TroubleDz commited on
Commit
184f291
1 Parent(s): 398f48b

Upload folder using huggingface_hub

Browse files
.gitattributes CHANGED
@@ -33,3 +33,10 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
 
 
 
 
 
 
 
 
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
36
+ demo.ipynb filter=lfs diff=lfs merge=lfs -text
37
+ figures/gradio_demo.png filter=lfs diff=lfs merge=lfs -text
38
+ figures/gradio_demo_controlnet.png filter=lfs diff=lfs merge=lfs -text
39
+ figures/gradio_demo_controlnet_img2img.png filter=lfs diff=lfs merge=lfs -text
40
+ figures/gradio_demo_img2img.png filter=lfs diff=lfs merge=lfs -text
41
+ figures/progressive_process.jpg filter=lfs diff=lfs merge=lfs -text
42
+ output_example.png filter=lfs diff=lfs merge=lfs -text
.idea/.gitignore ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ # 默认忽略的文件
2
+ /shelf/
3
+ /workspace.xml
.idea/DemoFusion-main.iml ADDED
@@ -0,0 +1,12 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ <?xml version="1.0" encoding="UTF-8"?>
2
+ <module type="PYTHON_MODULE" version="4">
3
+ <component name="NewModuleRootManager">
4
+ <content url="file://$MODULE_DIR$" />
5
+ <orderEntry type="jdk" jdkName="Ai_mode (4)" jdkType="Python SDK" />
6
+ <orderEntry type="sourceFolder" forTests="false" />
7
+ </component>
8
+ <component name="PyDocumentationSettings">
9
+ <option name="format" value="GOOGLE" />
10
+ <option name="myDocStringFormat" value="Google" />
11
+ </component>
12
+ </module>
.idea/inspectionProfiles/Project_Default.xml ADDED
@@ -0,0 +1,48 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ <component name="InspectionProjectProfileManager">
2
+ <profile version="1.0">
3
+ <option name="myName" value="Project Default" />
4
+ <inspection_tool class="PyPackageRequirementsInspection" enabled="true" level="WARNING" enabled_by_default="true">
5
+ <option name="ignoredPackages">
6
+ <value>
7
+ <list size="28">
8
+ <item index="0" class="java.lang.String" itemvalue="httpx" />
9
+ <item index="1" class="java.lang.String" itemvalue="gradio" />
10
+ <item index="2" class="java.lang.String" itemvalue="open-clip-torch" />
11
+ <item index="3" class="java.lang.String" itemvalue="PyYAML" />
12
+ <item index="4" class="java.lang.String" itemvalue="xformers" />
13
+ <item index="5" class="java.lang.String" itemvalue="numpy" />
14
+ <item index="6" class="java.lang.String" itemvalue="requests" />
15
+ <item index="7" class="java.lang.String" itemvalue="fsspec" />
16
+ <item index="8" class="java.lang.String" itemvalue="kornia" />
17
+ <item index="9" class="java.lang.String" itemvalue="gradio_client" />
18
+ <item index="10" class="java.lang.String" itemvalue="openai-clip" />
19
+ <item index="11" class="java.lang.String" itemvalue="sentencepiece" />
20
+ <item index="12" class="java.lang.String" itemvalue="wandb" />
21
+ <item index="13" class="java.lang.String" itemvalue="accelerate" />
22
+ <item index="14" class="java.lang.String" itemvalue="uvicorn" />
23
+ <item index="15" class="java.lang.String" itemvalue="urllib3" />
24
+ <item index="16" class="java.lang.String" itemvalue="triton" />
25
+ <item index="17" class="java.lang.String" itemvalue="timm" />
26
+ <item index="18" class="java.lang.String" itemvalue="opencv-python" />
27
+ <item index="19" class="java.lang.String" itemvalue="pandas" />
28
+ <item index="20" class="java.lang.String" itemvalue="tqdm" />
29
+ <item index="21" class="java.lang.String" itemvalue="pytorch-lightning" />
30
+ <item index="22" class="java.lang.String" itemvalue="fastapi" />
31
+ <item index="23" class="java.lang.String" itemvalue="einops-exts" />
32
+ <item index="24" class="java.lang.String" itemvalue="ninja" />
33
+ <item index="25" class="java.lang.String" itemvalue="matplotlib" />
34
+ <item index="26" class="java.lang.String" itemvalue="webdataset" />
35
+ <item index="27" class="java.lang.String" itemvalue="Pillow" />
36
+ </list>
37
+ </value>
38
+ </option>
39
+ </inspection_tool>
40
+ <inspection_tool class="PyUnresolvedReferencesInspection" enabled="true" level="WARNING" enabled_by_default="true">
41
+ <option name="ignoredIdentifiers">
42
+ <list>
43
+ <option value="JSX2511" />
44
+ </list>
45
+ </option>
46
+ </inspection_tool>
47
+ </profile>
48
+ </component>
.idea/inspectionProfiles/profiles_settings.xml ADDED
@@ -0,0 +1,6 @@
 
 
 
 
 
 
 
1
+ <component name="InspectionProjectProfileManager">
2
+ <settings>
3
+ <option name="USE_PROJECT_PROFILE" value="false" />
4
+ <version value="1.0" />
5
+ </settings>
6
+ </component>
.idea/misc.xml ADDED
@@ -0,0 +1,4 @@
 
 
 
 
 
1
+ <?xml version="1.0" encoding="UTF-8"?>
2
+ <project version="4">
3
+ <component name="ProjectRootManager" version="2" project-jdk-name="Ai_mode (4)" project-jdk-type="Python SDK" />
4
+ </project>
.idea/modules.xml ADDED
@@ -0,0 +1,8 @@
 
 
 
 
 
 
 
 
 
