND:
+            raise ValueError("PEFT backend is required for this method.")
+        set_adapter_layers(self, enabled=True)
+
+    def delete_adapters(self, adapter_names: Union[List[str], str]):
+        """
+        Delete an adapter's LoRA layers from the UNet.
+
+        Args:
+            adapter_names (`Union[List[str], str]`):
+                The names (single string or list of strings) of the adapter to delete.
+
+        Example:
+
+        ```py
+        from diffusers import AutoPipelineForText2Image
+        import torch
+
+        pipeline = AutoPipelineForText2Image.from_pretrained(
+            "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16
+        ).to("cuda")
+        pipeline.load_lora_weights(
+            "jbilcke-hf/sdxl-cinematic-1", weight_name="pytorch_lora_weights.safetensors", adapter_names="cinematic"
+        )
+        pipeline.delete_adapters("cinematic")
+        ```
+        """
+        if not USE_PEFT_BACKEND:
+            raise ValueError("PEFT backend is required for this method.")
+
+        if isinstance(adapter_names, str):
+            adapter_names = [adapter_names]
+
+        for adapter_name in adapter_names:
+            delete_adapter_layers(self, adapter_name)
+
+            # Pop also the corresponding adapter from the config
+            if hasattr(self, "peft_config"):
+                self.peft_config.pop(adapter_name, None)
+
+    def _convert_ip_adapter_image_proj_to_diffusers(self, state_dict):
+        updated_state_dict = {}
+        image_projection = None
+
+        if "proj.weight" in state_dict:
+            # IP-Adapter
+            num_image_text_embeds = 4
+            clip_embeddings_dim = state_dict["proj.weight"].shape[-1]
+            cross_attention_dim = state_dict["proj.weight"].shape[0] // 4
+
+            image_projection = ImageProjection(
+                cross_attention_dim=cross_attention_dim,
+                image_embed_dim=clip_embeddings_dim,
+                num_image_text_embeds=num_image_text_embeds,
+            )
+
+            for key, value in state_dict.items():
+                diffusers_name = key.replace("proj", "image_embeds")
+                updated_state_dict[diffusers_name] = value
+
+        elif "proj.3.weight" in state_dict:
+            # IP-Adapter Full
+            clip_embeddings_dim = state_dict["proj.0.weight"].shape[0]
+            cross_attention_dim = state_dict["proj.3.weight"].shape[0]
+
+            image_projection = MLPProjection(
+                cross_attention_dim=cross_attention_dim, image_embed_dim=clip_embeddings_dim
+            )
+
+            for key, value in state_dict.items():
+                diffusers_name = key.replace("proj.0", "ff.net.0.proj")
+                diffusers_name = diffusers_name.replace("proj.2", "ff.net.2")
+                diffusers_name = diffusers_name.replace("proj.3", "norm")
+                updated_state_dict[diffusers_name] = value
+
+        else:
+            # IP-Adapter Plus
+            num_image_text_embeds = state_dict["latents"].shape[1]
+            embed_dims = state_dict["proj_in.weight"].shape[1]
+            output_dims = state_dict["proj_out.weight"].shape[0]
+            hidden_dims = state_dict["latents"].shape[2]
+            heads = state_dict["layers.0.0.to_q.weight"].shape[0] // 64
+
+            image_projection = Resampler(
+                embed_dims=embed_dims,
+                output_dims=output_dims,
+                hidden_dims=hidden_dims,
+                heads=heads,
+                num_queries=num_image_text_embeds,
+            )
+
+            for key, value in state_dict.items():
+                diffusers_name = key.replace("0.to", "2.to")
+                diffusers_name = diffusers_name.replace("1.0.weight", "3.0.weight")
+                diffusers_name = diffusers_name.replace("1.0.bias", "3.0.bias")
+                diffusers_name = diffusers_name.replace("1.1.weight", "3.1.net.0.proj.weight")
+                diffusers_name = diffusers_name.replace("1.3.weight", "3.1.net.2.weight")
+
+                if "norm1" in diffusers_name:
+                    updated_state_dict[diffusers_name.replace("0.norm1", "0")] = value
+                elif "norm2" in diffusers_name:
+                    updated_state_dict[diffusers_name.replace("0.norm2", "1")] = value
+                elif "to_kv" in diffusers_name:
+                    v_chunk = value.chunk(2, dim=0)
+                    updated_state_dict[diffusers_name.replace("to_kv", "to_k")] = v_chunk[0]
+                    updated_state_dict[diffusers_name.replace("to_kv", "to_v")] = v_chunk[1]
+                elif "to_out" in diffusers_name:
+                    updated_state_dict[diffusers_name.replace("to_out", "to_out.0")] = value
+                else:
+                    updated_state_dict[diffusers_name] = value
+
+        image_projection.load_state_dict(updated_state_dict)
+        return image_projection
+
+    def _load_ip_adapter_weights(self, state_dict):
+        from ..models.attention_processor import (
+            AttnProcessor,
+            AttnProcessor2_0,
+            IPAdapterAttnProcessor,
+            IPAdapterAttnProcessor2_0,
+        )
+
+        if "proj.weight" in state_dict["image_proj"]:
+            # IP-Adapter
+            num_image_text_embeds = 4
+        elif "proj.3.weight" in state_dict["image_proj"]:
+            # IP-Adapter Full Face
+            num_image_text_embeds = 257  # 256 CLIP tokens + 1 CLS token
+        else:
+            # IP-Adapter Plus
+            num_image_text_embeds = state_dict["image_proj"]["latents"].shape[1]
+
+        # Set encoder_hid_proj after loading ip_adapter weights,
+        # because `Resampler` also has `attn_processors`.
+        self.encoder_hid_proj = None
+
+        # set ip-adapter cross-attention processors & load state_dict
+        attn_procs = {}
+        key_id = 1
+        for name in self.attn_processors.keys():
+            cross_attention_dim = None if name.endswith("attn1.processor") else self.config.cross_attention_dim
+            if name.startswith("mid_block"):
+                hidden_size = self.config.block_out_channels[-1]
+            elif name.startswith("up_blocks"):
+                block_id = int(name[len("up_blocks.")])
+                hidden_size = list(reversed(self.config.block_out_channels))[block_id]
+            elif name.startswith("down_blocks"):
+                block_id = int(name[len("down_blocks.")])
+                hidden_size = self.config.block_out_channels[block_id]
+            if cross_attention_dim is None or "motion_modules" in name:
+                attn_processor_class = (
+                    AttnProcessor2_0 if hasattr(F, "scaled_dot_product_attention") else AttnProcessor
+                )
+                attn_procs[name] = attn_processor_class()
+            else:
+                attn_processor_class = (
+                    IPAdapterAttnProcessor2_0 if hasattr(F, "scaled_dot_product_attention") else IPAdapterAttnProcessor
+                )
+                attn_procs[name] = attn_processor_class(
+                    hidden_size=hidden_size,
+                    cross_attention_dim=cross_attention_dim,
+                    scale=1.0,
+                    num_tokens=num_image_text_embeds,
+                ).to(dtype=self.dtype, device=self.device)
+
+                value_dict = {}
+                for k, w in attn_procs[name].state_dict().items():
+                    value_dict.update({f"{k}": state_dict["ip_adapter"][f"{key_id}.{k}"]})
+
+                attn_procs[name].load_state_dict(value_dict)
+                key_id += 2
+
+        self.set_attn_processor(attn_procs)
+
+        # convert IP-Adapter Image Projection layers to diffusers
+        image_projection = self._convert_ip_adapter_image_proj_to_diffusers(state_dict["image_proj"])
+
+        self.encoder_hid_proj = image_projection.to(device=self.device, dtype=self.dtype)
+        self.config.encoder_hid_dim_type = "ip_image_proj"

src/diffusers/loaders/utils.py

@@ -0,0 +1,59 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Dict
+
+import torch
+
+
+class AttnProcsLayers(torch.nn.Module):
+    def __init__(self, state_dict: Dict[str, torch.Tensor]):
+        super().__init__()
+        self.layers = torch.nn.ModuleList(state_dict.values())
+        self.mapping = dict(enumerate(state_dict.keys()))
+        self.rev_mapping = {v: k for k, v in enumerate(state_dict.keys())}
+
+        # .processor for unet, .self_attn for text encoder
+        self.split_keys = [".processor", ".self_attn"]
+
+        # we add a hook to state_dict() and load_state_dict() so that the
+        # naming fits with `unet.attn_processors`
+        def map_to(module, state_dict, *args, **kwargs):
+            new_state_dict = {}
+            for key, value in state_dict.items():
+                num = int(key.split(".")[1])  # 0 is always "layers"
+                new_key = key.replace(f"layers.{num}", module.mapping[num])
+                new_state_dict[new_key] = value
+
+            return new_state_dict
+
+        def remap_key(key, state_dict):
+            for k in self.split_keys:
+                if k in key:
+                    return key.split(k)[0] + k
+
+            raise ValueError(
+                f"There seems to be a problem with the state_dict: {set(state_dict.keys())}. {key} has to have one of {self.split_keys}."
+            )
+
+        def map_from(module, state_dict, *args, **kwargs):
+            all_keys = list(state_dict.keys())
+            for key in all_keys:
+                replace_key = remap_key(key, state_dict)
+                new_key = key.replace(replace_key, f"layers.{module.rev_mapping[replace_key]}")
+                state_dict[new_key] = state_dict[key]
+                del state_dict[key]
+
+        self._register_state_dict_hook(map_to)
+        self._register_load_state_dict_pre_hook(map_from, with_module=True)

src/diffusers/models/README.md

@@ -0,0 +1,3 @@
+# Models
+
+For more detail on the models, please refer to the [docs](https://huggingface.co/docs/diffusers/api/models/overview).

src/diffusers/models/__init__.py

@@ -0,0 +1,94 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import TYPE_CHECKING
+
+from ..utils import (
+    DIFFUSERS_SLOW_IMPORT,
+    _LazyModule,
+    is_flax_available,
+    is_torch_available,
+)
+
+
+_import_structure = {}
+
+if is_torch_available():
+    _import_structure["adapter"] = ["MultiAdapter", "T2IAdapter"]
+    _import_structure["autoencoders.autoencoder_asym_kl"] = ["AsymmetricAutoencoderKL"]
+    _import_structure["autoencoders.autoencoder_kl"] = ["AutoencoderKL"]
+    _import_structure["autoencoders.autoencoder_kl_temporal_decoder"] = ["AutoencoderKLTemporalDecoder"]
+    _import_structure["autoencoders.autoencoder_tiny"] = ["AutoencoderTiny"]
+    _import_structure["autoencoders.consistency_decoder_vae"] = ["ConsistencyDecoderVAE"]
+    _import_structure["controlnet"] = ["ControlNetModel"]
+    _import_structure["dual_transformer_2d"] = ["DualTransformer2DModel"]
+    _import_structure["embeddings"] = ["ImageProjection"]
+    _import_structure["modeling_utils"] = ["ModelMixin"]
+    _import_structure["prior_transformer"] = ["PriorTransformer"]
+    _import_structure["t5_film_transformer"] = ["T5FilmDecoder"]
+    _import_structure["transformer_2d"] = ["Transformer2DModel"]
+    _import_structure["transformer_temporal"] = ["TransformerTemporalModel"]
+    _import_structure["unet_1d"] = ["UNet1DModel"]
+    _import_structure["unet_2d"] = ["UNet2DModel"]
+    _import_structure["unet_2d_condition"] = ["UNet2DConditionModel"]
+    _import_structure["unet_3d_condition"] = ["UNet3DConditionModel"]
+    _import_structure["unet_kandinsky3"] = ["Kandinsky3UNet"]
+    _import_structure["unet_motion_model"] = ["MotionAdapter", "UNetMotionModel"]
+    _import_structure["unet_spatio_temporal_condition"] = ["UNetSpatioTemporalConditionModel"]
+    _import_structure["uvit_2d"] = ["UVit2DModel"]
+    _import_structure["vq_model"] = ["VQModel"]
+
+if is_flax_available():
+    _import_structure["controlnet_flax"] = ["FlaxControlNetModel"]
+    _import_structure["unet_2d_condition_flax"] = ["FlaxUNet2DConditionModel"]
+    _import_structure["vae_flax"] = ["FlaxAutoencoderKL"]
+
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+    if is_torch_available():
+        from .adapter import MultiAdapter, T2IAdapter
+        from .autoencoders import (
+            AsymmetricAutoencoderKL,
+            AutoencoderKL,
+            AutoencoderKLTemporalDecoder,
+            AutoencoderTiny,
+            ConsistencyDecoderVAE,
+        )
+        from .controlnet import ControlNetModel
+        from .dual_transformer_2d import DualTransformer2DModel
+        from .embeddings import ImageProjection
+        from .modeling_utils import ModelMixin
+        from .prior_transformer import PriorTransformer
+        from .t5_film_transformer import T5FilmDecoder
+        from .transformer_2d import Transformer2DModel
+        from .transformer_temporal import TransformerTemporalModel
+        from .unet_1d import UNet1DModel
+        from .unet_2d import UNet2DModel
+        from .unet_2d_condition import UNet2DConditionModel
+        from .unet_3d_condition import UNet3DConditionModel
+        from .unet_kandinsky3 import Kandinsky3UNet
+        from .unet_motion_model import MotionAdapter, UNetMotionModel
+        from .unet_spatio_temporal_condition import UNetSpatioTemporalConditionModel
+        from .uvit_2d import UVit2DModel
+        from .vq_model import VQModel
+
+    if is_flax_available():
+        from .controlnet_flax import FlaxControlNetModel
+        from .unet_2d_condition_flax import FlaxUNet2DConditionModel
+        from .vae_flax import FlaxAutoencoderKL
+
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)

src/diffusers/models/activations.py

@@ -0,0 +1,123 @@
+# coding=utf-8
+# Copyright 2023 HuggingFace Inc.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from ..utils import USE_PEFT_BACKEND
+from .lora import LoRACompatibleLinear
+
+
+ACTIVATION_FUNCTIONS = {
+    "swish": nn.SiLU(),
+    "silu": nn.SiLU(),
+    "mish": nn.Mish(),
+    "gelu": nn.GELU(),
+    "relu": nn.ReLU(),
+}
+
+
+def get_activation(act_fn: str) -> nn.Module:
+    """Helper function to get activation function from string.
+
+    Args:
+        act_fn (str): Name of activation function.
+
+    Returns:
+        nn.Module: Activation function.
+    """
+
+    act_fn = act_fn.lower()
+    if act_fn in ACTIVATION_FUNCTIONS:
+        return ACTIVATION_FUNCTIONS[act_fn]
+    else:
+        raise ValueError(f"Unsupported activation function: {act_fn}")
+
+
+class GELU(nn.Module):
+    r"""
+    GELU activation function with tanh approximation support with `approximate="tanh"`.
+
+    Parameters:
+        dim_in (`int`): The number of channels in the input.
+        dim_out (`int`): The number of channels in the output.
+        approximate (`str`, *optional*, defaults to `"none"`): If `"tanh"`, use tanh approximation.
+        bias (`bool`, defaults to True): Whether to use a bias in the linear layer.
+    """
+
+    def __init__(self, dim_in: int, dim_out: int, approximate: str = "none", bias: bool = True):
+        super().__init__()
+        self.proj = nn.Linear(dim_in, dim_out, bias=bias)
+        self.approximate = approximate
+
+    def gelu(self, gate: torch.Tensor) -> torch.Tensor:
+        if gate.device.type != "mps":
+            return F.gelu(gate, approximate=self.approximate)
+        # mps: gelu is not implemented for float16
+        return F.gelu(gate.to(dtype=torch.float32), approximate=self.approximate).to(dtype=gate.dtype)
+
+    def forward(self, hidden_states):
+        hidden_states = self.proj(hidden_states)
+        hidden_states = self.gelu(hidden_states)
+        return hidden_states
+
+
+class GEGLU(nn.Module):
+    r"""
+    A [variant](https://arxiv.org/abs/2002.05202) of the gated linear unit activation function.
+
+    Parameters:
+        dim_in (`int`): The number of channels in the input.
+        dim_out (`int`): The number of channels in the output.
+        bias (`bool`, defaults to True): Whether to use a bias in the linear layer.
+    """
+
+    def __init__(self, dim_in: int, dim_out: int, bias: bool = True):
+        super().__init__()
+        linear_cls = LoRACompatibleLinear if not USE_PEFT_BACKEND else nn.Linear
+
+        self.proj = linear_cls(dim_in, dim_out * 2, bias=bias)
+
+    def gelu(self, gate: torch.Tensor) -> torch.Tensor:
+        if gate.device.type != "mps":
+            return F.gelu(gate)
+        # mps: gelu is not implemented for float16
+        return F.gelu(gate.to(dtype=torch.float32)).to(dtype=gate.dtype)
+
+    def forward(self, hidden_states, scale: float = 1.0):
+        args = () if USE_PEFT_BACKEND else (scale,)
+        hidden_states, gate = self.proj(hidden_states, *args).chunk(2, dim=-1)
+        return hidden_states * self.gelu(gate)
+
+
+class ApproximateGELU(nn.Module):
+    r"""
+    The approximate form of the Gaussian Error Linear Unit (GELU). For more details, see section 2 of this
+    [paper](https://arxiv.org/abs/1606.08415).
+
+    Parameters:
+        dim_in (`int`): The number of channels in the input.
+        dim_out (`int`): The number of channels in the output.
+        bias (`bool`, defaults to True): Whether to use a bias in the linear layer.
+    """
+
+    def __init__(self, dim_in: int, dim_out: int, bias: bool = True):
+        super().__init__()
+        self.proj = nn.Linear(dim_in, dim_out, bias=bias)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.proj(x)
+        return x * torch.sigmoid(1.702 * x)

