Upload folder using huggingface_hub

.gitattributes

@@ -33,3 +33,10 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+demo.ipynb filter=lfs diff=lfs merge=lfs -text
+figures/gradio_demo.png filter=lfs diff=lfs merge=lfs -text
+figures/gradio_demo_controlnet.png filter=lfs diff=lfs merge=lfs -text
+figures/gradio_demo_controlnet_img2img.png filter=lfs diff=lfs merge=lfs -text
+figures/gradio_demo_img2img.png filter=lfs diff=lfs merge=lfs -text
+figures/progressive_process.jpg filter=lfs diff=lfs merge=lfs -text
+output_example.png filter=lfs diff=lfs merge=lfs -text

.idea/.gitignore

@@ -0,0 +1,3 @@
+# 默认忽略的文件
+/shelf/
+/workspace.xml

.idea/DemoFusion-main.iml

@@ -0,0 +1,12 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<module type="PYTHON_MODULE" version="4">
+  <component name="NewModuleRootManager">
+    <content url="file://$MODULE_DIR$" />
+    <orderEntry type="jdk" jdkName="Ai_mode (4)" jdkType="Python SDK" />
+    <orderEntry type="sourceFolder" forTests="false" />
+  </component>
+  <component name="PyDocumentationSettings">
+    <option name="format" value="GOOGLE" />
+    <option name="myDocStringFormat" value="Google" />
+  </component>
+</module>

.idea/inspectionProfiles/Project_Default.xml

@@ -0,0 +1,48 @@
+<component name="InspectionProjectProfileManager">
+  <profile version="1.0">
+    <option name="myName" value="Project Default" />
+    <inspection_tool class="PyPackageRequirementsInspection" enabled="true" level="WARNING" enabled_by_default="true">
+      <option name="ignoredPackages">
+        <value>
+          <list size="28">
+            <item index="0" class="java.lang.String" itemvalue="httpx" />
+            <item index="1" class="java.lang.String" itemvalue="gradio" />
+            <item index="2" class="java.lang.String" itemvalue="open-clip-torch" />
+            <item index="3" class="java.lang.String" itemvalue="PyYAML" />
+            <item index="4" class="java.lang.String" itemvalue="xformers" />
+            <item index="5" class="java.lang.String" itemvalue="numpy" />
+            <item index="6" class="java.lang.String" itemvalue="requests" />
+            <item index="7" class="java.lang.String" itemvalue="fsspec" />
+            <item index="8" class="java.lang.String" itemvalue="kornia" />
+            <item index="9" class="java.lang.String" itemvalue="gradio_client" />
+            <item index="10" class="java.lang.String" itemvalue="openai-clip" />
+            <item index="11" class="java.lang.String" itemvalue="sentencepiece" />
+            <item index="12" class="java.lang.String" itemvalue="wandb" />
+            <item index="13" class="java.lang.String" itemvalue="accelerate" />
+            <item index="14" class="java.lang.String" itemvalue="uvicorn" />
+            <item index="15" class="java.lang.String" itemvalue="urllib3" />
+            <item index="16" class="java.lang.String" itemvalue="triton" />
+            <item index="17" class="java.lang.String" itemvalue="timm" />
+            <item index="18" class="java.lang.String" itemvalue="opencv-python" />
+            <item index="19" class="java.lang.String" itemvalue="pandas" />
+            <item index="20" class="java.lang.String" itemvalue="tqdm" />
+            <item index="21" class="java.lang.String" itemvalue="pytorch-lightning" />
+            <item index="22" class="java.lang.String" itemvalue="fastapi" />
+            <item index="23" class="java.lang.String" itemvalue="einops-exts" />
+            <item index="24" class="java.lang.String" itemvalue="ninja" />
+            <item index="25" class="java.lang.String" itemvalue="matplotlib" />
+            <item index="26" class="java.lang.String" itemvalue="webdataset" />
+            <item index="27" class="java.lang.String" itemvalue="Pillow" />
+          </list>
+        </value>
+      </option>
+    </inspection_tool>
+    <inspection_tool class="PyUnresolvedReferencesInspection" enabled="true" level="WARNING" enabled_by_default="true">
+      <option name="ignoredIdentifiers">
+        <list>
+          <option value="JSX2511" />
+        </list>
+      </option>
+    </inspection_tool>
+  </profile>
+</component>

.idea/inspectionProfiles/profiles_settings.xml

@@ -0,0 +1,6 @@
+<component name="InspectionProjectProfileManager">
+  <settings>
+    <option name="USE_PROJECT_PROFILE" value="false" />
+    <version value="1.0" />
+  </settings>
+</component>

.idea/misc.xml

@@ -0,0 +1,4 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<project version="4">
+  <component name="ProjectRootManager" version="2" project-jdk-name="Ai_mode (4)" project-jdk-type="Python SDK" />
+</project>

.idea/modules.xml

@@ -0,0 +1,8 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<project version="4">
+  <component name="ProjectModuleManager">
+    <modules>
+      <module fileurl="file://$PROJECT_DIR$/.idea/DemoFusion-main.iml" filepath="$PROJECT_DIR$/.idea/DemoFusion-main.iml" />
+    </modules>
+  </component>
+</project>

.idea/workspace.xml

@@ -0,0 +1,71 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<project version="4">
+  <component name="AutoImportSettings">
+    <option name="autoReloadType" value="SELECTIVE" />
+  </component>
+  <component name="ChangeListManager">
+    <list default="true" id="cf13c2e0-fea8-4b36-9b87-d1a7a68e5205" name="更改" comment="" />
+    <option name="SHOW_DIALOG" value="false" />
+    <option name="HIGHLIGHT_CONFLICTS" value="true" />
+    <option name="HIGHLIGHT_NON_ACTIVE_CHANGELIST" value="false" />
+    <option name="LAST_RESOLUTION" value="IGNORE" />
+  </component>
+  <component name="MarkdownSettingsMigration">
+    <option name="stateVersion" value="1" />
+  </component>
+  <component name="ProjectColorInfo">{
+  &quot;associatedIndex&quot;: 2
+}</component>
+  <component name="ProjectId" id="2btF3w7MTy2zdOhszEbQETG3Qqf" />
+  <component name="ProjectViewState">
+    <option name="hideEmptyMiddlePackages" value="true" />
+    <option name="showLibraryContents" value="true" />
+  </component>
+  <component name="PropertiesComponent">{
+  &quot;keyToString&quot;: {
+    &quot;RunOnceActivity.OpenProjectViewOnStart&quot;: &quot;true&quot;,
+    &quot;RunOnceActivity.ShowReadmeOnStart&quot;: &quot;true&quot;,
+    &quot;last_opened_file_path&quot;: &quot;E:/AI/reduce_noise/DemoFusion-main&quot;,
+    &quot;settings.editor.selected.configurable&quot;: &quot;com.jetbrains.python.configuration.PyActiveSdkModuleConfigurable&quot;
+  }
+}</component>
+  <component name="RunManager">
+    <configuration name="gradio_demo_img2img" type="PythonConfigurationType" factoryName="Python" temporary="true" nameIsGenerated="true">
+      <module name="DemoFusion-main" />
+      <option name="INTERPRETER_OPTIONS" value="" />
+      <option name="PARENT_ENVS" value="true" />
+      <envs>
+        <env name="PYTHONUNBUFFERED" value="1" />
+      </envs>
+      <option name="SDK_HOME" value="" />
+      <option name="WORKING_DIRECTORY" value="$PROJECT_DIR$" />
+      <option name="IS_MODULE_SDK" value="true" />
+      <option name="ADD_CONTENT_ROOTS" value="true" />
+      <option name="ADD_SOURCE_ROOTS" value="true" />
+      <option name="SCRIPT_NAME" value="$PROJECT_DIR$/gradio_demo_img2img.py" />
+      <option name="PARAMETERS" value="" />
+      <option name="SHOW_COMMAND_LINE" value="false" />
+      <option name="EMULATE_TERMINAL" value="false" />
+      <option name="MODULE_MODE" value="false" />
+      <option name="REDIRECT_INPUT" value="false" />
+      <option name="INPUT_FILE" value="" />
+      <method v="2" />
+    </configuration>
+    <recent_temporary>
+      <list>
+        <item itemvalue="Python.gradio_demo_img2img" />
+      </list>
+    </recent_temporary>
+  </component>
+  <component name="SpellCheckerSettings" RuntimeDictionaries="0" Folders="0" CustomDictionaries="0" DefaultDictionary="应用程序级" UseSingleDictionary="true" transferred="true" />
+  <component name="TaskManager">
+    <task active="true" id="Default" summary="默认任务">
+      <changelist id="cf13c2e0-fea8-4b36-9b87-d1a7a68e5205" name="更改" comment="" />
+      <created>1707026272531</created>
+      <option name="number" value="Default" />
+      <option name="presentableId" value="Default" />
+      <updated>1707026272531</updated>
+    </task>
+    <servers />
+  </component>
+</project>

README.md

@@ -1,12 +1,158 @@
 ---
-title: Dzai
-emoji: 🐨
-colorFrom: green
-colorTo: gray
+title: dzai
+app_file: gradio_demo_img2img.py
 sdk: gradio
-sdk_version: 4.16.0
-app_file: app.py
-pinned: false
+sdk_version: 4.8.0
 ---
+# DemoFusion
+[![Project Page](https://img.shields.io/badge/Project-Page-green.svg)](https://ruoyidu.github.io/demofusion/demofusion.html)
+[![arXiv](https://img.shields.io/badge/arXiv-2311.16973-b31b1b.svg)](https://arxiv.org/pdf/2311.16973.pdf)
+[![Replicate](https://img.shields.io/badge/Demo-%F0%9F%9A%80%20Replicate-blue)](https://replicate.com/lucataco/demofusion)
+[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/camenduru/DemoFusion-colab/blob/main/DemoFusion_colab.ipynb)
+[![Hugging Face](https://img.shields.io/badge/i2i-%F0%9F%A4%97%20Hugging%20Face-blue)](https://huggingface.co/spaces/radames/Enhance-This-DemoFusion-SDXL)
+[![Page Views Count](https://badges.toozhao.com/badges/01HFMAPCVTA1T32KN2PASNYGYK/blue.svg)](https://badges.toozhao.com/stats/01HFMAPCVTA1T32KN2PASNYGYK "Get your own page views count badge on badges.toozhao.com")
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+Code release for "DemoFusion: Democratising High-Resolution Image Generation With No 💰" (arXiv 2023)
+
+<img src="figures/illustration.jpg" width="800"/>
+
+**Abstract**: High-resolution image generation with Generative Artificial Intelligence (GenAI) has immense potential but, due to the enormous capital investment required for training, it is increasingly centralised to a few large corporations, and hidden behind paywalls. This paper aims to democratise high-resolution GenAI by advancing the frontier of high-resolution generation while remaining accessible to a broad audience. We demonstrate that existing Latent Diffusion Models (LDMs) possess untapped potential for higher-resolution image generation. Our novel DemoFusion framework seamlessly extends open-source GenAI models, employing Progressive Upscaling, Skip Residual, and Dilated Sampling mechanisms to achieve higher-resolution image generation. The progressive nature of DemoFusion requires more passes, but the intermediate results can serve as "previews", facilitating rapid prompt iteration.
+
+# News
+- **2023.12.12**: ✨ DemoFusion with ControNet is availabe now! Check it out at `pipeline_demofusion_sdxl_controlnet`! The local [Gradio Demo](https://github.com/PRIS-CV/DemoFusion#DemoFusionControlNet-with-local-Gradio-demo) is also available.
+- **2023.12.10**: ✨ Image2Image is supported by `pipeline_demofusion_sdxl` now! The local [Gradio Demo](https://github.com/PRIS-CV/DemoFusion#Image2Image-with-local-Gradio-demo) is also available.
+- **2023.12.08**: 🚀 A HuggingFace Demo for Img2Img is now available! [![Hugging Face](https://img.shields.io/badge/i2i-%F0%9F%A4%97%20Hugging%20Face-blue)](https://huggingface.co/spaces/radames/Enhance-This-DemoFusion-SDXL) Thank [Radamés](https://github.com/radames) for the implementation and [![Hugging Face](https://img.shields.io/badge/Hugging%20Face-Diffusers-orange.svg)](https://huggingface.co/docs/diffusers/index) for the support!
+- **2023.12.07**: 🚀 Add Colab demo [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/camenduru/DemoFusion-colab/blob/main/DemoFusion_colab.ipynb). Check it out! Thank [camenduru](https://github.com/camenduru) for the implementation!
+- **2023.12.06**: ✨ The local [Gradio Demo](https://github.com/PRIS-CV/DemoFusion#Text2Image-with-local-Gradio-demo) is now available! Better interaction and presentation!
+- **2023.12.04**: ✨ A [low-vram version](https://github.com/PRIS-CV/DemoFusion#Text2Image-on-Windows-with-8-GB-of-VRAM) of DemoFusion is available! Thank [klimaleksus](https://github.com/klimaleksus) for the implementation!
+- **2023.12.01**: 🚀 Integrated to [Replicate](https://replicate.com/explore). Check out the online demo: [![Replicate](https://img.shields.io/badge/Demo-%F0%9F%9A%80%20Replicate-blue)](https://replicate.com/lucataco/demofusion) Thank [Luis C.](https://github.com/lucataco) for the implementation!
+- **2023.11.29**: 💰 `pipeline_demofusion_sdxl` is released.
+
+# Usage
+## A quick try with integrated demos
+- HuggingFace Space: Try Text2Image generation at [![Hugging Face](https://img.shields.io/badge/t2i-%F0%9F%A4%97%20Hugging%20Face-blue)](https://huggingface.co/spaces/fffiloni/DemoFusion) and Image2Image enhancement at [![Hugging Face](https://img.shields.io/badge/i2i-%F0%9F%A4%97%20Hugging%20Face-blue)](https://huggingface.co/spaces/radames/Enhance-This-DemoFusion-SDXL).
+- Colab: Try Text2Image generation at [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/camenduru/DemoFusion-colab/blob/main/DemoFusion_colab.ipynb) and Image2Image enhancement at [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/camenduru/DemoFusion-colab/blob/main/DemoFusion_img2img_colab.ipynb).
+- Replicate: Try Text2Image generation at [![Replicate](https://img.shields.io/badge/Demo-%F0%9F%9A%80%20Replicate-blue)](https://replicate.com/lucataco/demofusion) and Image2Image enhancement at [![Replicate](https://img.shields.io/badge/Demo-%F0%9F%9A%80%20Replicate-blue)](https://replicate.com/lucataco/demofusion-enhance).
+
+## Starting with our code
+### Hyper-parameters
+- `view_batch_size` (`int`, defaults to 16):
+  The batch size for multiple denoising paths. Typically, a larger batch size can result in higher efficiency but comes with increased GPU memory requirements.
+- `stride` (`int`, defaults to 64):
+  The stride of moving local patches. A smaller stride is better for alleviating seam issues, but it also introduces additional computational overhead and inference time.
+- `cosine_scale_1` (`float`, defaults to 3):
+  Control the decreasing rate of skip-residual. A smaller value results in better consistency with low-resolution results, but it may lead to more pronounced upsampling noise. Please refer to Appendix C in the DemoFusion paper.
+- `cosine_scale_2` (`float`, defaults to 1):
+  Control the decreasing rate of dilated sampling. A smaller value can better address the repetition issue, but it may lead to grainy images. For specific impacts, please refer to Appendix C in the DemoFusion paper.
+- `cosine_scale_3` (`float`, defaults to 1):
+  Control the decrease rate of the Gaussian filter. A smaller value results in less grainy images, but it may lead to over-smoothing images. Please refer to Appendix C in the DemoFusion paper.
+- `sigma` (`float`, defaults to 1):
+  The standard value of the Gaussian filter. A larger sigma promotes the global guidance of dilated sampling, but it has the potential of over-smoothing.
+- `multi_decoder` (`bool`, defaults to True):
+  Determine whether to use a tiled decoder. Generally, a tiled decoder becomes necessary when the resolution exceeds 3072*3072 on an RTX 3090 GPU.
+- `show_image` (`bool`, defaults to False):
+  Determine whether to show intermediate results during generation.
+
+### Text2Image (will take about 17 GB of VRAM)
+- Set up the dependencies as:
+```
+conda create -n demofusion python=3.9
+conda activate demofusion
+pip install -r requirements.txt
+```
+- Download `pipeline_demofusion_sdxl.py` and run it as follows. A use case can be found in `demo.ipynb`.
+```
+from pipeline_demofusion_sdxl import DemoFusionSDXLPipeline
+import torch
+
+model_ckpt = "stabilityai/stable-diffusion-xl-base-1.0"
+pipe = DemoFusionSDXLPipeline.from_pretrained(model_ckpt, torch_dtype=torch.float16)
+pipe = pipe.to("cuda")
+
+prompt = "Envision a portrait of an elderly woman, her face a canvas of time, framed by a headscarf with muted tones of rust and cream. Her eyes, blue like faded denim. Her attire, simple yet dignified."
+negative_prompt = "blurry, ugly, duplicate, poorly drawn, deformed, mosaic"
+
+images = pipe(prompt, negative_prompt=negative_prompt,
+              height=3072, width=3072, view_batch_size=16, stride=64,
+              num_inference_steps=50, guidance_scale=7.5,
+              cosine_scale_1=3, cosine_scale_2=1, cosine_scale_3=1, sigma=0.8,
+              multi_decoder=True, show_image=True
+             )
+
+for i, image in enumerate(images):
+    image.save('image_' + str(i) + '.png')
+```
+- ⚠️ When you have enough VRAM (e.g., generating 2048*2048 images on hardware with more than 18GB RAM), you can set `multi_decoder=False`, which can make the decoding process faster.
+- Please feel free to try different prompts and resolutions.
+- Default hyper-parameters are recommended, but they may not be optimal for all cases. For specific impacts of each hyper-parameter, please refer to Appendix C in the DemoFusion paper.
+- The code was cleaned before the release. If you encounter any issues, please contact us.
+
+### Text2Image on Windows with 8 GB of VRAM
+
+- Set up the environment as:
+
+```
+cmd
+git clone "https://github.com/PRIS-CV/DemoFusion"
+cd DemoFusion
+python -m venv venv
+venv\Scripts\activate
+pip install -U "xformers==0.0.22.post7+cu118" --index-url https://download.pytorch.org/whl/cu118
+pip install "diffusers==0.21.4" "matplotlib==3.8.2" "transformers==4.35.2" "accelerate==0.25.0"
+```
+
+- Launch DemoFusion as follows. The use case can be found in `demo_lowvram.py`.
+
+```
+python
+from pipeline_demofusion_sdxl import DemoFusionSDXLPipeline
+
+import torch
+from diffusers.models import AutoencoderKL
+vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=torch.float16)
+
+model_ckpt = "stabilityai/stable-diffusion-xl-base-1.0"
+pipe = DemoFusionSDXLPipeline.from_pretrained(model_ckpt, torch_dtype=torch.float16, vae=vae)
+pipe = pipe.to("cuda")
+
+prompt = "Envision a portrait of an elderly woman, her face a canvas of time, framed by a headscarf with muted tones of rust and cream. Her eyes, blue like faded denim. Her attire, simple yet dignified."
+negative_prompt = "blurry, ugly, duplicate, poorly drawn, deformed, mosaic"
+
+images = pipe(prompt, negative_prompt=negative_prompt,
+              height=2048, width=2048, view_batch_size=4, stride=64,
+              num_inference_steps=40, guidance_scale=7.5,
+              cosine_scale_1=3, cosine_scale_2=1, cosine_scale_3=1, sigma=0.8,
+              multi_decoder=True, show_image=False, lowvram=True
+             )
+
+for i, image in enumerate(images):
+    image.save('image_' + str(i) + '.png')
+```
+### Text2Image with local Gradio demo
+- Make sure you have installed `gradio` and `gradio_imageslider`.
+- Launch DemoFusion via Gradio demo now -- try `python gradio_demo.py`! Better Interaction and Presentation！
+<img src="figures/gradio_demo.png" width="600"/>
+
+### Image2Image with local Gradio demo
+- Make sure you have installed `gradio` and `gradio_imageslider`.
+- Launch DemoFusion Image2Image by `python gradio_demo_img2img.py`.
+<img src="figures/gradio_demo_img2img.png" width="600"/>
+- ⚠️ Please note that, as a tuning-free framework, DemoFusion's Image2Image capability is strongly correlated with the SDXL's training data distribution and will show a significant bias. An accurate prompt to describe the content and style of the input also significantly improves performance. Have fun and regard it as a side application of text+image based generation.
+
+### DemoFusion+ControlNet with local Gradio demo
+- Make sure you have installed `gradio` and `gradio_imageslider`.
+- Launch DemoFusion+ControNet Text2Image by `python gradio_demo.py`.
+- <img src="figures/gradio_demo_controlnet.png" width="600"/>
+- Launch DemoFusion+ControNet Image2Image by `python gradio_demo_img2img.py`.
+- <img src="figures/gradio_demo_controlnet_img2img.png" width="600"/>
+
+## Citation
+If you find this paper useful in your research, please consider citing:
+```
+@article{du2023demofusion,
+  title={DemoFusion: Democratising High-Resolution Image Generation With No $$$},
+  author={Du, Ruoyi and Chang, Dongliang and Hospedales, Timothy and Song, Yi-Zhe and Ma, Zhanyu},
+  journal={arXiv preprint arXiv:2311.16973},
+  year={2023}
+}
+```