1
+ <?xml version="1.0" encoding="UTF-8"?>
2
+ <project version="4">
3
+ <component name="ProjectModuleManager">
4
+ <modules>
5
+ <module fileurl="file://$PROJECT_DIR$/.idea/DemoFusion-main.iml" filepath="$PROJECT_DIR$/.idea/DemoFusion-main.iml" />
6
+ </modules>
7
+ </component>
8
+ </project>
.idea/workspace.xml ADDED
@@ -0,0 +1,71 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ <?xml version="1.0" encoding="UTF-8"?>
2
+ <project version="4">
3
+ <component name="AutoImportSettings">
4
+ <option name="autoReloadType" value="SELECTIVE" />
5
+ </component>
6
+ <component name="ChangeListManager">
7
+ <list default="true" id="cf13c2e0-fea8-4b36-9b87-d1a7a68e5205" name="更改" comment="" />
8
+ <option name="SHOW_DIALOG" value="false" />
9
+ <option name="HIGHLIGHT_CONFLICTS" value="true" />
10
+ <option name="HIGHLIGHT_NON_ACTIVE_CHANGELIST" value="false" />
11
+ <option name="LAST_RESOLUTION" value="IGNORE" />
12
+ </component>
13
+ <component name="MarkdownSettingsMigration">
14
+ <option name="stateVersion" value="1" />
15
+ </component>
16
+ <component name="ProjectColorInfo">{
17
+ &quot;associatedIndex&quot;: 2
18
+ }</component>
19
+ <component name="ProjectId" id="2btF3w7MTy2zdOhszEbQETG3Qqf" />
20
+ <component name="ProjectViewState">
21
+ <option name="hideEmptyMiddlePackages" value="true" />
22
+ <option name="showLibraryContents" value="true" />
23
+ </component>
24
+ <component name="PropertiesComponent">{
25
+ &quot;keyToString&quot;: {
26
+ &quot;RunOnceActivity.OpenProjectViewOnStart&quot;: &quot;true&quot;,
27
+ &quot;RunOnceActivity.ShowReadmeOnStart&quot;: &quot;true&quot;,
28
+ &quot;last_opened_file_path&quot;: &quot;E:/AI/reduce_noise/DemoFusion-main&quot;,
29
+ &quot;settings.editor.selected.configurable&quot;: &quot;com.jetbrains.python.configuration.PyActiveSdkModuleConfigurable&quot;
30
+ }
31
+ }</component>
32
+ <component name="RunManager">
33
+ <configuration name="gradio_demo_img2img" type="PythonConfigurationType" factoryName="Python" temporary="true" nameIsGenerated="true">
34
+ <module name="DemoFusion-main" />
35
+ <option name="INTERPRETER_OPTIONS" value="" />
36
+ <option name="PARENT_ENVS" value="true" />
37
+ <envs>
38
+ <env name="PYTHONUNBUFFERED" value="1" />
39
+ </envs>
40
+ <option name="SDK_HOME" value="" />
41
+ <option name="WORKING_DIRECTORY" value="$PROJECT_DIR$" />
42
+ <option name="IS_MODULE_SDK" value="true" />
43
+ <option name="ADD_CONTENT_ROOTS" value="true" />
44
+ <option name="ADD_SOURCE_ROOTS" value="true" />
45
+ <option name="SCRIPT_NAME" value="$PROJECT_DIR$/gradio_demo_img2img.py" />
46
+ <option name="PARAMETERS" value="" />
47
+ <option name="SHOW_COMMAND_LINE" value="false" />
48
+ <option name="EMULATE_TERMINAL" value="false" />
49
+ <option name="MODULE_MODE" value="false" />
50
+ <option name="REDIRECT_INPUT" value="false" />
51
+ <option name="INPUT_FILE" value="" />
52
+ <method v="2" />
53
+ </configuration>
54
+ <recent_temporary>
55
+ <list>
56
+ <item itemvalue="Python.gradio_demo_img2img" />
57
+ </list>
58
+ </recent_temporary>
59
+ </component>
60
+ <component name="SpellCheckerSettings" RuntimeDictionaries="0" Folders="0" CustomDictionaries="0" DefaultDictionary="应用程序级" UseSingleDictionary="true" transferred="true" />
61
+ <component name="TaskManager">
62
+ <task active="true" id="Default" summary="默认任务">
63
+ <changelist id="cf13c2e0-fea8-4b36-9b87-d1a7a68e5205" name="更改" comment="" />
64
+ <created>1707026272531</created>
65
+ <option name="number" value="Default" />
66
+ <option name="presentableId" value="Default" />
67
+ <updated>1707026272531</updated>
68
+ </task>
69
+ <servers />
70
+ </component>
71
+ </project>
README.md CHANGED
@@ -1,12 +1,158 @@
1
  ---
2
- title: Dzai
3
- emoji: 🐨
4
- colorFrom: green
5
- colorTo: gray
6
  sdk: gradio
7
- sdk_version: 4.16.0
8
- app_file: app.py
9
- pinned: false
10
  ---
 
 
 
 
 
 
 
11
 
12
- Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
+ title: dzai
3
+ app_file: gradio_demo_img2img.py
 
 
4
  sdk: gradio
5
+ sdk_version: 4.8.0
 
 
6
  ---
7
+ # DemoFusion
8
+ [![Project Page](https://img.shields.io/badge/Project-Page-green.svg)](https://ruoyidu.github.io/demofusion/demofusion.html)
9
+ [![arXiv](https://img.shields.io/badge/arXiv-2311.16973-b31b1b.svg)](https://arxiv.org/pdf/2311.16973.pdf)
10
+ [![Replicate](https://img.shields.io/badge/Demo-%F0%9F%9A%80%20Replicate-blue)](https://replicate.com/lucataco/demofusion)
11
+ [![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)
12
+ [![Hugging Face](https://img.shields.io/badge/i2i-%F0%9F%A4%97%20Hugging%20Face-blue)](https://huggingface.co/spaces/radames/Enhance-This-DemoFusion-SDXL)
13
+ [![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")
14
 