src/diffusers/models/adapter.py

@@ -0,0 +1,584 @@
+# Copyright 2022 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import os
+from typing import Callable, List, Optional, Union
+
+import torch
+import torch.nn as nn
+
+from ..configuration_utils import ConfigMixin, register_to_config
+from ..utils import logging
+from .modeling_utils import ModelMixin
+
+
+logger = logging.get_logger(__name__)
+
+
+class MultiAdapter(ModelMixin):
+    r"""
+    MultiAdapter is a wrapper model that contains multiple adapter models and merges their outputs according to
+    user-assigned weighting.
+
+    This model inherits from [`ModelMixin`]. Check the superclass documentation for the generic methods the library
+    implements for all the model (such as downloading or saving, etc.)
+
+    Parameters:
+        adapters (`List[T2IAdapter]`, *optional*, defaults to None):
+            A list of `T2IAdapter` model instances.
+    """
+
+    def __init__(self, adapters: List["T2IAdapter"]):
+        super(MultiAdapter, self).__init__()
+
+        self.num_adapter = len(adapters)
+        self.adapters = nn.ModuleList(adapters)
+
+        if len(adapters) == 0:
+            raise ValueError("Expecting at least one adapter")
+
+        if len(adapters) == 1:
+            raise ValueError("For a single adapter, please use the `T2IAdapter` class instead of `MultiAdapter`")
+
+        # The outputs from each adapter are added together with a weight.
+        # This means that the change in dimensions from downsampling must
+        # be the same for all adapters. Inductively, it also means the
+        # downscale_factor and total_downscale_factor must be the same for all
+        # adapters.
+        first_adapter_total_downscale_factor = adapters[0].total_downscale_factor
+        first_adapter_downscale_factor = adapters[0].downscale_factor
+        for idx in range(1, len(adapters)):
+            if (
+                adapters[idx].total_downscale_factor != first_adapter_total_downscale_factor
+                or adapters[idx].downscale_factor != first_adapter_downscale_factor
+            ):
+                raise ValueError(
+                    f"Expecting all adapters to have the same downscaling behavior, but got:\n"
+                    f"adapters[0].total_downscale_factor={first_adapter_total_downscale_factor}\n"
+                    f"adapters[0].downscale_factor={first_adapter_downscale_factor}\n"
+                    f"adapter[`{idx}`].total_downscale_factor={adapters[idx].total_downscale_factor}\n"
+                    f"adapter[`{idx}`].downscale_factor={adapters[idx].downscale_factor}"
+                )
+
+        self.total_downscale_factor = first_adapter_total_downscale_factor
+        self.downscale_factor = first_adapter_downscale_factor
+
+    def forward(self, xs: torch.Tensor, adapter_weights: Optional[List[float]] = None) -> List[torch.Tensor]:
+        r"""
+        Args:
+            xs (`torch.Tensor`):
+                (batch, channel, height, width) input images for multiple adapter models concated along dimension 1,
+                `channel` should equal to `num_adapter` * "number of channel of image".
+            adapter_weights (`List[float]`, *optional*, defaults to None):
+                List of floats representing the weight which will be multiply to each adapter's output before adding
+                them together.
+        """
+        if adapter_weights is None:
+            adapter_weights = torch.tensor([1 / self.num_adapter] * self.num_adapter)
+        else:
+            adapter_weights = torch.tensor(adapter_weights)
+
+        accume_state = None
+        for x, w, adapter in zip(xs, adapter_weights, self.adapters):
+            features = adapter(x)
+            if accume_state is None:
+                accume_state = features
+                for i in range(len(accume_state)):
+                    accume_state[i] = w * accume_state[i]
+            else:
+                for i in range(len(features)):
+                    accume_state[i] += w * features[i]
+        return accume_state
+
+    def save_pretrained(
+        self,
+        save_directory: Union[str, os.PathLike],
+        is_main_process: bool = True,
+        save_function: Callable = None,
+        safe_serialization: bool = True,
+        variant: Optional[str] = None,
+    ):
+        """
+        Save a model and its configuration file to a directory, so that it can be re-loaded using the
+        `[`~models.adapter.MultiAdapter.from_pretrained`]` class method.
+
+        Arguments:
+            save_directory (`str` or `os.PathLike`):
+                Directory to which to save. Will be created if it doesn't exist.
+            is_main_process (`bool`, *optional*, defaults to `True`):
+                Whether the process calling this is the main process or not. Useful when in distributed training like
+                TPUs and need to call this function on all processes. In this case, set `is_main_process=True` only on
+                the main process to avoid race conditions.
+            save_function (`Callable`):
+                The function to use to save the state dictionary. Useful on distributed training like TPUs when one
+                need to replace `torch.save` by another method. Can be configured with the environment variable
+                `DIFFUSERS_SAVE_MODE`.
+            safe_serialization (`bool`, *optional*, defaults to `True`):
+                Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).
+            variant (`str`, *optional*):
+                If specified, weights are saved in the format pytorch_model.<variant>.bin.
+        """
+        idx = 0
+        model_path_to_save = save_directory
+        for adapter in self.adapters:
+            adapter.save_pretrained(
+                model_path_to_save,
+                is_main_process=is_main_process,
+                save_function=save_function,
+                safe_serialization=safe_serialization,
+                variant=variant,
+            )
+
+            idx += 1
+            model_path_to_save = model_path_to_save + f"_{idx}"
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_path: Optional[Union[str, os.PathLike]], **kwargs):
+        r"""
+        Instantiate a pretrained MultiAdapter model from multiple pre-trained adapter models.
+
+        The model is set in evaluation mode by default using `model.eval()` (Dropout modules are deactivated). To train
+        the model, you should first set it back in training mode with `model.train()`.
+
+        The warning *Weights from XXX not initialized from pretrained model* means that the weights of XXX do not come
+        pretrained with the rest of the model. It is up to you to train those weights with a downstream fine-tuning
+        task.
+
+        The warning *Weights from XXX not used in YYY* means that the layer XXX is not used by YYY, therefore those
+        weights are discarded.
+
+        Parameters:
+            pretrained_model_path (`os.PathLike`):
+                A path to a *directory* containing model weights saved using
+                [`~diffusers.models.adapter.MultiAdapter.save_pretrained`], e.g., `./my_model_directory/adapter`.
+            torch_dtype (`str` or `torch.dtype`, *optional*):
+                Override the default `torch.dtype` and load the model under this dtype. If `"auto"` is passed the dtype
+                will be automatically derived from the model's weights.
+            output_loading_info(`bool`, *optional*, defaults to `False`):
+                Whether or not to also return a dictionary containing missing keys, unexpected keys and error messages.
+            device_map (`str` or `Dict[str, Union[int, str, torch.device]]`, *optional*):
+                A map that specifies where each submodule should go. It doesn't need to be refined to each
+                parameter/buffer name, once a given module name is inside, every submodule of it will be sent to the
+                same device.
+
+                To have Accelerate compute the most optimized `device_map` automatically, set `device_map="auto"`. For
+                more information about each option see [designing a device
+                map](https://hf.co/docs/accelerate/main/en/usage_guides/big_modeling#designing-a-device-map).
+            max_memory (`Dict`, *optional*):
+                A dictionary device identifier to maximum memory. Will default to the maximum memory available for each
+                GPU and the available CPU RAM if unset.
+            low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 else `False`):
+                Speed up model loading by not initializing the weights and only loading the pre-trained weights. This
+                also tries to not use more than 1x model size in CPU memory (including peak memory) while loading the
+                model. This is only supported when torch version >= 1.9.0. If you are using an older version of torch,
+                setting this argument to `True` will raise an error.
+            variant (`str`, *optional*):
+                If specified load weights from `variant` filename, *e.g.* pytorch_model.<variant>.bin. `variant` is
+                ignored when using `from_flax`.
+            use_safetensors (`bool`, *optional*, defaults to `None`):
+                If set to `None`, the `safetensors` weights will be downloaded if they're available **and** if the
+                `safetensors` library is installed. If set to `True`, the model will be forcibly loaded from
+                `safetensors` weights. If set to `False`, loading will *not* use `safetensors`.
+        """
+        idx = 0
+        adapters = []
+
+        # load adapter and append to list until no adapter directory exists anymore
+        # first adapter has to be saved under `./mydirectory/adapter` to be compliant with `DiffusionPipeline.from_pretrained`
+        # second, third, ... adapters have to be saved under `./mydirectory/adapter_1`, `./mydirectory/adapter_2`, ...
+        model_path_to_load = pretrained_model_path
+        while os.path.isdir(model_path_to_load):
+            adapter = T2IAdapter.from_pretrained(model_path_to_load, **kwargs)
+            adapters.append(adapter)
+
+            idx += 1
+            model_path_to_load = pretrained_model_path + f"_{idx}"
+
+        logger.info(f"{len(adapters)} adapters loaded from {pretrained_model_path}.")
+
+        if len(adapters) == 0:
+            raise ValueError(
+                f"No T2IAdapters found under {os.path.dirname(pretrained_model_path)}. Expected at least {pretrained_model_path + '_0'}."
+            )
+
+        return cls(adapters)
+
+
+class T2IAdapter(ModelMixin, ConfigMixin):
+    r"""
+    A simple ResNet-like model that accepts images containing control signals such as keyposes and depth. The model
+    generates multiple feature maps that are used as additional conditioning in [`UNet2DConditionModel`]. The model's
+    architecture follows the original implementation of
+    [Adapter](https://github.com/TencentARC/T2I-Adapter/blob/686de4681515662c0ac2ffa07bf5dda83af1038a/ldm/modules/encoders/adapter.py#L97)
+     and
+     [AdapterLight](https://github.com/TencentARC/T2I-Adapter/blob/686de4681515662c0ac2ffa07bf5dda83af1038a/ldm/modules/encoders/adapter.py#L235).
+
+    This model inherits from [`ModelMixin`]. Check the superclass documentation for the generic methods the library
+    implements for all the model (such as downloading or saving, etc.)
+
+    Parameters:
+        in_channels (`int`, *optional*, defaults to 3):
+            Number of channels of Aapter's input(*control image*). Set this parameter to 1 if you're using gray scale
+            image as *control image*.
+        channels (`List[int]`, *optional*, defaults to `(320, 640, 1280, 1280)`):
+            The number of channel of each downsample block's output hidden state. The `len(block_out_channels)` will
+            also determine the number of downsample blocks in the Adapter.
+        num_res_blocks (`int`, *optional*, defaults to 2):
+            Number of ResNet blocks in each downsample block.
+        downscale_factor (`int`, *optional*, defaults to 8):
+            A factor that determines the total downscale factor of the Adapter.
+        adapter_type (`str`, *optional*, defaults to `full_adapter`):
+            The type of Adapter to use. Choose either `full_adapter` or `full_adapter_xl` or `light_adapter`.
+    """
+
+    @register_to_config
+    def __init__(
+        self,
+        in_channels: int = 3,
+        channels: List[int] = [320, 640, 1280, 1280],
+        num_res_blocks: int = 2,
+        downscale_factor: int = 8,
+        adapter_type: str = "full_adapter",
+    ):
+        super().__init__()
+
+        if adapter_type == "full_adapter":
+            self.adapter = FullAdapter(in_channels, channels, num_res_blocks, downscale_factor)
+        elif adapter_type == "full_adapter_xl":
+            self.adapter = FullAdapterXL(in_channels, channels, num_res_blocks, downscale_factor)
+        elif adapter_type == "light_adapter":
+            self.adapter = LightAdapter(in_channels, channels, num_res_blocks, downscale_factor)
+        else:
+            raise ValueError(
+                f"Unsupported adapter_type: '{adapter_type}'. Choose either 'full_adapter' or "
+                "'full_adapter_xl' or 'light_adapter'."
+            )
+
+    def forward(self, x: torch.Tensor) -> List[torch.Tensor]:
+        r"""
+        This function processes the input tensor `x` through the adapter model and returns a list of feature tensors,
+        each representing information extracted at a different scale from the input. The length of the list is
+        determined by the number of downsample blocks in the Adapter, as specified by the `channels` and
+        `num_res_blocks` parameters during initialization.
+        """
+        return self.adapter(x)
+
+    @property
+    def total_downscale_factor(self):
+        return self.adapter.total_downscale_factor
+
+    @property
+    def downscale_factor(self):
+        """The downscale factor applied in the T2I-Adapter's initial pixel unshuffle operation. If an input image's dimensions are
+        not evenly divisible by the downscale_factor then an exception will be raised.
+        """
+        return self.adapter.unshuffle.downscale_factor
+
+
+# full adapter
+
+
+class FullAdapter(nn.Module):
+    r"""
+    See [`T2IAdapter`] for more information.
+    """
+
+    def __init__(
+        self,
+        in_channels: int = 3,
+        channels: List[int] = [320, 640, 1280, 1280],
+        num_res_blocks: int = 2,
+        downscale_factor: int = 8,
+    ):
+        super().__init__()
+
+        in_channels = in_channels * downscale_factor**2
+
+        self.unshuffle = nn.PixelUnshuffle(downscale_factor)
+        self.conv_in = nn.Conv2d(in_channels, channels[0], kernel_size=3, padding=1)
+
+        self.body = nn.ModuleList(
+            [
+                AdapterBlock(channels[0], channels[0], num_res_blocks),
+                *[
+                    AdapterBlock(channels[i - 1], channels[i], num_res_blocks, down=True)
+                    for i in range(1, len(channels))
+                ],
+            ]
+        )
+
+        self.total_downscale_factor = downscale_factor * 2 ** (len(channels) - 1)
+
+    def forward(self, x: torch.Tensor) -> List[torch.Tensor]:
+        r"""
+        This method processes the input tensor `x` through the FullAdapter model and performs operations including
+        pixel unshuffling, convolution, and a stack of AdapterBlocks. It returns a list of feature tensors, each
+        capturing information at a different stage of processing within the FullAdapter model. The number of feature
+        tensors in the list is determined by the number of downsample blocks specified during initialization.
+        """
+        x = self.unshuffle(x)
+        x = self.conv_in(x)
+
+        features = []
+
+        for block in self.body:
+            x = block(x)
+            features.append(x)
+
+        return features
+
+
+class FullAdapterXL(nn.Module):
+    r"""
+    See [`T2IAdapter`] for more information.
+    """
+
+    def __init__(
+        self,
+        in_channels: int = 3,
+        channels: List[int] = [320, 640, 1280, 1280],
+        num_res_blocks: int = 2,
+        downscale_factor: int = 16,
+    ):
+        super().__init__()
+
+        in_channels = in_channels * downscale_factor**2
+
+        self.unshuffle = nn.PixelUnshuffle(downscale_factor)
+        self.conv_in = nn.Conv2d(in_channels, channels[0], kernel_size=3, padding=1)
+
+        self.body = []
+        # blocks to extract XL features with dimensions of [320, 64, 64], [640, 64, 64], [1280, 32, 32], [1280, 32, 32]
+        for i in range(len(channels)):
+            if i == 1:
+                self.body.append(AdapterBlock(channels[i - 1], channels[i], num_res_blocks))
+            elif i == 2:
+                self.body.append(AdapterBlock(channels[i - 1], channels[i], num_res_blocks, down=True))
+            else:
+                self.body.append(AdapterBlock(channels[i], channels[i], num_res_blocks))
+
+        self.body = nn.ModuleList(self.body)
+        # XL has only one downsampling AdapterBlock.
+        self.total_downscale_factor = downscale_factor * 2
+
+    def forward(self, x: torch.Tensor) -> List[torch.Tensor]:
+        r"""
+        This method takes the tensor x as input and processes it through FullAdapterXL model. It consists of operations
+        including unshuffling pixels, applying convolution layer and appending each block into list of feature tensors.
+        """
+        x = self.unshuffle(x)
+        x = self.conv_in(x)
+
+        features = []
+
+        for block in self.body:
+            x = block(x)
+            features.append(x)
+
+        return features
+
+
+class AdapterBlock(nn.Module):
+    r"""
+    An AdapterBlock is a helper model that contains multiple ResNet-like blocks. It is used in the `FullAdapter` and
+    `FullAdapterXL` models.
+
+    Parameters:
+        in_channels (`int`):
+            Number of channels of AdapterBlock's input.
+        out_channels (`int`):
+            Number of channels of AdapterBlock's output.
+        num_res_blocks (`int`):
+            Number of ResNet blocks in the AdapterBlock.
+        down (`bool`, *optional*, defaults to `False`):
+            Whether to perform downsampling on AdapterBlock's input.
+    """
+
+    def __init__(self, in_channels: int, out_channels: int, num_res_blocks: int, down: bool = False):
+        super().__init__()
+
+        self.downsample = None
+        if down:
+            self.downsample = nn.AvgPool2d(kernel_size=2, stride=2, ceil_mode=True)
+
+        self.in_conv = None
+        if in_channels != out_channels:
+            self.in_conv = nn.Conv2d(in_channels, out_channels, kernel_size=1)
+
+        self.resnets = nn.Sequential(
+            *[AdapterResnetBlock(out_channels) for _ in range(num_res_blocks)],
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        r"""
+        This method takes tensor x as input and performs operations downsampling and convolutional layers if the
+        self.downsample and self.in_conv properties of AdapterBlock model are specified. Then it applies a series of
+        residual blocks to the input tensor.
+        """
+        if self.downsample is not None:
+            x = self.downsample(x)
+
+        if self.in_conv is not None:
+            x = self.in_conv(x)
+
+        x = self.resnets(x)
+
+        return x
+
+
+class AdapterResnetBlock(nn.Module):
+    r"""
+    An `AdapterResnetBlock` is a helper model that implements a ResNet-like block.
+
+    Parameters:
+        channels (`int`):
+            Number of channels of AdapterResnetBlock's input and output.
+    """
+
+    def __init__(self, channels: int):
+        super().__init__()
+        self.block1 = nn.Conv2d(channels, channels, kernel_size=3, padding=1)
+        self.act = nn.ReLU()
+        self.block2 = nn.Conv2d(channels, channels, kernel_size=1)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        r"""
+        This method takes input tensor x and applies a convolutional layer, ReLU activation, and another convolutional
+        layer on the input tensor. It returns addition with the input tensor.
+        """
+
+        h = self.act(self.block1(x))
+        h = self.block2(h)
+
+        return h + x
+
+
+# light adapter
+
+
+class LightAdapter(nn.Module):
+    r"""
+    See [`T2IAdapter`] for more information.
+    """
+
+    def __init__(
+        self,
+        in_channels: int = 3,
+        channels: List[int] = [320, 640, 1280],
+        num_res_blocks: int = 4,
+        downscale_factor: int = 8,
+    ):
+        super().__init__()
+
+        in_channels = in_channels * downscale_factor**2
+
+        self.unshuffle = nn.PixelUnshuffle(downscale_factor)
+
+        self.body = nn.ModuleList(
+            [
+                LightAdapterBlock(in_channels, channels[0], num_res_blocks),
+                *[
+                    LightAdapterBlock(channels[i], channels[i + 1], num_res_blocks, down=True)
+                    for i in range(len(channels) - 1)
+                ],
+                LightAdapterBlock(channels[-1], channels[-1], num_res_blocks, down=True),
+            ]
+        )
+
+        self.total_downscale_factor = downscale_factor * (2 ** len(channels))
+
+    def forward(self, x: torch.Tensor) -> List[torch.Tensor]:
+        r"""
+        This method takes the input tensor x and performs downscaling and appends it in list of feature tensors. Each
+        feature tensor corresponds to a different level of processing within the LightAdapter.
+        """
+        x = self.unshuffle(x)
+
+        features = []
+
+        for block in self.body:
+            x = block(x)
+            features.append(x)
+
+        return features
+
+
+class LightAdapterBlock(nn.Module):
+    r"""
+    A `LightAdapterBlock` is a helper model that contains multiple `LightAdapterResnetBlocks`. It is used in the
+    `LightAdapter` model.
+
+    Parameters:
+        in_channels (`int`):
+            Number of channels of LightAdapterBlock's input.
+        out_channels (`int`):
+            Number of channels of LightAdapterBlock's output.
+        num_res_blocks (`int`):
+            Number of LightAdapterResnetBlocks in the LightAdapterBlock.
+        down (`bool`, *optional*, defaults to `False`):
+            Whether to perform downsampling on LightAdapterBlock's input.
+    """
+
+    def __init__(self, in_channels: int, out_channels: int, num_res_blocks: int, down: bool = False):
+        super().__init__()
+        mid_channels = out_channels // 4
+
+        self.downsample = None
+        if down:
+            self.downsample = nn.AvgPool2d(kernel_size=2, stride=2, ceil_mode=True)
+
+        self.in_conv = nn.Conv2d(in_channels, mid_channels, kernel_size=1)
+        self.resnets = nn.Sequential(*[LightAdapterResnetBlock(mid_channels) for _ in range(num_res_blocks)])
+        self.out_conv = nn.Conv2d(mid_channels, out_channels, kernel_size=1)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        r"""
+        This method takes tensor x as input and performs downsampling if required. Then it applies in convolution
+        layer, a sequence of residual blocks, and out convolutional layer.
+        """
+        if self.downsample is not None:
+            x = self.downsample(x)
+
+        x = self.in_conv(x)
+        x = self.resnets(x)
+        x = self.out_conv(x)
+
+        return x
+
+
+class LightAdapterResnetBlock(nn.Module):
+    """
+    A `LightAdapterResnetBlock` is a helper model that implements a ResNet-like block with a slightly different
+    architecture than `AdapterResnetBlock`.
+
+    Parameters:
+        channels (`int`):
+            Number of channels of LightAdapterResnetBlock's input and output.
+    """
+
+    def __init__(self, channels: int):
+        super().__init__()
+        self.block1 = nn.Conv2d(channels, channels, kernel_size=3, padding=1)
+        self.act = nn.ReLU()
+        self.block2 = nn.Conv2d(channels, channels, kernel_size=3, padding=1)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        r"""
+        This function takes input tensor x and processes it through one convolutional layer, ReLU activation, and
+        another convolutional layer and adds it to input tensor.
+        """
+
+        h = self.act(self.block1(x))
+        h = self.block2(h)
+
+        return h + x