demo.ipynb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a6bbe553656b3d9c863a261a053722930b3b538d5b6b05eac66ff9ae83eaf976
+size 17016845

demo_lowvram.py

@@ -0,0 +1,34 @@
+
+'''
+Installation on Windows for GPU with 8 Gb of VRAM and xformers:
+
+git clone "https://github.com/PRIS-CV/DemoFusion"
+cd DemoFusion
+python -m venv venv
+venv\Scripts\activate
+pip install -U "xformers==0.0.22.post7+cu118" --index-url https://download.pytorch.org/whl/cu118
+pip install "diffusers==0.21.4" "matplotlib==3.8.2" "transformers==4.35.2" "accelerate==0.25.0"
+'''
+
+from pipeline_demofusion_sdxl import DemoFusionSDXLPipeline
+
+import torch
+from diffusers.models import AutoencoderKL
+vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=torch.float16)
+
+model_ckpt = "stabilityai/stable-diffusion-xl-base-1.0"
+pipe = DemoFusionSDXLPipeline.from_pretrained(model_ckpt, torch_dtype=torch.float16, vae=vae)
+pipe = pipe.to("cuda")
+
+prompt = "Envision a portrait of an elderly woman, her face a canvas of time, framed by a headscarf with muted tones of rust and cream. Her eyes, blue like faded denim. Her attire, simple yet dignified."
+negative_prompt = "blurry, ugly, duplicate, poorly drawn, deformed, mosaic"
+
+images = pipe(prompt, negative_prompt=negative_prompt,
+              height=2048, width=2048, view_batch_size=4, stride=64,
+              num_inference_steps=40, guidance_scale=7.5,
+              cosine_scale_1=3, cosine_scale_2=1, cosine_scale_3=1, sigma=0.8,
+              multi_decoder=True, show_image=False, lowvram=True
+             )
+
+for i, image in enumerate(images):
+    image.save('image_'+str(i)+'.png')

gradio_demo.py

@@ -0,0 +1,46 @@
+import gradio as gr
+from pipeline_demofusion_sdxl import DemoFusionSDXLPipeline
+from gradio_imageslider import ImageSlider
+import torch
+
+def generate_images(prompt, negative_prompt, height, width, num_inference_steps, guidance_scale, cosine_scale_1, cosine_scale_2, cosine_scale_3, sigma, view_batch_size, stride, seed):
+    model_ckpt = "stabilityai/stable-diffusion-xl-base-1.0"
+    pipe = DemoFusionSDXLPipeline.from_pretrained(model_ckpt, torch_dtype=torch.float16)
+    pipe = pipe.to("cuda")
+
+    generator = torch.Generator(device='cuda')
+    generator = generator.manual_seed(int(seed))
+
+    images = pipe(prompt, negative_prompt=negative_prompt, generator=generator,
+                  height=int(height), width=int(width), view_batch_size=int(view_batch_size), stride=int(stride),
+                  num_inference_steps=int(num_inference_steps), guidance_scale=guidance_scale,
+                  cosine_scale_1=cosine_scale_1, cosine_scale_2=cosine_scale_2, cosine_scale_3=cosine_scale_3, sigma=sigma,
+                  multi_decoder=True, show_image=False
+                 )
+
+    return (images[0], images[-1])
+
+iface = gr.Interface(
+    fn=generate_images,
+    inputs=[
+        gr.Textbox(label="Prompt"),
+        gr.Textbox(label="Negative Prompt", value="blurry, ugly, duplicate, poorly drawn, deformed, mosaic"),
+        gr.Slider(minimum=1024, maximum=4096, step=1024, value=2048, label="Height"),
+        gr.Slider(minimum=1024, maximum=4096, step=1024, value=2048, label="Width"),
+        gr.Slider(minimum=10, maximum=100, step=1, value=50, label="Num Inference Steps"),
+        gr.Slider(minimum=1, maximum=20, step=0.1, value=7.5, label="Guidance Scale"),
+        gr.Slider(minimum=0, maximum=5, step=0.1, value=3, label="Cosine Scale 1"),
+        gr.Slider(minimum=0, maximum=5, step=0.1, value=1, label="Cosine Scale 2"),
+        gr.Slider(minimum=0, maximum=5, step=0.1, value=1, label="Cosine Scale 3"),
+        gr.Slider(minimum=0.1, maximum=1, step=0.1, value=0.8, label="Sigma"),
+        gr.Slider(minimum=4, maximum=32, step=4, value=16, label="View Batch Size"),
+        gr.Slider(minimum=8, maximum=96, step=8, value=64, label="Stride"),
+        gr.Number(label="Seed", value=2013)
+    ],
+    # outputs=gr.Gallery(label="Generated Images"),
+    outputs=ImageSlider(label="Comparison of SDXL and DemoFusion"),
+    title="DemoFusion Gradio Demo",
+    description="Generate images with the DemoFusion SDXL Pipeline."
+)
+
+iface.launch()

gradio_demo_controlnet.py

@@ -0,0 +1,93 @@
+import gradio as gr
+from diffusers import ControlNetModel, AutoencoderKL
+from pipeline_demofusion_sdxl import DemoFusionSDXLPipeline
+from pipeline_demofusion_sdxl_controlnet import DemoFusionSDXLControlNetPipeline
+from gradio_imageslider import ImageSlider
+import torch, gc
+from torchvision import transforms
+from PIL import Image
+import numpy as np
+import cv2
+
+def load_and_process_image(pil_image):
+    transform = transforms.Compose(
+        [
+            transforms.Resize((1024, 1024)),
+            transforms.ToTensor(),
+            transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
+        ]
+    )
+    image = transform(pil_image)
+    image = image.unsqueeze(0).half()
+    return image
+
+
+def pad_image(image):
+    w, h = image.size
+    if w == h:
+        return image
+    elif w > h:
+        new_image = Image.new(image.mode, (w, w), (0, 0, 0))
+        pad_w = 0
+        pad_h = (w - h) // 2
+        new_image.paste(image, (0, pad_h))
+        return new_image
+    else:
+        new_image = Image.new(image.mode, (h, h), (0, 0, 0))
+        pad_w = (h - w) // 2
+        pad_h = 0
+        new_image.paste(image, (pad_w, 0))
+        return new_image
+
+def generate_images(prompt, negative_prompt, controlnet_conditioning_scale, height, width, num_inference_steps, guidance_scale, cosine_scale_1, cosine_scale_2, cosine_scale_3, sigma, view_batch_size, stride, seed, input_image):
+    padded_image = pad_image(input_image).resize((1024, 1024)).convert("RGB")
+    image_lr = load_and_process_image(padded_image).to('cuda')
+    controlnet = ControlNetModel.from_pretrained("diffusers/controlnet-canny-sdxl-1.0", torch_dtype=torch.float16)
+    vae = AutoencoderKL.from_pretrained("madebyollin/stable-diffusion-xl-base-1.0/vae-fix", torch_dtype=torch.float16)
+    pipe = DemoFusionSDXLControlNetPipeline.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", controlnet=controlnet, vae=vae, torch_dtype=torch.float16)
+    pipe = pipe.to("cuda")
+    generator = torch.Generator(device='cuda')
+    generator = generator.manual_seed(int(seed))
+    # get canny image
+    canny_image = np.array(padded_image)
+    canny_image = cv2.Canny(canny_image, 100, 200)
+    canny_image = canny_image[:, :, None]
+    canny_image = np.concatenate([canny_image, canny_image, canny_image], axis=2)
+    canny_image = Image.fromarray(canny_image)
+    images = pipe(prompt, negative_prompt=negative_prompt, controlnet_conditioning_scale=controlnet_conditioning_scale, 
+                  condition_image=canny_image, generator=generator,
+                  height=int(height), width=int(width), view_batch_size=int(view_batch_size), stride=int(stride),
+                  num_inference_steps=int(num_inference_steps), guidance_scale=guidance_scale,
+                  cosine_scale_1=cosine_scale_1, cosine_scale_2=cosine_scale_2, cosine_scale_3=cosine_scale_3, sigma=sigma,
+                  multi_decoder=True, show_image=False, lowvram=False
+                 )
+    for i, image in enumerate(images):
+      image.save('image_'+str(i)+'.png')
+    pipe = None
+    gc.collect()
+    torch.cuda.empty_cache()
+    return (canny_image, images[-1])
+
+with gr.Blocks(title=f"DemoFusion") as demo:
+    with gr.Column():
+      with gr.Row():
+        with gr.Group():
+          image_input = gr.Image(type="pil", label="Input Image")
+          prompt = gr.Textbox(label="Prompt", value="")
+          negative_prompt = gr.Textbox(label="Negative Prompt", value="blurry, ugly, duplicate, poorly drawn, deformed, mosaic")
+          controlnet_conditioning_scale = gr.Slider(minimum=0, maximum=1, step=0.1, value=0.5, label="ControlNet Conditioning Scale")
+          width = gr.Slider(minimum=1024, maximum=4096, step=1024, value=2048, label="Width")
+          height = gr.Slider(minimum=1024, maximum=4096, step=1024, value=2048, label="Height")
+          num_inference_steps = gr.Slider(minimum=10, maximum=100, step=1, value=50, label="Num Inference Steps")
+          guidance_scale = gr.Slider(minimum=1, maximum=20, step=0.1, value=7.5, label="Guidance Scale")
+          cosine_scale_1 = gr.Slider(minimum=0, maximum=5, step=0.1, value=3, label="Cosine Scale 1")
+          cosine_scale_2 = gr.Slider(minimum=0, maximum=5, step=0.1, value=1, label="Cosine Scale 2")
+          cosine_scale_3 = gr.Slider(minimum=0, maximum=5, step=0.1, value=1, label="Cosine Scale 3")
+          sigma = gr.Slider(minimum=0.1, maximum=1, step=0.1, value=0.8, label="Sigma")
+          view_batch_size = gr.Slider(minimum=4, maximum=32, step=4, value=16, label="View Batch Size")
+          stride = gr.Slider(minimum=8, maximum=96, step=8, value=64, label="Stride")
+          seed = gr.Number(label="Seed", value=2013)
+          button = gr.Button()
+        output_images = ImageSlider(show_label=False)
+    button.click(fn=generate_images, inputs=[prompt, negative_prompt, controlnet_conditioning_scale, height, width, num_inference_steps, guidance_scale, cosine_scale_1, cosine_scale_2, cosine_scale_3, sigma, view_batch_size, stride, seed, image_input], outputs=[output_images], show_progress=True)
+demo.queue().launch(inline=False, share=True, debug=True)

gradio_demo_controlnet_img2img.py

@@ -0,0 +1,93 @@
+import gradio as gr
+from diffusers import ControlNetModel, AutoencoderKL
+from pipeline_demofusion_sdxl import DemoFusionSDXLPipeline
+from pipeline_demofusion_sdxl_controlnet import DemoFusionSDXLControlNetPipeline
+from gradio_imageslider import ImageSlider
+import torch, gc
+from torchvision import transforms
+from PIL import Image
+import numpy as np
+import cv2
+
+def load_and_process_image(pil_image):
+    transform = transforms.Compose(
+        [
+            transforms.Resize((1024, 1024)),
+            transforms.ToTensor(),
+            transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
+        ]
+    )
+    image = transform(pil_image)
+    image = image.unsqueeze(0).half()
+    return image
+
+
+def pad_image(image):
+    w, h = image.size
+    if w == h:
+        return image
+    elif w > h:
+        new_image = Image.new(image.mode, (w, w), (0, 0, 0))
+        pad_w = 0
+        pad_h = (w - h) // 2
+        new_image.paste(image, (0, pad_h))
+        return new_image
+    else:
+        new_image = Image.new(image.mode, (h, h), (0, 0, 0))
+        pad_w = (h - w) // 2
+        pad_h = 0
+        new_image.paste(image, (pad_w, 0))
+        return new_image
+
+def generate_images(prompt, negative_prompt, controlnet_conditioning_scale, height, width, num_inference_steps, guidance_scale, cosine_scale_1, cosine_scale_2, cosine_scale_3, sigma, view_batch_size, stride, seed, input_image):
+    padded_image = pad_image(input_image).resize((1024, 1024)).convert("RGB")
+    image_lr = load_and_process_image(padded_image).to('cuda')
+    controlnet = ControlNetModel.from_pretrained("diffusers/controlnet-canny-sdxl-1.0", torch_dtype=torch.float16)
+    vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=torch.float16)
+    pipe = DemoFusionSDXLControlNetPipeline.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", controlnet=controlnet, vae=vae, torch_dtype=torch.float16)
+    pipe = pipe.to("cuda")
+    generator = torch.Generator(device='cuda')
+    generator = generator.manual_seed(int(seed))
+    # get canny image
+    canny_image = np.array(padded_image)
+    canny_image = cv2.Canny(canny_image, 100, 200)
+    canny_image = canny_image[:, :, None]
+    canny_image = np.concatenate([canny_image, canny_image, canny_image], axis=2)
+    canny_image = Image.fromarray(canny_image)
+    images = pipe(prompt, negative_prompt=negative_prompt, controlnet_conditioning_scale=controlnet_conditioning_scale, 
+                  image_lr=image_lr, condition_image=canny_image, generator=generator,
+                  height=int(height), width=int(width), view_batch_size=int(view_batch_size), stride=int(stride),
+                  num_inference_steps=int(num_inference_steps), guidance_scale=guidance_scale,
+                  cosine_scale_1=cosine_scale_1, cosine_scale_2=cosine_scale_2, cosine_scale_3=cosine_scale_3, sigma=sigma,
+                  multi_decoder=True, show_image=False, lowvram=False
+                 )
+    for i, image in enumerate(images):
+      image.save('image_'+str(i)+'.png')
+    pipe = None
+    gc.collect()
+    torch.cuda.empty_cache()
+    return (images[0], images[-1])
+
+with gr.Blocks(title=f"DemoFusion") as demo:
+    with gr.Column():
+      with gr.Row():
+        with gr.Group():
+          image_input = gr.Image(type="pil", label="Input Image")
+          prompt = gr.Textbox(label="Prompt (Note: an accurate prompt to describe the content and style of the input will significantly improve performance.)", value="8k high definition, high details")
+          negative_prompt = gr.Textbox(label="Negative Prompt", value="blurry, ugly, duplicate, poorly drawn, deformed, mosaic")
+          controlnet_conditioning_scale = gr.Slider(minimum=0, maximum=1, step=0.1, value=0.5, label="ControlNet Conditioning Scale")
+          width = gr.Slider(minimum=1024, maximum=4096, step=1024, value=2048, label="Width")
+          height = gr.Slider(minimum=1024, maximum=4096, step=1024, value=2048, label="Height")
+          num_inference_steps = gr.Slider(minimum=10, maximum=100, step=1, value=50, label="Num Inference Steps")
+          guidance_scale = gr.Slider(minimum=1, maximum=20, step=0.1, value=7.5, label="Guidance Scale")
+          cosine_scale_1 = gr.Slider(minimum=0, maximum=5, step=0.1, value=3, label="Cosine Scale 1")
+          cosine_scale_2 = gr.Slider(minimum=0, maximum=5, step=0.1, value=1, label="Cosine Scale 2")
+          cosine_scale_3 = gr.Slider(minimum=0, maximum=5, step=0.1, value=1, label="Cosine Scale 3")
+          sigma = gr.Slider(minimum=0.1, maximum=1, step=0.1, value=0.8, label="Sigma")
+          view_batch_size = gr.Slider(minimum=4, maximum=32, step=4, value=16, label="View Batch Size")
+          stride = gr.Slider(minimum=8, maximum=96, step=8, value=64, label="Stride")
+          seed = gr.Number(label="Seed", value=2013)
+          button = gr.Button()
+        output_images = ImageSlider(show_label=False)
+    button.click(fn=generate_images, inputs=[prompt, negative_prompt, controlnet_conditioning_scale, height, width, num_inference_steps, guidance_scale, cosine_scale_1, cosine_scale_2, cosine_scale_3, sigma, view_batch_size, stride, seed, image_input], outputs=[output_images], show_progress=True)
+demo.queue().launch(inline=False, share=True, debug=True)