15
+ Code release for "DemoFusion: Democratising High-Resolution Image Generation With No 💰" (arXiv 2023)
16
+
17
+ <img src="figures/illustration.jpg" width="800"/>
18
+
19
+ **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.
20
+
21
+ # News
22
+ - **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.
23
+ - **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.
24
+ - **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!
25
+ - **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!
26
+ - **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!
27
+ - **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!
28
+ - **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!
29
+ - **2023.11.29**: 💰 `pipeline_demofusion_sdxl` is released.
30
+
31
+ # Usage
32
+ ## A quick try with integrated demos
33
+ - 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).
34
+ - 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).
35
+ - 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).
36
+
37
+ ## Starting with our code
38
+ ### Hyper-parameters
39
+ - `view_batch_size` (`int`, defaults to 16):
40
+ The batch size for multiple denoising paths. Typically, a larger batch size can result in higher efficiency but comes with increased GPU memory requirements.
41
+ - `stride` (`int`, defaults to 64):
42
+ 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.
43
+ - `cosine_scale_1` (`float`, defaults to 3):
44
+ 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.
45
+ - `cosine_scale_2` (`float`, defaults to 1):
46
+ 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.
47
+ - `cosine_scale_3` (`float`, defaults to 1):
48
+ 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.
49
+ - `sigma` (`float`, defaults to 1):
50
+ 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.
51
+ - `multi_decoder` (`bool`, defaults to True):
52
+ Determine whether to use a tiled decoder. Generally, a tiled decoder becomes necessary when the resolution exceeds 3072*3072 on an RTX 3090 GPU.
53
+ - `show_image` (`bool`, defaults to False):
54
+ Determine whether to show intermediate results during generation.
55
+
56
+ ### Text2Image (will take about 17 GB of VRAM)
57
+ - Set up the dependencies as:
58
+ ```
59
+ conda create -n demofusion python=3.9
60
+ conda activate demofusion
61
+ pip install -r requirements.txt
62
+ ```
63
+ - Download `pipeline_demofusion_sdxl.py` and run it as follows. A use case can be found in `demo.ipynb`.
64
+ ```
65
+ from pipeline_demofusion_sdxl import DemoFusionSDXLPipeline
66
+ import torch
67
+
68
+ model_ckpt = "stabilityai/stable-diffusion-xl-base-1.0"
69
+ pipe = DemoFusionSDXLPipeline.from_pretrained(model_ckpt, torch_dtype=torch.float16)
70
+ pipe = pipe.to("cuda")
71
+
72
+ 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."
73
+ negative_prompt = "blurry, ugly, duplicate, poorly drawn, deformed, mosaic"
74
+
75
+ images = pipe(prompt, negative_prompt=negative_prompt,
76
+ height=3072, width=3072, view_batch_size=16, stride=64,
77
+ num_inference_steps=50, guidance_scale=7.5,
78
+ cosine_scale_1=3, cosine_scale_2=1, cosine_scale_3=1, sigma=0.8,
79
+ multi_decoder=True, show_image=True
80
+ )
81
+
82
+ for i, image in enumerate(images):
83
+ image.save('image_' + str(i) + '.png')
84
+ ```
85
+ - ⚠️ 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.
86
+ - Please feel free to try different prompts and resolutions.
87
+ - 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.
88
+ - The code was cleaned before the release. If you encounter any issues, please contact us.
89
+
90
+ ### Text2Image on Windows with 8 GB of VRAM
91
+
92
+ - Set up the environment as:
93
+
94
+ ```
95
+ cmd
96
+ git clone "https://github.com/PRIS-CV/DemoFusion"
97
+ cd DemoFusion
98
+ python -m venv venv
99
+ venv\Scripts\activate
100
+ pip install -U "xformers==0.0.22.post7+cu118" --index-url https://download.pytorch.org/whl/cu118
101
+ pip install "diffusers==0.21.4" "matplotlib==3.8.2" "transformers==4.35.2" "accelerate==0.25.0"
102
+ ```
103
+
104
+ - Launch DemoFusion as follows. The use case can be found in `demo_lowvram.py`.
105
+
106
+ ```
107
+ python
108
+ from pipeline_demofusion_sdxl import DemoFusionSDXLPipeline
109
+
110
+ import torch
111
+ from diffusers.models import AutoencoderKL
112
+ vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=torch.float16)
113
+
114
+ model_ckpt = "stabilityai/stable-diffusion-xl-base-1.0"
115
+ pipe = DemoFusionSDXLPipeline.from_pretrained(model_ckpt, torch_dtype=torch.float16, vae=vae)
116
+ pipe = pipe.to("cuda")
117
+
118
+ 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."
119
+ negative_prompt = "blurry, ugly, duplicate, poorly drawn, deformed, mosaic"
120
+
121
+ images = pipe(prompt, negative_prompt=negative_prompt,
122
+ height=2048, width=2048, view_batch_size=4, stride=64,
123
+ num_inference_steps=40, guidance_scale=7.5,
124
+ cosine_scale_1=3, cosine_scale_2=1, cosine_scale_3=1, sigma=0.8,
125
+ multi_decoder=True, show_image=False, lowvram=True
126
+ )
127
+
128
+ for i, image in enumerate(images):
129
+ image.save('image_' + str(i) + '.png')
130
+ ```
131
+ ### Text2Image with local Gradio demo
132
+ - Make sure you have installed `gradio` and `gradio_imageslider`.
133
+ - Launch DemoFusion via Gradio demo now -- try `python gradio_demo.py`! Better Interaction and Presentation!
134
+ <img src="figures/gradio_demo.png" width="600"/>
135
+
136
+ ### Image2Image with local Gradio demo
137
+ - Make sure you have installed `gradio` and `gradio_imageslider`.
138
+ - Launch DemoFusion Image2Image by `python gradio_demo_img2img.py`.
139
+ <img src="figures/gradio_demo_img2img.png" width="600"/>
140
+ - ⚠️ 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.
141
+
142
+ ### DemoFusion+ControlNet with local Gradio demo
143
+ - Make sure you have installed `gradio` and `gradio_imageslider`.
144
+ - Launch DemoFusion+ControNet Text2Image by `python gradio_demo.py`.
145
+ - <img src="figures/gradio_demo_controlnet.png" width="600"/>
146
+ - Launch DemoFusion+ControNet Image2Image by `python gradio_demo_img2img.py`.
147
+ - <img src="figures/gradio_demo_controlnet_img2img.png" width="600"/>
148
+
149
+ ## Citation
150
+ If you find this paper useful in your research, please consider citing:
151
+ ```
152
+ @article{du2023demofusion,
153
+ title={DemoFusion: Democratising High-Resolution Image Generation With No $$$},
154
+ author={Du, Ruoyi and Chang, Dongliang and Hospedales, Timothy and Song, Yi-Zhe and Ma, Zhanyu},
155
+ journal={arXiv preprint arXiv:2311.16973},
156
+ year={2023}
157
+ }
158
+ ```
__pycache__/pipeline_demofusion_sdxl.cpython-311.pyc ADDED
Binary file (77.2 kB). View file
 