src/diffusers/models/attention.py

@@ -0,0 +1,668 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import Any, Dict, Optional
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from ..utils import USE_PEFT_BACKEND
+from ..utils.torch_utils import maybe_allow_in_graph
+from .activations import GEGLU, GELU, ApproximateGELU
+from .attention_processor import Attention
+from .embeddings import SinusoidalPositionalEmbedding
+from .lora import LoRACompatibleLinear
+from .normalization import AdaLayerNorm, AdaLayerNormContinuous, AdaLayerNormZero, RMSNorm
+
+
+def _chunked_feed_forward(
+    ff: nn.Module, hidden_states: torch.Tensor, chunk_dim: int, chunk_size: int, lora_scale: Optional[float] = None
+):
+    # "feed_forward_chunk_size" can be used to save memory
+    if hidden_states.shape[chunk_dim] % chunk_size != 0:
+        raise ValueError(
+            f"`hidden_states` dimension to be chunked: {hidden_states.shape[chunk_dim]} has to be divisible by chunk size: {chunk_size}. Make sure to set an appropriate `chunk_size` when calling `unet.enable_forward_chunking`."
+        )
+
+    num_chunks = hidden_states.shape[chunk_dim] // chunk_size
+    if lora_scale is None:
+        ff_output = torch.cat(
+            [ff(hid_slice) for hid_slice in hidden_states.chunk(num_chunks, dim=chunk_dim)],
+            dim=chunk_dim,
+        )
+    else:
+        # TOOD(Patrick): LoRA scale can be removed once PEFT refactor is complete
+        ff_output = torch.cat(
+            [ff(hid_slice, scale=lora_scale) for hid_slice in hidden_states.chunk(num_chunks, dim=chunk_dim)],
+            dim=chunk_dim,
+        )
+
+    return ff_output
+
+
+@maybe_allow_in_graph
+class GatedSelfAttentionDense(nn.Module):
+    r"""
+    A gated self-attention dense layer that combines visual features and object features.
+
+    Parameters:
+        query_dim (`int`): The number of channels in the query.
+        context_dim (`int`): The number of channels in the context.
+        n_heads (`int`): The number of heads to use for attention.
+        d_head (`int`): The number of channels in each head.
+    """
+
+    def __init__(self, query_dim: int, context_dim: int, n_heads: int, d_head: int):
+        super().__init__()
+
+        # we need a linear projection since we need cat visual feature and obj feature
+        self.linear = nn.Linear(context_dim, query_dim)
+
+        self.attn = Attention(query_dim=query_dim, heads=n_heads, dim_head=d_head)
+        self.ff = FeedForward(query_dim, activation_fn="geglu")
+
+        self.norm1 = nn.LayerNorm(query_dim)
+        self.norm2 = nn.LayerNorm(query_dim)
+
+        self.register_parameter("alpha_attn", nn.Parameter(torch.tensor(0.0)))
+        self.register_parameter("alpha_dense", nn.Parameter(torch.tensor(0.0)))
+
+        self.enabled = True
+
+    def forward(self, x: torch.Tensor, objs: torch.Tensor) -> torch.Tensor:
+        if not self.enabled:
+            return x
+
+        n_visual = x.shape[1]
+        objs = self.linear(objs)
+
+        x = x + self.alpha_attn.tanh() * self.attn(self.norm1(torch.cat([x, objs], dim=1)))[:, :n_visual, :]
+        x = x + self.alpha_dense.tanh() * self.ff(self.norm2(x))
+
+        return x
+
+
+@maybe_allow_in_graph
+class BasicTransformerBlock(nn.Module):
+    r"""
+    A basic Transformer block.
+
+    Parameters:
+        dim (`int`): The number of channels in the input and output.
+        num_attention_heads (`int`): The number of heads to use for multi-head attention.
+        attention_head_dim (`int`): The number of channels in each head.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention.
+        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
+        num_embeds_ada_norm (:
+            obj: `int`, *optional*): The number of diffusion steps used during training. See `Transformer2DModel`.
+        attention_bias (:
+            obj: `bool`, *optional*, defaults to `False`): Configure if the attentions should contain a bias parameter.
+        only_cross_attention (`bool`, *optional*):
+            Whether to use only cross-attention layers. In this case two cross attention layers are used.
+        double_self_attention (`bool`, *optional*):
+            Whether to use two self-attention layers. In this case no cross attention layers are used.
+        upcast_attention (`bool`, *optional*):
+            Whether to upcast the attention computation to float32. This is useful for mixed precision training.
+        norm_elementwise_affine (`bool`, *optional*, defaults to `True`):
+            Whether to use learnable elementwise affine parameters for normalization.
+        norm_type (`str`, *optional*, defaults to `"layer_norm"`):
+            The normalization layer to use. Can be `"layer_norm"`, `"ada_norm"` or `"ada_norm_zero"`.
+        final_dropout (`bool` *optional*, defaults to False):
+            Whether to apply a final dropout after the last feed-forward layer.
+        attention_type (`str`, *optional*, defaults to `"default"`):
+            The type of attention to use. Can be `"default"` or `"gated"` or `"gated-text-image"`.
+        positional_embeddings (`str`, *optional*, defaults to `None`):
+            The type of positional embeddings to apply to.
+        num_positional_embeddings (`int`, *optional*, defaults to `None`):
+            The maximum number of positional embeddings to apply.
+    """
+
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        dropout=0.0,
+        cross_attention_dim: Optional[int] = None,
+        activation_fn: str = "geglu",
+        num_embeds_ada_norm: Optional[int] = None,
+        attention_bias: bool = False,
+        only_cross_attention: bool = False,
+        double_self_attention: bool = False,
+        upcast_attention: bool = False,
+        norm_elementwise_affine: bool = True,
+        norm_type: str = "layer_norm",  # 'layer_norm', 'ada_norm', 'ada_norm_zero', 'ada_norm_single'
+        norm_eps: float = 1e-5,
+        final_dropout: bool = False,
+        attention_type: str = "default",
+        positional_embeddings: Optional[str] = None,
+        num_positional_embeddings: Optional[int] = None,
+        ada_norm_continous_conditioning_embedding_dim: Optional[int] = None,
+        ada_norm_bias: Optional[int] = None,
+        ff_inner_dim: Optional[int] = None,
+        ff_bias: bool = True,
+        attention_out_bias: bool = True,
+    ):
+        super().__init__()
+        self.only_cross_attention = only_cross_attention
+
+        self.use_ada_layer_norm_zero = (num_embeds_ada_norm is not None) and norm_type == "ada_norm_zero"
+        self.use_ada_layer_norm = (num_embeds_ada_norm is not None) and norm_type == "ada_norm"
+        self.use_ada_layer_norm_single = norm_type == "ada_norm_single"
+        self.use_layer_norm = norm_type == "layer_norm"
+        self.use_ada_layer_norm_continuous = norm_type == "ada_norm_continuous"
+
+        if norm_type in ("ada_norm", "ada_norm_zero") and num_embeds_ada_norm is None:
+            raise ValueError(
+                f"`norm_type` is set to {norm_type}, but `num_embeds_ada_norm` is not defined. Please make sure to"
+                f" define `num_embeds_ada_norm` if setting `norm_type` to {norm_type}."
+            )
+
+        if positional_embeddings and (num_positional_embeddings is None):
+            raise ValueError(
+                "If `positional_embedding` type is defined, `num_positition_embeddings` must also be defined."
+            )
+
+        if positional_embeddings == "sinusoidal":
+            self.pos_embed = SinusoidalPositionalEmbedding(dim, max_seq_length=num_positional_embeddings)
+        else:
+            self.pos_embed = None
+
+        # Define 3 blocks. Each block has its own normalization layer.
+        # 1. Self-Attn
+        if self.use_ada_layer_norm:
+            self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm)
+        elif self.use_ada_layer_norm_zero:
+            self.norm1 = AdaLayerNormZero(dim, num_embeds_ada_norm)
+        elif self.use_ada_layer_norm_continuous:
+            self.norm1 = AdaLayerNormContinuous(
+                dim,
+                ada_norm_continous_conditioning_embedding_dim,
+                norm_elementwise_affine,
+                norm_eps,
+                ada_norm_bias,
+                "rms_norm",
+            )
+        else:
+            self.norm1 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps)
+
+        self.attn1 = Attention(
+            query_dim=dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            dropout=dropout,
+            bias=attention_bias,
+            cross_attention_dim=cross_attention_dim if only_cross_attention else None,
+            upcast_attention=upcast_attention,
+            out_bias=attention_out_bias,
+        )
+
+        # 2. Cross-Attn
+        if cross_attention_dim is not None or double_self_attention:
+            # We currently only use AdaLayerNormZero for self attention where there will only be one attention block.
+            # I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during
+            # the second cross attention block.
+            if self.use_ada_layer_norm:
+                self.norm2 = AdaLayerNorm(dim, num_embeds_ada_norm)
+            elif self.use_ada_layer_norm_continuous:
+                self.norm2 = AdaLayerNormContinuous(
+                    dim,
+                    ada_norm_continous_conditioning_embedding_dim,
+                    norm_elementwise_affine,
+                    norm_eps,
+                    ada_norm_bias,
+                    "rms_norm",
+                )
+            else:
+                self.norm2 = nn.LayerNorm(dim, norm_eps, norm_elementwise_affine)
+
+            self.attn2 = Attention(
+                query_dim=dim,
+                cross_attention_dim=cross_attention_dim if not double_self_attention else None,
+                heads=num_attention_heads,
+                dim_head=attention_head_dim,
+                dropout=dropout,
+                bias=attention_bias,
+                upcast_attention=upcast_attention,
+                out_bias=attention_out_bias,
+            )  # is self-attn if encoder_hidden_states is none
+        else:
+            self.norm2 = None
+            self.attn2 = None
+
+        # 3. Feed-forward
+        if self.use_ada_layer_norm_continuous:
+            self.norm3 = AdaLayerNormContinuous(
+                dim,
+                ada_norm_continous_conditioning_embedding_dim,
+                norm_elementwise_affine,
+                norm_eps,
+                ada_norm_bias,
+                "layer_norm",
+            )
+        elif not self.use_ada_layer_norm_single:
+            self.norm3 = nn.LayerNorm(dim, norm_eps, norm_elementwise_affine)
+
+        self.ff = FeedForward(
+            dim,
+            dropout=dropout,
+            activation_fn=activation_fn,
+            final_dropout=final_dropout,
+            inner_dim=ff_inner_dim,
+            bias=ff_bias,
+        )
+
+        # 4. Fuser
+        if attention_type == "gated" or attention_type == "gated-text-image":
+            self.fuser = GatedSelfAttentionDense(dim, cross_attention_dim, num_attention_heads, attention_head_dim)
+
+        # 5. Scale-shift for PixArt-Alpha.
+        if self.use_ada_layer_norm_single:
+            self.scale_shift_table = nn.Parameter(torch.randn(6, dim) / dim**0.5)
+
+        # let chunk size default to None
+        self._chunk_size = None
+        self._chunk_dim = 0
+
+    def set_chunk_feed_forward(self, chunk_size: Optional[int], dim: int = 0):
+        # Sets chunk feed-forward
+        self._chunk_size = chunk_size
+        self._chunk_dim = dim
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        timestep: Optional[torch.LongTensor] = None,
+        cross_attention_kwargs: Dict[str, Any] = None,
+        class_labels: Optional[torch.LongTensor] = None,
+        added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,
+    ) -> torch.FloatTensor:
+        # Notice that normalization is always applied before the real computation in the following blocks.
+        # 0. Self-Attention
+        batch_size = hidden_states.shape[0]
+
+        if self.use_ada_layer_norm:
+            norm_hidden_states = self.norm1(hidden_states, timestep)
+        elif self.use_ada_layer_norm_zero:
+            norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1(
+                hidden_states, timestep, class_labels, hidden_dtype=hidden_states.dtype
+            )
+        elif self.use_layer_norm:
+            norm_hidden_states = self.norm1(hidden_states)
+        elif self.use_ada_layer_norm_continuous:
+            norm_hidden_states = self.norm1(hidden_states, added_cond_kwargs["pooled_text_emb"])
+        elif self.use_ada_layer_norm_single:
+            shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (
+                self.scale_shift_table[None] + timestep.reshape(batch_size, 6, -1)
+            ).chunk(6, dim=1)
+            norm_hidden_states = self.norm1(hidden_states)
+            norm_hidden_states = norm_hidden_states * (1 + scale_msa) + shift_msa
+            norm_hidden_states = norm_hidden_states.squeeze(1)
+        else:
+            raise ValueError("Incorrect norm used")
+
+        if self.pos_embed is not None:
+            norm_hidden_states = self.pos_embed(norm_hidden_states)
+
+        # 1. Retrieve lora scale.
+        lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
+
+        # 2. Prepare GLIGEN inputs
+        cross_attention_kwargs = cross_attention_kwargs.copy() if cross_attention_kwargs is not None else {}
+        gligen_kwargs = cross_attention_kwargs.pop("gligen", None)
+
+        attn_output = self.attn1(
+            norm_hidden_states,
+            encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None,
+            attention_mask=attention_mask,
+            **cross_attention_kwargs,
+        )
+        if self.use_ada_layer_norm_zero:
+            attn_output = gate_msa.unsqueeze(1) * attn_output
+        elif self.use_ada_layer_norm_single:
+            attn_output = gate_msa * attn_output
+
+        hidden_states = attn_output + hidden_states
+        if hidden_states.ndim == 4:
+            hidden_states = hidden_states.squeeze(1)
+
+        # 2.5 GLIGEN Control
+        if gligen_kwargs is not None:
+            hidden_states = self.fuser(hidden_states, gligen_kwargs["objs"])
+
+        # 3. Cross-Attention
+        if self.attn2 is not None:
+            if self.use_ada_layer_norm:
+                norm_hidden_states = self.norm2(hidden_states, timestep)
+            elif self.use_ada_layer_norm_zero or self.use_layer_norm:
+                norm_hidden_states = self.norm2(hidden_states)
+            elif self.use_ada_layer_norm_single:
+                # For PixArt norm2 isn't applied here:
+                # https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L70C1-L76C103
+                norm_hidden_states = hidden_states
+            elif self.use_ada_layer_norm_continuous:
+                norm_hidden_states = self.norm2(hidden_states, added_cond_kwargs["pooled_text_emb"])
+            else:
+                raise ValueError("Incorrect norm")
+
+            if self.pos_embed is not None and self.use_ada_layer_norm_single is False:
+                norm_hidden_states = self.pos_embed(norm_hidden_states)
+
+            attn_output = self.attn2(
+                norm_hidden_states,
+                encoder_hidden_states=encoder_hidden_states,
+                attention_mask=encoder_attention_mask,
+                **cross_attention_kwargs,
+            )
+            hidden_states = attn_output + hidden_states
+
+        # 4. Feed-forward
+        if self.use_ada_layer_norm_continuous:
+            norm_hidden_states = self.norm3(hidden_states, added_cond_kwargs["pooled_text_emb"])
+        elif not self.use_ada_layer_norm_single:
+            norm_hidden_states = self.norm3(hidden_states)
+
+        if self.use_ada_layer_norm_zero:
+            norm_hidden_states = norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None]
+
+        if self.use_ada_layer_norm_single:
+            norm_hidden_states = self.norm2(hidden_states)
+            norm_hidden_states = norm_hidden_states * (1 + scale_mlp) + shift_mlp
+
+        if self._chunk_size is not None:
+            # "feed_forward_chunk_size" can be used to save memory
+            ff_output = _chunked_feed_forward(
+                self.ff, norm_hidden_states, self._chunk_dim, self._chunk_size, lora_scale=lora_scale
+            )
+        else:
+            ff_output = self.ff(norm_hidden_states, scale=lora_scale)
+
+        if self.use_ada_layer_norm_zero:
+            ff_output = gate_mlp.unsqueeze(1) * ff_output
+        elif self.use_ada_layer_norm_single:
+            ff_output = gate_mlp * ff_output
+
+        hidden_states = ff_output + hidden_states
+        if hidden_states.ndim == 4:
+            hidden_states = hidden_states.squeeze(1)
+
+        return hidden_states
+
+
+@maybe_allow_in_graph
+class TemporalBasicTransformerBlock(nn.Module):
+    r"""
+    A basic Transformer block for video like data.
+
+    Parameters:
+        dim (`int`): The number of channels in the input and output.
+        time_mix_inner_dim (`int`): The number of channels for temporal attention.
+        num_attention_heads (`int`): The number of heads to use for multi-head attention.
+        attention_head_dim (`int`): The number of channels in each head.
+        cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention.
+    """
+
+    def __init__(
+        self,
+        dim: int,
+        time_mix_inner_dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        cross_attention_dim: Optional[int] = None,
+    ):
+        super().__init__()
+        self.is_res = dim == time_mix_inner_dim
+
+        self.norm_in = nn.LayerNorm(dim)
+
+        # Define 3 blocks. Each block has its own normalization layer.
+        # 1. Self-Attn
+        self.norm_in = nn.LayerNorm(dim)
+        self.ff_in = FeedForward(
+            dim,
+            dim_out=time_mix_inner_dim,
+            activation_fn="geglu",
+        )
+
+        self.norm1 = nn.LayerNorm(time_mix_inner_dim)
+        self.attn1 = Attention(
+            query_dim=time_mix_inner_dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            cross_attention_dim=None,
+        )
+
+        # 2. Cross-Attn
+        if cross_attention_dim is not None:
+            # We currently only use AdaLayerNormZero for self attention where there will only be one attention block.
+            # I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during
+            # the second cross attention block.
+            self.norm2 = nn.LayerNorm(time_mix_inner_dim)
+            self.attn2 = Attention(
+                query_dim=time_mix_inner_dim,
+                cross_attention_dim=cross_attention_dim,
+                heads=num_attention_heads,
+                dim_head=attention_head_dim,
+            )  # is self-attn if encoder_hidden_states is none
+        else:
+            self.norm2 = None
+            self.attn2 = None
+
+        # 3. Feed-forward
+        self.norm3 = nn.LayerNorm(time_mix_inner_dim)
+        self.ff = FeedForward(time_mix_inner_dim, activation_fn="geglu")
+
+        # let chunk size default to None
+        self._chunk_size = None
+        self._chunk_dim = None
+
+    def set_chunk_feed_forward(self, chunk_size: Optional[int], **kwargs):
+        # Sets chunk feed-forward
+        self._chunk_size = chunk_size
+        # chunk dim should be hardcoded to 1 to have better speed vs. memory trade-off
+        self._chunk_dim = 1
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        num_frames: int,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+    ) -> torch.FloatTensor:
+        # Notice that normalization is always applied before the real computation in the following blocks.
+        # 0. Self-Attention
+        batch_size = hidden_states.shape[0]
+
+        batch_frames, seq_length, channels = hidden_states.shape
+        batch_size = batch_frames // num_frames
+
+        hidden_states = hidden_states[None, :].reshape(batch_size, num_frames, seq_length, channels)
+        hidden_states = hidden_states.permute(0, 2, 1, 3)
+        hidden_states = hidden_states.reshape(batch_size * seq_length, num_frames, channels)
+
+        residual = hidden_states
+        hidden_states = self.norm_in(hidden_states)
+
+        if self._chunk_size is not None:
+            hidden_states = _chunked_feed_forward(self.ff_in, hidden_states, self._chunk_dim, self._chunk_size)
+        else:
+            hidden_states = self.ff_in(hidden_states)
+
+        if self.is_res:
+            hidden_states = hidden_states + residual
+
+        norm_hidden_states = self.norm1(hidden_states)
+        attn_output = self.attn1(norm_hidden_states, encoder_hidden_states=None)
+        hidden_states = attn_output + hidden_states
+
+        # 3. Cross-Attention
+        if self.attn2 is not None:
+            norm_hidden_states = self.norm2(hidden_states)
+            attn_output = self.attn2(norm_hidden_states, encoder_hidden_states=encoder_hidden_states)
+            hidden_states = attn_output + hidden_states
+
+        # 4. Feed-forward
+        norm_hidden_states = self.norm3(hidden_states)
+
+        if self._chunk_size is not None:
+            ff_output = _chunked_feed_forward(self.ff, norm_hidden_states, self._chunk_dim, self._chunk_size)
+        else:
+            ff_output = self.ff(norm_hidden_states)
+
+        if self.is_res:
+            hidden_states = ff_output + hidden_states
+        else:
+            hidden_states = ff_output
+
+        hidden_states = hidden_states[None, :].reshape(batch_size, seq_length, num_frames, channels)
+        hidden_states = hidden_states.permute(0, 2, 1, 3)
+        hidden_states = hidden_states.reshape(batch_size * num_frames, seq_length, channels)
+
+        return hidden_states
+
+
+class SkipFFTransformerBlock(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        kv_input_dim: int,
+        kv_input_dim_proj_use_bias: bool,
+        dropout=0.0,
+        cross_attention_dim: Optional[int] = None,
+        attention_bias: bool = False,
+        attention_out_bias: bool = True,
+    ):
+        super().__init__()
+        if kv_input_dim != dim:
+            self.kv_mapper = nn.Linear(kv_input_dim, dim, kv_input_dim_proj_use_bias)
+        else:
+            self.kv_mapper = None
+
+        self.norm1 = RMSNorm(dim, 1e-06)
+
+        self.attn1 = Attention(
+            query_dim=dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            dropout=dropout,
+            bias=attention_bias,
+            cross_attention_dim=cross_attention_dim,
+            out_bias=attention_out_bias,
+        )
+
+        self.norm2 = RMSNorm(dim, 1e-06)
+
+        self.attn2 = Attention(
+            query_dim=dim,
+            cross_attention_dim=cross_attention_dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            dropout=dropout,
+            bias=attention_bias,
+            out_bias=attention_out_bias,
+        )
+
+    def forward(self, hidden_states, encoder_hidden_states, cross_attention_kwargs):
+        cross_attention_kwargs = cross_attention_kwargs.copy() if cross_attention_kwargs is not None else {}
+
+        if self.kv_mapper is not None:
+            encoder_hidden_states = self.kv_mapper(F.silu(encoder_hidden_states))
+
+        norm_hidden_states = self.norm1(hidden_states)
+
+        attn_output = self.attn1(
+            norm_hidden_states,
+            encoder_hidden_states=encoder_hidden_states,
+            **cross_attention_kwargs,
+        )
+
+        hidden_states = attn_output + hidden_states
+
+        norm_hidden_states = self.norm2(hidden_states)
+
+        attn_output = self.attn2(
+            norm_hidden_states,
+            encoder_hidden_states=encoder_hidden_states,
+            **cross_attention_kwargs,
+        )
+
+        hidden_states = attn_output + hidden_states
+
+        return hidden_states
+
+
+class FeedForward(nn.Module):
+    r"""
+    A feed-forward layer.
+
+    Parameters:
+        dim (`int`): The number of channels in the input.
+        dim_out (`int`, *optional*): The number of channels in the output. If not given, defaults to `dim`.
+        mult (`int`, *optional*, defaults to 4): The multiplier to use for the hidden dimension.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
+        final_dropout (`bool` *optional*, defaults to False): Apply a final dropout.
+        bias (`bool`, defaults to True): Whether to use a bias in the linear layer.
+    """
+
+    def __init__(
+        self,
+        dim: int,
+        dim_out: Optional[int] = None,
+        mult: int = 4,
+        dropout: float = 0.0,
+        activation_fn: str = "geglu",
+        final_dropout: bool = False,
+        inner_dim=None,
+        bias: bool = True,
+    ):
+        super().__init__()
+        if inner_dim is None:
+            inner_dim = int(dim * mult)
+        dim_out = dim_out if dim_out is not None else dim
+        linear_cls = LoRACompatibleLinear if not USE_PEFT_BACKEND else nn.Linear
+
+        if activation_fn == "gelu":
+            act_fn = GELU(dim, inner_dim, bias=bias)
+        if activation_fn == "gelu-approximate":
+            act_fn = GELU(dim, inner_dim, approximate="tanh", bias=bias)
+        elif activation_fn == "geglu":
+            act_fn = GEGLU(dim, inner_dim, bias=bias)
+        elif activation_fn == "geglu-approximate":
+            act_fn = ApproximateGELU(dim, inner_dim, bias=bias)
+
+        self.net = nn.ModuleList([])
+        # project in
+        self.net.append(act_fn)
+        # project dropout
+        self.net.append(nn.Dropout(dropout))
+        # project out
+        self.net.append(linear_cls(inner_dim, dim_out, bias=bias))
+        # FF as used in Vision Transformer, MLP-Mixer, etc. have a final dropout
+        if final_dropout:
+            self.net.append(nn.Dropout(dropout))
+
+    def forward(self, hidden_states: torch.Tensor, scale: float = 1.0) -> torch.Tensor:
+        compatible_cls = (GEGLU,) if USE_PEFT_BACKEND else (GEGLU, LoRACompatibleLinear)
+        for module in self.net:
+            if isinstance(module, compatible_cls):
+                hidden_states = module(hidden_states, scale)
+            else:
+                hidden_states = module(hidden_states)
+        return hidden_states