gradio_demo_img2img.py

@@ -0,0 +1,81 @@
+import gradio as gr
+from diffusers import AutoencoderKL
+from pipeline_demofusion_sdxl import DemoFusionSDXLPipeline
+from gradio_imageslider import ImageSlider
+import torch, gc
+from torchvision import transforms
+from PIL import Image
+
+def load_and_process_image(pil_image):
+    transform = transforms.Compose(
+        [
+            transforms.Resize((1024, 1024)),
+            transforms.ToTensor(),
+            transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
+        ]
+    )
+    image = transform(pil_image)
+    image = image.unsqueeze(0).half()
+    return image
+
+
+def pad_image(image):
+    w, h = image.size
+    if w == h:
+        return image
+    elif w > h:
+        new_image = Image.new(image.mode, (w, w), (0, 0, 0))
+        pad_w = 0
+        pad_h = (w - h) // 2
+        new_image.paste(image, (0, pad_h))
+        return new_image
+    else:
+        new_image = Image.new(image.mode, (h, h), (0, 0, 0))
+        pad_w = (h - w) // 2
+        pad_h = 0
+        new_image.paste(image, (pad_w, 0))
+        return new_image
+
+def generate_images(prompt, negative_prompt, height, width, num_inference_steps, guidance_scale, cosine_scale_1, cosine_scale_2, cosine_scale_3, sigma, view_batch_size, stride, seed, input_image):
+    padded_image = pad_image(input_image).resize((1024, 1024)).convert("RGB")
+    image_lr = load_and_process_image(padded_image).to('cuda')
+    vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=torch.float16)
+    pipe = DemoFusionSDXLPipeline.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", vae=vae, torch_dtype=torch.float16)
+    pipe = pipe.to("cuda")
+    generator = torch.Generator(device='cuda')
+    generator = generator.manual_seed(int(seed))
+    images = pipe(prompt, negative_prompt=negative_prompt, generator=generator,
+                  height=int(height), width=int(width), view_batch_size=int(view_batch_size), stride=int(stride),
+                  num_inference_steps=int(num_inference_steps), guidance_scale=guidance_scale,
+                  cosine_scale_1=cosine_scale_1, cosine_scale_2=cosine_scale_2, cosine_scale_3=cosine_scale_3, sigma=sigma,
+                  multi_decoder=True, show_image=False, lowvram=False, image_lr=image_lr
+                 )
+    for i, image in enumerate(images):
+      image.save('image_'+str(i)+'.png')
+    pipe = None
+    gc.collect()
+    torch.cuda.empty_cache()
+    return (images[0], images[-1])
+
+with gr.Blocks(title=f"DemoFusion") as demo:
+    with gr.Column():
+      with gr.Row():
+        with gr.Group():
+          image_input = gr.Image(type="pil", label="Input Image")
+          prompt = gr.Textbox(label="Prompt (Note: an accurate prompt to describe the content and style of the input will significantly improve performance.)", value="8k high definition, high details")
+          negative_prompt = gr.Textbox(label="Negative Prompt", value="blurry, ugly, duplicate, poorly drawn, deformed, mosaic")
+          width = gr.Slider(minimum=1024, maximum=4096, step=1024, value=2048, label="Width")
+          height = gr.Slider(minimum=1024, maximum=4096, step=1024, value=2048, label="Height")
+          num_inference_steps = gr.Slider(minimum=5, maximum=100, step=1, value=50, label="Num Inference Steps")
+          guidance_scale = gr.Slider(minimum=1, maximum=20, step=0.1, value=7.5, label="Guidance Scale")
+          cosine_scale_1 = gr.Slider(minimum=0, maximum=5, step=0.1, value=3, label="Cosine Scale 1")
+          cosine_scale_2 = gr.Slider(minimum=0, maximum=5, step=0.1, value=1, label="Cosine Scale 2")
+          cosine_scale_3 = gr.Slider(minimum=0, maximum=5, step=0.1, value=1, label="Cosine Scale 3")
+          sigma = gr.Slider(minimum=0.1, maximum=1, step=0.1, value=0.8, label="Sigma")
+          view_batch_size = gr.Slider(minimum=4, maximum=32, step=4, value=16, label="View Batch Size")
+          stride = gr.Slider(minimum=8, maximum=96, step=8, value=64, label="Stride")
+          seed = gr.Number(label="Seed", value=2013)
+          button = gr.Button()
+        output_images = ImageSlider(show_label=False)
+    button.click(fn=generate_images, inputs=[prompt, negative_prompt, height, width, num_inference_steps, guidance_scale, cosine_scale_1, cosine_scale_2, cosine_scale_3, sigma, view_batch_size, stride, seed, image_input], outputs=[output_images], show_progress=True)
+demo.queue().launch(inline=False, share=True, debug=True)