demo.ipynb ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a6bbe553656b3d9c863a261a053722930b3b538d5b6b05eac66ff9ae83eaf976
3
+ size 17016845
demo_lowvram.py ADDED
@@ -0,0 +1,34 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ '''
3
+ Installation on Windows for GPU with 8 Gb of VRAM and xformers:
4
+
5
+ git clone "https://github.com/PRIS-CV/DemoFusion"
6
+ cd DemoFusion
7
+ python -m venv venv
8
+ venv\Scripts\activate
9
+ pip install -U "xformers==0.0.22.post7+cu118" --index-url https://download.pytorch.org/whl/cu118
10
+ pip install "diffusers==0.21.4" "matplotlib==3.8.2" "transformers==4.35.2" "accelerate==0.25.0"
11
+ '''
12
+
13
+ from pipeline_demofusion_sdxl import DemoFusionSDXLPipeline
14
+
15
+ import torch
16
+ from diffusers.models import AutoencoderKL
17
+ vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=torch.float16)
18
+
19
+ model_ckpt = "stabilityai/stable-diffusion-xl-base-1.0"
20
+ pipe = DemoFusionSDXLPipeline.from_pretrained(model_ckpt, torch_dtype=torch.float16, vae=vae)
21
+ pipe = pipe.to("cuda")
22
+
23
+ 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."
24
+ negative_prompt = "blurry, ugly, duplicate, poorly drawn, deformed, mosaic"
25
+
26
+ images = pipe(prompt, negative_prompt=negative_prompt,
27
+ height=2048, width=2048, view_batch_size=4, stride=64,
28
+ num_inference_steps=40, guidance_scale=7.5,
29
+ cosine_scale_1=3, cosine_scale_2=1, cosine_scale_3=1, sigma=0.8,
30
+ multi_decoder=True, show_image=False, lowvram=True
31
+ )
32
+
33
+ for i, image in enumerate(images):
34
+ image.save('image_'+str(i)+'.png')
figures/gradio_demo.png ADDED

Git LFS Details

  • SHA256: 705de2ac8f94dc7e361323d930c9bb669a99d7015caa13e9919201abd7b9bddf
  • Pointer size: 132 Bytes
  • Size of remote file: 1.03 MB
figures/gradio_demo_controlnet.png ADDED

Git LFS Details

  • SHA256: 4775087b4a33289b001ed4fbc12a98727a70499c25ca624ca5db004821945473
  • Pointer size: 132 Bytes
  • Size of remote file: 3.81 MB
figures/gradio_demo_controlnet_img2img.png ADDED

Git LFS Details

  • SHA256: 839f11104086fa21ab80109f2bc22d6cd7919b9df5a93f80af4c1463e94385b0
  • Pointer size: 132 Bytes
  • Size of remote file: 4.44 MB
figures/gradio_demo_img2img.png ADDED

Git LFS Details

  • SHA256: 1a60e3bfdede9a8b80855e893b528737a4a854c0ddc17be0444a2d6bd75b0599
  • Pointer size: 132 Bytes
  • Size of remote file: 5.41 MB
figures/illustration.jpg ADDED
figures/progressive_process.jpg ADDED