src/diffusers/models/attention_flax.py

@@ -0,0 +1,494 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import functools
+import math
+
+import flax.linen as nn
+import jax
+import jax.numpy as jnp
+
+
+def _query_chunk_attention(query, key, value, precision, key_chunk_size: int = 4096):
+    """Multi-head dot product attention with a limited number of queries."""
+    num_kv, num_heads, k_features = key.shape[-3:]
+    v_features = value.shape[-1]
+    key_chunk_size = min(key_chunk_size, num_kv)
+    query = query / jnp.sqrt(k_features)
+
+    @functools.partial(jax.checkpoint, prevent_cse=False)
+    def summarize_chunk(query, key, value):
+        attn_weights = jnp.einsum("...qhd,...khd->...qhk", query, key, precision=precision)
+
+        max_score = jnp.max(attn_weights, axis=-1, keepdims=True)
+        max_score = jax.lax.stop_gradient(max_score)
+        exp_weights = jnp.exp(attn_weights - max_score)
+
+        exp_values = jnp.einsum("...vhf,...qhv->...qhf", value, exp_weights, precision=precision)
+        max_score = jnp.einsum("...qhk->...qh", max_score)
+
+        return (exp_values, exp_weights.sum(axis=-1), max_score)
+
+    def chunk_scanner(chunk_idx):
+        # julienne key array
+        key_chunk = jax.lax.dynamic_slice(
+            operand=key,
+            start_indices=[0] * (key.ndim - 3) + [chunk_idx, 0, 0],  # [...,k,h,d]
+            slice_sizes=list(key.shape[:-3]) + [key_chunk_size, num_heads, k_features],  # [...,k,h,d]
+        )
+
+        # julienne value array
+        value_chunk = jax.lax.dynamic_slice(
+            operand=value,
+            start_indices=[0] * (value.ndim - 3) + [chunk_idx, 0, 0],  # [...,v,h,d]
+            slice_sizes=list(value.shape[:-3]) + [key_chunk_size, num_heads, v_features],  # [...,v,h,d]
+        )
+
+        return summarize_chunk(query, key_chunk, value_chunk)
+
+    chunk_values, chunk_weights, chunk_max = jax.lax.map(f=chunk_scanner, xs=jnp.arange(0, num_kv, key_chunk_size))
+
+    global_max = jnp.max(chunk_max, axis=0, keepdims=True)
+    max_diffs = jnp.exp(chunk_max - global_max)
+
+    chunk_values *= jnp.expand_dims(max_diffs, axis=-1)
+    chunk_weights *= max_diffs
+
+    all_values = chunk_values.sum(axis=0)
+    all_weights = jnp.expand_dims(chunk_weights, -1).sum(axis=0)
+
+    return all_values / all_weights
+
+
+def jax_memory_efficient_attention(
+    query, key, value, precision=jax.lax.Precision.HIGHEST, query_chunk_size: int = 1024, key_chunk_size: int = 4096
+):
+    r"""
+    Flax Memory-efficient multi-head dot product attention. https://arxiv.org/abs/2112.05682v2
+    https://github.com/AminRezaei0x443/memory-efficient-attention
+
+    Args:
+        query (`jnp.ndarray`): (batch..., query_length, head, query_key_depth_per_head)
+        key (`jnp.ndarray`): (batch..., key_value_length, head, query_key_depth_per_head)
+        value (`jnp.ndarray`): (batch..., key_value_length, head, value_depth_per_head)
+        precision (`jax.lax.Precision`, *optional*, defaults to `jax.lax.Precision.HIGHEST`):
+            numerical precision for computation
+        query_chunk_size (`int`, *optional*, defaults to 1024):
+            chunk size to divide query array value must divide query_length equally without remainder
+        key_chunk_size (`int`, *optional*, defaults to 4096):
+            chunk size to divide key and value array value must divide key_value_length equally without remainder
+
+    Returns:
+        (`jnp.ndarray`) with shape of (batch..., query_length, head, value_depth_per_head)
+    """
+    num_q, num_heads, q_features = query.shape[-3:]
+
+    def chunk_scanner(chunk_idx, _):
+        # julienne query array
+        query_chunk = jax.lax.dynamic_slice(
+            operand=query,
+            start_indices=([0] * (query.ndim - 3)) + [chunk_idx, 0, 0],  # [...,q,h,d]
+            slice_sizes=list(query.shape[:-3]) + [min(query_chunk_size, num_q), num_heads, q_features],  # [...,q,h,d]
+        )
+
+        return (
+            chunk_idx + query_chunk_size,  # unused ignore it
+            _query_chunk_attention(
+                query=query_chunk, key=key, value=value, precision=precision, key_chunk_size=key_chunk_size
+            ),
+        )
+
+    _, res = jax.lax.scan(
+        f=chunk_scanner,
+        init=0,
+        xs=None,
+        length=math.ceil(num_q / query_chunk_size),  # start counter  # stop counter
+    )
+
+    return jnp.concatenate(res, axis=-3)  # fuse the chunked result back
+
+
+class FlaxAttention(nn.Module):
+    r"""
+    A Flax multi-head attention module as described in: https://arxiv.org/abs/1706.03762
+
+    Parameters:
+        query_dim (:obj:`int`):
+            Input hidden states dimension
+        heads (:obj:`int`, *optional*, defaults to 8):
+            Number of heads
+        dim_head (:obj:`int`, *optional*, defaults to 64):
+            Hidden states dimension inside each head
+        dropout (:obj:`float`, *optional*, defaults to 0.0):
+            Dropout rate
+        use_memory_efficient_attention (`bool`, *optional*, defaults to `False`):
+            enable memory efficient attention https://arxiv.org/abs/2112.05682
+        split_head_dim (`bool`, *optional*, defaults to `False`):
+            Whether to split the head dimension into a new axis for the self-attention computation. In most cases,
+            enabling this flag should speed up the computation for Stable Diffusion 2.x and Stable Diffusion XL.
+        dtype (:obj:`jnp.dtype`, *optional*, defaults to jnp.float32):
+            Parameters `dtype`
+
+    """
+
+    query_dim: int
+    heads: int = 8
+    dim_head: int = 64
+    dropout: float = 0.0
+    use_memory_efficient_attention: bool = False
+    split_head_dim: bool = False
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self):
+        inner_dim = self.dim_head * self.heads
+        self.scale = self.dim_head**-0.5
+
+        # Weights were exported with old names {to_q, to_k, to_v, to_out}
+        self.query = nn.Dense(inner_dim, use_bias=False, dtype=self.dtype, name="to_q")
+        self.key = nn.Dense(inner_dim, use_bias=False, dtype=self.dtype, name="to_k")
+        self.value = nn.Dense(inner_dim, use_bias=False, dtype=self.dtype, name="to_v")
+
+        self.proj_attn = nn.Dense(self.query_dim, dtype=self.dtype, name="to_out_0")
+        self.dropout_layer = nn.Dropout(rate=self.dropout)
+
+    def reshape_heads_to_batch_dim(self, tensor):
+        batch_size, seq_len, dim = tensor.shape
+        head_size = self.heads
+        tensor = tensor.reshape(batch_size, seq_len, head_size, dim // head_size)
+        tensor = jnp.transpose(tensor, (0, 2, 1, 3))
+        tensor = tensor.reshape(batch_size * head_size, seq_len, dim // head_size)
+        return tensor
+
+    def reshape_batch_dim_to_heads(self, tensor):
+        batch_size, seq_len, dim = tensor.shape
+        head_size = self.heads
+        tensor = tensor.reshape(batch_size // head_size, head_size, seq_len, dim)
+        tensor = jnp.transpose(tensor, (0, 2, 1, 3))
+        tensor = tensor.reshape(batch_size // head_size, seq_len, dim * head_size)
+        return tensor
+
+    def __call__(self, hidden_states, context=None, deterministic=True):
+        context = hidden_states if context is None else context
+
+        query_proj = self.query(hidden_states)
+        key_proj = self.key(context)
+        value_proj = self.value(context)
+
+        if self.split_head_dim:
+            b = hidden_states.shape[0]
+            query_states = jnp.reshape(query_proj, (b, -1, self.heads, self.dim_head))
+            key_states = jnp.reshape(key_proj, (b, -1, self.heads, self.dim_head))
+            value_states = jnp.reshape(value_proj, (b, -1, self.heads, self.dim_head))
+        else:
+            query_states = self.reshape_heads_to_batch_dim(query_proj)
+            key_states = self.reshape_heads_to_batch_dim(key_proj)
+            value_states = self.reshape_heads_to_batch_dim(value_proj)
+
+        if self.use_memory_efficient_attention:
+            query_states = query_states.transpose(1, 0, 2)
+            key_states = key_states.transpose(1, 0, 2)
+            value_states = value_states.transpose(1, 0, 2)
+
+            # this if statement create a chunk size for each layer of the unet
+            # the chunk size is equal to the query_length dimension of the deepest layer of the unet
+
+            flatten_latent_dim = query_states.shape[-3]
+            if flatten_latent_dim % 64 == 0:
+                query_chunk_size = int(flatten_latent_dim / 64)
+            elif flatten_latent_dim % 16 == 0:
+                query_chunk_size = int(flatten_latent_dim / 16)
+            elif flatten_latent_dim % 4 == 0:
+                query_chunk_size = int(flatten_latent_dim / 4)
+            else:
+                query_chunk_size = int(flatten_latent_dim)
+
+            hidden_states = jax_memory_efficient_attention(
+                query_states, key_states, value_states, query_chunk_size=query_chunk_size, key_chunk_size=4096 * 4
+            )
+
+            hidden_states = hidden_states.transpose(1, 0, 2)
+        else:
+            # compute attentions
+            if self.split_head_dim:
+                attention_scores = jnp.einsum("b t n h, b f n h -> b n f t", key_states, query_states)
+            else:
+                attention_scores = jnp.einsum("b i d, b j d->b i j", query_states, key_states)
+
+            attention_scores = attention_scores * self.scale
+            attention_probs = nn.softmax(attention_scores, axis=-1 if self.split_head_dim else 2)
+
+            # attend to values
+            if self.split_head_dim:
+                hidden_states = jnp.einsum("b n f t, b t n h -> b f n h", attention_probs, value_states)
+                b = hidden_states.shape[0]
+                hidden_states = jnp.reshape(hidden_states, (b, -1, self.heads * self.dim_head))
+            else:
+                hidden_states = jnp.einsum("b i j, b j d -> b i d", attention_probs, value_states)
+                hidden_states = self.reshape_batch_dim_to_heads(hidden_states)
+
+        hidden_states = self.proj_attn(hidden_states)
+        return self.dropout_layer(hidden_states, deterministic=deterministic)
+
+
+class FlaxBasicTransformerBlock(nn.Module):
+    r"""
+    A Flax transformer block layer with `GLU` (Gated Linear Unit) activation function as described in:
+    https://arxiv.org/abs/1706.03762
+
+
+    Parameters:
+        dim (:obj:`int`):
+            Inner hidden states dimension
+        n_heads (:obj:`int`):
+            Number of heads
+        d_head (:obj:`int`):
+            Hidden states dimension inside each head
+        dropout (:obj:`float`, *optional*, defaults to 0.0):
+            Dropout rate
+        only_cross_attention (`bool`, defaults to `False`):
+            Whether to only apply cross attention.
+        dtype (:obj:`jnp.dtype`, *optional*, defaults to jnp.float32):
+            Parameters `dtype`
+        use_memory_efficient_attention (`bool`, *optional*, defaults to `False`):
+            enable memory efficient attention https://arxiv.org/abs/2112.05682
+        split_head_dim (`bool`, *optional*, defaults to `False`):
+            Whether to split the head dimension into a new axis for the self-attention computation. In most cases,
+            enabling this flag should speed up the computation for Stable Diffusion 2.x and Stable Diffusion XL.
+    """
+
+    dim: int
+    n_heads: int
+    d_head: int
+    dropout: float = 0.0
+    only_cross_attention: bool = False
+    dtype: jnp.dtype = jnp.float32
+    use_memory_efficient_attention: bool = False
+    split_head_dim: bool = False
+
+    def setup(self):
+        # self attention (or cross_attention if only_cross_attention is True)
+        self.attn1 = FlaxAttention(
+            self.dim,
+            self.n_heads,
+            self.d_head,
+            self.dropout,
+            self.use_memory_efficient_attention,
+            self.split_head_dim,
+            dtype=self.dtype,
+        )
+        # cross attention
+        self.attn2 = FlaxAttention(
+            self.dim,
+            self.n_heads,
+            self.d_head,
+            self.dropout,
+            self.use_memory_efficient_attention,
+            self.split_head_dim,
+            dtype=self.dtype,
+        )
+        self.ff = FlaxFeedForward(dim=self.dim, dropout=self.dropout, dtype=self.dtype)
+        self.norm1 = nn.LayerNorm(epsilon=1e-5, dtype=self.dtype)
+        self.norm2 = nn.LayerNorm(epsilon=1e-5, dtype=self.dtype)
+        self.norm3 = nn.LayerNorm(epsilon=1e-5, dtype=self.dtype)
+        self.dropout_layer = nn.Dropout(rate=self.dropout)
+
+    def __call__(self, hidden_states, context, deterministic=True):
+        # self attention
+        residual = hidden_states
+        if self.only_cross_attention:
+            hidden_states = self.attn1(self.norm1(hidden_states), context, deterministic=deterministic)
+        else:
+            hidden_states = self.attn1(self.norm1(hidden_states), deterministic=deterministic)
+        hidden_states = hidden_states + residual
+
+        # cross attention
+        residual = hidden_states
+        hidden_states = self.attn2(self.norm2(hidden_states), context, deterministic=deterministic)
+        hidden_states = hidden_states + residual
+
+        # feed forward
+        residual = hidden_states
+        hidden_states = self.ff(self.norm3(hidden_states), deterministic=deterministic)
+        hidden_states = hidden_states + residual
+
+        return self.dropout_layer(hidden_states, deterministic=deterministic)
+
+
+class FlaxTransformer2DModel(nn.Module):
+    r"""
+    A Spatial Transformer layer with Gated Linear Unit (GLU) activation function as described in:
+    https://arxiv.org/pdf/1506.02025.pdf
+
+
+    Parameters:
+        in_channels (:obj:`int`):
+            Input number of channels
+        n_heads (:obj:`int`):
+            Number of heads
+        d_head (:obj:`int`):
+            Hidden states dimension inside each head
+        depth (:obj:`int`, *optional*, defaults to 1):
+            Number of transformers block
+        dropout (:obj:`float`, *optional*, defaults to 0.0):
+            Dropout rate
+        use_linear_projection (`bool`, defaults to `False`): tbd
+        only_cross_attention (`bool`, defaults to `False`): tbd
+        dtype (:obj:`jnp.dtype`, *optional*, defaults to jnp.float32):
+            Parameters `dtype`
+        use_memory_efficient_attention (`bool`, *optional*, defaults to `False`):
+            enable memory efficient attention https://arxiv.org/abs/2112.05682
+        split_head_dim (`bool`, *optional*, defaults to `False`):
+            Whether to split the head dimension into a new axis for the self-attention computation. In most cases,
+            enabling this flag should speed up the computation for Stable Diffusion 2.x and Stable Diffusion XL.
+    """
+
+    in_channels: int
+    n_heads: int
+    d_head: int
+    depth: int = 1
+    dropout: float = 0.0
+    use_linear_projection: bool = False
+    only_cross_attention: bool = False
+    dtype: jnp.dtype = jnp.float32
+    use_memory_efficient_attention: bool = False
+    split_head_dim: bool = False
+
+    def setup(self):
+        self.norm = nn.GroupNorm(num_groups=32, epsilon=1e-5)
+
+        inner_dim = self.n_heads * self.d_head
+        if self.use_linear_projection:
+            self.proj_in = nn.Dense(inner_dim, dtype=self.dtype)
+        else:
+            self.proj_in = nn.Conv(
+                inner_dim,
+                kernel_size=(1, 1),
+                strides=(1, 1),
+                padding="VALID",
+                dtype=self.dtype,
+            )
+
+        self.transformer_blocks = [
+            FlaxBasicTransformerBlock(
+                inner_dim,
+                self.n_heads,
+                self.d_head,
+                dropout=self.dropout,
+                only_cross_attention=self.only_cross_attention,
+                dtype=self.dtype,
+                use_memory_efficient_attention=self.use_memory_efficient_attention,
+                split_head_dim=self.split_head_dim,
+            )
+            for _ in range(self.depth)
+        ]
+
+        if self.use_linear_projection:
+            self.proj_out = nn.Dense(inner_dim, dtype=self.dtype)
+        else:
+            self.proj_out = nn.Conv(
+                inner_dim,
+                kernel_size=(1, 1),
+                strides=(1, 1),
+                padding="VALID",
+                dtype=self.dtype,
+            )
+
+        self.dropout_layer = nn.Dropout(rate=self.dropout)
+
+    def __call__(self, hidden_states, context, deterministic=True):
+        batch, height, width, channels = hidden_states.shape
+        residual = hidden_states
+        hidden_states = self.norm(hidden_states)
+        if self.use_linear_projection:
+            hidden_states = hidden_states.reshape(batch, height * width, channels)
+            hidden_states = self.proj_in(hidden_states)
+        else:
+            hidden_states = self.proj_in(hidden_states)
+            hidden_states = hidden_states.reshape(batch, height * width, channels)
+
+        for transformer_block in self.transformer_blocks:
+            hidden_states = transformer_block(hidden_states, context, deterministic=deterministic)
+
+        if self.use_linear_projection:
+            hidden_states = self.proj_out(hidden_states)
+            hidden_states = hidden_states.reshape(batch, height, width, channels)
+        else:
+            hidden_states = hidden_states.reshape(batch, height, width, channels)
+            hidden_states = self.proj_out(hidden_states)
+
+        hidden_states = hidden_states + residual
+        return self.dropout_layer(hidden_states, deterministic=deterministic)
+
+
+class FlaxFeedForward(nn.Module):
+    r"""
+    Flax module that encapsulates two Linear layers separated by a non-linearity. It is the counterpart of PyTorch's
+    [`FeedForward`] class, with the following simplifications:
+    - The activation function is currently hardcoded to a gated linear unit from:
+    https://arxiv.org/abs/2002.05202
+    - `dim_out` is equal to `dim`.
+    - The number of hidden dimensions is hardcoded to `dim * 4` in [`FlaxGELU`].
+
+    Parameters:
+        dim (:obj:`int`):
+            Inner hidden states dimension
+        dropout (:obj:`float`, *optional*, defaults to 0.0):
+            Dropout rate
+        dtype (:obj:`jnp.dtype`, *optional*, defaults to jnp.float32):
+            Parameters `dtype`
+    """
+
+    dim: int
+    dropout: float = 0.0
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self):
+        # The second linear layer needs to be called
+        # net_2 for now to match the index of the Sequential layer
+        self.net_0 = FlaxGEGLU(self.dim, self.dropout, self.dtype)
+        self.net_2 = nn.Dense(self.dim, dtype=self.dtype)
+
+    def __call__(self, hidden_states, deterministic=True):
+        hidden_states = self.net_0(hidden_states, deterministic=deterministic)
+        hidden_states = self.net_2(hidden_states)
+        return hidden_states
+
+
+class FlaxGEGLU(nn.Module):
+    r"""
+    Flax implementation of a Linear layer followed by the variant of the gated linear unit activation function from
+    https://arxiv.org/abs/2002.05202.
+
+    Parameters:
+        dim (:obj:`int`):
+            Input hidden states dimension
+        dropout (:obj:`float`, *optional*, defaults to 0.0):
+            Dropout rate
+        dtype (:obj:`jnp.dtype`, *optional*, defaults to jnp.float32):
+            Parameters `dtype`
+    """
+
+    dim: int
+    dropout: float = 0.0
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self):
+        inner_dim = self.dim * 4
+        self.proj = nn.Dense(inner_dim * 2, dtype=self.dtype)
+        self.dropout_layer = nn.Dropout(rate=self.dropout)
+
+    def __call__(self, hidden_states, deterministic=True):
+        hidden_states = self.proj(hidden_states)
+        hidden_linear, hidden_gelu = jnp.split(hidden_states, 2, axis=2)
+        return self.dropout_layer(hidden_linear * nn.gelu(hidden_gelu), deterministic=deterministic)

