2410

README.md

@@ -7,14 +7,18 @@ pipeline_tag: text-to-image
 
 *XS Size, Excess Quality*
 
-![promo](promo.png)
+![promo](media/promo.png)
 
 At AiArtLab, we strive to create a free, compact and fast model that can be trained on consumer graphics cards.
 
-- 1.5b UNet
-- Qwen3-0.6b text encoder
-- 16ch Simple VAE, which preserves details and anatomy.
-- The model was trained ~3 month on 4xRTX5090 on approximately 1+ million images with various resolutions and styles, including anime and realistic photos.
+- Model: 1.5b parameters
+- Text encoder: Qwen3 (0.6B parameters)
+- VAE: 8x16ch, [Simple VAE](https://huggingface.co/AiArtLab/simplevae)
+
+
+The model was trained for ~3 months on 4xRTX 5090 GPUs using over 1 million images of various resolutions and styles, primarily anime and illustrations.
+
+[Gradio](https://sdxs.aiartlab.org/)
 
 ### Example
 
@@ -69,4 +73,4 @@ BTC: 3JHv9Hb8kEW8zMAccdgCdZGfrHeMhH1rpN
 
 ## More examples
 
-![result_grid](result_grid.png)
+![result_grid](media/result_grid.jpg)

pipeline_sdxs.py

@@ -102,9 +102,9 @@ class SdxsPipeline(DiffusionPipeline):
     def generate_latents(
         self,
         text_embeddings,
-        height: int = 576,
-        width: int = 576,
-        num_inference_steps: int = 40,
+        height: int = 640,
+        width: int = 640,
+        num_inference_steps: int = 50,
         guidance_scale: float = 5.0,
         latent_channels: int = 16,
         batch_size: int = 1,
@@ -192,9 +192,9 @@ class SdxsPipeline(DiffusionPipeline):
     def __call__(
         self,
         prompt: Optional[Union[str, List[str]]] = None,
-        height: int = 576,
-        width: int = 576,
-        num_inference_steps: int = 40,
+        height: int = 640,
+        width: int = 640,
+        num_inference_steps: int = 50,
         guidance_scale: float = 5.0,
         latent_channels: int = 16,
         output_type: str = "pil",

test.ipynb

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:79e3d9e6dd5f0879cb0aa79d4cd97dd23352c9d807e9f4a593ebc234322e668c
-size 5563216
+oid sha256:3da8c81f59608e0732fe9708b4cc2becf77ea6fa215601d0f30c057110e85b3c
+size 5563222

train.py

@@ -31,7 +31,7 @@ project = "unet"
 batch_size = 48
 base_learning_rate = 5e-5
 min_learning_rate = 1e-5
-num_epochs = 50
+num_epochs = 40
 # samples/save per epoch
 sample_interval_share = 3
 use_wandb = True

unet/diffusion_pytorch_model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:1443f2abd54e2b73c333d8556c0dc008a2f08f7a5e2823a553347f1eebc3cd61
+oid sha256:0fd75e6dcdd3c39a84c5927b5b77f6cb8591b8feb0a9996aeca9f336b84f9d89
 size 6184944280