pipeline_demofusion_sdxl.py

@@ -0,0 +1,1446 @@
+# 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 typing import Any, Callable, Dict, List, Optional, Tuple, Union
+import matplotlib.pyplot as plt
+
+import torch
+import torch.nn.functional as F
+import numpy as np
+import random
+import warnings
+from transformers import CLIPTextModel, CLIPTextModelWithProjection, CLIPTokenizer
+
+from diffusers.image_processor import VaeImageProcessor
+from diffusers.loaders import (
+    FromSingleFileMixin,
+    LoraLoaderMixin,
+    TextualInversionLoaderMixin,
+)
+from diffusers.models import AutoencoderKL, UNet2DConditionModel
+from diffusers.models.attention_processor import (
+    AttnProcessor2_0,
+    LoRAAttnProcessor2_0,
+    LoRAXFormersAttnProcessor,
+    XFormersAttnProcessor,
+)
+from diffusers.models.lora import adjust_lora_scale_text_encoder
+from diffusers.schedulers import KarrasDiffusionSchedulers
+from diffusers.utils import (
+    is_accelerate_available,
+    is_accelerate_version,
+    is_invisible_watermark_available,
+    logging,
+    replace_example_docstring,
+)
+from diffusers.utils.torch_utils import randn_tensor
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline
+from diffusers.pipelines.stable_diffusion_xl import StableDiffusionXLPipelineOutput
+
+
+if is_invisible_watermark_available():
+    from .watermark import StableDiffusionXLWatermarker
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```py
+        >>> import torch
+        >>> from diffusers import StableDiffusionXLPipeline
+
+        >>> pipe = StableDiffusionXLPipeline.from_pretrained(
+        ...     "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16
+        ... )
+        >>> pipe = pipe.to("cuda")
+
+        >>> prompt = "a photo of an astronaut riding a horse on mars"
+        >>> image = pipe(prompt).images[0]
+        ```
+"""
+
+def gaussian_kernel(kernel_size=3, sigma=1.0, channels=3):
+    x_coord = torch.arange(kernel_size)
+    gaussian_1d = torch.exp(-(x_coord - (kernel_size - 1) / 2) ** 2 / (2 * sigma ** 2))
+    gaussian_1d = gaussian_1d / gaussian_1d.sum()
+    gaussian_2d = gaussian_1d[:, None] * gaussian_1d[None, :]
+    kernel = gaussian_2d[None, None, :, :].repeat(channels, 1, 1, 1)
+    
+    return kernel
+
+def gaussian_filter(latents, kernel_size=3, sigma=1.0):
+    channels = latents.shape[1]
+    kernel = gaussian_kernel(kernel_size, sigma, channels).to(latents.device, latents.dtype)
+    blurred_latents = F.conv2d(latents, kernel, padding=kernel_size//2, groups=channels)
+    
+    return blurred_latents
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg
+def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
+    """
+    Rescale `noise_cfg` according to `guidance_rescale`. Based on findings of [Common Diffusion Noise Schedules and
+    Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf). See Section 3.4
+    """
+    std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True)
+    std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True)
+    # rescale the results from guidance (fixes overexposure)
+    noise_pred_rescaled = noise_cfg * (std_text / std_cfg)
+    # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images
+    noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg
+    return noise_cfg
+
+
+class DemoFusionSDXLPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoaderMixin, TextualInversionLoaderMixin):
+    """
+    Pipeline for text-to-image generation using Stable Diffusion XL.
+
+    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+    library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+    In addition the pipeline inherits the following loading methods:
+        - *LoRA*: [`StableDiffusionXLPipeline.load_lora_weights`]
+        - *Ckpt*: [`loaders.FromSingleFileMixin.from_single_file`]
+
+    as well as the following saving methods:
+        - *LoRA*: [`loaders.StableDiffusionXLPipeline.save_lora_weights`]
+
+    Args:
+        vae ([`AutoencoderKL`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+        text_encoder ([`CLIPTextModel`]):
+            Frozen text-encoder. Stable Diffusion XL uses the text portion of
+            [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
+            the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
+        text_encoder_2 ([` CLIPTextModelWithProjection`]):
+            Second frozen text-encoder. Stable Diffusion XL uses the text and pool portion of
+            [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection),
+            specifically the
+            [laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k)
+            variant.
+        tokenizer (`CLIPTokenizer`):
+            Tokenizer of class
+            [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+        tokenizer_2 (`CLIPTokenizer`):
+            Second Tokenizer of class
+            [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+        unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
+        scheduler ([`SchedulerMixin`]):
+            A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+            [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+        force_zeros_for_empty_prompt (`bool`, *optional*, defaults to `"True"`):
+            Whether the negative prompt embeddings shall be forced to always be set to 0. Also see the config of
+            `stabilityai/stable-diffusion-xl-base-1-0`.
+        add_watermarker (`bool`, *optional*):
+            Whether to use the [invisible_watermark library](https://github.com/ShieldMnt/invisible-watermark/) to
+            watermark output images. If not defined, it will default to True if the package is installed, otherwise no
+            watermarker will be used.
+    """
+    model_cpu_offload_seq = "text_encoder->text_encoder_2->unet->vae"
+
+    def __init__(
+        self,
+        vae: AutoencoderKL,
+        text_encoder: CLIPTextModel,
+        text_encoder_2: CLIPTextModelWithProjection,
+        tokenizer: CLIPTokenizer,
+        tokenizer_2: CLIPTokenizer,
+        unet: UNet2DConditionModel,
+        scheduler: KarrasDiffusionSchedulers,
+        force_zeros_for_empty_prompt: bool = True,
+        add_watermarker: Optional[bool] = None,
+    ):
+        super().__init__()
+
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            text_encoder_2=text_encoder_2,
+            tokenizer=tokenizer,
+            tokenizer_2=tokenizer_2,
+            unet=unet,
+            scheduler=scheduler,
+        )
+        self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt)
+        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+        self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+        self.default_sample_size = self.unet.config.sample_size
+
+        add_watermarker = add_watermarker if add_watermarker is not None else is_invisible_watermark_available()
+
+        if add_watermarker:
+            self.watermark = StableDiffusionXLWatermarker()
+        else:
+            self.watermark = None
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_slicing
+    def enable_vae_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.vae.enable_slicing()
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_slicing
+    def disable_vae_slicing(self):
+        r"""
+        Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_slicing()
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_tiling
+    def enable_vae_tiling(self):
+        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.vae.enable_tiling()
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_tiling
+    def disable_vae_tiling(self):
+        r"""
+        Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_tiling()
+
+    def encode_prompt(
+        self,
+        prompt: str,
+        prompt_2: Optional[str] = None,
+        device: Optional[torch.device] = None,
+        num_images_per_prompt: int = 1,
+        do_classifier_free_guidance: bool = True,
+        negative_prompt: Optional[str] = None,
+        negative_prompt_2: Optional[str] = None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        lora_scale: Optional[float] = None,
+    ):
+        r"""
+        Encodes the prompt into text encoder hidden states.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                prompt to be encoded
+            prompt_2 (`str` or `List[str]`, *optional*):
+                The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
+                used in both text-encoders
+            device: (`torch.device`):
+                torch device
+            num_images_per_prompt (`int`):
+                number of images that should be generated per prompt
+            do_classifier_free_guidance (`bool`):
+                whether to use classifier free guidance or not
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            negative_prompt_2 (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
+                `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders
+            prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
+                If not provided, pooled text embeddings will be generated from `prompt` input argument.
+            negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
+                input argument.
+            lora_scale (`float`, *optional*):
+                A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+        """
+        device = device or self._execution_device
+
+        # set lora scale so that monkey patched LoRA
+        # function of text encoder can correctly access it
+        if lora_scale is not None and isinstance(self, LoraLoaderMixin):
+            self._lora_scale = lora_scale
+
+            # dynamically adjust the LoRA scale
+            adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
+            adjust_lora_scale_text_encoder(self.text_encoder_2, lora_scale)
+
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        # Define tokenizers and text encoders
+        tokenizers = [self.tokenizer, self.tokenizer_2] if self.tokenizer is not None else [self.tokenizer_2]
+        text_encoders = (
+            [self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2]
+        )
+
+        if prompt_embeds is None:
+            prompt_2 = prompt_2 or prompt
+            # textual inversion: procecss multi-vector tokens if necessary
+            prompt_embeds_list = []
+            prompts = [prompt, prompt_2]
+            for prompt, tokenizer, text_encoder in zip(prompts, tokenizers, text_encoders):
+                if isinstance(self, TextualInversionLoaderMixin):
+                    prompt = self.maybe_convert_prompt(prompt, tokenizer)
+
+                text_inputs = tokenizer(
+                    prompt,
+                    padding="max_length",
+                    max_length=tokenizer.model_max_length,
+                    truncation=True,
+                    return_tensors="pt",
+                )
+
+                text_input_ids = text_inputs.input_ids
+                untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+                if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+                    text_input_ids, untruncated_ids
+                ):
+                    removed_text = tokenizer.batch_decode(untruncated_ids[:, tokenizer.model_max_length - 1 : -1])
+                    logger.warning(
+                        "The following part of your input was truncated because CLIP can only handle sequences up to"
+                        f" {tokenizer.model_max_length} tokens: {removed_text}"
+                    )
+
+                prompt_embeds = text_encoder(
+                    text_input_ids.to(device),
+                    output_hidden_states=True,
+                )
+
+                # We are only ALWAYS interested in the pooled output of the final text encoder
+                pooled_prompt_embeds = prompt_embeds[0]
+                prompt_embeds = prompt_embeds.hidden_states[-2]
+
+                prompt_embeds_list.append(prompt_embeds)
+
+            prompt_embeds = torch.concat(prompt_embeds_list, dim=-1)
+
+        # get unconditional embeddings for classifier free guidance
+        zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt
+        if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt:
+            negative_prompt_embeds = torch.zeros_like(prompt_embeds)
+            negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds)
+        elif do_classifier_free_guidance and negative_prompt_embeds is None:
+            negative_prompt = negative_prompt or ""
+            negative_prompt_2 = negative_prompt_2 or negative_prompt
+
+            uncond_tokens: List[str]
+            if prompt is not None and type(prompt) is not type(negative_prompt):
+                raise TypeError(
+                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+                    f" {type(prompt)}."
+                )
+            elif isinstance(negative_prompt, str):
+                uncond_tokens = [negative_prompt, negative_prompt_2]
+            elif batch_size != len(negative_prompt):
+                raise ValueError(
+                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                    " the batch size of `prompt`."
+                )
+            else:
+                uncond_tokens = [negative_prompt, negative_prompt_2]
+
+            negative_prompt_embeds_list = []
+            for negative_prompt, tokenizer, text_encoder in zip(uncond_tokens, tokenizers, text_encoders):
+                if isinstance(self, TextualInversionLoaderMixin):
+                    negative_prompt = self.maybe_convert_prompt(negative_prompt, tokenizer)
+
+                max_length = prompt_embeds.shape[1]
+                uncond_input = tokenizer(
+                    negative_prompt,
+                    padding="max_length",
+                    max_length=max_length,
+                    truncation=True,
+                    return_tensors="pt",
+                )
+
+                negative_prompt_embeds = text_encoder(
+                    uncond_input.input_ids.to(device),
+                    output_hidden_states=True,
+                )
+                # We are only ALWAYS interested in the pooled output of the final text encoder
+                negative_pooled_prompt_embeds = negative_prompt_embeds[0]
+                negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2]
+
+                negative_prompt_embeds_list.append(negative_prompt_embeds)
+
+            negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1)
+
+        prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device)
+        bs_embed, seq_len, _ = prompt_embeds.shape
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+        if do_classifier_free_guidance:
+            # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+            seq_len = negative_prompt_embeds.shape[1]
+            negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device)
+            negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+            negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+        pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
+            bs_embed * num_images_per_prompt, -1
+        )
+        if do_classifier_free_guidance:
+            negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
+                bs_embed * num_images_per_prompt, -1
+            )
+
+        return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+    def prepare_extra_step_kwargs(self, generator, eta):
+        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+        # and should be between [0, 1]
+
+        accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        extra_step_kwargs = {}
+        if accepts_eta:
+            extra_step_kwargs["eta"] = eta
+
+        # check if the scheduler accepts generator
+        accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        if accepts_generator:
+            extra_step_kwargs["generator"] = generator
+        return extra_step_kwargs
+
+    def check_inputs(
+        self,
+        prompt,
+        prompt_2,
+        height,
+        width,
+        callback_steps,
+        negative_prompt=None,
+        negative_prompt_2=None,
+        prompt_embeds=None,
+        negative_prompt_embeds=None,
+        pooled_prompt_embeds=None,
+        negative_pooled_prompt_embeds=None,
+        num_images_per_prompt=None,
+    ):
+        if height % 8 != 0 or width % 8 != 0:
+            raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+        if (callback_steps is None) or (
+            callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+        ):
+            raise ValueError(
+                f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+                f" {type(callback_steps)}."
+            )
+
+        if prompt is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt_2 is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt is None and prompt_embeds is None:
+            raise ValueError(
+                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+            )
+        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+        elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)):
+            raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}")
+
+        if negative_prompt is not None and negative_prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+            )
+        elif negative_prompt_2 is not None and negative_prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:"
+                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+            )
+
+        if prompt_embeds is not None and negative_prompt_embeds is not None:
+            if prompt_embeds.shape != negative_prompt_embeds.shape:
+                raise ValueError(
+                    "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+                    f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+                    f" {negative_prompt_embeds.shape}."
+                )
+
+        if prompt_embeds is not None and pooled_prompt_embeds is None:
+            raise ValueError(
+                "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`."
+            )
+
+        if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None:
+            raise ValueError(
+                "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`."
+            )
+
+        # DemoFusion specific checks
+        if max(height, width) % 1024 != 0:
+            raise ValueError(f"the larger one of `height` and `width` has to be divisible by 1024 but are {height} and {width}.")
+
+        if num_images_per_prompt != 1:
+            warnings.warn("num_images_per_prompt != 1 is not supported by DemoFusion and will be ignored.")
+            num_images_per_prompt = 1
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
+    def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
+        shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
+        if isinstance(generator, list) and len(generator) != batch_size:
+            raise ValueError(
+                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+            )
+
+        if latents is None:
+            latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+        else:
+            latents = latents.to(device)
+
+        # scale the initial noise by the standard deviation required by the scheduler
+        latents = latents * self.scheduler.init_noise_sigma
+        return latents
+
+    def _get_add_time_ids(self, original_size, crops_coords_top_left, target_size, dtype):
+        add_time_ids = list(original_size + crops_coords_top_left + target_size)
+
+        passed_add_embed_dim = (
+            self.unet.config.addition_time_embed_dim * len(add_time_ids) + self.text_encoder_2.config.projection_dim
+        )
+        expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features
+
+        if expected_add_embed_dim != passed_add_embed_dim:
+            raise ValueError(
+                f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`."
+            )
+
+        add_time_ids = torch.tensor([add_time_ids], dtype=dtype)
+        return add_time_ids
+
+    def get_views(self, height, width, window_size=128, stride=64, random_jitter=False):
+        # Here, we define the mappings F_i (see Eq. 7 in the MultiDiffusion paper https://arxiv.org/abs/2302.08113)
+        # if panorama's height/width < window_size, num_blocks of height/width should return 1
+        height //= self.vae_scale_factor
+        width //= self.vae_scale_factor
+        num_blocks_height = int((height - window_size) / stride - 1e-6) + 2 if height > window_size else 1
+        num_blocks_width = int((width - window_size) / stride - 1e-6) + 2 if width > window_size else 1
+        total_num_blocks = int(num_blocks_height * num_blocks_width)
+        views = []
+        for i in range(total_num_blocks):
+            h_start = int((i // num_blocks_width) * stride)
+            h_end = h_start + window_size
+            w_start = int((i % num_blocks_width) * stride)
+            w_end = w_start + window_size
+
+            if h_end > height:
+                h_start = int(h_start + height - h_end)
+                h_end = int(height)
+            if w_end > width:
+                w_start = int(w_start + width - w_end)
+                w_end = int(width)
+            if h_start < 0:
+                h_end = int(h_end - h_start)
+                h_start = 0
+            if w_start < 0:
+                w_end = int(w_end - w_start)
+                w_start = 0
+
+            if random_jitter:
+                jitter_range = (window_size - stride) // 4
+                w_jitter = 0
+                h_jitter = 0
+                if (w_start != 0) and (w_end != width):
+                    w_jitter = random.randint(-jitter_range, jitter_range)
+                elif (w_start == 0) and (w_end != width):
+                    w_jitter = random.randint(-jitter_range, 0)
+                elif (w_start != 0) and (w_end == width):
+                    w_jitter = random.randint(0, jitter_range)
+                if (h_start != 0) and (h_end != height):
+                    h_jitter = random.randint(-jitter_range, jitter_range)
+                elif (h_start == 0) and (h_end != height):
+                    h_jitter = random.randint(-jitter_range, 0)
+                elif (h_start != 0) and (h_end == height):
+                    h_jitter = random.randint(0, jitter_range)
+                h_start += (h_jitter + jitter_range)
+                h_end += (h_jitter + jitter_range)
+                w_start += (w_jitter + jitter_range)
+                w_end += (w_jitter + jitter_range)
+            
+            views.append((h_start, h_end, w_start, w_end))
+        return views
+
+    def tiled_decode(self, latents, current_height, current_width):
+        sample_size = self.unet.config.sample_size
+        core_size = self.unet.config.sample_size // 4
+        core_stride = core_size
+        pad_size = self.unet.config.sample_size // 8 * 3
+        decoder_view_batch_size = 1
+        
+        if self.lowvram:
+            core_stride = core_size // 2
+            pad_size = core_size
+
+        views = self.get_views(current_height, current_width, stride=core_stride, window_size=core_size)
+        views_batch = [views[i : i + decoder_view_batch_size] for i in range(0, len(views), decoder_view_batch_size)]
+        latents_ = F.pad(latents, (pad_size, pad_size, pad_size, pad_size), 'constant', 0)
+        image = torch.zeros(latents.size(0), 3, current_height, current_width).to(latents.device)
+        count = torch.zeros_like(image).to(latents.device)
+        # get the latents corresponding to the current view coordinates
+        with self.progress_bar(total=len(views_batch)) as progress_bar:
+            for j, batch_view in enumerate(views_batch):
+                vb_size = len(batch_view)
+                latents_for_view = torch.cat(
+                    [
+                        latents_[:, :, h_start:h_end+pad_size*2, w_start:w_end+pad_size*2]
+                        for h_start, h_end, w_start, w_end in batch_view
+                    ]
+                ).to(self.vae.device)
+                image_patch = self.vae.decode(latents_for_view / self.vae.config.scaling_factor, return_dict=False)[0]
+                h_start, h_end, w_start, w_end = views[j]
+                h_start, h_end, w_start, w_end = h_start * self.vae_scale_factor, h_end * self.vae_scale_factor, w_start * self.vae_scale_factor, w_end * self.vae_scale_factor
+                p_h_start, p_h_end, p_w_start, p_w_end = pad_size * self.vae_scale_factor, image_patch.size(2) - pad_size * self.vae_scale_factor, pad_size * self.vae_scale_factor, image_patch.size(3) - pad_size * self.vae_scale_factor
+                image[:, :, h_start:h_end, w_start:w_end] += image_patch[:, :, p_h_start:p_h_end, p_w_start:p_w_end].to(latents.device)
+                count[:, :, h_start:h_end, w_start:w_end] += 1
+                progress_bar.update()
+        image = image / count
+        
+        return image
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_upscale.StableDiffusionUpscalePipeline.upcast_vae
+    def upcast_vae(self):
+        dtype = self.vae.dtype
+        self.vae.to(dtype=torch.float32)
+        use_torch_2_0_or_xformers = isinstance(
+            self.vae.decoder.mid_block.attentions[0].processor,
+            (
+                AttnProcessor2_0,
+                XFormersAttnProcessor,
+                LoRAXFormersAttnProcessor,
+                LoRAAttnProcessor2_0,
+            ),
+        )
+        # if xformers or torch_2_0 is used attention block does not need
+        # to be in float32 which can save lots of memory
+        if use_torch_2_0_or_xformers:
+            self.vae.post_quant_conv.to(dtype)
+            self.vae.decoder.conv_in.to(dtype)
+            self.vae.decoder.mid_block.to(dtype)
+
+    @torch.no_grad()
+    @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+        self,
+        prompt: Union[str, List[str]] = None,
+        prompt_2: Optional[Union[str, List[str]]] = None,
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        num_inference_steps: int = 50,
+        denoising_end: Optional[float] = None,
+        guidance_scale: float = 5.0,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        negative_prompt_2: Optional[Union[str, List[str]]] = None,
+        num_images_per_prompt: Optional[int] = 1,
+        eta: float = 0.0,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = False,
+        callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+        callback_steps: int = 1,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        guidance_rescale: float = 0.0,
+        original_size: Optional[Tuple[int, int]] = None,
+        crops_coords_top_left: Tuple[int, int] = (0, 0),
+        target_size: Optional[Tuple[int, int]] = None,
+        negative_original_size: Optional[Tuple[int, int]] = None,
+        negative_crops_coords_top_left: Tuple[int, int] = (0, 0),
+        negative_target_size: Optional[Tuple[int, int]] = None,
+        ################### DemoFusion specific parameters ####################
+        image_lr: Optional[torch.FloatTensor] = None,
+        view_batch_size: int = 16,
+        multi_decoder: bool = True,
+        stride: Optional[int] = 64,
+        cosine_scale_1: Optional[float] = 3.,
+        cosine_scale_2: Optional[float] = 1.,
+        cosine_scale_3: Optional[float] = 1.,
+        sigma: Optional[float] = 1.0,
+        show_image: bool = False,
+        lowvram: bool = False,
+    ):
+        r"""
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+                instead.
+            prompt_2 (`str` or `List[str]`, *optional*):
+                The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
+                used in both text-encoders
+            height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+                The height in pixels of the generated image. This is set to 1024 by default for the best results.
+                Anything below 512 pixels won't work well for
+                [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0)
+                and checkpoints that are not specifically fine-tuned on low resolutions.
+            width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+                The width in pixels of the generated image. This is set to 1024 by default for the best results.
+                Anything below 512 pixels won't work well for
+                [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0)
+                and checkpoints that are not specifically fine-tuned on low resolutions.
+            num_inference_steps (`int`, *optional*, defaults to 50):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            denoising_end (`float`, *optional*):
+                When specified, determines the fraction (between 0.0 and 1.0) of the total denoising process to be