src/diffusers/models/autoencoders/__init__.py

@@ -0,0 +1,5 @@
+from .autoencoder_asym_kl import AsymmetricAutoencoderKL
+from .autoencoder_kl import AutoencoderKL
+from .autoencoder_kl_temporal_decoder import AutoencoderKLTemporalDecoder
+from .autoencoder_tiny import AutoencoderTiny
+from .consistency_decoder_vae import ConsistencyDecoderVAE

src/diffusers/models/autoencoders/autoencoder_asym_kl.py

@@ -0,0 +1,186 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...utils.accelerate_utils import apply_forward_hook
+from ..modeling_outputs import AutoencoderKLOutput
+from ..modeling_utils import ModelMixin
+from .vae import DecoderOutput, DiagonalGaussianDistribution, Encoder, MaskConditionDecoder
+
+
+class AsymmetricAutoencoderKL(ModelMixin, ConfigMixin):
+    r"""
+    Designing a Better Asymmetric VQGAN for StableDiffusion https://arxiv.org/abs/2306.04632 . A VAE model with KL loss
+    for encoding images into latents and decoding latent representations into images.
+
+    This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented
+    for all models (such as downloading or saving).
+
+    Parameters:
+        in_channels (int, *optional*, defaults to 3): Number of channels in the input image.
+        out_channels (int,  *optional*, defaults to 3): Number of channels in the output.
+        down_block_types (`Tuple[str]`, *optional*, defaults to `("DownEncoderBlock2D",)`):
+            Tuple of downsample block types.
+        down_block_out_channels (`Tuple[int]`, *optional*, defaults to `(64,)`):
+            Tuple of down block output channels.
+        layers_per_down_block (`int`, *optional*, defaults to `1`):
+            Number layers for down block.
+        up_block_types (`Tuple[str]`, *optional*, defaults to `("UpDecoderBlock2D",)`):
+            Tuple of upsample block types.
+        up_block_out_channels (`Tuple[int]`, *optional*, defaults to `(64,)`):
+            Tuple of up block output channels.
+        layers_per_up_block (`int`, *optional*, defaults to `1`):
+            Number layers for up block.
+        act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use.
+        latent_channels (`int`, *optional*, defaults to 4): Number of channels in the latent space.
+        sample_size (`int`, *optional*, defaults to `32`): Sample input size.
+        norm_num_groups (`int`, *optional*, defaults to `32`):
+            Number of groups to use for the first normalization layer in ResNet blocks.
+        scaling_factor (`float`, *optional*, defaults to 0.18215):
+            The component-wise standard deviation of the trained latent space computed using the first batch of the
+            training set. This is used to scale the latent space to have unit variance when training the diffusion
+            model. The latents are scaled with the formula `z = z * scaling_factor` before being passed to the
+            diffusion model. When decoding, the latents are scaled back to the original scale with the formula: `z = 1
+            / scaling_factor * z`. For more details, refer to sections 4.3.2 and D.1 of the [High-Resolution Image
+            Synthesis with Latent Diffusion Models](https://arxiv.org/abs/2112.10752) paper.
+    """
+
+    @register_to_config
+    def __init__(
+        self,
+        in_channels: int = 3,
+        out_channels: int = 3,
+        down_block_types: Tuple[str, ...] = ("DownEncoderBlock2D",),
+        down_block_out_channels: Tuple[int, ...] = (64,),
+        layers_per_down_block: int = 1,
+        up_block_types: Tuple[str, ...] = ("UpDecoderBlock2D",),
+        up_block_out_channels: Tuple[int, ...] = (64,),
+        layers_per_up_block: int = 1,
+        act_fn: str = "silu",
+        latent_channels: int = 4,
+        norm_num_groups: int = 32,
+        sample_size: int = 32,
+        scaling_factor: float = 0.18215,
+    ) -> None:
+        super().__init__()
+
+        # pass init params to Encoder
+        self.encoder = Encoder(
+            in_channels=in_channels,
+            out_channels=latent_channels,
+            down_block_types=down_block_types,
+            block_out_channels=down_block_out_channels,
+            layers_per_block=layers_per_down_block,
+            act_fn=act_fn,
+            norm_num_groups=norm_num_groups,
+            double_z=True,
+        )
+
+        # pass init params to Decoder
+        self.decoder = MaskConditionDecoder(
+            in_channels=latent_channels,
+            out_channels=out_channels,
+            up_block_types=up_block_types,
+            block_out_channels=up_block_out_channels,
+            layers_per_block=layers_per_up_block,
+            act_fn=act_fn,
+            norm_num_groups=norm_num_groups,
+        )
+
+        self.quant_conv = nn.Conv2d(2 * latent_channels, 2 * latent_channels, 1)
+        self.post_quant_conv = nn.Conv2d(latent_channels, latent_channels, 1)
+
+        self.use_slicing = False
+        self.use_tiling = False
+
+        self.register_to_config(block_out_channels=up_block_out_channels)
+        self.register_to_config(force_upcast=False)
+
+    @apply_forward_hook
+    def encode(
+        self, x: torch.FloatTensor, return_dict: bool = True
+    ) -> Union[AutoencoderKLOutput, Tuple[torch.FloatTensor]]:
+        h = self.encoder(x)
+        moments = self.quant_conv(h)
+        posterior = DiagonalGaussianDistribution(moments)
+
+        if not return_dict:
+            return (posterior,)
+
+        return AutoencoderKLOutput(latent_dist=posterior)
+
+    def _decode(
+        self,
+        z: torch.FloatTensor,
+        image: Optional[torch.FloatTensor] = None,
+        mask: Optional[torch.FloatTensor] = None,
+        return_dict: bool = True,
+    ) -> Union[DecoderOutput, Tuple[torch.FloatTensor]]:
+        z = self.post_quant_conv(z)
+        dec = self.decoder(z, image, mask)
+
+        if not return_dict:
+            return (dec,)
+
+        return DecoderOutput(sample=dec)
+
+    @apply_forward_hook
+    def decode(
+        self,
+        z: torch.FloatTensor,
+        generator: Optional[torch.Generator] = None,
+        image: Optional[torch.FloatTensor] = None,
+        mask: Optional[torch.FloatTensor] = None,
+        return_dict: bool = True,
+    ) -> Union[DecoderOutput, Tuple[torch.FloatTensor]]:
+        decoded = self._decode(z, image, mask).sample
+
+        if not return_dict:
+            return (decoded,)
+
+        return DecoderOutput(sample=decoded)
+
+    def forward(
+        self,
+        sample: torch.FloatTensor,
+        mask: Optional[torch.FloatTensor] = None,
+        sample_posterior: bool = False,
+        return_dict: bool = True,
+        generator: Optional[torch.Generator] = None,
+    ) -> Union[DecoderOutput, Tuple[torch.FloatTensor]]:
+        r"""
+        Args:
+            sample (`torch.FloatTensor`): Input sample.
+            mask (`torch.FloatTensor`, *optional*, defaults to `None`): Optional inpainting mask.
+            sample_posterior (`bool`, *optional*, defaults to `False`):
+                Whether to sample from the posterior.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`DecoderOutput`] instead of a plain tuple.
+        """
+        x = sample
+        posterior = self.encode(x).latent_dist
+        if sample_posterior:
+            z = posterior.sample(generator=generator)
+        else:
+            z = posterior.mode()
+        dec = self.decode(z, sample, mask).sample
+
+        if not return_dict:
+            return (dec,)
+
+        return DecoderOutput(sample=dec)