Git LFS Details

  • SHA256: 14888eb1f7c01fef168e43f1be3aeaeb44bc30c2ff4ee8d394f9eac802a95f7c
  • Pointer size: 132 Bytes
  • Size of remote file: 1.91 MB
gradio_demo.py ADDED
@@ -0,0 +1,46 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import gradio as gr
2
+ from pipeline_demofusion_sdxl import DemoFusionSDXLPipeline
3
+ from gradio_imageslider import ImageSlider
4
+ import torch
5
+
6
+ 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):
7
+ model_ckpt = "stabilityai/stable-diffusion-xl-base-1.0"
8
+ pipe = DemoFusionSDXLPipeline.from_pretrained(model_ckpt, torch_dtype=torch.float16)
9
+ pipe = pipe.to("cuda")
10
+
11
+ generator = torch.Generator(device='cuda')
12
+ generator = generator.manual_seed(int(seed))
13
+
14
+ images = pipe(prompt, negative_prompt=negative_prompt, generator=generator,
15
+ height=int(height), width=int(width), view_batch_size=int(view_batch_size), stride=int(stride),
16
+ num_inference_steps=int(num_inference_steps), guidance_scale=guidance_scale,
17
+ cosine_scale_1=cosine_scale_1, cosine_scale_2=cosine_scale_2, cosine_scale_3=cosine_scale_3, sigma=sigma,
18
+ multi_decoder=True, show_image=False
19
+ )
20
+
21
+ return (images[0], images[-1])
22
+
23
+ iface = gr.Interface(
24
+ fn=generate_images,
25
+ inputs=[
26
+ gr.Textbox(label="Prompt"),
27
+ gr.Textbox(label="Negative Prompt", value="blurry, ugly, duplicate, poorly drawn, deformed, mosaic"),
28
+ gr.Slider(minimum=1024, maximum=4096, step=1024, value=2048, label="Height"),
29
+ gr.Slider(minimum=1024, maximum=4096, step=1024, value=2048, label="Width"),
30
+ gr.Slider(minimum=10, maximum=100, step=1, value=50, label="Num Inference Steps"),
31
+ gr.Slider(minimum=1, maximum=20, step=0.1, value=7.5, label="Guidance Scale"),
32
+ gr.Slider(minimum=0, maximum=5, step=0.1, value=3, label="Cosine Scale 1"),
33
+ gr.Slider(minimum=0, maximum=5, step=0.1, value=1, label="Cosine Scale 2"),
34
+ gr.Slider(minimum=0, maximum=5, step=0.1, value=1, label="Cosine Scale 3"),
35
+ gr.Slider(minimum=0.1, maximum=1, step=0.1, value=0.8, label="Sigma"),
36
+ gr.Slider(minimum=4, maximum=32, step=4, value=16, label="View Batch Size"),
37
+ gr.Slider(minimum=8, maximum=96, step=8, value=64, label="Stride"),
38
+ gr.Number(label="Seed", value=2013)
39
+ ],
40
+ # outputs=gr.Gallery(label="Generated Images"),
41
+ outputs=ImageSlider(label="Comparison of SDXL and DemoFusion"),
42
+ title="DemoFusion Gradio Demo",
43
+ description="Generate images with the DemoFusion SDXL Pipeline."
44
+ )
45
+
46
+ iface.launch()
gradio_demo_controlnet.py ADDED
@@ -0,0 +1,93 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import gradio as gr
2
+ from diffusers import ControlNetModel, AutoencoderKL
3
+ from pipeline_demofusion_sdxl import DemoFusionSDXLPipeline
4
+ from pipeline_demofusion_sdxl_controlnet import DemoFusionSDXLControlNetPipeline
5
+ from gradio_imageslider import ImageSlider
6
+ import torch, gc
7
+ from torchvision import transforms
8
+ from PIL import Image
9
+ import numpy as np
10
+ import cv2
11
+
12
+ def load_and_process_image(pil_image):
13
+ transform = transforms.Compose(
14
+ [
15
+ transforms.Resize((1024, 1024)),
16
+ transforms.ToTensor(),
17
+ transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
18
+ ]
19
+ )
20
+ image = transform(pil_image)
21
+ image = image.unsqueeze(0).half()
22
+ return image
23
+
24
+
25
+ def pad_image(image):
26
+ w, h = image.size
27
+ if w == h:
28
+ return image
29
+ elif w > h:
30
+ new_image = Image.new(image.mode, (w, w), (0, 0, 0))
31
+ pad_w = 0
32
+ pad_h = (w - h) // 2
33
+ new_image.paste(image, (0, pad_h))
34
+ return new_image
35
+ else:
36
+ new_image = Image.new(image.mode, (h, h), (0, 0, 0))
37
+ pad_w = (h - w) // 2
38
+ pad_h = 0
39
+ new_image.paste(image, (pad_w, 0))
40
+ return new_image
41
+
42
+ 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):
43
+ padded_image = pad_image(input_image).resize((1024, 1024)).convert("RGB")
44
+ image_lr = load_and_process_image(padded_image).to('cuda')
45
+ controlnet = ControlNetModel.from_pretrained("diffusers/controlnet-canny-sdxl-1.0", torch_dtype=torch.float16)
46
+ vae = AutoencoderKL.from_pretrained("madebyollin/stable-diffusion-xl-base-1.0/vae-fix", torch_dtype=torch.float16)
47
+ pipe = DemoFusionSDXLControlNetPipeline.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", controlnet=controlnet, vae=vae, torch_dtype=torch.float16)
48
+ pipe = pipe.to("cuda")
49
+ generator = torch.Generator(device='cuda')
50
+ generator = generator.manual_seed(int(seed))
51
+ # get canny image
52
+ canny_image = np.array(padded_image)
53
+ canny_image = cv2.Canny(canny_image, 100, 200)
54
+ canny_image = canny_image[:, :, None]
55
+ canny_image = np.concatenate([canny_image, canny_image, canny_image], axis=2)
56
+ canny_image = Image.fromarray(canny_image)
57
+ images = pipe(prompt, negative_prompt=negative_prompt, controlnet_conditioning_scale=controlnet_conditioning_scale,
58
+ condition_image=canny_image, generator=generator,
59
+ height=int(height), width=int(width), view_batch_size=int(view_batch_size), stride=int(stride),
60
+ num_inference_steps=int(num_inference_steps), guidance_scale=guidance_scale,
61
+ cosine_scale_1=cosine_scale_1, cosine_scale_2=cosine_scale_2, cosine_scale_3=cosine_scale_3, sigma=sigma,