+                completed before it is intentionally prematurely terminated. As a result, the returned sample will
+                still retain a substantial amount of noise as determined by the discrete timesteps selected by the
+                scheduler. The denoising_end parameter should ideally be utilized when this pipeline forms a part of a
+                "Mixture of Denoisers" multi-pipeline setup, as elaborated in [**Refining the Image
+                Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output)
+            guidance_scale (`float`, *optional*, defaults to 5.0):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+                usually at the expense of lower image quality.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            negative_prompt_2 (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
+                `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders
+            num_images_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            eta (`float`, *optional*, defaults to 0.0):
+                Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
+                [`schedulers.DDIMScheduler`], will be ignored for others.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+                to make generation deterministic.
+            latents (`torch.FloatTensor`, *optional*):
+                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor will ge generated by sampling using the supplied random `generator`.
+            prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
+                If not provided, pooled text embeddings will be generated from `prompt` input argument.
+            negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
+                input argument.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generate image. Choose between
+                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] instead
+                of a plain tuple.
+            callback (`Callable`, *optional*):
+                A function that will be called every `callback_steps` steps during inference. The function will be
+                called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+            callback_steps (`int`, *optional*, defaults to 1):
+                The frequency at which the `callback` function will be called. If not specified, the callback will be
+                called at every step.
+            cross_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            guidance_rescale (`float`, *optional*, defaults to 0.7):
+                Guidance rescale factor proposed by [Common Diffusion Noise Schedules and Sample Steps are
+                Flawed](https://arxiv.org/pdf/2305.08891.pdf) `guidance_scale` is defined as `φ` in equation 16. of
+                [Common Diffusion Noise Schedules and Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf).
+                Guidance rescale factor should fix overexposure when using zero terminal SNR.
+            original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+                If `original_size` is not the same as `target_size` the image will appear to be down- or upsampled.
+                `original_size` defaults to `(width, height)` if not specified. Part of SDXL's micro-conditioning as
+                explained in section 2.2 of
+                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+            crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
+                `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position
+                `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting
+                `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of
+                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+            target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+                For most cases, `target_size` should be set to the desired height and width of the generated image. If
+                not specified it will default to `(width, height)`. Part of SDXL's micro-conditioning as explained in
+                section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+            negative_original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+                To negatively condition the generation process based on a specific image resolution. Part of SDXL's
+                micro-conditioning as explained in section 2.2 of
+                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
+                information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
+            negative_crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
+                To negatively condition the generation process based on a specific crop coordinates. Part of SDXL's
+                micro-conditioning as explained in section 2.2 of
+                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
+                information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
+            negative_target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+                To negatively condition the generation process based on a target image resolution. It should be as same
+                as the `target_size` for most cases. Part of SDXL's micro-conditioning as explained in section 2.2 of
+                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
+                information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
+            ################### DemoFusion specific parameters ####################
+            image_lr (`torch.FloatTensor`, *optional*, , defaults to None):
+                Low-resolution image input for upscaling. If provided, DemoFusion will encode it as the initial latent representation.
+            view_batch_size (`int`, defaults to 16):
+                The batch size for multiple denoising paths. Typically, a larger batch size can result in higher 
+                efficiency but comes with increased GPU memory requirements.
+            multi_decoder (`bool`, defaults to True):
+                Determine whether to use a tiled decoder. Generally, when the resolution exceeds 3072x3072, 
+                a tiled decoder becomes necessary.
+            stride (`int`, defaults to 64):
+                The stride of moving local patches. A smaller stride is better for alleviating seam issues,
+                but it also introduces additional computational overhead and inference time.
+            cosine_scale_1 (`float`, defaults to 3):
+                Control the strength of skip-residual. For specific impacts, please refer to Appendix C
+                in the DemoFusion paper.
+            cosine_scale_2 (`float`, defaults to 1):
+                Control the strength of dilated sampling. For specific impacts, please refer to Appendix C
+                in the DemoFusion paper.
+            cosine_scale_3 (`float`, defaults to 1):
+                Control the strength of the gaussion filter. For specific impacts, please refer to Appendix C
+                in the DemoFusion paper.
+            sigma (`float`, defaults to 1):
+                The standard value of the gaussian filter.
+            show_image (`bool`, defaults to False):
+                Determine whether to show intermediate results during generation.
+            lowvram (`bool`, defaults to False):
+                Try to fit in 8 Gb of VRAM, with xformers installed.
+
+        Examples:
+
+        Returns:
+            a `list` with the generated images at each phase.
+        """
+        
+        # 0. Default height and width to unet
+        height = height or self.default_sample_size * self.vae_scale_factor
+        width = width or self.default_sample_size * self.vae_scale_factor
+
+        x1_size = self.default_sample_size * self.vae_scale_factor
+
+        height_scale = height / x1_size
+        width_scale = width / x1_size
+        scale_num = int(max(height_scale, width_scale))
+        aspect_ratio = min(height_scale, width_scale) / max(height_scale, width_scale)
+
+        original_size = original_size or (height, width)
+        target_size = target_size or (height, width)
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(
+            prompt,
+            prompt_2,
+            height,
+            width,
+            callback_steps,
+            negative_prompt,
+            negative_prompt_2,
+            prompt_embeds,
+            negative_prompt_embeds,
+            pooled_prompt_embeds,
+            negative_pooled_prompt_embeds,
+            num_images_per_prompt,
+        )
+
+        # 2. Define call parameters
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        device = self._execution_device
+        self.lowvram = lowvram
+        if self.lowvram:
+            self.vae.cpu()
+            self.unet.cpu()
+            self.text_encoder.to(device)
+            self.text_encoder_2.to(device)
+            image_lr.cpu()
+
+        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+        # corresponds to doing no classifier free guidance.
+        do_classifier_free_guidance = guidance_scale > 1.0
+
+        # 3. Encode input prompt
+        text_encoder_lora_scale = (
+            cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None
+        )
+        (
+            prompt_embeds,
+            negative_prompt_embeds,
+            pooled_prompt_embeds,
+            negative_pooled_prompt_embeds,
+        ) = self.encode_prompt(
+            prompt=prompt,
+            prompt_2=prompt_2,
+            device=device,
+            num_images_per_prompt=num_images_per_prompt,
+            do_classifier_free_guidance=do_classifier_free_guidance,
+            negative_prompt=negative_prompt,
+            negative_prompt_2=negative_prompt_2,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            pooled_prompt_embeds=pooled_prompt_embeds,
+            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
+            lora_scale=text_encoder_lora_scale,
+        )
+
+        # 4. Prepare timesteps
+        self.scheduler.set_timesteps(num_inference_steps, device=device)
+
+        timesteps = self.scheduler.timesteps
+
+        # 5. Prepare latent variables
+        num_channels_latents = self.unet.config.in_channels
+        latents = self.prepare_latents(
+            batch_size * num_images_per_prompt,
+            num_channels_latents,
+            height // scale_num,
+            width // scale_num,
+            prompt_embeds.dtype,
+            device,
+            generator,
+            latents,
+        )
+
+        # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+        # 7. Prepare added time ids & embeddings
+        add_text_embeds = pooled_prompt_embeds
+        add_time_ids = self._get_add_time_ids(
+            original_size, crops_coords_top_left, target_size, dtype=prompt_embeds.dtype
+        )
+        if negative_original_size is not None and negative_target_size is not None:
+            negative_add_time_ids = self._get_add_time_ids(
+                negative_original_size,
+                negative_crops_coords_top_left,
+                negative_target_size,
+                dtype=prompt_embeds.dtype,
+            )
+        else:
+            negative_add_time_ids = add_time_ids
+
+        if do_classifier_free_guidance:
+            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
+            add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0)
+            add_time_ids = torch.cat([negative_add_time_ids, add_time_ids], dim=0)
+        del negative_prompt_embeds, negative_pooled_prompt_embeds, negative_add_time_ids
+
+        prompt_embeds = prompt_embeds.to(device)
+        add_text_embeds = add_text_embeds.to(device)
+        add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1)
+
+        # 8. Denoising loop
+        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
+
+        # 7.1 Apply denoising_end
+        if denoising_end is not None and isinstance(denoising_end, float) and denoising_end > 0 and denoising_end < 1:
+            discrete_timestep_cutoff = int(
+                round(
+                    self.scheduler.config.num_train_timesteps
+                    - (denoising_end * self.scheduler.config.num_train_timesteps)
+                )
+            )
+            num_inference_steps = len(list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps)))
+            timesteps = timesteps[:num_inference_steps]
+
+        output_images = []
+        
+    ###################################################### Phase Initialization ########################################################
+
+        if self.lowvram:
+            self.text_encoder.cpu()
+            self.text_encoder_2.cpu()
+
+        if image_lr == None:
+            print("### Phase 1 Denoising ###")
+            with self.progress_bar(total=num_inference_steps) as progress_bar:
+                for i, t in enumerate(timesteps):
+    
+                    if self.lowvram:
+                        self.vae.cpu()
+                        self.unet.to(device)
+    
+                    latents_for_view = latents
+    
+                    # expand the latents if we are doing classifier free guidance
+                    latent_model_input = (
+                        latents.repeat_interleave(2, dim=0)
+                        if do_classifier_free_guidance
+                        else latents
+                    )
+                    latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+    
+                    # predict the noise residual
+                    added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids}
+                    noise_pred = self.unet(
+                        latent_model_input,
+                        t,
+                        encoder_hidden_states=prompt_embeds,
+                        cross_attention_kwargs=cross_attention_kwargs,
+                        added_cond_kwargs=added_cond_kwargs,
+                        return_dict=False,
+                    )[0]
+    
+                    # perform guidance
+                    if do_classifier_free_guidance:
+                        noise_pred_uncond, noise_pred_text = noise_pred[::2], noise_pred[1::2]
+                        noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+    
+                    if do_classifier_free_guidance and guidance_rescale > 0.0:
+                        # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
+                        noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=guidance_rescale)
+    
+                    # compute the previous noisy sample x_t -> x_t-1
+                    latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
+    
+                    # call the callback, if provided
+                    if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                        progress_bar.update()
+                        if callback is not None and i % callback_steps == 0:
+                            step_idx = i // getattr(self.scheduler, "order", 1)
+                            callback(step_idx, t, latents)
+            del latents_for_view, latent_model_input, noise_pred, noise_pred_text, noise_pred_uncond
+        else:
+            print("### Encoding Real Image ###")
+            latents = self.vae.encode(image_lr)
+            latents = latents.latent_dist.sample() * self.vae.config.scaling_factor
+                        
+        anchor_mean = latents.mean()
+        anchor_std = latents.std()
+        if self.lowvram:
+            latents = latents.cpu()
+            torch.cuda.empty_cache()
+        if not output_type == "latent":
+            # make sure the VAE is in float32 mode, as it overflows in float16
+            needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
+            
+            if self.lowvram:
+                needs_upcasting = False # use madebyollin/sdxl-vae-fp16-fix in lowvram mode!
+                self.unet.cpu()
+                self.vae.to(device)
+
+            if needs_upcasting:
+                self.upcast_vae()
+                latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
+            if self.lowvram and multi_decoder:
+                current_width_height = self.unet.config.sample_size * self.vae_scale_factor
+                image = self.tiled_decode(latents, current_width_height, current_width_height)
+            else:
+                image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+            # cast back to fp16 if needed
+            if needs_upcasting:
+                self.vae.to(dtype=torch.float16)
+    
+        image = self.image_processor.postprocess(image, output_type=output_type)
+        if show_image:
+            plt.figure(figsize=(10, 10))
+            plt.imshow(image[0])
+            plt.axis('off')  # Turn off axis numbers and ticks
+            plt.show()
+        output_images.append(image[0])
+                        
+    ####################################################### Phase Upscaling #####################################################
+        if image_lr == None:
+            starting_scale = 2
+        else:
+            starting_scale = 1
+        for current_scale_num in range(starting_scale, scale_num + 1):
+            if self.lowvram:
+                latents = latents.to(device)
+                self.unet.to(device)
+                torch.cuda.empty_cache()
+            print("### Phase {} Denoising ###".format(current_scale_num))
+            current_height = self.unet.config.sample_size * self.vae_scale_factor * current_scale_num
+            current_width = self.unet.config.sample_size * self.vae_scale_factor * current_scale_num
+            if height > width:
+                current_width = int(current_width * aspect_ratio)
+            else:
+                current_height = int(current_height * aspect_ratio)
+        
+            latents = F.interpolate(latents.to(device), size=(int(current_height / self.vae_scale_factor), int(current_width / self.vae_scale_factor)), mode='bicubic')
+
+            noise_latents = []
+            noise = torch.randn_like(latents)
+            for timestep in timesteps:
+                noise_latent = self.scheduler.add_noise(latents, noise, timestep.unsqueeze(0))
+                noise_latents.append(noise_latent)
+            latents = noise_latents[0]
+    
+            with self.progress_bar(total=num_inference_steps) as progress_bar:
+                for i, t in enumerate(timesteps):
+                    count = torch.zeros_like(latents)
+                    value = torch.zeros_like(latents)
+                    cosine_factor = 0.5 * (1 + torch.cos(torch.pi * (self.scheduler.config.num_train_timesteps - t) / self.scheduler.config.num_train_timesteps)).cpu()
+
+                    c1 = cosine_factor ** cosine_scale_1
+                    latents = latents * (1 - c1) + noise_latents[i] * c1
+                    
+                    ############################################# MultiDiffusion #############################################
+                    
+                    views = self.get_views(current_height, current_width, stride=stride, window_size=self.unet.config.sample_size, random_jitter=True)
+                    views_batch = [views[i : i + view_batch_size] for i in range(0, len(views), view_batch_size)]
+
+                    jitter_range = (self.unet.config.sample_size - stride) // 4
+                    latents_ = F.pad(latents, (jitter_range, jitter_range, jitter_range, jitter_range), 'constant', 0)
+
+                    count_local = torch.zeros_like(latents_)
+                    value_local = torch.zeros_like(latents_)
+                    
+                    for j, batch_view in enumerate(views_batch):
+                        vb_size = len(batch_view)
+
+                        # get the latents corresponding to the current view coordinates
+                        latents_for_view = torch.cat(
+                            [
+                                latents_[:, :, h_start:h_end, w_start:w_end]
+                                for h_start, h_end, w_start, w_end in batch_view
+                            ]
+                        )
+
+                        # expand the latents if we are doing classifier free guidance
+                        latent_model_input = latents_for_view
+                        latent_model_input = (
+                            latent_model_input.repeat_interleave(2, dim=0)
+                            if do_classifier_free_guidance
+                            else latent_model_input
+                        )
+                        latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+                        prompt_embeds_input = torch.cat([prompt_embeds] * vb_size)
+                        add_text_embeds_input = torch.cat([add_text_embeds] * vb_size)
+                        add_time_ids_input = []
+                        for h_start, h_end, w_start, w_end in batch_view:
+                            add_time_ids_ = add_time_ids.clone()
+                            add_time_ids_[:, 2] = h_start * self.vae_scale_factor
+                            add_time_ids_[:, 3] = w_start * self.vae_scale_factor
+                            add_time_ids_input.append(add_time_ids_)
+                        add_time_ids_input = torch.cat(add_time_ids_input)
+
+                        # predict the noise residual
+                        added_cond_kwargs = {"text_embeds": add_text_embeds_input, "time_ids": add_time_ids_input}
+                        noise_pred = self.unet(
+                            latent_model_input,
+                            t,
+                            encoder_hidden_states=prompt_embeds_input,
+                            cross_attention_kwargs=cross_attention_kwargs,
+                            added_cond_kwargs=added_cond_kwargs,
+                            return_dict=False,
+                        )[0]
+
+                        if do_classifier_free_guidance:
+                            noise_pred_uncond, noise_pred_text = noise_pred[::2], noise_pred[1::2]
+                            noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+                        if do_classifier_free_guidance and guidance_rescale > 0.0:
+                            # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
+                            noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=guidance_rescale)
+
+                        # compute the previous noisy sample x_t -> x_t-1
+                        self.scheduler._init_step_index(t)
+                        latents_denoised_batch = self.scheduler.step(
+                            noise_pred, t, latents_for_view, **extra_step_kwargs, return_dict=False)[0]
+
+                        # extract value from batch
+                        for latents_view_denoised, (h_start, h_end, w_start, w_end) in zip(
+                            latents_denoised_batch.chunk(vb_size), batch_view
+                        ):
+                            value_local[:, :, h_start:h_end, w_start:w_end] += latents_view_denoised
+                            count_local[:, :, h_start:h_end, w_start:w_end] += 1
+
+                    value_local = value_local[: ,:, jitter_range: jitter_range + current_height // self.vae_scale_factor, jitter_range: jitter_range + current_width // self.vae_scale_factor]
+                    count_local = count_local[: ,:, jitter_range: jitter_range + current_height // self.vae_scale_factor, jitter_range: jitter_range + current_width // self.vae_scale_factor]
+                    
+                    c2 = cosine_factor ** cosine_scale_2
+
+                    value += value_local / count_local * (1 - c2)
+                    count += torch.ones_like(value_local) * (1 - c2)
+                        
+                    ############################################# Dilated Sampling #############################################
+
+                    views = [[h, w] for h in range(current_scale_num) for w in range(current_scale_num)]
+                    views_batch = [views[i : i + view_batch_size] for i in range(0, len(views), view_batch_size)]
+                    
+                    h_pad = (current_scale_num - (latents.size(2) % current_scale_num)) % current_scale_num
+                    w_pad = (current_scale_num - (latents.size(3) % current_scale_num)) % current_scale_num
+                    latents_ = F.pad(latents, (w_pad, 0, h_pad, 0), 'constant', 0)
+                    
+                    count_global = torch.zeros_like(latents_)
+                    value_global = torch.zeros_like(latents_)
+
+                    c3 = 0.99 * cosine_factor ** cosine_scale_3 + 1e-2
+                    std_, mean_ = latents_.std(), latents_.mean()
+                    latents_gaussian = gaussian_filter(latents_, kernel_size=(2*current_scale_num-1), sigma=sigma*c3)
+                    latents_gaussian = (latents_gaussian - latents_gaussian.mean()) / latents_gaussian.std() * std_ + mean_
+
+                    for j, batch_view in enumerate(views_batch):
+                        latents_for_view = torch.cat(
+                            [
+                                latents_[:, :, h::current_scale_num, w::current_scale_num]
+                                for h, w in batch_view
+                            ]
+                        )
+                        latents_for_view_gaussian = torch.cat(
+                            [
+                                latents_gaussian[:, :, h::current_scale_num, w::current_scale_num]
+                                for h, w in batch_view
+                            ]
+                        )
+                    
+                        vb_size = latents_for_view.size(0)
+    
+                        # expand the latents if we are doing classifier free guidance
+                        latent_model_input = latents_for_view_gaussian
+                        latent_model_input = (
+                            latent_model_input.repeat_interleave(2, dim=0)
+                            if do_classifier_free_guidance
+                            else latent_model_input
+                        )
+                        latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+    
+                        prompt_embeds_input = torch.cat([prompt_embeds] * vb_size)
+                        add_text_embeds_input = torch.cat([add_text_embeds] * vb_size)
+                        add_time_ids_input = torch.cat([add_time_ids] * vb_size)
+    
+                        # predict the noise residual
+                        added_cond_kwargs = {"text_embeds": add_text_embeds_input, "time_ids": add_time_ids_input}
+                        noise_pred = self.unet(
+                            latent_model_input,
+                            t,
+                            encoder_hidden_states=prompt_embeds_input,
+                            cross_attention_kwargs=cross_attention_kwargs,
+                            added_cond_kwargs=added_cond_kwargs,
+                            return_dict=False,
+                        )[0]
+    
+                        if do_classifier_free_guidance:
+                            noise_pred_uncond, noise_pred_text = noise_pred[::2], noise_pred[1::2]
+                            noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+    
+                        if do_classifier_free_guidance and guidance_rescale > 0.0:
+                            # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
+                            noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=guidance_rescale)
+
+                        # compute the previous noisy sample x_t -> x_t-1
+                        self.scheduler._init_step_index(t)
+                        latents_denoised_batch = self.scheduler.step(
+                            noise_pred, t, latents_for_view, **extra_step_kwargs, return_dict=False)[0]
+    
+                        # extract value from batch
+                        for latents_view_denoised, (h, w) in zip(
+                            latents_denoised_batch.chunk(vb_size), batch_view
+                        ):
+                            value_global[:, :, h::current_scale_num, w::current_scale_num] += latents_view_denoised
+                            count_global[:, :, h::current_scale_num, w::current_scale_num] += 1
+
+                    c2 = cosine_factor ** cosine_scale_2
+
+                    value_global = value_global[: ,:, h_pad:, w_pad:]
+    
+                    value += value_global * c2
+                    count += torch.ones_like(value_global) * c2
+                    
+                           ###########################################################   
+                
+                    latents = torch.where(count > 0, value / count, value)
+                
+                    # call the callback, if provided
+                    if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                        progress_bar.update()
+                        if callback is not None and i % callback_steps == 0:
+                            step_idx = i // getattr(self.scheduler, "order", 1)
+                            callback(step_idx, t, latents)
+
+    #########################################################################################################################################
+
+                latents = (latents - latents.mean()) / latents.std() * anchor_std + anchor_mean
+                if self.lowvram:
+                    latents = latents.cpu()
+                    torch.cuda.empty_cache()
+                if not output_type == "latent":
+                    # make sure the VAE is in float32 mode, as it overflows in float16
+                    needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
+        
+                    if self.lowvram:
+                        needs_upcasting = False # use madebyollin/sdxl-vae-fp16-fix in lowvram mode!
+                        self.unet.cpu()
+                        self.vae.to(device)
+                    
+                    if needs_upcasting:
+                        self.upcast_vae()
+                        latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
+                    
+                    print("### Phase {} Decoding ###".format(current_scale_num))
+                    if multi_decoder:
+                        image = self.tiled_decode(latents, current_height, current_width)
+                    else:
+                        image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+        
+                    # cast back to fp16 if needed
+                    if needs_upcasting:
+                        self.vae.to(dtype=torch.float16)
+                else:
+                    image = latents
+        
+                if not output_type == "latent":
+                    image = self.image_processor.postprocess(image, output_type=output_type)
+                    if show_image:
+                        plt.figure(figsize=(10, 10))
+                        plt.imshow(image[0])
+                        plt.axis('off')  # Turn off axis numbers and ticks
+                        plt.show()
+                    output_images.append(image[0])
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        return output_images
+
+    # 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]], **kwargs):
+        # 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.
+
+        # Remove any existing hooks.
+        if is_accelerate_available() and is_accelerate_version(">=", "0.17.0.dev0"):
+            from accelerate.hooks import AlignDevicesHook, CpuOffload, remove_hook_from_module
+        else:
+            raise ImportError("Offloading requires `accelerate v0.17.0` or higher.")
+
+        is_model_cpu_offload = False
+        is_sequential_cpu_offload = False
+        recursive = False
+        for _, component in self.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."
+                    )
+                    recursive = is_sequential_cpu_offload
+                    remove_hook_from_module(component, recurse=recursive)
+        state_dict, network_alphas = self.lora_state_dict(
+            pretrained_model_name_or_path_or_dict,
+            unet_config=self.unet.config,
+            **kwargs,
+        )
+        self.load_lora_into_unet(state_dict, network_alphas=network_alphas, unet=self.unet)
+
+        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,
+            )
+
+        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,
+            )
+
+        # Offload back.
+        if is_model_cpu_offload:
+            self.enable_model_cpu_offload()
+        elif is_sequential_cpu_offload:
+            self.enable_sequential_cpu_offload()
+
+    @classmethod
+    def save_lora_weights(
+        self,
+        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,
+    ):
+        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"))
+
+        self.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):
+        self._remove_text_encoder_monkey_patch_classmethod(self.text_encoder)
+        self._remove_text_encoder_monkey_patch_classmethod(self.text_encoder_2)

pipeline_demofusion_sdxl_controlnet.py

@@ -0,0 +1,1795 @@
+# 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 typing import Any, Callable, Dict, List, Optional, Tuple, Union
+import matplotlib.pyplot as plt
+
+import numpy as np
+import PIL.Image
+import torch
+import torch.nn.functional as F
+import random
+import warnings
+from transformers import CLIPTextModel, CLIPTextModelWithProjection, CLIPTokenizer
+
+from diffusers.utils.import_utils import is_invisible_watermark_available
+
+from diffusers.image_processor import PipelineImageInput, VaeImageProcessor
+from diffusers.loaders import FromSingleFileMixin, LoraLoaderMixin, TextualInversionLoaderMixin
+from diffusers.models import AutoencoderKL, ControlNetModel, UNet2DConditionModel
+from diffusers.models.attention_processor import (
+    AttnProcessor2_0,
+    LoRAAttnProcessor2_0,
+    LoRAXFormersAttnProcessor,
+    XFormersAttnProcessor,
+)
+from diffusers.models.lora import adjust_lora_scale_text_encoder
+from diffusers.schedulers import KarrasDiffusionSchedulers
+from diffusers.utils import (
+    is_accelerate_available,
+    is_accelerate_version,
+    logging,
+    replace_example_docstring,
+)
+from diffusers.utils.torch_utils import is_compiled_module, randn_tensor
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline
+from diffusers.pipelines.stable_diffusion_xl import StableDiffusionXLPipelineOutput
+
+
+if is_invisible_watermark_available():
+    from diffusers.pipelines.stable_diffusion_xl.watermark import StableDiffusionXLWatermarker
+
+from diffusers.pipelines.controlnet.multicontrolnet import MultiControlNetModel
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+"""
+
+def gaussian_kernel(kernel_size=3, sigma=1.0, channels=3):
+    x_coord = torch.arange(kernel_size)
+    gaussian_1d = torch.exp(-(x_coord - (kernel_size - 1) / 2) ** 2 / (2 * sigma ** 2))
+    gaussian_1d = gaussian_1d / gaussian_1d.sum()
+    gaussian_2d = gaussian_1d[:, None] * gaussian_1d[None, :]
+    kernel = gaussian_2d[None, None, :, :].repeat(channels, 1, 1, 1)
+    
+    return kernel
+
+def gaussian_filter(latents, kernel_size=3, sigma=1.0):
+    channels = latents.shape[1]
+    kernel = gaussian_kernel(kernel_size, sigma, channels).to(latents.device, latents.dtype)
+    blurred_latents = F.conv2d(latents, kernel, padding=kernel_size//2, groups=channels)
+    
+    return blurred_latents
+
+class DemoFusionSDXLControlNetPipeline(
+    DiffusionPipeline, TextualInversionLoaderMixin, LoraLoaderMixin, FromSingleFileMixin
+):
+    r"""
+    Pipeline for text-to-image generation using Stable Diffusion XL with ControlNet guidance.
+
+    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.).
+
+    The pipeline also inherits the following loading methods:
+        - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+        - [`loaders.LoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+        - [`loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files
+
+    Args:
+        vae ([`AutoencoderKL`]):
+            Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
+        text_encoder ([`~transformers.CLIPTextModel`]):
+            Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
+        text_encoder_2 ([`~transformers.CLIPTextModelWithProjection`]):
+            Second frozen text-encoder
+            ([laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k)).
+        tokenizer ([`~transformers.CLIPTokenizer`]):
+            A `CLIPTokenizer` to tokenize text.
+        tokenizer_2 ([`~transformers.CLIPTokenizer`]):
+            A `CLIPTokenizer` to tokenize text.
+        unet ([`UNet2DConditionModel`]):
+            A `UNet2DConditionModel` to denoise the encoded image latents.
+        controlnet ([`ControlNetModel`] or `List[ControlNetModel]`):
+            Provides additional conditioning to the `unet` during the denoising process. If you set multiple
+            ControlNets as a list, the outputs from each ControlNet are added together to create one combined
+            additional conditioning.
+        scheduler ([`SchedulerMixin`]):
+            A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+            [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+        force_zeros_for_empty_prompt (`bool`, *optional*, defaults to `"True"`):
+            Whether the negative prompt embeddings should always be set to 0. Also see the config of
+            `stabilityai/stable-diffusion-xl-base-1-0`.
+        add_watermarker (`bool`, *optional*):
+            Whether to use the [invisible_watermark](https://github.com/ShieldMnt/invisible-watermark/) library to
+            watermark output images. If not defined, it defaults to `True` if the package is installed; otherwise no
+            watermarker is used.
+    """
+    model_cpu_offload_seq = (
+        "text_encoder->text_encoder_2->unet->vae"  # leave controlnet out on purpose because it iterates with unet
+    )
+
+    def __init__(
+        self,
+        vae: AutoencoderKL,
+        text_encoder: CLIPTextModel,
+        text_encoder_2: CLIPTextModelWithProjection,
+        tokenizer: CLIPTokenizer,
+        tokenizer_2: CLIPTokenizer,
+        unet: UNet2DConditionModel,
+        controlnet: Union[ControlNetModel, List[ControlNetModel], Tuple[ControlNetModel], MultiControlNetModel],
+        scheduler: KarrasDiffusionSchedulers,
+        force_zeros_for_empty_prompt: bool = True,
+        add_watermarker: Optional[bool] = None,
+    ):
+        super().__init__()
+
+        if isinstance(controlnet, (list, tuple)):
+            controlnet = MultiControlNetModel(controlnet)
+
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            text_encoder_2=text_encoder_2,
+            tokenizer=tokenizer,
+            tokenizer_2=tokenizer_2,
+            unet=unet,
+            controlnet=controlnet,
+            scheduler=scheduler,
+        )
+        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+        self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True)
+        self.control_image_processor = VaeImageProcessor(
+            vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True, do_normalize=False
+        )
+        add_watermarker = add_watermarker if add_watermarker is not None else is_invisible_watermark_available()
+
+        if add_watermarker:
+            self.watermark = StableDiffusionXLWatermarker()
+        else:
+            self.watermark = None
+
+        self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt)
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_slicing
+    def enable_vae_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.vae.enable_slicing()
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_slicing
+    def disable_vae_slicing(self):
+        r"""
+        Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_slicing()
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_tiling
+    def enable_vae_tiling(self):
+        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.vae.enable_tiling()
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_tiling
+    def disable_vae_tiling(self):
+        r"""
+        Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_tiling()
+
+    # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.encode_prompt
+    def encode_prompt(
+        self,
+        prompt: str,
+        prompt_2: Optional[str] = None,
+        device: Optional[torch.device] = None,
+        num_images_per_prompt: int = 1,
+        do_classifier_free_guidance: bool = True,
+        negative_prompt: Optional[str] = None,
+        negative_prompt_2: Optional[str] = None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        lora_scale: Optional[float] = None,
+    ):
+        r"""
+        Encodes the prompt into text encoder hidden states.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                prompt to be encoded
+            prompt_2 (`str` or `List[str]`, *optional*):
+                The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
+                used in both text-encoders
+            device: (`torch.device`):
+                torch device
+            num_images_per_prompt (`int`):
+                number of images that should be generated per prompt
+            do_classifier_free_guidance (`bool`):
+                whether to use classifier free guidance or not
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            negative_prompt_2 (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
+                `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders
+            prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
+                If not provided, pooled text embeddings will be generated from `prompt` input argument.
+            negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
+                input argument.
+            lora_scale (`float`, *optional*):
+                A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+        """
+        device = device or self._execution_device
+
+        # set lora scale so that monkey patched LoRA
+        # function of text encoder can correctly access it
+        if lora_scale is not None and isinstance(self, LoraLoaderMixin):
+            self._lora_scale = lora_scale
+
+            # dynamically adjust the LoRA scale
+            adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
+            adjust_lora_scale_text_encoder(self.text_encoder_2, lora_scale)
+
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        # Define tokenizers and text encoders
+        tokenizers = [self.tokenizer, self.tokenizer_2] if self.tokenizer is not None else [self.tokenizer_2]
+        text_encoders = (
+            [self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2]
+        )
+
+        if prompt_embeds is None:
+            prompt_2 = prompt_2 or prompt
+            # textual inversion: procecss multi-vector tokens if necessary
+            prompt_embeds_list = []
+            prompts = [prompt, prompt_2]
+            for prompt, tokenizer, text_encoder in zip(prompts, tokenizers, text_encoders):
+                if isinstance(self, TextualInversionLoaderMixin):
+                    prompt = self.maybe_convert_prompt(prompt, tokenizer)
+
+                text_inputs = tokenizer(
+                    prompt,
+                    padding="max_length",
+                    max_length=tokenizer.model_max_length,
+                    truncation=True,
+                    return_tensors="pt",
+                )
+
+                text_input_ids = text_inputs.input_ids
+                untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+                if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+                    text_input_ids, untruncated_ids
+                ):
+                    removed_text = tokenizer.batch_decode(untruncated_ids[:, tokenizer.model_max_length - 1 : -1])
+                    logger.warning(
+                        "The following part of your input was truncated because CLIP can only handle sequences up to"
+                        f" {tokenizer.model_max_length} tokens: {removed_text}"
+                    )
+
+                prompt_embeds = text_encoder(
+                    text_input_ids.to(device),
+                    output_hidden_states=True,
+                )
+
+                # We are only ALWAYS interested in the pooled output of the final text encoder
+                pooled_prompt_embeds = prompt_embeds[0]
+                prompt_embeds = prompt_embeds.hidden_states[-2]
+
+                prompt_embeds_list.append(prompt_embeds)
+
+            prompt_embeds = torch.concat(prompt_embeds_list, dim=-1)
+
+        # get unconditional embeddings for classifier free guidance
+        zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt
+        if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt:
+            negative_prompt_embeds = torch.zeros_like(prompt_embeds)
+            negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds)
+        elif do_classifier_free_guidance and negative_prompt_embeds is None:
+            negative_prompt = negative_prompt or ""
+            negative_prompt_2 = negative_prompt_2 or negative_prompt
+
+            uncond_tokens: List[str]
+            if prompt is not None and type(prompt) is not type(negative_prompt):
+                raise TypeError(
+                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+                    f" {type(prompt)}."
+                )
+            elif isinstance(negative_prompt, str):
+                uncond_tokens = [negative_prompt, negative_prompt_2]
+            elif batch_size != len(negative_prompt):
+                raise ValueError(
+                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                    " the batch size of `prompt`."
+                )
+            else:
+                uncond_tokens = [negative_prompt, negative_prompt_2]
+
+            negative_prompt_embeds_list = []
+            for negative_prompt, tokenizer, text_encoder in zip(uncond_tokens, tokenizers, text_encoders):
+                if isinstance(self, TextualInversionLoaderMixin):
+                    negative_prompt = self.maybe_convert_prompt(negative_prompt, tokenizer)
+
+                max_length = prompt_embeds.shape[1]
+                uncond_input = tokenizer(
+                    negative_prompt,
+                    padding="max_length",
+                    max_length=max_length,
+                    truncation=True,
+                    return_tensors="pt",
+                )
+
+                negative_prompt_embeds = text_encoder(
+                    uncond_input.input_ids.to(device),
+                    output_hidden_states=True,
+                )
+                # We are only ALWAYS interested in the pooled output of the final text encoder
+                negative_pooled_prompt_embeds = negative_prompt_embeds[0]
+                negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2]
+
+                negative_prompt_embeds_list.append(negative_prompt_embeds)
+
+            negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1)
+
+        prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device)
+        bs_embed, seq_len, _ = prompt_embeds.shape
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+        if do_classifier_free_guidance:
+            # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+            seq_len = negative_prompt_embeds.shape[1]
+            negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device)
+            negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+            negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+        pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
+            bs_embed * num_images_per_prompt, -1
+        )
+        if do_classifier_free_guidance:
+            negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
+                bs_embed * num_images_per_prompt, -1
+            )
+
+        return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+    def prepare_extra_step_kwargs(self, generator, eta):
+        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+        # and should be between [0, 1]
+
+        accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        extra_step_kwargs = {}
+        if accepts_eta:
+            extra_step_kwargs["eta"] = eta
+
+        # check if the scheduler accepts generator
+        accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        if accepts_generator:
+            extra_step_kwargs["generator"] = generator
+        return extra_step_kwargs
+
+    def check_inputs(
+        self,
+        prompt,
+        prompt_2,
+        image,
+        callback_steps,
+        negative_prompt=None,
+        negative_prompt_2=None,
+        prompt_embeds=None,
+        negative_prompt_embeds=None,
+        pooled_prompt_embeds=None,
+        negative_pooled_prompt_embeds=None,
+        controlnet_conditioning_scale=1.0,
+        control_guidance_start=0.0,
+        control_guidance_end=1.0,
+    ):
+        if (callback_steps is None) or (
+            callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+        ):
+            raise ValueError(
+                f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+                f" {type(callback_steps)}."
+            )
+
+        if prompt is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt_2 is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt is None and prompt_embeds is None:
+            raise ValueError(
+                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+            )
+        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+        elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)):
+            raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}")
+
+        if negative_prompt is not None and negative_prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+            )
+        elif negative_prompt_2 is not None and negative_prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:"
+                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+            )
+
+        if prompt_embeds is not None and negative_prompt_embeds is not None:
+            if prompt_embeds.shape != negative_prompt_embeds.shape:
+                raise ValueError(
+                    "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+                    f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+                    f" {negative_prompt_embeds.shape}."
+                )
+
+        if prompt_embeds is not None and pooled_prompt_embeds is None:
+            raise ValueError(
+                "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`."
+            )
+
+        if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None:
+            raise ValueError(
+                "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`."
+            )
+
+        # `prompt` needs more sophisticated handling when there are multiple
+        # conditionings.
+        if isinstance(self.controlnet, MultiControlNetModel):
+            if isinstance(prompt, list):
+                logger.warning(
+                    f"You have {len(self.controlnet.nets)} ControlNets and you have passed {len(prompt)}"
+                    " prompts. The conditionings will be fixed across the prompts."
+                )
+
+        # Check `image`
+        is_compiled = hasattr(F, "scaled_dot_product_attention") and isinstance(
+            self.controlnet, torch._dynamo.eval_frame.OptimizedModule
+        )
+        if (
+            isinstance(self.controlnet, ControlNetModel)
+            or is_compiled
+            and isinstance(self.controlnet._orig_mod, ControlNetModel)
+        ):
+            self.check_image(image, prompt, prompt_embeds)
+        elif (
+            isinstance(self.controlnet, MultiControlNetModel)
+            or is_compiled
+            and isinstance(self.controlnet._orig_mod, MultiControlNetModel)
+        ):
+            if not isinstance(image, list):
+                raise TypeError("For multiple controlnets: `image` must be type `list`")
+
+            # When `image` is a nested list:
+            # (e.g. [[canny_image_1, pose_image_1], [canny_image_2, pose_image_2]])
+            elif any(isinstance(i, list) for i in image):
+                raise ValueError("A single batch of multiple conditionings are supported at the moment.")
+            elif len(image) != len(self.controlnet.nets):
+                raise ValueError(
+                    f"For multiple controlnets: `image` must have the same length as the number of controlnets, but got {len(image)} images and {len(self.controlnet.nets)} ControlNets."
+                )
+
+            for image_ in image:
+                self.check_image(image_, prompt, prompt_embeds)
+        else:
+            assert False
+
+        # Check `controlnet_conditioning_scale`
+        if (
+            isinstance(self.controlnet, ControlNetModel)
+            or is_compiled
+            and isinstance(self.controlnet._orig_mod, ControlNetModel)
+        ):
+            if not isinstance(controlnet_conditioning_scale, float):
+                raise TypeError("For single controlnet: `controlnet_conditioning_scale` must be type `float`.")
+        elif (
+            isinstance(self.controlnet, MultiControlNetModel)
+            or is_compiled
+            and isinstance(self.controlnet._orig_mod, MultiControlNetModel)
+        ):
+            if isinstance(controlnet_conditioning_scale, list):
+                if any(isinstance(i, list) for i in controlnet_conditioning_scale):
+                    raise ValueError("A single batch of multiple conditionings are supported at the moment.")
+            elif isinstance(controlnet_conditioning_scale, list) and len(controlnet_conditioning_scale) != len(
+                self.controlnet.nets
+            ):
+                raise ValueError(
+                    "For multiple controlnets: When `controlnet_conditioning_scale` is specified as `list`, it must have"
+                    " the same length as the number of controlnets"
+                )
+        else:
+            assert False
+
+        if not isinstance(control_guidance_start, (tuple, list)):
+            control_guidance_start = [control_guidance_start]
+
+        if not isinstance(control_guidance_end, (tuple, list)):
+            control_guidance_end = [control_guidance_end]
+
+        if len(control_guidance_start) != len(control_guidance_end):
+            raise ValueError(
+                f"`control_guidance_start` has {len(control_guidance_start)} elements, but `control_guidance_end` has {len(control_guidance_end)} elements. Make sure to provide the same number of elements to each list."
+            )
+
+        if isinstance(self.controlnet, MultiControlNetModel):
+            if len(control_guidance_start) != len(self.controlnet.nets):
+                raise ValueError(
+                    f"`control_guidance_start`: {control_guidance_start} has {len(control_guidance_start)} elements but there are {len(self.controlnet.nets)} controlnets available. Make sure to provide {len(self.controlnet.nets)}."
+                )
+
+        for start, end in zip(control_guidance_start, control_guidance_end):
+            if start >= end:
+                raise ValueError(
+                    f"control guidance start: {start} cannot be larger or equal to control guidance end: {end}."
+                )
+            if start < 0.0:
+                raise ValueError(f"control guidance start: {start} can't be smaller than 0.")
+            if end > 1.0:
+                raise ValueError(f"control guidance end: {end} can't be larger than 1.0.")
+
+    # Copied from diffusers.pipelines.controlnet.pipeline_controlnet.StableDiffusionControlNetPipeline.check_image
+    def check_image(self, image, prompt, prompt_embeds):
+        image_is_pil = isinstance(image, PIL.Image.Image)
+        image_is_tensor = isinstance(image, torch.Tensor)
+        image_is_np = isinstance(image, np.ndarray)
+        image_is_pil_list = isinstance(image, list) and isinstance(image[0], PIL.Image.Image)
+        image_is_tensor_list = isinstance(image, list) and isinstance(image[0], torch.Tensor)
+        image_is_np_list = isinstance(image, list) and isinstance(image[0], np.ndarray)
+
+        if (
+            not image_is_pil
+            and not image_is_tensor
+            and not image_is_np
+            and not image_is_pil_list
+            and not image_is_tensor_list
+            and not image_is_np_list
+        ):
+            raise TypeError(
+                f"image must be passed and be one of PIL image, numpy array, torch tensor, list of PIL images, list of numpy arrays or list of torch tensors, but is {type(image)}"
+            )
+
+        if image_is_pil:
+            image_batch_size = 1
+        else:
+            image_batch_size = len(image)
+
+        if prompt is not None and isinstance(prompt, str):
+            prompt_batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            prompt_batch_size = len(prompt)
+        elif prompt_embeds is not None:
+            prompt_batch_size = prompt_embeds.shape[0]
+
+        if image_batch_size != 1 and image_batch_size != prompt_batch_size:
+            raise ValueError(
+                f"If image batch size is not 1, image batch size must be same as prompt batch size. image batch size: {image_batch_size}, prompt batch size: {prompt_batch_size}"
+            )
+
+    # Copied from diffusers.pipelines.controlnet.pipeline_controlnet.StableDiffusionControlNetPipeline.prepare_image
+    def prepare_image(
+        self,
+        image,
+        width,
+        height,
+        batch_size,
+        num_images_per_prompt,
+        device,
+        dtype,
+        do_classifier_free_guidance=False,
+        guess_mode=False,
+    ):
+        image = self.control_image_processor.preprocess(image, height=height, width=width).to(dtype=torch.float32)