src/diffusers/models/autoencoders/autoencoder_kl.py

@@ -0,0 +1,489 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import Dict, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...loaders import FromOriginalVAEMixin
+from ...utils.accelerate_utils import apply_forward_hook
+from ..attention_processor import (
+    ADDED_KV_ATTENTION_PROCESSORS,
+    CROSS_ATTENTION_PROCESSORS,
+    Attention,
+    AttentionProcessor,
+    AttnAddedKVProcessor,
+    AttnProcessor,
+)
+from ..modeling_outputs import AutoencoderKLOutput
+from ..modeling_utils import ModelMixin
+from .vae import Decoder, DecoderOutput, DiagonalGaussianDistribution, Encoder
+
+
+class AutoencoderKL(ModelMixin, ConfigMixin, FromOriginalVAEMixin):
+    r"""
+    A VAE model with KL loss for encoding images into latents and decoding latent representations into images.
+
+    This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented
+    for all models (such as downloading or saving).
+
+    Parameters:
+        in_channels (int, *optional*, defaults to 3): Number of channels in the input image.
+        out_channels (int,  *optional*, defaults to 3): Number of channels in the output.
+        down_block_types (`Tuple[str]`, *optional*, defaults to `("DownEncoderBlock2D",)`):
+            Tuple of downsample block types.
+        up_block_types (`Tuple[str]`, *optional*, defaults to `("UpDecoderBlock2D",)`):
+            Tuple of upsample block types.
+        block_out_channels (`Tuple[int]`, *optional*, defaults to `(64,)`):
+            Tuple of block output channels.
+        act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use.
+        latent_channels (`int`, *optional*, defaults to 4): Number of channels in the latent space.
+        sample_size (`int`, *optional*, defaults to `32`): Sample input size.
+        scaling_factor (`float`, *optional*, defaults to 0.18215):
+            The component-wise standard deviation of the trained latent space computed using the first batch of the
+            training set. This is used to scale the latent space to have unit variance when training the diffusion
+            model. The latents are scaled with the formula `z = z * scaling_factor` before being passed to the
+            diffusion model. When decoding, the latents are scaled back to the original scale with the formula: `z = 1
+            / scaling_factor * z`. For more details, refer to sections 4.3.2 and D.1 of the [High-Resolution Image
+            Synthesis with Latent Diffusion Models](https://arxiv.org/abs/2112.10752) paper.
+        force_upcast (`bool`, *optional*, default to `True`):
+            If enabled it will force the VAE to run in float32 for high image resolution pipelines, such as SD-XL. VAE
+            can be fine-tuned / trained to a lower range without loosing too much precision in which case
+            `force_upcast` can be set to `False` - see: https://huggingface.co/madebyollin/sdxl-vae-fp16-fix
+    """
+
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        in_channels: int = 3,
+        out_channels: int = 3,
+        down_block_types: Tuple[str] = ("DownEncoderBlock2D",),
+        up_block_types: Tuple[str] = ("UpDecoderBlock2D",),
+        block_out_channels: Tuple[int] = (64,),
+        layers_per_block: int = 1,
+        act_fn: str = "silu",
+        latent_channels: int = 4,
+        norm_num_groups: int = 32,
+        sample_size: int = 32,
+        scaling_factor: float = 0.18215,
+        force_upcast: float = True,
+    ):
+        super().__init__()
+
+        # pass init params to Encoder
+        self.encoder = Encoder(
+            in_channels=in_channels,
+            out_channels=latent_channels,
+            down_block_types=down_block_types,
+            block_out_channels=block_out_channels,
+            layers_per_block=layers_per_block,
+            act_fn=act_fn,
+            norm_num_groups=norm_num_groups,
+            double_z=True,
+        )
+
+        # pass init params to Decoder
+        self.decoder = Decoder(
+            in_channels=latent_channels,
+            out_channels=out_channels,
+            up_block_types=up_block_types,
+            block_out_channels=block_out_channels,
+            layers_per_block=layers_per_block,
+            norm_num_groups=norm_num_groups,
+            act_fn=act_fn,
+        )
+
+        self.quant_conv = nn.Conv2d(2 * latent_channels, 2 * latent_channels, 1)
+        self.post_quant_conv = nn.Conv2d(latent_channels, latent_channels, 1)
+
+        self.use_slicing = False
+        self.use_tiling = False
+
+        # only relevant if vae tiling is enabled
+        self.tile_sample_min_size = self.config.sample_size
+        sample_size = (
+            self.config.sample_size[0]
+            if isinstance(self.config.sample_size, (list, tuple))
+            else self.config.sample_size
+        )
+        self.tile_latent_min_size = int(sample_size / (2 ** (len(self.config.block_out_channels) - 1)))
+        self.tile_overlap_factor = 0.25
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if isinstance(module, (Encoder, Decoder)):
+            module.gradient_checkpointing = value
+
+    def enable_tiling(self, use_tiling: bool = True):
+        r"""
+        Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
+        compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
+        processing larger images.
+        """
+        self.use_tiling = use_tiling
+
+    def disable_tiling(self):
+        r"""
+        Disable tiled VAE decoding. If `enable_tiling` was previously enabled, this method will go back to computing
+        decoding in one step.
+        """
+        self.enable_tiling(False)
+
+    def enable_slicing(self):
+        r"""
+        Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
+        compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
+        """
+        self.use_slicing = True
+
+    def disable_slicing(self):
+        r"""
+        Disable sliced VAE decoding. If `enable_slicing` was previously enabled, this method will go back to computing
+        decoding in one step.
+        """
+        self.use_slicing = False
+
+    @property
+    # Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.attn_processors
+    def attn_processors(self) -> Dict[str, AttentionProcessor]:
+        r"""
+        Returns:
+            `dict` of attention processors: A dictionary containing all attention processors used in the model with
+            indexed by its weight name.
+        """
+        # set recursively
+        processors = {}
+
+        def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
+            if hasattr(module, "get_processor"):
+                processors[f"{name}.processor"] = module.get_processor(return_deprecated_lora=True)
+
+            for sub_name, child in module.named_children():
+                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
+
+            return processors
+
+        for name, module in self.named_children():
+            fn_recursive_add_processors(name, module, processors)
+
+        return processors
+
+    # Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.set_attn_processor
+    def set_attn_processor(
+        self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]], _remove_lora=False
+    ):
+        r"""
+        Sets the attention processor to use to compute attention.
+
+        Parameters:
+            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
+                The instantiated processor class or a dictionary of processor classes that will be set as the processor
+                for **all** `Attention` layers.
+
+                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
+                processor. This is strongly recommended when setting trainable attention processors.
+
+        """
+        count = len(self.attn_processors.keys())
+
+        if isinstance(processor, dict) and len(processor) != count:
+            raise ValueError(
+                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
+                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
+            )
+
+        def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
+            if hasattr(module, "set_processor"):
+                if not isinstance(processor, dict):
+                    module.set_processor(processor, _remove_lora=_remove_lora)
+                else:
+                    module.set_processor(processor.pop(f"{name}.processor"), _remove_lora=_remove_lora)
+
+            for sub_name, child in module.named_children():
+                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
+
+        for name, module in self.named_children():
+            fn_recursive_attn_processor(name, module, processor)
+
+    # Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.set_default_attn_processor
+    def set_default_attn_processor(self):
+        """
+        Disables custom attention processors and sets the default attention implementation.
+        """
+        if all(proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+            processor = AttnAddedKVProcessor()
+        elif all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+            processor = AttnProcessor()
+        else:
+            raise ValueError(
+                f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
+            )
+
+        self.set_attn_processor(processor, _remove_lora=True)
+
+    @apply_forward_hook
+    def encode(
+        self, x: torch.FloatTensor, return_dict: bool = True
+    ) -> Union[AutoencoderKLOutput, Tuple[DiagonalGaussianDistribution]]:
+        """
+        Encode a batch of images into latents.
+
+        Args:
+            x (`torch.FloatTensor`): Input batch of images.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether to return a [`~models.autoencoder_kl.AutoencoderKLOutput`] instead of a plain tuple.
+
+        Returns:
+                The latent representations of the encoded images. If `return_dict` is True, a
+                [`~models.autoencoder_kl.AutoencoderKLOutput`] is returned, otherwise a plain `tuple` is returned.
+        """
+        if self.use_tiling and (x.shape[-1] > self.tile_sample_min_size or x.shape[-2] > self.tile_sample_min_size):
+            return self.tiled_encode(x, return_dict=return_dict)
+
+        if self.use_slicing and x.shape[0] > 1:
+            encoded_slices = [self.encoder(x_slice) for x_slice in x.split(1)]
+            h = torch.cat(encoded_slices)
+        else:
+            h = self.encoder(x)
+
+        moments = self.quant_conv(h)
+        posterior = DiagonalGaussianDistribution(moments)
+
+        if not return_dict:
+            return (posterior,)
+
+        return AutoencoderKLOutput(latent_dist=posterior)
+
+    def _decode(self, z: torch.FloatTensor, return_dict: bool = True) -> Union[DecoderOutput, torch.FloatTensor]:
+        if self.use_tiling and (z.shape[-1] > self.tile_latent_min_size or z.shape[-2] > self.tile_latent_min_size):
+            return self.tiled_decode(z, return_dict=return_dict)
+
+        z = self.post_quant_conv(z)
+        dec = self.decoder(z)
+
+        if not return_dict:
+            return (dec,)
+
+        return DecoderOutput(sample=dec)
+
+    @apply_forward_hook
+    def decode(
+        self, z: torch.FloatTensor, return_dict: bool = True, generator=None
+    ) -> Union[DecoderOutput, torch.FloatTensor]:
+        """
+        Decode a batch of images.
+
+        Args:
+            z (`torch.FloatTensor`): Input batch of latent vectors.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether to return a [`~models.vae.DecoderOutput`] instead of a plain tuple.
+
+        Returns:
+            [`~models.vae.DecoderOutput`] or `tuple`:
+                If return_dict is True, a [`~models.vae.DecoderOutput`] is returned, otherwise a plain `tuple` is
+                returned.
+
+        """
+        if self.use_slicing and z.shape[0] > 1:
+            decoded_slices = [self._decode(z_slice).sample for z_slice in z.split(1)]
+            decoded = torch.cat(decoded_slices)
+        else:
+            decoded = self._decode(z).sample
+
+        if not return_dict:
+            return (decoded,)
+
+        return DecoderOutput(sample=decoded)
+
+    def blend_v(self, a: torch.Tensor, b: torch.Tensor, blend_extent: int) -> torch.Tensor:
+        blend_extent = min(a.shape[2], b.shape[2], blend_extent)
+        for y in range(blend_extent):
+            b[:, :, y, :] = a[:, :, -blend_extent + y, :] * (1 - y / blend_extent) + b[:, :, y, :] * (y / blend_extent)
+        return b
+
+    def blend_h(self, a: torch.Tensor, b: torch.Tensor, blend_extent: int) -> torch.Tensor:
+        blend_extent = min(a.shape[3], b.shape[3], blend_extent)
+        for x in range(blend_extent):
+            b[:, :, :, x] = a[:, :, :, -blend_extent + x] * (1 - x / blend_extent) + b[:, :, :, x] * (x / blend_extent)
+        return b
+
+    def tiled_encode(self, x: torch.FloatTensor, return_dict: bool = True) -> AutoencoderKLOutput:
+        r"""Encode a batch of images using a tiled encoder.
+
+        When this option is enabled, the VAE will split the input tensor into tiles to compute encoding in several
+        steps. This is useful to keep memory use constant regardless of image size. The end result of tiled encoding is
+        different from non-tiled encoding because each tile uses a different encoder. To avoid tiling artifacts, the
+        tiles overlap and are blended together to form a smooth output. You may still see tile-sized changes in the
+        output, but they should be much less noticeable.
+
+        Args:
+            x (`torch.FloatTensor`): Input batch of images.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~models.autoencoder_kl.AutoencoderKLOutput`] instead of a plain tuple.
+
+        Returns:
+            [`~models.autoencoder_kl.AutoencoderKLOutput`] or `tuple`:
+                If return_dict is True, a [`~models.autoencoder_kl.AutoencoderKLOutput`] is returned, otherwise a plain
+                `tuple` is returned.
+        """
+        overlap_size = int(self.tile_sample_min_size * (1 - self.tile_overlap_factor))
+        blend_extent = int(self.tile_latent_min_size * self.tile_overlap_factor)
+        row_limit = self.tile_latent_min_size - blend_extent
+
+        # Split the image into 512x512 tiles and encode them separately.
+        rows = []
+        for i in range(0, x.shape[2], overlap_size):
+            row = []
+            for j in range(0, x.shape[3], overlap_size):
+                tile = x[:, :, i : i + self.tile_sample_min_size, j : j + self.tile_sample_min_size]
+                tile = self.encoder(tile)
+                tile = self.quant_conv(tile)
+                row.append(tile)
+            rows.append(row)
+        result_rows = []
+        for i, row in enumerate(rows):
+            result_row = []
+            for j, tile in enumerate(row):
+                # blend the above tile and the left tile
+                # to the current tile and add the current tile to the result row
+                if i > 0:
+                    tile = self.blend_v(rows[i - 1][j], tile, blend_extent)
+                if j > 0:
+                    tile = self.blend_h(row[j - 1], tile, blend_extent)
+                result_row.append(tile[:, :, :row_limit, :row_limit])
+            result_rows.append(torch.cat(result_row, dim=3))
+
+        moments = torch.cat(result_rows, dim=2)
+        posterior = DiagonalGaussianDistribution(moments)
+
+        if not return_dict:
+            return (posterior,)
+
+        return AutoencoderKLOutput(latent_dist=posterior)
+
+    def tiled_decode(self, z: torch.FloatTensor, return_dict: bool = True) -> Union[DecoderOutput, torch.FloatTensor]:
+        r"""
+        Decode a batch of images using a tiled decoder.
+
+        Args:
+            z (`torch.FloatTensor`): Input batch of latent vectors.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~models.vae.DecoderOutput`] instead of a plain tuple.
+
+        Returns:
+            [`~models.vae.DecoderOutput`] or `tuple`:
+                If return_dict is True, a [`~models.vae.DecoderOutput`] is returned, otherwise a plain `tuple` is
+                returned.
+        """
+        overlap_size = int(self.tile_latent_min_size * (1 - self.tile_overlap_factor))
+        blend_extent = int(self.tile_sample_min_size * self.tile_overlap_factor)
+        row_limit = self.tile_sample_min_size - blend_extent
+
+        # Split z into overlapping 64x64 tiles and decode them separately.
+        # The tiles have an overlap to avoid seams between tiles.
+        rows = []
+        for i in range(0, z.shape[2], overlap_size):
+            row = []
+            for j in range(0, z.shape[3], overlap_size):
+                tile = z[:, :, i : i + self.tile_latent_min_size, j : j + self.tile_latent_min_size]
+                tile = self.post_quant_conv(tile)
+                decoded = self.decoder(tile)
+                row.append(decoded)
+            rows.append(row)
+        result_rows = []
+        for i, row in enumerate(rows):
+            result_row = []
+            for j, tile in enumerate(row):
+                # blend the above tile and the left tile
+                # to the current tile and add the current tile to the result row
+                if i > 0:
+                    tile = self.blend_v(rows[i - 1][j], tile, blend_extent)
+                if j > 0:
+                    tile = self.blend_h(row[j - 1], tile, blend_extent)
+                result_row.append(tile[:, :, :row_limit, :row_limit])
+            result_rows.append(torch.cat(result_row, dim=3))
+
+        dec = torch.cat(result_rows, dim=2)
+        if not return_dict:
+            return (dec,)
+
+        return DecoderOutput(sample=dec)
+
+    def forward(
+        self,
+        sample: torch.FloatTensor,
+        sample_posterior: bool = False,
+        return_dict: bool = True,
+        generator: Optional[torch.Generator] = None,
+    ) -> Union[DecoderOutput, torch.FloatTensor]:
+        r"""
+        Args:
+            sample (`torch.FloatTensor`): Input sample.
+            sample_posterior (`bool`, *optional*, defaults to `False`):
+                Whether to sample from the posterior.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`DecoderOutput`] instead of a plain tuple.
+        """
+        x = sample
+        posterior = self.encode(x).latent_dist
+        if sample_posterior:
+            z = posterior.sample(generator=generator)
+        else:
+            z = posterior.mode()
+        dec = self.decode(z).sample
+
+        if not return_dict:
+            return (dec,)
+
+        return DecoderOutput(sample=dec)
+
+    # Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.fuse_qkv_projections
+    def fuse_qkv_projections(self):
+        """
+        Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query,
+        key, value) are fused. For cross-attention modules, key and value projection matrices are fused.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+        """
+        self.original_attn_processors = None
+
+        for _, attn_processor in self.attn_processors.items():
+            if "Added" in str(attn_processor.__class__.__name__):
+                raise ValueError("`fuse_qkv_projections()` is not supported for models having added KV projections.")
+
+        self.original_attn_processors = self.attn_processors
+
+        for module in self.modules():
+            if isinstance(module, Attention):
+                module.fuse_projections(fuse=True)
+
+    # Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.unfuse_qkv_projections
+    def unfuse_qkv_projections(self):
+        """Disables the fused QKV projection if enabled.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+
+        """
+        if self.original_attn_processors is not None:
+            self.set_attn_processor(self.original_attn_processors)