62
+ multi_decoder=True, show_image=False, lowvram=False
63
+ )
64
+ for i, image in enumerate(images):
65
+ image.save('image_'+str(i)+'.png')
66
+ pipe = None
67
+ gc.collect()
68
+ torch.cuda.empty_cache()
69
+ return (canny_image, images[-1])
70
+
71
+ with gr.Blocks(title=f"DemoFusion") as demo:
72
+ with gr.Column():
73
+ with gr.Row():
74
+ with gr.Group():
75
+ image_input = gr.Image(type="pil", label="Input Image")
76
+ prompt = gr.Textbox(label="Prompt", value="")
77
+ negative_prompt = gr.Textbox(label="Negative Prompt", value="blurry, ugly, duplicate, poorly drawn, deformed, mosaic")
78
+ controlnet_conditioning_scale = gr.Slider(minimum=0, maximum=1, step=0.1, value=0.5, label="ControlNet Conditioning Scale")
79
+ width = gr.Slider(minimum=1024, maximum=4096, step=1024, value=2048, label="Width")
80
+ height = gr.Slider(minimum=1024, maximum=4096, step=1024, value=2048, label="Height")
81
+ num_inference_steps = gr.Slider(minimum=10, maximum=100, step=1, value=50, label="Num Inference Steps")
82
+ guidance_scale = gr.Slider(minimum=1, maximum=20, step=0.1, value=7.5, label="Guidance Scale")
83
+ cosine_scale_1 = gr.Slider(minimum=0, maximum=5, step=0.1, value=3, label="Cosine Scale 1")
84
+ cosine_scale_2 = gr.Slider(minimum=0, maximum=5, step=0.1, value=1, label="Cosine Scale 2")
85
+ cosine_scale_3 = gr.Slider(minimum=0, maximum=5, step=0.1, value=1, label="Cosine Scale 3")
86
+ sigma = gr.Slider(minimum=0.1, maximum=1, step=0.1, value=0.8, label="Sigma")
87
+ view_batch_size = gr.Slider(minimum=4, maximum=32, step=4, value=16, label="View Batch Size")
88
+ stride = gr.Slider(minimum=8, maximum=96, step=8, value=64, label="Stride")
89
+ seed = gr.Number(label="Seed", value=2013)
90
+ button = gr.Button()
91
+ output_images = ImageSlider(show_label=False)
92
+ 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)
93
+ demo.queue().launch(inline=False, share=True, debug=True)
gradio_demo_controlnet_img2img.py ADDED
@@ -0,0 +1,93 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import gradio as gr
2
+ from diffusers import ControlNetModel, AutoencoderKL
3
+ from pipeline_demofusion_sdxl import DemoFusionSDXLPipeline
4
+ from pipeline_demofusion_sdxl_controlnet import DemoFusionSDXLControlNetPipeline
5
+ from gradio_imageslider import ImageSlider
6
+ import torch, gc
7
+ from torchvision import transforms
8
+ from PIL import Image
9
+ import numpy as np
10
+ import cv2
11
+
12
+ def load_and_process_image(pil_image):
13
+ transform = transforms.Compose(
14
+ [
15
+ transforms.Resize((1024, 1024)),
16
+ transforms.ToTensor(),
17
+ transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
18
+ ]
19
+ )
20
+ image = transform(pil_image)
21
+ image = image.unsqueeze(0).half()
22
+ return image
23
+
24
+
25
+ def pad_image(image):
26
+ w, h = image.size
27
+ if w == h:
28
+ return image
29
+ elif w > h:
30
+ new_image = Image.new(image.mode, (w, w), (0, 0, 0))
31
+ pad_w = 0
32
+ pad_h = (w - h) // 2
33
+ new_image.paste(image, (0, pad_h))
34
+ return new_image
35
+ else:
36
+ new_image = Image.new(image.mode, (h, h), (0, 0, 0))
37
+ pad_w = (h - w) // 2
38
+ pad_h = 0
39
+ new_image.paste(image, (pad_w, 0))
40
+ return new_image
41
+
42
+ 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):
43
+ padded_image = pad_image(input_image).resize((1024, 1024)).convert("RGB")
44
+ image_lr = load_and_process_image(padded_image).to('cuda')
45
+ controlnet = ControlNetModel.from_pretrained("diffusers/controlnet-canny-sdxl-1.0", torch_dtype=torch.float16)
46
+ vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=torch.float16)
47
+ pipe = DemoFusionSDXLControlNetPipeline.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", controlnet=controlnet, vae=vae, torch_dtype=torch.float16)
48
+ pipe = pipe.to("cuda")
49
+ generator = torch.Generator(device='cuda')
50
+ generator = generator.manual_seed(int(seed))
51
+ # get canny image
52
+ canny_image = np.array(padded_image)
53
+ canny_image = cv2.Canny(canny_image, 100, 200)
54
+ canny_image = canny_image[:, :, None]
55
+ canny_image = np.concatenate([canny_image, canny_image, canny_image], axis=2)
56
+ canny_image = Image.fromarray(canny_image)
57
+ images = pipe(prompt, negative_prompt=negative_prompt, controlnet_conditioning_scale=controlnet_conditioning_scale,
58
+ image_lr=image_lr, condition_image=canny_image, generator=generator,
59
+ height=int(height), width=int(width), view_batch_size=int(view_batch_size), stride=int(stride),
60
+ num_inference_steps=int(num_inference_steps), guidance_scale=guidance_scale,
61
+ cosine_scale_1=cosine_scale_1, cosine_scale_2=cosine_scale_2, cosine_scale_3=cosine_scale_3, sigma=sigma,
62
+ multi_decoder=True, show_image=False, lowvram=False
63
+ )
64
+ for i, image in enumerate(images):
65
+ image.save('image_'+str(i)+'.png')
66
+ pipe = None
67
+ gc.collect()
68
+ torch.cuda.empty_cache()
69
+ return (images[0], images[-1])
70
+
71
+ with gr.Blocks(title=f"DemoFusion") as demo:
72
+ with gr.Column():
73
+ with gr.Row():
74
+ with gr.Group():
75
+ image_input = gr.Image(type="pil", label="Input Image")
76
+ 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")
77
+ negative_prompt = gr.Textbox(label="Negative Prompt", value="blurry, ugly, duplicate, poorly drawn, deformed, mosaic")