+        image_batch_size = image.shape[0]
+
+        if image_batch_size == 1:
+            repeat_by = batch_size
+        else:
+            # image batch size is the same as prompt batch size
+            repeat_by = num_images_per_prompt
+
+        image = image.repeat_interleave(repeat_by, dim=0)
+
+        image = image.to(device=device, dtype=dtype)
+
+        if do_classifier_free_guidance and not guess_mode:
+            image = torch.cat([image] * 2)
+
+        return image
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
+    def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
+        shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
+        if isinstance(generator, list) and len(generator) != batch_size:
+            raise ValueError(
+                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+            )
+
+        if latents is None:
+            latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+        else:
+            latents = latents.to(device)
+
+        # scale the initial noise by the standard deviation required by the scheduler
+        latents = latents * self.scheduler.init_noise_sigma
+        return latents
+
+    # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline._get_add_time_ids
+    def _get_add_time_ids(self, original_size, crops_coords_top_left, target_size, dtype):
+        add_time_ids = list(original_size + crops_coords_top_left + target_size)
+
+        passed_add_embed_dim = (
+            self.unet.config.addition_time_embed_dim * len(add_time_ids) + self.text_encoder_2.config.projection_dim
+        )
+        expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features
+
+        if expected_add_embed_dim != passed_add_embed_dim:
+            raise ValueError(
+                f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`."
+            )
+
+        add_time_ids = torch.tensor([add_time_ids], dtype=dtype)
+        return add_time_ids
+
+    def get_views(self, height, width, window_size=128, stride=64, random_jitter=False):
+        # Here, we define the mappings F_i (see Eq. 7 in the MultiDiffusion paper https://arxiv.org/abs/2302.08113)
+        # if panorama's height/width < window_size, num_blocks of height/width should return 1
+        height //= self.vae_scale_factor
+        width //= self.vae_scale_factor
+        num_blocks_height = int((height - window_size) / stride - 1e-6) + 2 if height > window_size else 1
+        num_blocks_width = int((width - window_size) / stride - 1e-6) + 2 if width > window_size else 1
+        total_num_blocks = int(num_blocks_height * num_blocks_width)
+        views = []
+        for i in range(total_num_blocks):
+            h_start = int((i // num_blocks_width) * stride)
+            h_end = h_start + window_size
+            w_start = int((i % num_blocks_width) * stride)
+            w_end = w_start + window_size
+
+            if h_end > height:
+                h_start = int(h_start + height - h_end)
+                h_end = int(height)
+            if w_end > width:
+                w_start = int(w_start + width - w_end)
+                w_end = int(width)
+            if h_start < 0:
+                h_end = int(h_end - h_start)
+                h_start = 0
+            if w_start < 0:
+                w_end = int(w_end - w_start)
+                w_start = 0
+
+            if random_jitter:
+                jitter_range = (window_size - stride) // 4
+                w_jitter = 0
+                h_jitter = 0
+                if (w_start != 0) and (w_end != width):
+                    w_jitter = random.randint(-jitter_range, jitter_range)
+                elif (w_start == 0) and (w_end != width):
+                    w_jitter = random.randint(-jitter_range, 0)
+                elif (w_start != 0) and (w_end == width):
+                    w_jitter = random.randint(0, jitter_range)
+                if (h_start != 0) and (h_end != height):
+                    h_jitter = random.randint(-jitter_range, jitter_range)
+                elif (h_start == 0) and (h_end != height):
+                    h_jitter = random.randint(-jitter_range, 0)
+                elif (h_start != 0) and (h_end == height):
+                    h_jitter = random.randint(0, jitter_range)
+                h_start += (h_jitter + jitter_range)
+                h_end += (h_jitter + jitter_range)
+                w_start += (w_jitter + jitter_range)
+                w_end += (w_jitter + jitter_range)
+            
+            views.append((h_start, h_end, w_start, w_end))
+        return views
+
+    def tiled_decode(self, latents, current_height, current_width):
+        sample_size = self.unet.config.sample_size
+        core_size = self.unet.config.sample_size // 4
+        core_stride = core_size
+        pad_size = self.unet.config.sample_size // 8 * 3
+        decoder_view_batch_size = 1
+        
+        if self.lowvram:
+            core_stride = core_size // 2
+            pad_size = core_size
+
+        views = self.get_views(current_height, current_width, stride=core_stride, window_size=core_size)
+        views_batch = [views[i : i + decoder_view_batch_size] for i in range(0, len(views), decoder_view_batch_size)]
+        latents_ = F.pad(latents, (pad_size, pad_size, pad_size, pad_size), 'constant', 0)
+        image = torch.zeros(latents.size(0), 3, current_height, current_width).to(latents.device)
+        count = torch.zeros_like(image).to(latents.device)
+        # get the latents corresponding to the current view coordinates
+        with self.progress_bar(total=len(views_batch)) as progress_bar:
+            for j, batch_view in enumerate(views_batch):
+                vb_size = len(batch_view)
+                latents_for_view = torch.cat(
+                    [
+                        latents_[:, :, h_start:h_end+pad_size*2, w_start:w_end+pad_size*2]
+                        for h_start, h_end, w_start, w_end in batch_view
+                    ]
+                ).to(self.vae.device)
+                image_patch = self.vae.decode(latents_for_view / self.vae.config.scaling_factor, return_dict=False)[0]
+                h_start, h_end, w_start, w_end = views[j]
+                h_start, h_end, w_start, w_end = h_start * self.vae_scale_factor, h_end * self.vae_scale_factor, w_start * self.vae_scale_factor, w_end * self.vae_scale_factor
+                p_h_start, p_h_end, p_w_start, p_w_end = pad_size * self.vae_scale_factor, image_patch.size(2) - pad_size * self.vae_scale_factor, pad_size * self.vae_scale_factor, image_patch.size(3) - pad_size * self.vae_scale_factor
+                image[:, :, h_start:h_end, w_start:w_end] += image_patch[:, :, p_h_start:p_h_end, p_w_start:p_w_end].to(latents.device)
+                count[:, :, h_start:h_end, w_start:w_end] += 1
+                progress_bar.update()
+        image = image / count
+        
+        return image
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_upscale.StableDiffusionUpscalePipeline.upcast_vae
+    def upcast_vae(self):
+        dtype = self.vae.dtype
+        self.vae.to(dtype=torch.float32)
+        use_torch_2_0_or_xformers = isinstance(
+            self.vae.decoder.mid_block.attentions[0].processor,
+            (
+                AttnProcessor2_0,
+                XFormersAttnProcessor,
+                LoRAXFormersAttnProcessor,
+                LoRAAttnProcessor2_0,
+            ),
+        )
+        # if xformers or torch_2_0 is used attention block does not need
+        # to be in float32 which can save lots of memory
+        if use_torch_2_0_or_xformers:
+            self.vae.post_quant_conv.to(dtype)
+            self.vae.decoder.conv_in.to(dtype)
+            self.vae.decoder.mid_block.to(dtype)
+
+    @torch.no_grad()
+    @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+        self,
+        prompt: Union[str, List[str]] = None,
+        prompt_2: Optional[Union[str, List[str]]] = None,
+        condition_image: PipelineImageInput = None,
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        num_inference_steps: int = 50,
+        guidance_scale: float = 5.0,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        negative_prompt_2: Optional[Union[str, List[str]]] = None,
+        num_images_per_prompt: Optional[int] = 1,
+        eta: float = 0.0,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+        callback_steps: int = 1,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        controlnet_conditioning_scale: Union[float, List[float]] = 1.0,
+        guess_mode: bool = False,
+        control_guidance_start: Union[float, List[float]] = 0.0,
+        control_guidance_end: Union[float, List[float]] = 1.0,
+        original_size: Tuple[int, int] = None,
+        crops_coords_top_left: Tuple[int, int] = (0, 0),
+        target_size: Tuple[int, int] = None,
+        negative_original_size: Optional[Tuple[int, int]] = None,
+        negative_crops_coords_top_left: Tuple[int, int] = (0, 0),
+        negative_target_size: Optional[Tuple[int, int]] = None,
+        ################### DemoFusion specific parameters ####################
+        image_lr: Optional[torch.FloatTensor] = None,
+        view_batch_size: int = 16,
+        multi_decoder: bool = True,
+        stride: Optional[int] = 64,
+        cosine_scale_1: Optional[float] = 3.,
+        cosine_scale_2: Optional[float] = 1.,
+        cosine_scale_3: Optional[float] = 1.,
+        sigma: Optional[float] = 1.0,
+        show_image: bool = False,
+        lowvram: bool = False,
+    ):
+        r"""
+        The call function to the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
+            prompt_2 (`str` or `List[str]`, *optional*):
+                The prompt or prompts to be sent to `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
+                used in both text-encoders.
+            image (`torch.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.FloatTensor]`, `List[PIL.Image.Image]`, `List[np.ndarray]`,:
+                    `List[List[torch.FloatTensor]]`, `List[List[np.ndarray]]` or `List[List[PIL.Image.Image]]`):
+                The ControlNet input condition to provide guidance to the `unet` for generation. If the type is
+                specified as `torch.FloatTensor`, it is passed to ControlNet as is. `PIL.Image.Image` can also be
+                accepted as an image. The dimensions of the output image defaults to `image`'s dimensions. If height
+                and/or width are passed, `image` is resized accordingly. If multiple ControlNets are specified in
+                `init`, images must be passed as a list such that each element of the list can be correctly batched for
+                input to a single ControlNet.
+            height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+                The height in pixels of the generated image. Anything below 512 pixels won't work well for
+                [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0)
+                and checkpoints that are not specifically fine-tuned on low resolutions.
+            width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+                The width in pixels of the generated image. Anything below 512 pixels won't work well for
+                [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0)
+                and checkpoints that are not specifically fine-tuned on low resolutions.
+            num_inference_steps (`int`, *optional*, defaults to 50):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            guidance_scale (`float`, *optional*, defaults to 5.0):
+                A higher guidance scale value encourages the model to generate images closely linked to the text
+                `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide what to not include in image generation. If not defined, you need to
+                pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+            negative_prompt_2 (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide what to not include in image generation. This is sent to `tokenizer_2`
+                and `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders.
+            num_images_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            eta (`float`, *optional*, defaults to 0.0):
+                Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+                to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+                generation deterministic.
+            latents (`torch.FloatTensor`, *optional*):
+                Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor is generated by sampling using the supplied random `generator`.
+            prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+                provided, text embeddings are generated from the `prompt` input argument.
+            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
+                not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+            pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated pooled text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
+                not provided, pooled text embeddings are generated from `prompt` input argument.
+            negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs (prompt
+                weighting). If not provided, pooled `negative_prompt_embeds` are generated from `negative_prompt` input
+                argument.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+                plain tuple.
+            callback (`Callable`, *optional*):
+                A function that calls every `callback_steps` steps during inference. The function is called with the
+                following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+            callback_steps (`int`, *optional*, defaults to 1):
+                The frequency at which the `callback` function is called. If not specified, the callback is called at
+                every step.
+            cross_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
+                [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            controlnet_conditioning_scale (`float` or `List[float]`, *optional*, defaults to 1.0):
+                The outputs of the ControlNet are multiplied by `controlnet_conditioning_scale` before they are added
+                to the residual in the original `unet`. If multiple ControlNets are specified in `init`, you can set
+                the corresponding scale as a list.
+            guess_mode (`bool`, *optional*, defaults to `False`):
+                The ControlNet encoder tries to recognize the content of the input image even if you remove all
+                prompts. A `guidance_scale` value between 3.0 and 5.0 is recommended.
+            control_guidance_start (`float` or `List[float]`, *optional*, defaults to 0.0):
+                The percentage of total steps at which the ControlNet starts applying.
+            control_guidance_end (`float` or `List[float]`, *optional*, defaults to 1.0):
+                The percentage of total steps at which the ControlNet stops applying.
+            original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+                If `original_size` is not the same as `target_size` the image will appear to be down- or upsampled.
+                `original_size` defaults to `(width, height)` if not specified. Part of SDXL's micro-conditioning as
+                explained in section 2.2 of
+                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+            crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
+                `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position
+                `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting
+                `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of
+                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+            target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+                For most cases, `target_size` should be set to the desired height and width of the generated image. If
+                not specified it will default to `(width, height)`. Part of SDXL's micro-conditioning as explained in
+                section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+            negative_original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+                To negatively condition the generation process based on a specific image resolution. Part of SDXL's
+                micro-conditioning as explained in section 2.2 of
+                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
+                information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
+            negative_crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
+                To negatively condition the generation process based on a specific crop coordinates. Part of SDXL's
+                micro-conditioning as explained in section 2.2 of
+                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
+                information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
+            negative_target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+                To negatively condition the generation process based on a target image resolution. It should be as same
+                as the `target_size` for most cases. Part of SDXL's micro-conditioning as explained in section 2.2 of
+                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
+                information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
+            ################### DemoFusion specific parameters ####################
+            image_lr (`torch.FloatTensor`, *optional*, , defaults to None):
+                Low-resolution image input for upscaling. If provided, DemoFusion will encode it as the initial latent representation.
+            view_batch_size (`int`, defaults to 16):
+                The batch size for multiple denoising paths. Typically, a larger batch size can result in higher 
+                efficiency but comes with increased GPU memory requirements.
+            multi_decoder (`bool`, defaults to True):
+                Determine whether to use a tiled decoder. Generally, when the resolution exceeds 3072x3072, 
+                a tiled decoder becomes necessary.
+            stride (`int`, defaults to 64):
+                The stride of moving local patches. A smaller stride is better for alleviating seam issues,
+                but it also introduces additional computational overhead and inference time.
+            cosine_scale_1 (`float`, defaults to 3):
+                Control the strength of skip-residual. For specific impacts, please refer to Appendix C
+                in the DemoFusion paper.
+            cosine_scale_2 (`float`, defaults to 1):
+                Control the strength of dilated sampling. For specific impacts, please refer to Appendix C
+                in the DemoFusion paper.
+            cosine_scale_3 (`float`, defaults to 1):
+                Control the strength of the gaussion filter. For specific impacts, please refer to Appendix C
+                in the DemoFusion paper.
+            sigma (`float`, defaults to 1):
+                The standard value of the gaussian filter.
+            show_image (`bool`, defaults to False):
+                Determine whether to show intermediate results during generation.
+            lowvram (`bool`, defaults to False):
+                Try to fit in 8 Gb of VRAM, with xformers installed.
+        Examples:
+
+        Returns:
+            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+                If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned,
+                otherwise a `tuple` is returned containing the output images.
+        """
+        controlnet = self.controlnet._orig_mod if is_compiled_module(self.controlnet) else self.controlnet
+
+        # align format for control guidance
+        if not isinstance(control_guidance_start, list) and isinstance(control_guidance_end, list):
+            control_guidance_start = len(control_guidance_end) * [control_guidance_start]
+        elif not isinstance(control_guidance_end, list) and isinstance(control_guidance_start, list):
+            control_guidance_end = len(control_guidance_start) * [control_guidance_end]
+        elif not isinstance(control_guidance_start, list) and not isinstance(control_guidance_end, list):
+            mult = len(controlnet.nets) if isinstance(controlnet, MultiControlNetModel) else 1
+            control_guidance_start, control_guidance_end = mult * [control_guidance_start], mult * [
+                control_guidance_end
+            ]
+
+        # 0. Default height and width to unet
+        height = height or self.unet.config.sample_size * self.vae_scale_factor
+        width = width or self.unet.config.sample_size * self.vae_scale_factor
+
+        x1_size = self.unet.config.sample_size * self.vae_scale_factor
+
+        height_scale = height / x1_size
+        width_scale = width / x1_size
+        scale_num = int(max(height_scale, width_scale))
+        aspect_ratio = min(height_scale, width_scale) / max(height_scale, width_scale)
+
+        original_size = original_size or (height, width)
+        target_size = target_size or (height, width)
+        
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(
+            prompt,
+            prompt_2,
+            condition_image,
+            callback_steps,
+            negative_prompt,
+            negative_prompt_2,
+            prompt_embeds,
+            negative_prompt_embeds,
+            pooled_prompt_embeds,
+            negative_pooled_prompt_embeds,
+            controlnet_conditioning_scale,
+            control_guidance_start,
+            control_guidance_end,
+        )
+
+        # 2. Define call parameters
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        device = self._execution_device
+        self.lowvram = lowvram
+        if self.lowvram:
+            self.vae.cpu()
+            self.unet.cpu()
+            self.text_encoder.to(device)
+            self.text_encoder_2.to(device)
+            image_lr.cpu()
+        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+        # corresponds to doing no classifier free guidance.
+        do_classifier_free_guidance = guidance_scale > 1.0
+
+        if isinstance(controlnet, MultiControlNetModel) and isinstance(controlnet_conditioning_scale, float):
+            controlnet_conditioning_scale = [controlnet_conditioning_scale] * len(controlnet.nets)
+
+        global_pool_conditions = (
+            controlnet.config.global_pool_conditions
+            if isinstance(controlnet, ControlNetModel)
+            else controlnet.nets[0].config.global_pool_conditions
+        )
+        guess_mode = guess_mode or global_pool_conditions
+
+        # 3. Encode input prompt
+        text_encoder_lora_scale = (
+            cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None
+        )
+        (
+            prompt_embeds,
+            negative_prompt_embeds,
+            pooled_prompt_embeds,
+            negative_pooled_prompt_embeds,
+        ) = self.encode_prompt(
+            prompt,
+            prompt_2,
+            device,
+            num_images_per_prompt,
+            do_classifier_free_guidance,
+            negative_prompt,
+            negative_prompt_2,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            pooled_prompt_embeds=pooled_prompt_embeds,
+            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
+            lora_scale=text_encoder_lora_scale,
+        )
+
+        # 4. Prepare image
+        if isinstance(controlnet, ControlNetModel):
+            condition_image = self.prepare_image(
+                image=condition_image,
+                width=width // scale_num,
+                height=height // scale_num,
+                batch_size=batch_size * num_images_per_prompt,
+                num_images_per_prompt=num_images_per_prompt,
+                device=device,
+                dtype=controlnet.dtype,
+                do_classifier_free_guidance=do_classifier_free_guidance,
+                guess_mode=guess_mode,
+            )
+            # height, width = condition_image.shape[-2:]
+            # condition_image.shape ([2, 3, 1024, 1024])
+        elif isinstance(controlnet, MultiControlNetModel):
+            condition_images = []
+
+            for image_ in condition_image:
+                image_ = self.prepare_image(
+                    image=image_,
+                    width=width // scale_num,
+                    height=height // scale_num,
+                    batch_size=batch_size * num_images_per_prompt,
+                    num_images_per_prompt=num_images_per_prompt,
+                    device=device,
+                    dtype=controlnet.dtype,
+                    do_classifier_free_guidance=do_classifier_free_guidance,
+                    guess_mode=guess_mode,
+                )
+
+                condition_images.append(image_)
+
+            condition_image = condition_images
+            # height, width = condition_image[0].shape[-2:]
+        else:
+            assert False
+
+        # 5. Prepare timesteps
+        self.scheduler.set_timesteps(num_inference_steps, device=device)
+        timesteps = self.scheduler.timesteps
+
+        # 6. Prepare latent variables
+        num_channels_latents = self.unet.config.in_channels
+        latents = self.prepare_latents(
+            batch_size * num_images_per_prompt,
+            num_channels_latents,
+            height // scale_num,
+            width // scale_num,
+            prompt_embeds.dtype,
+            device,
+            generator,
+            latents,
+        )
+
+        # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+        # 7.1 Create tensor stating which controlnets to keep
+        controlnet_keep = []
+        for i in range(len(timesteps)):
+            keeps = [
+                1.0 - float(i / len(timesteps) < s or (i + 1) / len(timesteps) > e)
+                for s, e in zip(control_guidance_start, control_guidance_end)
+            ]
+            controlnet_keep.append(keeps[0] if isinstance(controlnet, ControlNetModel) else keeps)
+
+        # 7.2 Prepare added time ids & embeddings
+        if isinstance(condition_image, list):
+            original_size = original_size or condition_image[0].shape[-2:]
+        else:
+            original_size = original_size or condition_image.shape[-2:]
+        target_size = target_size or (height, width)
+
+        add_text_embeds = pooled_prompt_embeds
+        add_time_ids = self._get_add_time_ids(
+            original_size, crops_coords_top_left, target_size, dtype=prompt_embeds.dtype
+        )
+
+        if negative_original_size is not None and negative_target_size is not None:
+            negative_add_time_ids = self._get_add_time_ids(
+                negative_original_size,
+                negative_crops_coords_top_left,
+                negative_target_size,
+                dtype=prompt_embeds.dtype,
+            )
+        else:
+            negative_add_time_ids = add_time_ids
+
+        if do_classifier_free_guidance:
+            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
+            add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0)
+            add_time_ids = torch.cat([negative_add_time_ids, add_time_ids], dim=0)
+
+        prompt_embeds = prompt_embeds.to(device)
+        add_text_embeds = add_text_embeds.to(device)
+        add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1)
+
+        
+
+        # 8. Denoising loop
+        num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+
+        output_images = []
+        
+    ###################################################### Phase Initialization ########################################################
+
+        if self.lowvram:
+            self.text_encoder.cpu()
+            self.text_encoder_2.cpu()
+        
+        if image_lr == None:
+            print("### Phase 1 Denoising ###")
+            with self.progress_bar(total=num_inference_steps) as progress_bar:
+                for i, t in enumerate(timesteps):
+    
+                    if self.lowvram:
+                        self.vae.cpu()
+                        self.unet.to(device)
+                    
+                    latents_for_view = latents
+    
+                    # expand the latents if we are doing classifier free guidance
+                    latent_model_input = (
+                        latents.repeat_interleave(2, dim=0)
+                        if do_classifier_free_guidance
+                        else latents
+                    )
+                    latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+    
+                    
+                    added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids}
+    
+                    # controlnet(s) inference
+                    if guess_mode and do_classifier_free_guidance:
+                        # Infer ControlNet only for the conditional batch.
