Merge pull request #24 from longgui0318/main

* diffusers 0.26 support . from https://github.com/levihsu/OOTDiffusion/issues/44#issuecomment-1968365131

---------

Co-authored-by: gubei <>

README.md

@@ -45,6 +45,8 @@ conda install cuda-nvcc -c nvidia
 ```
 
 ![](./assets/tutorial_windows_install.png)
+![](./assets/tutorial_windows_install_1.png)
+![](./assets/tutorial_windows_install_2.png)
 ![](./assets/tutorial_windows_install_3.png)
 
 根据自己系统选择 Windows 10 SDK / Windows 11 SDK.
@@ -64,38 +66,15 @@ vcvars64.bat
 
 编译完成，成功启动。
 
-## diffusers 版本
-
-`main` 分支锁定 diffusers==0.24
-
-`diffusers-0.26` 分支锁定 diffusers==0.26.x
-
-要切换分支，请使用下面命令：
-
-```
-git switch diffusers-0.26
-```
-
-并重新安装依赖：
-
-```
-pip install --force-reinstall -r custom_nodes/ComfyUI-OOTDiffusion/requirements.txt
-```
-
 ## FAQ 常见错误
 
-```
-fatal error: cuda_runtime.h: No such file or directory compilation terminated.
-ninja: build stopped: subcommand failed.
-```
+> fatal error: cuda_runtime.h: No such file or directory compilation terminated. ninja: build stopped: subcommand failed.
+>
+> 解决办法：`conda install cuda-toolkit=12.1 -c nvidia`
 
-解决办法：`conda install cuda-toolkit=12.1 -c nvidia` 并覆写 `CUDA_HOME` `CUDA_PATH` 环境变量
-
-```
-subprocess.CalledProcessError: Command '['where', 'cl']' returned non-zero exit status 1.
-```
-
-解决办法：仅在 Windows 下出现，根据 [Windows 配置教程](#Windows-指南)
+> subprocess.CalledProcessError: Command '['where', 'cl']' returned non-zero exit status 1.
+>
+> 解决办法：仅在 Windows 下出现，根据 [Windows 配置教程](#Windows-指南)
 
 ## Node 节点
 
@@ -105,8 +84,6 @@ Load OOTDiffusion from Hub: 从 huggingface 自动下载并加载 OOTDiffusion P
 
 OOTDiffusion Generate: 生成图像
 
-参数：
-
     cfg: 输出图像和输入衣服的贴合程度
 
 ## Example image 示例图片
@@ -117,16 +94,16 @@ Full body 全身: [模特](./assets/model_fullbody_1.png) [裤子](./assets/clot
 
 Full body 裙子: [模特](./assets/model_dress_1.png) [裙子](./assets/cloth_dress_1.jpg)
 
+## Detail 细节
+
+目前此项目只是对 OOTDiffusion 的功能做了个简单的迁移。
+OOTDiffusion 本体依赖于 `diffusers==0.24.0` 实现，所以假如有其他节点的依赖冲突是没办法解决的（本就不该依赖 diffusers）。
+靠 vendor 也能解决，所以也不是大问题。
+
+在 `Ubuntu 22.02` / `Python 3.10.x` 下可以正常运行。Windows 没有测试过。
+
 ## 更新日志 Release Note
 
-2024-03-14:
-
-添加 `diffusers-0.26` 分支
-
-2024-03-10:
-
-添加 humanparsing onnx 支持
-
 2024-03-04:
 
 添加 Full body 模型

pipelines_ootd/transformer_garm_2d.py

@@ -26,7 +26,7 @@ from diffusers.configuration_utils import ConfigMixin, register_to_config
 from diffusers.models.embeddings import ImagePositionalEmbeddings
 from diffusers.utils import USE_PEFT_BACKEND, BaseOutput, deprecate
 # from diffusers.models.attention import BasicTransformerBlock
-from diffusers.models.embeddings import CaptionProjection, PatchEmbed
+from diffusers.models.embeddings import PixArtAlphaTextProjection, PatchEmbed
 from diffusers.models.lora import LoRACompatibleConv, LoRACompatibleLinear
 from diffusers.models.modeling_utils import ModelMixin
 from diffusers.models.normalization import AdaLayerNormSingle
@@ -237,7 +237,7 @@ class Transformer2DModel(ModelMixin, ConfigMixin):
 
         self.caption_projection = None
         if caption_channels is not None:
-            self.caption_projection = CaptionProjection(in_features=caption_channels, hidden_size=inner_dim)
+            self.caption_projection = PixArtAlphaTextProjection(in_features=caption_channels, hidden_size=inner_dim)
 
         self.gradient_checkpointing = False
 

pipelines_ootd/transformer_vton_2d.py

@@ -26,7 +26,7 @@ from diffusers.configuration_utils import ConfigMixin, register_to_config
 from diffusers.models.embeddings import ImagePositionalEmbeddings
 from diffusers.utils import USE_PEFT_BACKEND, BaseOutput, deprecate
 # from diffusers.models.attention import BasicTransformerBlock
-from diffusers.models.embeddings import CaptionProjection, PatchEmbed
+from diffusers.models.embeddings import PixArtAlphaTextProjection, PatchEmbed
 from diffusers.models.lora import LoRACompatibleConv, LoRACompatibleLinear
 from diffusers.models.modeling_utils import ModelMixin
 from diffusers.models.normalization import AdaLayerNormSingle
@@ -237,7 +237,7 @@ class Transformer2DModel(ModelMixin, ConfigMixin):
 
         self.caption_projection = None
         if caption_channels is not None:
-            self.caption_projection = CaptionProjection(in_features=caption_channels, hidden_size=inner_dim)
+            self.caption_projection = PixArtAlphaTextProjection(in_features=caption_channels, hidden_size=inner_dim)
 
         self.gradient_checkpointing = False
 

pipelines_ootd/unet_garm_2d_condition.py

@@ -44,7 +44,7 @@ from diffusers.models.embeddings import (
     ImageHintTimeEmbedding,
     ImageProjection,
     ImageTimeEmbedding,
-    PositionNet,
+    GLIGENTextBoundingboxProjection,
     TextImageProjection,
     TextImageTimeEmbedding,
     TextTimeEmbedding,
@@ -624,7 +624,7 @@ class UNetGarm2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMi
                 positive_len = cross_attention_dim[0]
 
             feature_type = "text-only" if attention_type == "gated" else "text-image"
-            self.position_net = PositionNet(
+            self.position_net = GLIGENTextBoundingboxProjection(
                 positive_len=positive_len, out_dim=cross_attention_dim, feature_type=feature_type
             )
 

pipelines_ootd/unet_vton_2d_condition.py

@@ -44,7 +44,7 @@ from diffusers.models.embeddings import (
     ImageHintTimeEmbedding,
     ImageProjection,
     ImageTimeEmbedding,
-    PositionNet,
+    GLIGENTextBoundingboxProjection,
     TextImageProjection,
     TextImageTimeEmbedding,
     TextTimeEmbedding,
@@ -624,7 +624,7 @@ class UNetVton2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMi
                 positive_len = cross_attention_dim[0]
 
             feature_type = "text-only" if attention_type == "gated" else "text-image"
-            self.position_net = PositionNet(
+            self.position_net = GLIGENTextBoundingboxProjection(
                 positive_len=positive_len, out_dim=cross_attention_dim, feature_type=feature_type
             )
 

requirements.txt

@@ -5,7 +5,7 @@ scipy
 scikit-image
 opencv-python
 pillow
-diffusers==0.24.0
+diffusers>=0.26.0
 transformers
 accelerate
 matplotlib