78
+ controlnet_conditioning_scale = gr.Slider(minimum=0, maximum=1, step=0.1, value=0.5, label="ControlNet Conditioning Scale")
79
+ width = gr.Slider(minimum=1024, maximum=4096, step=1024, value=2048, label="Width")
80
+ height = gr.Slider(minimum=1024, maximum=4096, step=1024, value=2048, label="Height")
81
+ num_inference_steps = gr.Slider(minimum=10, maximum=100, step=1, value=50, label="Num Inference Steps")
82
+ guidance_scale = gr.Slider(minimum=1, maximum=20, step=0.1, value=7.5, label="Guidance Scale")
83
+ cosine_scale_1 = gr.Slider(minimum=0, maximum=5, step=0.1, value=3, label="Cosine Scale 1")
84
+ cosine_scale_2 = gr.Slider(minimum=0, maximum=5, step=0.1, value=1, label="Cosine Scale 2")
85
+ cosine_scale_3 = gr.Slider(minimum=0, maximum=5, step=0.1, value=1, label="Cosine Scale 3")
86
+ sigma = gr.Slider(minimum=0.1, maximum=1, step=0.1, value=0.8, label="Sigma")
87
+ view_batch_size = gr.Slider(minimum=4, maximum=32, step=4, value=16, label="View Batch Size")
88
+ stride = gr.Slider(minimum=8, maximum=96, step=8, value=64, label="Stride")
89
+ seed = gr.Number(label="Seed", value=2013)
90
+ button = gr.Button()
91
+ output_images = ImageSlider(show_label=False)
92
+ 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)
93
+ demo.queue().launch(inline=False, share=True, debug=True)
gradio_demo_img2img.py ADDED
@@ -0,0 +1,81 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import gradio as gr
2
+ from diffusers import AutoencoderKL
3
+ from pipeline_demofusion_sdxl import DemoFusionSDXLPipeline
4
+ from gradio_imageslider import ImageSlider
5
+ import torch, gc
6
+ from torchvision import transforms
7
+ from PIL import Image
8
+
9
+ def load_and_process_image(pil_image):
10
+ transform = transforms.Compose(
11
+ [
12
+ transforms.Resize((1024, 1024)),
13
+ transforms.ToTensor(),
14
+ transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
15
+ ]
16
+ )
17
+ image = transform(pil_image)
18
+ image = image.unsqueeze(0).half()
19
+ return image
20
+
21
+
22
+ def pad_image(image):
23
+ w, h = image.size
24
+ if w == h:
25
+ return image
26
+ elif w > h:
27
+ new_image = Image.new(image.mode, (w, w), (0, 0, 0))
28
+ pad_w = 0
29
+ pad_h = (w - h) // 2
30
+ new_image.paste(image, (0, pad_h))
31
+ return new_image
32
+ else:
33
+ new_image = Image.new(image.mode, (h, h), (0, 0, 0))
34
+ pad_w = (h - w) // 2
35
+ pad_h = 0
36
+ new_image.paste(image, (pad_w, 0))
37
+ return new_image
38
+
39
+ 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):
40
+ padded_image = pad_image(input_image).resize((1024, 1024)).convert("RGB")
41
+ image_lr = load_and_process_image(padded_image).to('cuda')
42
+ vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=torch.float16)
43
+ pipe = DemoFusionSDXLPipeline.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", vae=vae, torch_dtype=torch.float16)
44
+ pipe = pipe.to("cuda")
45
+ generator = torch.Generator(device='cuda')
46
+ generator = generator.manual_seed(int(seed))
47
+ images = pipe(prompt, negative_prompt=negative_prompt, generator=generator,
48
+ height=int(height), width=int(width), view_batch_size=int(view_batch_size), stride=int(stride),
49
+ num_inference_steps=int(num_inference_steps), guidance_scale=guidance_scale,
50
+ cosine_scale_1=cosine_scale_1, cosine_scale_2=cosine_scale_2, cosine_scale_3=cosine_scale_3, sigma=sigma,
51
+ multi_decoder=True, show_image=False, lowvram=False, image_lr=image_lr
52
+ )
53
+ for i, image in enumerate(images):
54
+ image.save('image_'+str(i)+'.png')
55
+ pipe = None
56
+ gc.collect()
57
+ torch.cuda.empty_cache()
58
+ return (images[0], images[-1])
59
+
60
+ with gr.Blocks(title=f"DemoFusion") as demo:
61
+ with gr.Column():
62
+ with gr.Row():
63
+ with gr.Group():
64
+ image_input = gr.Image(type="pil", label="Input Image")
65
+ 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")
66
+ negative_prompt = gr.Textbox(label="Negative Prompt", value="blurry, ugly, duplicate, poorly drawn, deformed, mosaic")
67
+ width = gr.Slider(minimum=1024, maximum=4096, step=1024, value=2048, label="Width")
68
+ height = gr.Slider(minimum=1024, maximum=4096, step=1024, value=2048, label="Height")
69
+ num_inference_steps = gr.Slider(minimum=5, maximum=100, step=1, value=50, label="Num Inference Steps")
70
+ guidance_scale = gr.Slider(minimum=1, maximum=20, step=0.1, value=7.5, label="Guidance Scale")
71
+ cosine_scale_1 = gr.Slider(minimum=0, maximum=5, step=0.1, value=3, label="Cosine Scale 1")
72
+ cosine_scale_2 = gr.Slider(minimum=0, maximum=5, step=0.1, value=1, label="Cosine Scale 2")
73
+ cosine_scale_3 = gr.Slider(minimum=0, maximum=5, step=0.1, value=1, label="Cosine Scale 3")
74
+ sigma = gr.Slider(minimum=0.1, maximum=1, step=0.1, value=0.8, label="Sigma")
75
+ view_batch_size = gr.Slider(minimum=4, maximum=32, step=4, value=16, label="View Batch Size")
76
+ stride = gr.Slider(minimum=8, maximum=96, step=8, value=64, label="Stride")
77
+ seed = gr.Number(label="Seed", value=2013)
78
+ button = gr.Button()
79
+ output_images = ImageSlider(show_label=False)
80
+ 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)
81
+ demo.queue().launch(inline=False, share=True, debug=True)
output_example.png ADDED