+                        control_model_input = latents
+                        control_model_input = self.scheduler.scale_model_input(control_model_input, t)
+                        controlnet_prompt_embeds = prompt_embeds.chunk(2)[1]
+                        controlnet_added_cond_kwargs = {
+                            "text_embeds": add_text_embeds.chunk(2)[1],
+                            "time_ids": add_time_ids.chunk(2)[1],
+                        }
+                    else:
+                        control_model_input = latent_model_input
+                        controlnet_prompt_embeds = prompt_embeds
+                        controlnet_added_cond_kwargs = added_cond_kwargs
+    
+                    if isinstance(controlnet_keep[i], list):
+                        cond_scale = [c * s for c, s in zip(controlnet_conditioning_scale, controlnet_keep[i])]
+                    else:
+                        controlnet_cond_scale = controlnet_conditioning_scale
+                        if isinstance(controlnet_cond_scale, list):
+                            controlnet_cond_scale = controlnet_cond_scale[0]
+                        cond_scale = controlnet_cond_scale * controlnet_keep[i]
+    
+                    # print(condition_image.shape, control_model_input.shape, controlnet_prompt_embeds.shape, t, cond_scale, guess_mode)
+                    # print(controlnet_added_cond_kwargs["text_embeds"].shape, controlnet_added_cond_kwargs["time_ids"].shape)
+                    down_block_res_samples, mid_block_res_sample = self.controlnet(
+                        control_model_input,
+                        t,
+                        encoder_hidden_states=controlnet_prompt_embeds,
+                        controlnet_cond=condition_image,
+                        conditioning_scale=cond_scale,
+                        guess_mode=guess_mode,
+                        added_cond_kwargs=controlnet_added_cond_kwargs,
+                        return_dict=False,
+                    )
+    
+                    if guess_mode and do_classifier_free_guidance:
+                        # Infered ControlNet only for the conditional batch.
+                        # To apply the output of ControlNet to both the unconditional and conditional batches,
+                        # add 0 to the unconditional batch to keep it unchanged.
+                        down_block_res_samples = [torch.cat([torch.zeros_like(d), d]) for d in down_block_res_samples]
+                        mid_block_res_sample = torch.cat([torch.zeros_like(mid_block_res_sample), mid_block_res_sample])
+    
+                    # predict the noise residual
+                    noise_pred = self.unet(
+                        latent_model_input,
+                        t,
+                        encoder_hidden_states=prompt_embeds,
+                        cross_attention_kwargs=cross_attention_kwargs,
+                        down_block_additional_residuals=down_block_res_samples,
+                        mid_block_additional_residual=mid_block_res_sample,
+                        added_cond_kwargs=added_cond_kwargs,
+                        return_dict=False,
+                    )[0]
+    
+                    # perform guidance
+                    if do_classifier_free_guidance:
+                        noise_pred_uncond, noise_pred_text = noise_pred[::2], noise_pred[1::2]
+                        noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+    
+                    # compute the previous noisy sample x_t -> x_t-1
+                    latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
+    
+                    # call the callback, if provided
+                    if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                        progress_bar.update()
+                        if callback is not None and i % callback_steps == 0:
+                            step_idx = i // getattr(self.scheduler, "order", 1)
+                            callback(step_idx, t, latents)
+        else:
+            print("### Encoding Real Image ###")
+            latents = self.vae.encode(image_lr)
+            latents = latents.latent_dist.sample() * self.vae.config.scaling_factor
+            
+        anchor_mean = latents.mean()
+        anchor_std = latents.std()
+        if self.lowvram:
+            latents = latents.cpu()
+            torch.cuda.empty_cache()
+        if not output_type == "latent":
+            # make sure the VAE is in float32 mode, as it overflows in float16
+            needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
+
+            if self.lowvram:
+                needs_upcasting = False # use madebyollin/sdxl-vae-fp16-fix in lowvram mode!
+                self.unet.cpu()
+                self.vae.to(device)
+
+            if needs_upcasting:
+                self.upcast_vae()
+                latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
+            if self.lowvram and multi_decoder:
+                current_width_height = self.unet.config.sample_size * self.vae_scale_factor
+                image = self.tiled_decode(latents, current_width_height, current_width_height)
+            else:
+                image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+            # cast back to fp16 if needed
+            if needs_upcasting:
+                self.vae.to(dtype=torch.float16)
+
+        image = self.image_processor.postprocess(image, output_type=output_type)
+        if show_image:
+            plt.figure(figsize=(10, 10))
+            plt.imshow(image[0])
+            plt.axis('off')  # Turn off axis numbers and ticks
+            plt.show()
+        output_images.append(image[0])
+                        
+    ####################################################### Phase Upscaling #####################################################
+        if image_lr == None:
+            starting_scale = 2
+        else:
+            starting_scale = 1
+        for current_scale_num in range(starting_scale, scale_num + 1):
+            if self.lowvram:
+                latents = latents.to(device)
+                self.unet.to(device)
+                torch.cuda.empty_cache()
+            print("### Phase {} Denoising ###".format(current_scale_num))
+            current_height = self.unet.config.sample_size * self.vae_scale_factor * current_scale_num
+            current_width = self.unet.config.sample_size * self.vae_scale_factor * current_scale_num
+            if height > width:
+                current_width = int(current_width * aspect_ratio)
+            else:
+                current_height = int(current_height * aspect_ratio)
+        
+            latents = F.interpolate(latents, size=(int(current_height / self.vae_scale_factor), int(current_width / self.vae_scale_factor)), mode='bicubic')
+            condition_image = F.interpolate(condition_image, size=(current_height, current_width), mode='bicubic')
+    
+            noise_latents = []
+            noise = torch.randn_like(latents)
+            for timestep in timesteps:
+                noise_latent = self.scheduler.add_noise(latents, noise, timestep.unsqueeze(0))
+                noise_latents.append(noise_latent)
+            latents = noise_latents[0]
+    
+            with self.progress_bar(total=num_inference_steps) as progress_bar:
+                for i, t in enumerate(timesteps):
+                    count = torch.zeros_like(latents)
+                    value = torch.zeros_like(latents)
+                    cosine_factor = 0.5 * (1 + torch.cos(torch.pi * (self.scheduler.config.num_train_timesteps - t) / self.scheduler.config.num_train_timesteps)).cpu()
+
+                    c1 = cosine_factor ** cosine_scale_1
+                    latents = latents * (1 - c1) + noise_latents[i] * c1
+                    
+                    ############################################# MultiDiffusion #############################################
+                    
+                    views = self.get_views(current_height, current_width, stride=stride, window_size=self.unet.config.sample_size, random_jitter=True)
+                    views_batch = [views[i : i + view_batch_size] for i in range(0, len(views), view_batch_size)]
+
+                    jitter_range = (self.unet.config.sample_size - stride) // 4
+                    latents_ = F.pad(latents, (jitter_range, jitter_range, jitter_range, jitter_range), 'constant', 0)
+                    condition_image_ = F.pad(condition_image, (jitter_range * self.vae_scale_factor, jitter_range * self.vae_scale_factor, jitter_range * self.vae_scale_factor, jitter_range * self.vae_scale_factor), 'constant', 0)
+                  
+                    count_local = torch.zeros_like(latents_)
+                    value_local = torch.zeros_like(latents_)
+                    
+                    for j, batch_view in enumerate(views_batch):
+                        vb_size = len(batch_view)
+
+                        # get the latents corresponding to the current view coordinates
+                        latents_for_view = torch.cat(
+                            [
+                                latents_[:, :, h_start:h_end, w_start:w_end]
+                                for h_start, h_end, w_start, w_end in batch_view
+                            ]
+                        )
+                        condition_image_for_view = torch.cat(
+                            [
+                                condition_image_[0:1, :, h_start * self.vae_scale_factor:h_end * self.vae_scale_factor, w_start * self.vae_scale_factor:w_end * self.vae_scale_factor]
+                                for h_start, h_end, w_start, w_end in batch_view
+                            ]
+                        )
+
+                        # expand the latents if we are doing classifier free guidance
+                        latent_model_input = latents_for_view
+                        latent_model_input = (
+                            latent_model_input.repeat_interleave(2, dim=0)
+                            if do_classifier_free_guidance
+                            else latent_model_input
+                        )
+                        latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+                        condition_image_input = condition_image_for_view
+                        condition_image_input = (
+                            condition_image_input.repeat_interleave(2, dim=0)
+                            if do_classifier_free_guidance
+                            else condition_image_input
+                        )
+
+                        prompt_embeds_input = torch.cat([prompt_embeds] * vb_size)
+                        add_text_embeds_input = torch.cat([add_text_embeds] * vb_size)
+                        add_time_ids_input = []
+                        for h_start, h_end, w_start, w_end in batch_view:
+                            add_time_ids_ = add_time_ids.clone()
+                            add_time_ids_[:, 2] = h_start * self.vae_scale_factor
+                            add_time_ids_[:, 3] = w_start * self.vae_scale_factor
+                            add_time_ids_input.append(add_time_ids_)
+                        add_time_ids_input = torch.cat(add_time_ids_input)
+
+                        added_cond_kwargs = {"text_embeds": add_text_embeds_input, "time_ids": add_time_ids_input}
+
+                        # controlnet(s) inference
+                        if guess_mode and do_classifier_free_guidance:
+                            # Infer ControlNet only for the conditional batch.
+                            control_model_input = latent_model_input
+                            control_model_input = self.scheduler.scale_model_input(control_model_input, t)
+                            controlnet_prompt_embeds = prompt_embeds_input.chunk(2)[1]
+                            controlnet_added_cond_kwargs = {
+                                "text_embeds": add_text_embeds_input.chunk(2)[1],
+                                "time_ids": add_time_ids_input.chunk(2)[1],
+                            }
+                        else:
+                            control_model_input = latent_model_input
+                            controlnet_prompt_embeds = prompt_embeds_input
+                            controlnet_added_cond_kwargs = added_cond_kwargs
+        
+                        if isinstance(controlnet_keep[i], list):
+                            cond_scale = [c * s for c, s in zip(controlnet_conditioning_scale, controlnet_keep[i])]
+                        else:
+                            controlnet_cond_scale = controlnet_conditioning_scale
+                            if isinstance(controlnet_cond_scale, list):
+                                controlnet_cond_scale = controlnet_cond_scale[0]
+                            cond_scale = controlnet_cond_scale * controlnet_keep[i]
+
+                        down_block_res_samples, mid_block_res_sample = self.controlnet(
+                            control_model_input,
+                            t,
+                            encoder_hidden_states=controlnet_prompt_embeds,
+                            controlnet_cond=condition_image_input,
+                            conditioning_scale=cond_scale,
+                            guess_mode=guess_mode,
+                            added_cond_kwargs=controlnet_added_cond_kwargs,
+                            return_dict=False,
+                        )
+        
+                        if guess_mode and do_classifier_free_guidance:
+                            # Infered ControlNet only for the conditional batch.
+                            # To apply the output of ControlNet to both the unconditional and conditional batches,
+                            # add 0 to the unconditional batch to keep it unchanged.
+                            down_block_res_samples = [torch.cat([torch.zeros_like(d), d]) for d in down_block_res_samples]
+                            mid_block_res_sample = torch.cat([torch.zeros_like(mid_block_res_sample), mid_block_res_sample])
+
+                        # predict the noise residual
+                        noise_pred = self.unet(
+                            latent_model_input,
+                            t,
+                            encoder_hidden_states=prompt_embeds_input,
+                            cross_attention_kwargs=cross_attention_kwargs,
+                            down_block_additional_residuals=down_block_res_samples,
+                            mid_block_additional_residual=mid_block_res_sample,
+                            added_cond_kwargs=added_cond_kwargs,
+                            return_dict=False,
+                        )[0]
+
+                        if do_classifier_free_guidance:
+                            noise_pred_uncond, noise_pred_text = noise_pred[::2], noise_pred[1::2]
+                            noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) * 1
+
+                        # compute the previous noisy sample x_t -> x_t-1
+                        self.scheduler._init_step_index(t)
+                        latents_denoised_batch = self.scheduler.step(
+                            noise_pred, t, latents_for_view, **extra_step_kwargs, return_dict=False)[0]
+
+                        # extract value from batch
+                        for latents_view_denoised, (h_start, h_end, w_start, w_end) in zip(
+                            latents_denoised_batch.chunk(vb_size), batch_view
+                        ):
+                            value_local[:, :, h_start:h_end, w_start:w_end] += latents_view_denoised
+                            count_local[:, :, h_start:h_end, w_start:w_end] += 1
+
+                    value_local = value_local[: ,:, jitter_range: jitter_range + current_height // self.vae_scale_factor, jitter_range: jitter_range + current_width // self.vae_scale_factor]
+                    count_local = count_local[: ,:, jitter_range: jitter_range + current_height // self.vae_scale_factor, jitter_range: jitter_range + current_width // self.vae_scale_factor]
+                    
+                    c2 = cosine_factor ** cosine_scale_2
+
+                    value += value_local / count_local * (1 - c2)
+                    count += torch.ones_like(value_local) * (1 - c2)
+                        
+                    ############################################# Dilated Sampling #############################################
+
+                    h_pad = (current_scale_num - (latents.size(2) % current_scale_num)) % current_scale_num
+                    w_pad = (current_scale_num - (latents.size(3) % current_scale_num)) % current_scale_num
+                    latents_ = F.pad(latents, (w_pad, 0, h_pad, 0), 'constant', 0)
+                    
+                    count_global = torch.zeros_like(latents_)
+                    value_global = torch.zeros_like(latents_)
+                            
+                    c3 = 0.99 * cosine_factor ** cosine_scale_3 + 1e-2
+                    std_, mean_ = latents_.std(), latents_.mean()
+                    latents_gaussian = gaussian_filter(latents_, kernel_size=(2*current_scale_num-1), sigma=sigma*c3)
+                    latents_gaussian = (latents_gaussian - latents_gaussian.mean()) / latents_gaussian.std() * std_ + mean_
+                            
+                    latents_for_view = []
+                    for h in range(current_scale_num):
+                        for w in range(current_scale_num):
+                            latents_for_view.append(latents_[:, :, h::current_scale_num, w::current_scale_num])
+                    latents_for_view = torch.cat(latents_for_view)
+                        
+                    latents_for_view_gaussian = []
+                    for h in range(current_scale_num):
+                        for w in range(current_scale_num):
+                            latents_for_view_gaussian.append(latents_gaussian[:, :, h::current_scale_num, w::current_scale_num])
+                    latents_for_view_gaussian = torch.cat(latents_for_view_gaussian)
+
+                    condition_image_for_view = []
+                    for h in range(current_scale_num):
+                        for w in range(current_scale_num):
+                            condition_image_ = F.pad(condition_image, (w_pad * self.vae_scale_factor, w * self.vae_scale_factor, h_pad * self.vae_scale_factor, h * self.vae_scale_factor), 'constant', 0)
+                            condition_image_for_view.append(condition_image_[0:1, :, h * self.vae_scale_factor::current_scale_num, w * self.vae_scale_factor::current_scale_num])
+                    condition_image_for_view = torch.cat(condition_image_for_view)
+                    
+                    vb_size = latents_for_view.size(0)
+
+                    # expand the latents if we are doing classifier free guidance
+                    latent_model_input = latents_for_view_gaussian
+                    latent_model_input = (
+                        latent_model_input.repeat_interleave(2, dim=0)
+                        if do_classifier_free_guidance
+                        else latent_model_input
+                    )
+                    latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+                    condition_image_input = condition_image_for_view
+                    condition_image_input = (
+                        condition_image_input.repeat_interleave(2, dim=0)
+                        if do_classifier_free_guidance
+                        else condition_image_input
+                    )
+
+                    prompt_embeds_input = torch.cat([prompt_embeds] * vb_size)
+                    add_text_embeds_input = torch.cat([add_text_embeds] * vb_size)
+                    add_time_ids_input = torch.cat([add_time_ids] * vb_size)
+
+                    added_cond_kwargs = {"text_embeds": add_text_embeds_input, "time_ids": add_time_ids_input}
+
+                    # controlnet(s) inference
+                    if guess_mode and do_classifier_free_guidance:
+                        # Infer ControlNet only for the conditional batch.
+                        control_model_input = latent_model_input
+                        control_model_input = self.scheduler.scale_model_input(control_model_input, t)
+                        controlnet_prompt_embeds = prompt_embeds_input.chunk(2)[1]
+                        controlnet_added_cond_kwargs = {
+                            "text_embeds": add_text_embeds_input.chunk(2)[1],
+                            "time_ids": add_time_ids_input.chunk(2)[1],
+                        }
+                    else:
+                        control_model_input = latent_model_input
+                        controlnet_prompt_embeds = prompt_embeds_input
+                        controlnet_added_cond_kwargs = added_cond_kwargs
+    
+                    if isinstance(controlnet_keep[i], list):
+                        cond_scale = [c * s for c, s in zip(controlnet_conditioning_scale, controlnet_keep[i])]
+                    else:
+                        controlnet_cond_scale = controlnet_conditioning_scale
+                        if isinstance(controlnet_cond_scale, list):
+                            controlnet_cond_scale = controlnet_cond_scale[0]
+                        cond_scale = controlnet_cond_scale * controlnet_keep[i]
+    
+                    down_block_res_samples, mid_block_res_sample = self.controlnet(
+                        control_model_input,
+                        t,
+                        encoder_hidden_states=controlnet_prompt_embeds,
+                        controlnet_cond=condition_image_input,
+                        conditioning_scale=cond_scale,
+                        guess_mode=guess_mode,
+                        added_cond_kwargs=controlnet_added_cond_kwargs,
+                        return_dict=False,
+                    )
+    
+                    if guess_mode and do_classifier_free_guidance:
+                        # Infered ControlNet only for the conditional batch.
+                        # To apply the output of ControlNet to both the unconditional and conditional batches,
+                        # add 0 to the unconditional batch to keep it unchanged.
+                        down_block_res_samples = [torch.cat([torch.zeros_like(d), d]) for d in down_block_res_samples]
+                        mid_block_res_sample = torch.cat([torch.zeros_like(mid_block_res_sample), mid_block_res_sample])
+
+                    # predict the noise residual
+                    noise_pred = self.unet(
+                        latent_model_input,
+                        t,
+                        encoder_hidden_states=prompt_embeds_input,
+                        cross_attention_kwargs=cross_attention_kwargs,
+                        down_block_additional_residuals=down_block_res_samples,
+                        mid_block_additional_residual=mid_block_res_sample,
+                        added_cond_kwargs=added_cond_kwargs,
+                        return_dict=False,
+                    )[0]
+
+                    if do_classifier_free_guidance:
+                        noise_pred_uncond, noise_pred_text = noise_pred[::2], noise_pred[1::2]
+                        noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+                    # extract value from batch
+                    for h in range(current_scale_num):
+                        for w in range(current_scale_num):
+                            noise_pred_ = noise_pred.chunk(vb_size)[h*current_scale_num+w]
+                            value_global[:, :, h::current_scale_num, w::current_scale_num] += noise_pred_
+                            count_global[:, :, h::current_scale_num, w::current_scale_num] += 1
+                    
+                    # compute the previous noisy sample x_t -> x_t-1
+                    self.scheduler._init_step_index(t)
+                    value_global = self.scheduler.step(
+                        value_global, t, latents_, **extra_step_kwargs, return_dict=False)[0]
+
+                    c2 = cosine_factor ** cosine_scale_2
+
+                    value_global = value_global[: ,:, h_pad:, w_pad:]
+    
+                    value += value_global * c2
+                    count += torch.ones_like(value_global) * c2
+                    
+                           ###########################################################   
+                
+                    latents = torch.where(count > 0, value / count, value)
+                
+                    # call the callback, if provided
+                    if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                        progress_bar.update()
+                        if callback is not None and i % callback_steps == 0:
+                            step_idx = i // getattr(self.scheduler, "order", 1)
+                            callback(step_idx, t, latents)
+
+    #########################################################################################################################################
+        
+                latents = (latents - latents.mean()) / latents.std() * anchor_std + anchor_mean
+                if self.lowvram:
+                    latents = latents.cpu()
+                    torch.cuda.empty_cache()
+                if not output_type == "latent":
+                    # make sure the VAE is in float32 mode, as it overflows in float16
+                    needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
+
+                    if self.lowvram:
+                        needs_upcasting = False # use madebyollin/sdxl-vae-fp16-fix in lowvram mode!
+                        self.unet.cpu()
+                        self.vae.to(device)
+                    
+                    if needs_upcasting:
+                        self.upcast_vae()
+                        latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
+                    
+                    print("### Phase {} Decoding ###".format(current_scale_num))
+                    if multi_decoder:
+                        image = self.tiled_decode(latents, current_height, current_width)
+                    else:
+                        image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+        
+                    # cast back to fp16 if needed
+                    if needs_upcasting:
+                        self.vae.to(dtype=torch.float16)
+                else:
+                    image = latents
+        
+                if not output_type == "latent":
+                    image = self.image_processor.postprocess(image, output_type=output_type)
+                    if show_image:
+                        plt.figure(figsize=(10, 10))
+                        plt.imshow(image[0])
+                        plt.axis('off')  # Turn off axis numbers and ticks
+                        plt.show()
+                    output_images.append(image[0])
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        return output_images
+
+    # 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]], **kwargs):
+        # 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.
+
+        # Remove any existing hooks.
+        if is_accelerate_available() and is_accelerate_version(">=", "0.17.0.dev0"):
+            from accelerate.hooks import AlignDevicesHook, CpuOffload, remove_hook_from_module
+        else:
+            raise ImportError("Offloading requires `accelerate v0.17.0` or higher.")
+
+        is_model_cpu_offload = False
+        is_sequential_cpu_offload = False
+        recursive = False
+        for _, component in self.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."
+                    )
+                    recursive = is_sequential_cpu_offload
+                    remove_hook_from_module(component, recurse=recursive)
+        state_dict, network_alphas = self.lora_state_dict(
+            pretrained_model_name_or_path_or_dict,
+            unet_config=self.unet.config,
+            **kwargs,
+        )
+        self.load_lora_into_unet(state_dict, network_alphas=network_alphas, unet=self.unet)
+
+        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,
+            )
+
+        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,
+            )
+
+        # Offload back.
+        if is_model_cpu_offload:
+            self.enable_model_cpu_offload()
+        elif is_sequential_cpu_offload:
+            self.enable_sequential_cpu_offload()
+
+    @classmethod
+    def save_lora_weights(
+        self,
+        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,
+    ):
+        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"))
+
+        self.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):
+        self._remove_text_encoder_monkey_patch_classmethod(self.text_encoder)
+        self._remove_text_encoder_monkey_patch_classmethod(self.text_encoder_2)

requirements.txt

@@ -0,0 +1,11 @@
+diffusers~=0.21.4
+torch~=2.1.0
+scipy~=1.11.3
+omegaconf~=2.3.0
+accelerate~=0.23.0
+transformers~=4.34.0
+tqdm
+einops
+matplotlib
+gradio
+gradio_imageslider