src/diffusers/models/autoencoders/autoencoder_kl_temporal_decoder.py

@@ -0,0 +1,402 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import Dict, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...loaders import FromOriginalVAEMixin
+from ...utils import is_torch_version
+from ...utils.accelerate_utils import apply_forward_hook
+from ..attention_processor import CROSS_ATTENTION_PROCESSORS, AttentionProcessor, AttnProcessor
+from ..modeling_outputs import AutoencoderKLOutput
+from ..modeling_utils import ModelMixin
+from ..unet_3d_blocks import MidBlockTemporalDecoder, UpBlockTemporalDecoder
+from .vae import DecoderOutput, DiagonalGaussianDistribution, Encoder
+
+
+class TemporalDecoder(nn.Module):
+    def __init__(
+        self,
+        in_channels: int = 4,
+        out_channels: int = 3,
+        block_out_channels: Tuple[int] = (128, 256, 512, 512),
+        layers_per_block: int = 2,
+    ):
+        super().__init__()
+        self.layers_per_block = layers_per_block
+
+        self.conv_in = nn.Conv2d(in_channels, block_out_channels[-1], kernel_size=3, stride=1, padding=1)
+        self.mid_block = MidBlockTemporalDecoder(
+            num_layers=self.layers_per_block,
+            in_channels=block_out_channels[-1],
+            out_channels=block_out_channels[-1],
+            attention_head_dim=block_out_channels[-1],
+        )
+
+        # up
+        self.up_blocks = nn.ModuleList([])
+        reversed_block_out_channels = list(reversed(block_out_channels))
+        output_channel = reversed_block_out_channels[0]
+        for i in range(len(block_out_channels)):
+            prev_output_channel = output_channel
+            output_channel = reversed_block_out_channels[i]
+
+            is_final_block = i == len(block_out_channels) - 1
+            up_block = UpBlockTemporalDecoder(
+                num_layers=self.layers_per_block + 1,
+                in_channels=prev_output_channel,
+                out_channels=output_channel,
+                add_upsample=not is_final_block,
+            )
+            self.up_blocks.append(up_block)
+            prev_output_channel = output_channel
+
+        self.conv_norm_out = nn.GroupNorm(num_channels=block_out_channels[0], num_groups=32, eps=1e-6)
+
+        self.conv_act = nn.SiLU()
+        self.conv_out = torch.nn.Conv2d(
+            in_channels=block_out_channels[0],
+            out_channels=out_channels,
+            kernel_size=3,
+            padding=1,
+        )
+
+        conv_out_kernel_size = (3, 1, 1)
+        padding = [int(k // 2) for k in conv_out_kernel_size]
+        self.time_conv_out = torch.nn.Conv3d(
+            in_channels=out_channels,
+            out_channels=out_channels,
+            kernel_size=conv_out_kernel_size,
+            padding=padding,
+        )
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        sample: torch.FloatTensor,
+        image_only_indicator: torch.FloatTensor,
+        num_frames: int = 1,
+    ) -> torch.FloatTensor:
+        r"""The forward method of the `Decoder` class."""
+
+        sample = self.conv_in(sample)
+
+        upscale_dtype = next(iter(self.up_blocks.parameters())).dtype
+        if self.training and self.gradient_checkpointing:
+
+            def create_custom_forward(module):
+                def custom_forward(*inputs):
+                    return module(*inputs)
+
+                return custom_forward
+
+            if is_torch_version(">=", "1.11.0"):
+                # middle
+                sample = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(self.mid_block),
+                    sample,
+                    image_only_indicator,
+                    use_reentrant=False,
+                )
+                sample = sample.to(upscale_dtype)
+
+                # up
+                for up_block in self.up_blocks:
+                    sample = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(up_block),
+                        sample,
+                        image_only_indicator,
+                        use_reentrant=False,
+                    )
+            else:
+                # middle
+                sample = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(self.mid_block),
+                    sample,
+                    image_only_indicator,
+                )
+                sample = sample.to(upscale_dtype)
+
+                # up
+                for up_block in self.up_blocks:
+                    sample = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(up_block),
+                        sample,
+                        image_only_indicator,
+                    )
+        else:
+            # middle
+            sample = self.mid_block(sample, image_only_indicator=image_only_indicator)
+            sample = sample.to(upscale_dtype)
+
+            # up
+            for up_block in self.up_blocks:
+                sample = up_block(sample, image_only_indicator=image_only_indicator)
+
+        # post-process
+        sample = self.conv_norm_out(sample)
+        sample = self.conv_act(sample)
+        sample = self.conv_out(sample)
+
+        batch_frames, channels, height, width = sample.shape
+        batch_size = batch_frames // num_frames
+        sample = sample[None, :].reshape(batch_size, num_frames, channels, height, width).permute(0, 2, 1, 3, 4)
+        sample = self.time_conv_out(sample)
+
+        sample = sample.permute(0, 2, 1, 3, 4).reshape(batch_frames, channels, height, width)
+
+        return sample
+
+
+class AutoencoderKLTemporalDecoder(ModelMixin, ConfigMixin, FromOriginalVAEMixin):
+    r"""
+    A VAE model with KL loss for encoding images into latents and decoding latent representations into images.
+
+    This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented
+    for all models (such as downloading or saving).
+
+    Parameters:
+        in_channels (int, *optional*, defaults to 3): Number of channels in the input image.
+        out_channels (int,  *optional*, defaults to 3): Number of channels in the output.
+        down_block_types (`Tuple[str]`, *optional*, defaults to `("DownEncoderBlock2D",)`):
+            Tuple of downsample block types.
+        block_out_channels (`Tuple[int]`, *optional*, defaults to `(64,)`):
+            Tuple of block output channels.
+        layers_per_block: (`int`, *optional*, defaults to 1): Number of layers per block.
+        latent_channels (`int`, *optional*, defaults to 4): Number of channels in the latent space.
+        sample_size (`int`, *optional*, defaults to `32`): Sample input size.
+        scaling_factor (`float`, *optional*, defaults to 0.18215):
+            The component-wise standard deviation of the trained latent space computed using the first batch of the
+            training set. This is used to scale the latent space to have unit variance when training the diffusion
+            model. The latents are scaled with the formula `z = z * scaling_factor` before being passed to the
+            diffusion model. When decoding, the latents are scaled back to the original scale with the formula: `z = 1
+            / scaling_factor * z`. For more details, refer to sections 4.3.2 and D.1 of the [High-Resolution Image
+            Synthesis with Latent Diffusion Models](https://arxiv.org/abs/2112.10752) paper.
+        force_upcast (`bool`, *optional*, default to `True`):
+            If enabled it will force the VAE to run in float32 for high image resolution pipelines, such as SD-XL. VAE
+            can be fine-tuned / trained to a lower range without loosing too much precision in which case
+            `force_upcast` can be set to `False` - see: https://huggingface.co/madebyollin/sdxl-vae-fp16-fix
+    """
+
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        in_channels: int = 3,
+        out_channels: int = 3,
+        down_block_types: Tuple[str] = ("DownEncoderBlock2D",),
+        block_out_channels: Tuple[int] = (64,),
+        layers_per_block: int = 1,
+        latent_channels: int = 4,
+        sample_size: int = 32,
+        scaling_factor: float = 0.18215,
+        force_upcast: float = True,
+    ):
+        super().__init__()
+
+        # pass init params to Encoder
+        self.encoder = Encoder(
+            in_channels=in_channels,
+            out_channels=latent_channels,
+            down_block_types=down_block_types,
+            block_out_channels=block_out_channels,
+            layers_per_block=layers_per_block,
+            double_z=True,
+        )
+
+        # pass init params to Decoder
+        self.decoder = TemporalDecoder(
+            in_channels=latent_channels,
+            out_channels=out_channels,
+            block_out_channels=block_out_channels,
+            layers_per_block=layers_per_block,
+        )
+
+        self.quant_conv = nn.Conv2d(2 * latent_channels, 2 * latent_channels, 1)
+
+        sample_size = (
+            self.config.sample_size[0]
+            if isinstance(self.config.sample_size, (list, tuple))
+            else self.config.sample_size
+        )
+        self.tile_latent_min_size = int(sample_size / (2 ** (len(self.config.block_out_channels) - 1)))
+        self.tile_overlap_factor = 0.25
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if isinstance(module, (Encoder, TemporalDecoder)):
+            module.gradient_checkpointing = value
+
+    @property
+    # Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.attn_processors
+    def attn_processors(self) -> Dict[str, AttentionProcessor]:
+        r"""
+        Returns:
+            `dict` of attention processors: A dictionary containing all attention processors used in the model with
+            indexed by its weight name.
+        """
+        # set recursively
+        processors = {}
+
+        def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
+            if hasattr(module, "get_processor"):
+                processors[f"{name}.processor"] = module.get_processor(return_deprecated_lora=True)
+
+            for sub_name, child in module.named_children():
+                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
+
+            return processors
+
+        for name, module in self.named_children():
+            fn_recursive_add_processors(name, module, processors)
+
+        return processors
+
+    # Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.set_attn_processor
+    def set_attn_processor(
+        self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]], _remove_lora=False
+    ):
+        r"""
+        Sets the attention processor to use to compute attention.
+
+        Parameters:
+            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
+                The instantiated processor class or a dictionary of processor classes that will be set as the processor
+                for **all** `Attention` layers.
+
+                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
+                processor. This is strongly recommended when setting trainable attention processors.
+
+        """
+        count = len(self.attn_processors.keys())
+
+        if isinstance(processor, dict) and len(processor) != count:
+            raise ValueError(
+                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
+                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
+            )
+
+        def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
+            if hasattr(module, "set_processor"):
+                if not isinstance(processor, dict):
+                    module.set_processor(processor, _remove_lora=_remove_lora)
+                else:
+                    module.set_processor(processor.pop(f"{name}.processor"), _remove_lora=_remove_lora)
+
+            for sub_name, child in module.named_children():
+                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
+
+        for name, module in self.named_children():
+            fn_recursive_attn_processor(name, module, processor)
+
+    def set_default_attn_processor(self):
+        """
+        Disables custom attention processors and sets the default attention implementation.
+        """
+        if all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+            processor = AttnProcessor()
+        else:
+            raise ValueError(
+                f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
+            )
+
+        self.set_attn_processor(processor, _remove_lora=True)
+
+    @apply_forward_hook
+    def encode(
+        self, x: torch.FloatTensor, return_dict: bool = True
+    ) -> Union[AutoencoderKLOutput, Tuple[DiagonalGaussianDistribution]]:
+        """
+        Encode a batch of images into latents.
+
+        Args:
+            x (`torch.FloatTensor`): Input batch of images.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether to return a [`~models.autoencoder_kl.AutoencoderKLOutput`] instead of a plain tuple.
+
+        Returns:
+                The latent representations of the encoded images. If `return_dict` is True, a
+                [`~models.autoencoder_kl.AutoencoderKLOutput`] is returned, otherwise a plain `tuple` is returned.
+        """
+        h = self.encoder(x)
+        moments = self.quant_conv(h)
+        posterior = DiagonalGaussianDistribution(moments)
+
+        if not return_dict:
+            return (posterior,)
+
+        return AutoencoderKLOutput(latent_dist=posterior)
+
+    @apply_forward_hook
+    def decode(
+        self,
+        z: torch.FloatTensor,
+        num_frames: int,
+        return_dict: bool = True,
+    ) -> Union[DecoderOutput, torch.FloatTensor]:
+        """
+        Decode a batch of images.
+
+        Args:
+            z (`torch.FloatTensor`): Input batch of latent vectors.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether to return a [`~models.vae.DecoderOutput`] instead of a plain tuple.
+
+        Returns:
+            [`~models.vae.DecoderOutput`] or `tuple`:
+                If return_dict is True, a [`~models.vae.DecoderOutput`] is returned, otherwise a plain `tuple` is
+                returned.
+
+        """
+        batch_size = z.shape[0] // num_frames
+        image_only_indicator = torch.zeros(batch_size, num_frames, dtype=z.dtype, device=z.device)
+        decoded = self.decoder(z, num_frames=num_frames, image_only_indicator=image_only_indicator)
+
+        if not return_dict:
+            return (decoded,)
+
+        return DecoderOutput(sample=decoded)
+
+    def forward(
+        self,
+        sample: torch.FloatTensor,
+        sample_posterior: bool = False,
+        return_dict: bool = True,
+        generator: Optional[torch.Generator] = None,
+        num_frames: int = 1,
+    ) -> Union[DecoderOutput, torch.FloatTensor]:
+        r"""
+        Args:
+            sample (`torch.FloatTensor`): Input sample.
+            sample_posterior (`bool`, *optional*, defaults to `False`):
+                Whether to sample from the posterior.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`DecoderOutput`] instead of a plain tuple.
+        """
+        x = sample
+        posterior = self.encode(x).latent_dist
+        if sample_posterior:
+            z = posterior.sample(generator=generator)
+        else:
+            z = posterior.mode()
+
+        dec = self.decode(z, num_frames=num_frames).sample
+
+        if not return_dict:
+            return (dec,)
+
+        return DecoderOutput(sample=dec)

src/diffusers/models/autoencoders/autoencoder_tiny.py

@@ -0,0 +1,345 @@
+# Copyright 2023 Ollin Boer Bohan and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from dataclasses import dataclass
+from typing import Optional, Tuple, Union
+
+import torch
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...utils import BaseOutput
+from ...utils.accelerate_utils import apply_forward_hook
+from ..modeling_utils import ModelMixin
+from .vae import DecoderOutput, DecoderTiny, EncoderTiny
+
+
+@dataclass
+class AutoencoderTinyOutput(BaseOutput):
+    """
+    Output of AutoencoderTiny encoding method.
+
+    Args:
+        latents (`torch.Tensor`): Encoded outputs of the `Encoder`.
+
+    """
+
+    latents: torch.Tensor
+
+
+class AutoencoderTiny(ModelMixin, ConfigMixin):
+    r"""
+    A tiny distilled VAE model for encoding images into latents and decoding latent representations into images.
+
+    [`AutoencoderTiny`] is a wrapper around the original implementation of `TAESD`.
+
+    This model inherits from [`ModelMixin`]. Check the superclass documentation for its generic methods implemented for
+    all models (such as downloading or saving).
+
+    Parameters:
+        in_channels (`int`, *optional*, defaults to 3): Number of channels in the input image.
+        out_channels (`int`,  *optional*, defaults to 3): Number of channels in the output.
+        encoder_block_out_channels (`Tuple[int]`, *optional*, defaults to `(64, 64, 64, 64)`):
+            Tuple of integers representing the number of output channels for each encoder block. The length of the
+            tuple should be equal to the number of encoder blocks.
+        decoder_block_out_channels (`Tuple[int]`, *optional*, defaults to `(64, 64, 64, 64)`):
+            Tuple of integers representing the number of output channels for each decoder block. The length of the
+            tuple should be equal to the number of decoder blocks.
+        act_fn (`str`, *optional*, defaults to `"relu"`):
+            Activation function to be used throughout the model.
+        latent_channels (`int`, *optional*, defaults to 4):
+            Number of channels in the latent representation. The latent space acts as a compressed representation of
+            the input image.
+        upsampling_scaling_factor (`int`, *optional*, defaults to 2):
+            Scaling factor for upsampling in the decoder. It determines the size of the output image during the
+            upsampling process.
+        num_encoder_blocks (`Tuple[int]`, *optional*, defaults to `(1, 3, 3, 3)`):
+            Tuple of integers representing the number of encoder blocks at each stage of the encoding process. The
+            length of the tuple should be equal to the number of stages in the encoder. Each stage has a different
+            number of encoder blocks.
+        num_decoder_blocks (`Tuple[int]`, *optional*, defaults to `(3, 3, 3, 1)`):
+            Tuple of integers representing the number of decoder blocks at each stage of the decoding process. The
+            length of the tuple should be equal to the number of stages in the decoder. Each stage has a different
+            number of decoder blocks.
+        latent_magnitude (`float`, *optional*, defaults to 3.0):
+            Magnitude of the latent representation. This parameter scales the latent representation values to control
+            the extent of information preservation.
+        latent_shift (float, *optional*, defaults to 0.5):
+            Shift applied to the latent representation. This parameter controls the center of the latent space.
+        scaling_factor (`float`, *optional*, defaults to 1.0):
+            The component-wise standard deviation of the trained latent space computed using the first batch of the
+            training set. This is used to scale the latent space to have unit variance when training the diffusion
+            model. The latents are scaled with the formula `z = z * scaling_factor` before being passed to the
+            diffusion model. When decoding, the latents are scaled back to the original scale with the formula: `z = 1
+            / scaling_factor * z`. For more details, refer to sections 4.3.2 and D.1 of the [High-Resolution Image
+            Synthesis with Latent Diffusion Models](https://arxiv.org/abs/2112.10752) paper. For this Autoencoder,
+            however, no such scaling factor was used, hence the value of 1.0 as the default.
+        force_upcast (`bool`, *optional*, default to `False`):
+            If enabled it will force the VAE to run in float32 for high image resolution pipelines, such as SD-XL. VAE
+            can be fine-tuned / trained to a lower range without losing too much precision, in which case
+            `force_upcast` can be set to `False` (see this fp16-friendly
+            [AutoEncoder](https://huggingface.co/madebyollin/sdxl-vae-fp16-fix)).
+    """
+
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        in_channels: int = 3,
+        out_channels: int = 3,
+        encoder_block_out_channels: Tuple[int, ...] = (64, 64, 64, 64),
+        decoder_block_out_channels: Tuple[int, ...] = (64, 64, 64, 64),
+        act_fn: str = "relu",
+        latent_channels: int = 4,
+        upsampling_scaling_factor: int = 2,
+        num_encoder_blocks: Tuple[int, ...] = (1, 3, 3, 3),
+        num_decoder_blocks: Tuple[int, ...] = (3, 3, 3, 1),
+        latent_magnitude: int = 3,
+        latent_shift: float = 0.5,
+        force_upcast: bool = False,
+        scaling_factor: float = 1.0,
+    ):
+        super().__init__()
+
+        if len(encoder_block_out_channels) != len(num_encoder_blocks):
+            raise ValueError("`encoder_block_out_channels` should have the same length as `num_encoder_blocks`.")
+        if len(decoder_block_out_channels) != len(num_decoder_blocks):
+            raise ValueError("`decoder_block_out_channels` should have the same length as `num_decoder_blocks`.")
+
+        self.encoder = EncoderTiny(
+            in_channels=in_channels,
+            out_channels=latent_channels,
+            num_blocks=num_encoder_blocks,
+            block_out_channels=encoder_block_out_channels,
+            act_fn=act_fn,
+        )
+
+        self.decoder = DecoderTiny(
+            in_channels=latent_channels,
+            out_channels=out_channels,
+            num_blocks=num_decoder_blocks,
+            block_out_channels=decoder_block_out_channels,
+            upsampling_scaling_factor=upsampling_scaling_factor,
+            act_fn=act_fn,
+        )
+
+        self.latent_magnitude = latent_magnitude
+        self.latent_shift = latent_shift
+        self.scaling_factor = scaling_factor
+
+        self.use_slicing = False
+        self.use_tiling = False
+
+        # only relevant if vae tiling is enabled
+        self.spatial_scale_factor = 2**out_channels
+        self.tile_overlap_factor = 0.125
+        self.tile_sample_min_size = 512
+        self.tile_latent_min_size = self.tile_sample_min_size // self.spatial_scale_factor
+
+        self.register_to_config(block_out_channels=decoder_block_out_channels)
+        self.register_to_config(force_upcast=False)
+
+    def _set_gradient_checkpointing(self, module, value: bool = False) -> None:
+        if isinstance(module, (EncoderTiny, DecoderTiny)):
+            module.gradient_checkpointing = value
+
+    def scale_latents(self, x: torch.FloatTensor) -> torch.FloatTensor:
+        """raw latents -> [0, 1]"""
+        return x.div(2 * self.latent_magnitude).add(self.latent_shift).clamp(0, 1)
+
+    def unscale_latents(self, x: torch.FloatTensor) -> torch.FloatTensor:
+        """[0, 1] -> raw latents"""
+        return x.sub(self.latent_shift).mul(2 * self.latent_magnitude)
+
+    def enable_slicing(self) -> None:
+        r"""
+        Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
+        compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
+        """
+        self.use_slicing = True
+
+    def disable_slicing(self) -> None:
+        r"""
+        Disable sliced VAE decoding. If `enable_slicing` was previously enabled, this method will go back to computing
+        decoding in one step.
+        """
+        self.use_slicing = False
+
+    def enable_tiling(self, use_tiling: bool = True) -> None:
+        r"""
+        Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
+        compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
+        processing larger images.
+        """
+        self.use_tiling = use_tiling
+
+    def disable_tiling(self) -> None:
+        r"""
+        Disable tiled VAE decoding. If `enable_tiling` was previously enabled, this method will go back to computing
+        decoding in one step.
+        """
+        self.enable_tiling(False)
+
+    def _tiled_encode(self, x: torch.FloatTensor) -> torch.FloatTensor:
+        r"""Encode a batch of images using a tiled encoder.
+
+        When this option is enabled, the VAE will split the input tensor into tiles to compute encoding in several
+        steps. This is useful to keep memory use constant regardless of image size. To avoid tiling artifacts, the
+        tiles overlap and are blended together to form a smooth output.
+
+        Args:
+            x (`torch.FloatTensor`): Input batch of images.
+
+        Returns:
+            `torch.FloatTensor`: Encoded batch of images.
+        """
+        # scale of encoder output relative to input
+        sf = self.spatial_scale_factor
+        tile_size = self.tile_sample_min_size
+
+        # number of pixels to blend and to traverse between tile
+        blend_size = int(tile_size * self.tile_overlap_factor)
+        traverse_size = tile_size - blend_size
+
+        # tiles index (up/left)
+        ti = range(0, x.shape[-2], traverse_size)
+        tj = range(0, x.shape[-1], traverse_size)
+
+        # mask for blending
+        blend_masks = torch.stack(
+            torch.meshgrid([torch.arange(tile_size / sf) / (blend_size / sf - 1)] * 2, indexing="ij")
+        )
+        blend_masks = blend_masks.clamp(0, 1).to(x.device)
+
+        # output array
+        out = torch.zeros(x.shape[0], 4, x.shape[-2] // sf, x.shape[-1] // sf, device=x.device)
+        for i in ti:
+            for j in tj:
+                tile_in = x[..., i : i + tile_size, j : j + tile_size]
+                # tile result
+                tile_out = out[..., i // sf : (i + tile_size) // sf, j // sf : (j + tile_size) // sf]
+                tile = self.encoder(tile_in)
+                h, w = tile.shape[-2], tile.shape[-1]
+                # blend tile result into output
+                blend_mask_i = torch.ones_like(blend_masks[0]) if i == 0 else blend_masks[0]
+                blend_mask_j = torch.ones_like(blend_masks[1]) if j == 0 else blend_masks[1]
+                blend_mask = blend_mask_i * blend_mask_j
+                tile, blend_mask = tile[..., :h, :w], blend_mask[..., :h, :w]
+                tile_out.copy_(blend_mask * tile + (1 - blend_mask) * tile_out)
+        return out
+
+    def _tiled_decode(self, x: torch.FloatTensor) -> torch.FloatTensor:
+        r"""Encode a batch of images using a tiled encoder.
+
+        When this option is enabled, the VAE will split the input tensor into tiles to compute encoding in several
+        steps. This is useful to keep memory use constant regardless of image size. To avoid tiling artifacts, the
+        tiles overlap and are blended together to form a smooth output.
+
+        Args:
+            x (`torch.FloatTensor`): Input batch of images.
+
+        Returns:
+            `torch.FloatTensor`: Encoded batch of images.
+        """
+        # scale of decoder output relative to input
+        sf = self.spatial_scale_factor
+        tile_size = self.tile_latent_min_size
+
+        # number of pixels to blend and to traverse between tiles
+        blend_size = int(tile_size * self.tile_overlap_factor)
+        traverse_size = tile_size - blend_size
+
+        # tiles index (up/left)
+        ti = range(0, x.shape[-2], traverse_size)
+        tj = range(0, x.shape[-1], traverse_size)
+
+        # mask for blending
+        blend_masks = torch.stack(
+            torch.meshgrid([torch.arange(tile_size * sf) / (blend_size * sf - 1)] * 2, indexing="ij")
+        )
+        blend_masks = blend_masks.clamp(0, 1).to(x.device)
+
+        # output array
+        out = torch.zeros(x.shape[0], 3, x.shape[-2] * sf, x.shape[-1] * sf, device=x.device)
+        for i in ti:
+            for j in tj:
+                tile_in = x[..., i : i + tile_size, j : j + tile_size]
+                # tile result
+                tile_out = out[..., i * sf : (i + tile_size) * sf, j * sf : (j + tile_size) * sf]
+                tile = self.decoder(tile_in)
+                h, w = tile.shape[-2], tile.shape[-1]
+                # blend tile result into output
+                blend_mask_i = torch.ones_like(blend_masks[0]) if i == 0 else blend_masks[0]
+                blend_mask_j = torch.ones_like(blend_masks[1]) if j == 0 else blend_masks[1]
+                blend_mask = (blend_mask_i * blend_mask_j)[..., :h, :w]
+                tile_out.copy_(blend_mask * tile + (1 - blend_mask) * tile_out)
+        return out
+
+    @apply_forward_hook
+    def encode(
+        self, x: torch.FloatTensor, return_dict: bool = True
+    ) -> Union[AutoencoderTinyOutput, Tuple[torch.FloatTensor]]:
+        if self.use_slicing and x.shape[0] > 1:
+            output = [self._tiled_encode(x_slice) if self.use_tiling else self.encoder(x) for x_slice in x.split(1)]
+            output = torch.cat(output)
+        else:
+            output = self._tiled_encode(x) if self.use_tiling else self.encoder(x)
+
+        if not return_dict:
+            return (output,)
+
+        return AutoencoderTinyOutput(latents=output)
+
+    @apply_forward_hook
+    def decode(
+        self, x: torch.FloatTensor, generator: Optional[torch.Generator] = None, return_dict: bool = True
+    ) -> Union[DecoderOutput, Tuple[torch.FloatTensor]]:
+        if self.use_slicing and x.shape[0] > 1:
+            output = [self._tiled_decode(x_slice) if self.use_tiling else self.decoder(x) for x_slice in x.split(1)]
+            output = torch.cat(output)
+        else:
+            output = self._tiled_decode(x) if self.use_tiling else self.decoder(x)
+
+        if not return_dict:
+            return (output,)
+
+        return DecoderOutput(sample=output)
+
+    def forward(
+        self,
+        sample: torch.FloatTensor,
+        return_dict: bool = True,
+    ) -> Union[DecoderOutput, Tuple[torch.FloatTensor]]:
+        r"""
+        Args:
+            sample (`torch.FloatTensor`): Input sample.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`DecoderOutput`] instead of a plain tuple.
+        """
+        enc = self.encode(sample).latents
+
+        # scale latents to be in [0, 1], then quantize latents to a byte tensor,
+        # as if we were storing the latents in an RGBA uint8 image.
+        scaled_enc = self.scale_latents(enc).mul_(255).round_().byte()
+
+        # unquantize latents back into [0, 1], then unscale latents back to their original range,
+        # as if we were loading the latents from an RGBA uint8 image.
+        unscaled_enc = self.unscale_latents(scaled_enc / 255.0)
+
+        dec = self.decode(unscaled_enc)
+
+        if not return_dict:
+            return (dec,)
+        return DecoderOutput(sample=dec)

src/diffusers/models/autoencoders/consistency_decoder_vae.py

@@ -0,0 +1,437 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from dataclasses import dataclass
+from typing import Dict, Optional, Tuple, Union
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...schedulers import ConsistencyDecoderScheduler
+from ...utils import BaseOutput
+from ...utils.accelerate_utils import apply_forward_hook
+from ...utils.torch_utils import randn_tensor
+from ..attention_processor import (
+    ADDED_KV_ATTENTION_PROCESSORS,
+    CROSS_ATTENTION_PROCESSORS,
+    AttentionProcessor,
+    AttnAddedKVProcessor,
+    AttnProcessor,
+)
+from ..modeling_utils import ModelMixin
+from ..unet_2d import UNet2DModel
+from .vae import DecoderOutput, DiagonalGaussianDistribution, Encoder
+
+
+@dataclass
+class ConsistencyDecoderVAEOutput(BaseOutput):
+    """
+    Output of encoding method.
+
+    Args:
+        latent_dist (`DiagonalGaussianDistribution`):
+            Encoded outputs of `Encoder` represented as the mean and logvar of `DiagonalGaussianDistribution`.
+            `DiagonalGaussianDistribution` allows for sampling latents from the distribution.
+    """
+
+    latent_dist: "DiagonalGaussianDistribution"
+
+
+class ConsistencyDecoderVAE(ModelMixin, ConfigMixin):
+    r"""
+    The consistency decoder used with DALL-E 3.
+
+    Examples:
+        ```py
+        >>> import torch
+        >>> from diffusers import StableDiffusionPipeline, ConsistencyDecoderVAE
+
+        >>> vae = ConsistencyDecoderVAE.from_pretrained("openai/consistency-decoder", torch_dtype=torch.float16)
+        >>> pipe = StableDiffusionPipeline.from_pretrained(
+        ...     "runwayml/stable-diffusion-v1-5", vae=vae, torch_dtype=torch.float16
+        ... ).to("cuda")
+
+        >>> pipe("horse", generator=torch.manual_seed(0)).images
+        ```
+    """
+
+    @register_to_config
+    def __init__(
+        self,
+        scaling_factor: float = 0.18215,
+        latent_channels: int = 4,
+        encoder_act_fn: str = "silu",
+        encoder_block_out_channels: Tuple[int, ...] = (128, 256, 512, 512),
+        encoder_double_z: bool = True,
+        encoder_down_block_types: Tuple[str, ...] = (
+            "DownEncoderBlock2D",
+            "DownEncoderBlock2D",
+            "DownEncoderBlock2D",
+            "DownEncoderBlock2D",
+        ),
+        encoder_in_channels: int = 3,
+        encoder_layers_per_block: int = 2,
+        encoder_norm_num_groups: int = 32,
+        encoder_out_channels: int = 4,
+        decoder_add_attention: bool = False,
+        decoder_block_out_channels: Tuple[int, ...] = (320, 640, 1024, 1024),
+        decoder_down_block_types: Tuple[str, ...] = (
+            "ResnetDownsampleBlock2D",
+            "ResnetDownsampleBlock2D",
+            "ResnetDownsampleBlock2D",
+            "ResnetDownsampleBlock2D",
+        ),
+        decoder_downsample_padding: int = 1,
+        decoder_in_channels: int = 7,
+        decoder_layers_per_block: int = 3,
+        decoder_norm_eps: float = 1e-05,
+        decoder_norm_num_groups: int = 32,
+        decoder_num_train_timesteps: int = 1024,
+        decoder_out_channels: int = 6,
+        decoder_resnet_time_scale_shift: str = "scale_shift",
+        decoder_time_embedding_type: str = "learned",
+        decoder_up_block_types: Tuple[str, ...] = (
+            "ResnetUpsampleBlock2D",
+            "ResnetUpsampleBlock2D",
+            "ResnetUpsampleBlock2D",
+            "ResnetUpsampleBlock2D",
+        ),
+    ):
+        super().__init__()
+        self.encoder = Encoder(
+            act_fn=encoder_act_fn,
+            block_out_channels=encoder_block_out_channels,
+            double_z=encoder_double_z,
+            down_block_types=encoder_down_block_types,
+            in_channels=encoder_in_channels,
+            layers_per_block=encoder_layers_per_block,
+            norm_num_groups=encoder_norm_num_groups,
+            out_channels=encoder_out_channels,
+        )
+
+        self.decoder_unet = UNet2DModel(
+            add_attention=decoder_add_attention,
+            block_out_channels=decoder_block_out_channels,
+            down_block_types=decoder_down_block_types,
+            downsample_padding=decoder_downsample_padding,
+            in_channels=decoder_in_channels,
+            layers_per_block=decoder_layers_per_block,
+            norm_eps=decoder_norm_eps,
+            norm_num_groups=decoder_norm_num_groups,
+            num_train_timesteps=decoder_num_train_timesteps,
+            out_channels=decoder_out_channels,
+            resnet_time_scale_shift=decoder_resnet_time_scale_shift,
+            time_embedding_type=decoder_time_embedding_type,
+            up_block_types=decoder_up_block_types,
+        )
+        self.decoder_scheduler = ConsistencyDecoderScheduler()
+        self.register_to_config(block_out_channels=encoder_block_out_channels)
+        self.register_to_config(force_upcast=False)
+        self.register_buffer(
+            "means",
+            torch.tensor([0.38862467, 0.02253063, 0.07381133, -0.0171294])[None, :, None, None],
+            persistent=False,
+        )
+        self.register_buffer(
+            "stds", torch.tensor([0.9654121, 1.0440036, 0.76147926, 0.77022034])[None, :, None, None], persistent=False
+        )
+
+        self.quant_conv = nn.Conv2d(2 * latent_channels, 2 * latent_channels, 1)
+
+        self.use_slicing = False
+        self.use_tiling = False
+
+    # Copied from diffusers.models.autoencoders.autoencoder_kl.AutoencoderKL.enable_tiling
+    def enable_tiling(self, use_tiling: bool = True):
+        r"""
+        Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
+        compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
+        processing larger images.
+        """
+        self.use_tiling = use_tiling
+
+    # Copied from diffusers.models.autoencoders.autoencoder_kl.AutoencoderKL.disable_tiling
+    def disable_tiling(self):
+        r"""
+        Disable tiled VAE decoding. If `enable_tiling` was previously enabled, this method will go back to computing
+        decoding in one step.
+        """
+        self.enable_tiling(False)
+
+    # Copied from diffusers.models.autoencoders.autoencoder_kl.AutoencoderKL.enable_slicing
+    def enable_slicing(self):
+        r"""
+        Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
+        compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
+        """
+        self.use_slicing = True
+
+    # Copied from diffusers.models.autoencoders.autoencoder_kl.AutoencoderKL.disable_slicing
+    def disable_slicing(self):
+        r"""
+        Disable sliced VAE decoding. If `enable_slicing` was previously enabled, this method will go back to computing
+        decoding in one step.
+        """
+        self.use_slicing = False
+
+    @property
+    # Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.attn_processors
+    def attn_processors(self) -> Dict[str, AttentionProcessor]:
+        r"""
+        Returns:
+            `dict` of attention processors: A dictionary containing all attention processors used in the model with
+            indexed by its weight name.
+        """
+        # set recursively
+        processors = {}
+
+        def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
+            if hasattr(module, "get_processor"):
+                processors[f"{name}.processor"] = module.get_processor(return_deprecated_lora=True)
+
+            for sub_name, child in module.named_children():
+                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
+
+            return processors
+
+        for name, module in self.named_children():
+            fn_recursive_add_processors(name, module, processors)
+
+        return processors
+
+    # Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.set_attn_processor
+    def set_attn_processor(
+        self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]], _remove_lora=False
+    ):
+        r"""
+        Sets the attention processor to use to compute attention.
+
+        Parameters:
+            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
+                The instantiated processor class or a dictionary of processor classes that will be set as the processor
+                for **all** `Attention` layers.
+
+                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
+                processor. This is strongly recommended when setting trainable attention processors.
+
+        """
+        count = len(self.attn_processors.keys())
+
+        if isinstance(processor, dict) and len(processor) != count:
+            raise ValueError(
+                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
+                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
+            )
+
+        def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
+            if hasattr(module, "set_processor"):
+                if not isinstance(processor, dict):
+                    module.set_processor(processor, _remove_lora=_remove_lora)
+                else:
+                    module.set_processor(processor.pop(f"{name}.processor"), _remove_lora=_remove_lora)
+
+            for sub_name, child in module.named_children():
+                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
+
+        for name, module in self.named_children():
+            fn_recursive_attn_processor(name, module, processor)
+
+    # Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.set_default_attn_processor
+    def set_default_attn_processor(self):
+        """
+        Disables custom attention processors and sets the default attention implementation.
+        """
+        if all(proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+            processor = AttnAddedKVProcessor()
+        elif all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+            processor = AttnProcessor()
+        else:
+            raise ValueError(
+                f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
+            )
+
+        self.set_attn_processor(processor, _remove_lora=True)
+
+    @apply_forward_hook
+    def encode(
+        self, x: torch.FloatTensor, return_dict: bool = True
+    ) -> Union[ConsistencyDecoderVAEOutput, Tuple[DiagonalGaussianDistribution]]:
+        """
+        Encode a batch of images into latents.
+
+        Args:
+            x (`torch.FloatTensor`): Input batch of images.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether to return a [`~models.consistecy_decoder_vae.ConsistencyDecoderOoutput`] instead of a plain
+                tuple.
+
+        Returns:
+                The latent representations of the encoded images. If `return_dict` is True, a
+                [`~models.consistency_decoder_vae.ConsistencyDecoderVAEOutput`] is returned, otherwise a plain `tuple`
+                is returned.
+        """
+        if self.use_tiling and (x.shape[-1] > self.tile_sample_min_size or x.shape[-2] > self.tile_sample_min_size):
+            return self.tiled_encode(x, return_dict=return_dict)
+
+        if self.use_slicing and x.shape[0] > 1:
+            encoded_slices = [self.encoder(x_slice) for x_slice in x.split(1)]
+            h = torch.cat(encoded_slices)
+        else:
+            h = self.encoder(x)
+
+        moments = self.quant_conv(h)
+        posterior = DiagonalGaussianDistribution(moments)
+
+        if not return_dict:
+            return (posterior,)
+
+        return ConsistencyDecoderVAEOutput(latent_dist=posterior)
+
+    @apply_forward_hook
+    def decode(
+        self,
+        z: torch.FloatTensor,
+        generator: Optional[torch.Generator] = None,
+        return_dict: bool = True,
+        num_inference_steps: int = 2,
+    ) -> Union[DecoderOutput, Tuple[torch.FloatTensor]]:
+        z = (z * self.config.scaling_factor - self.means) / self.stds
+
+        scale_factor = 2 ** (len(self.config.block_out_channels) - 1)
+        z = F.interpolate(z, mode="nearest", scale_factor=scale_factor)
+
+        batch_size, _, height, width = z.shape
+
+        self.decoder_scheduler.set_timesteps(num_inference_steps, device=self.device)
+
+        x_t = self.decoder_scheduler.init_noise_sigma * randn_tensor(
+            (batch_size, 3, height, width), generator=generator, dtype=z.dtype, device=z.device
+        )
+
+        for t in self.decoder_scheduler.timesteps:
+            model_input = torch.concat([self.decoder_scheduler.scale_model_input(x_t, t), z], dim=1)
+            model_output = self.decoder_unet(model_input, t).sample[:, :3, :, :]
+            prev_sample = self.decoder_scheduler.step(model_output, t, x_t, generator).prev_sample
+            x_t = prev_sample
+
+        x_0 = x_t
+
+        if not return_dict:
+            return (x_0,)
+
+        return DecoderOutput(sample=x_0)
+
+    # Copied from diffusers.models.autoencoders.autoencoder_kl.AutoencoderKL.blend_v
+    def blend_v(self, a: torch.Tensor, b: torch.Tensor, blend_extent: int) -> torch.Tensor:
+        blend_extent = min(a.shape[2], b.shape[2], blend_extent)
+        for y in range(blend_extent):
+            b[:, :, y, :] = a[:, :, -blend_extent + y, :] * (1 - y / blend_extent) + b[:, :, y, :] * (y / blend_extent)
+        return b
+
+    # Copied from diffusers.models.autoencoders.autoencoder_kl.AutoencoderKL.blend_h
+    def blend_h(self, a: torch.Tensor, b: torch.Tensor, blend_extent: int) -> torch.Tensor:
+        blend_extent = min(a.shape[3], b.shape[3], blend_extent)
+        for x in range(blend_extent):
+            b[:, :, :, x] = a[:, :, :, -blend_extent + x] * (1 - x / blend_extent) + b[:, :, :, x] * (x / blend_extent)
+        return b
+
+    def tiled_encode(self, x: torch.FloatTensor, return_dict: bool = True) -> ConsistencyDecoderVAEOutput:
+        r"""Encode a batch of images using a tiled encoder.
+
+        When this option is enabled, the VAE will split the input tensor into tiles to compute encoding in several
+        steps. This is useful to keep memory use constant regardless of image size. The end result of tiled encoding is
+        different from non-tiled encoding because each tile uses a different encoder. To avoid tiling artifacts, the
+        tiles overlap and are blended together to form a smooth output. You may still see tile-sized changes in the
+        output, but they should be much less noticeable.
+
+        Args:
+            x (`torch.FloatTensor`): Input batch of images.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~models.consistency_decoder_vae.ConsistencyDecoderVAEOutput`] instead of a
+                plain tuple.
+
+        Returns:
+            [`~models.consistency_decoder_vae.ConsistencyDecoderVAEOutput`] or `tuple`:
+                If return_dict is True, a [`~models.consistency_decoder_vae.ConsistencyDecoderVAEOutput`] is returned,
+                otherwise a plain `tuple` is returned.
+        """
+        overlap_size = int(self.tile_sample_min_size * (1 - self.tile_overlap_factor))
+        blend_extent = int(self.tile_latent_min_size * self.tile_overlap_factor)
+        row_limit = self.tile_latent_min_size - blend_extent
+
+        # Split the image into 512x512 tiles and encode them separately.
+        rows = []
+        for i in range(0, x.shape[2], overlap_size):
+            row = []
+            for j in range(0, x.shape[3], overlap_size):
+                tile = x[:, :, i : i + self.tile_sample_min_size, j : j + self.tile_sample_min_size]
+                tile = self.encoder(tile)
+                tile = self.quant_conv(tile)
+                row.append(tile)
+            rows.append(row)
+        result_rows = []
+        for i, row in enumerate(rows):
+            result_row = []
+            for j, tile in enumerate(row):
+                # blend the above tile and the left tile
+                # to the current tile and add the current tile to the result row
+                if i > 0:
+                    tile = self.blend_v(rows[i - 1][j], tile, blend_extent)
+                if j > 0:
+                    tile = self.blend_h(row[j - 1], tile, blend_extent)
+                result_row.append(tile[:, :, :row_limit, :row_limit])
+            result_rows.append(torch.cat(result_row, dim=3))
+
+        moments = torch.cat(result_rows, dim=2)
+        posterior = DiagonalGaussianDistribution(moments)
+
+        if not return_dict:
+            return (posterior,)
+
+        return ConsistencyDecoderVAEOutput(latent_dist=posterior)
+
+    def forward(
+        self,
+        sample: torch.FloatTensor,
+        sample_posterior: bool = False,
+        return_dict: bool = True,
+        generator: Optional[torch.Generator] = None,
+    ) -> Union[DecoderOutput, Tuple[torch.FloatTensor]]:
+        r"""
+        Args:
+            sample (`torch.FloatTensor`): Input sample.
+            sample_posterior (`bool`, *optional*, defaults to `False`):
+                Whether to sample from the posterior.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`DecoderOutput`] instead of a plain tuple.
+            generator (`torch.Generator`, *optional*, defaults to `None`):
+                Generator to use for sampling.
+
+        Returns:
+            [`DecoderOutput`] or `tuple`:
+                If return_dict is True, a [`DecoderOutput`] is returned, otherwise a plain `tuple` is returned.
+        """
+        x = sample
+        posterior = self.encode(x).latent_dist
+        if sample_posterior:
+            z = posterior.sample(generator=generator)
+        else:
+            z = posterior.mode()
+        dec = self.decode(z, generator=generator).sample
+
+        if not return_dict:
+            return (dec,)
+
+        return DecoderOutput(sample=dec)