Git LFS Details

  • SHA256: f07ec95e64c728eaf49df80921d0740fed21e07a71aa7be37307b8d94c210aaa
  • Pointer size: 133 Bytes
  • Size of remote file: 10.5 MB
pipeline_demofusion_sdxl.py ADDED
@@ -0,0 +1,1446 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright 2023 The HuggingFace Team. All rights reserved.
2
+ #
3
+ # Licensed under the Apache License, Version 2.0 (the "License");
4
+ # you may not use this file except in compliance with the License.
5
+ # You may obtain a copy of the License at
6
+ #
7
+ # http://www.apache.org/licenses/LICENSE-2.0
8
+ #
9
+ # Unless required by applicable law or agreed to in writing, software
10
+ # distributed under the License is distributed on an "AS IS" BASIS,
11
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12
+ # See the License for the specific language governing permissions and
13
+ # limitations under the License.
14
+
15
+ import inspect
16
+ import os
17
+ from typing import Any, Callable, Dict, List, Optional, Tuple, Union
18
+ import matplotlib.pyplot as plt
19
+
20
+ import torch
21
+ import torch.nn.functional as F
22
+ import numpy as np
23
+ import random
24
+ import warnings
25
+ from transformers import CLIPTextModel, CLIPTextModelWithProjection, CLIPTokenizer
26
+
27
+ from diffusers.image_processor import VaeImageProcessor
28
+ from diffusers.loaders import (
29
+ FromSingleFileMixin,
30
+ LoraLoaderMixin,
31
+ TextualInversionLoaderMixin,
32
+ )
33
+ from diffusers.models import AutoencoderKL, UNet2DConditionModel
34
+ from diffusers.models.attention_processor import (
35
+ AttnProcessor2_0,
36
+ LoRAAttnProcessor2_0,
37
+ LoRAXFormersAttnProcessor,
38
+ XFormersAttnProcessor,
39
+ )
40
+ from diffusers.models.lora import adjust_lora_scale_text_encoder
41
+ from diffusers.schedulers import KarrasDiffusionSchedulers
42
+ from diffusers.utils import (
43
+ is_accelerate_available,
44
+ is_accelerate_version,
45
+ is_invisible_watermark_available,
46
+ logging,
47
+ replace_example_docstring,
48
+ )
49
+ from diffusers.utils.torch_utils import randn_tensor
50
+ from diffusers.pipelines.pipeline_utils import DiffusionPipeline
51
+ from diffusers.pipelines.stable_diffusion_xl import StableDiffusionXLPipelineOutput
52
+
53
+
54
+ if is_invisible_watermark_available():
55
+ from .watermark import StableDiffusionXLWatermarker
56
+
57
+
58
+ logger = logging.get_logger(__name__) # pylint: disable=invalid-name
59
+
60
+ EXAMPLE_DOC_STRING = """
61
+ Examples:
62
+ ```py
63
+ >>> import torch
64
+ >>> from diffusers import StableDiffusionXLPipeline
65
+
66
+ >>> pipe = StableDiffusionXLPipeline.from_pretrained(
67
+ ... "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16
68
+ ... )
69
+ >>> pipe = pipe.to("cuda")
70
+
71
+ >>> prompt = "a photo of an astronaut riding a horse on mars"
72
+ >>> image = pipe(prompt).images[0]
73
+ ```
74
+ """
75
+
76
+ def gaussian_kernel(kernel_size=3, sigma=1.0, channels=3):
77
+ x_coord = torch.arange(kernel_size)
78
+ gaussian_1d = torch.exp(-(x_coord - (kernel_size - 1) / 2) ** 2 / (2 * sigma ** 2))
79
+ gaussian_1d = gaussian_1d / gaussian_1d.sum()
80
+ gaussian_2d = gaussian_1d[:, None] * gaussian_1d[None, :]
81
+ kernel = gaussian_2d[None, None, :, :].repeat(channels, 1, 1, 1)
82
+
83
+ return kernel
84
+
85
+ def gaussian_filter(latents, kernel_size=3, sigma=1.0):
86
+ channels = latents.shape[1]
87
+ kernel = gaussian_kernel(kernel_size, sigma, channels).to(latents.device, latents.dtype)
88
+ blurred_latents = F.conv2d(latents, kernel, padding=kernel_size//2, groups=channels)
89
+
90
+ return blurred_latents
91
+
92
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg
93
+ def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
94
+ """
95
+ Rescale `noise_cfg` according to `guidance_rescale`. Based on findings of [Common Diffusion Noise Schedules and
96
+ Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf). See Section 3.4
97
+ """
98
+ std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True)
99
+ std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True)
100
+ # rescale the results from guidance (fixes overexposure)
101
+ noise_pred_rescaled = noise_cfg * (std_text / std_cfg)
102
+ # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images
103
+ noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg
104
+ return noise_cfg
105
+
106
+
107
+ class DemoFusionSDXLPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoaderMixin, TextualInversionLoaderMixin):
108
+ """
109
+ Pipeline for text-to-image generation using Stable Diffusion XL.
110
+
111
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
112
+ library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
113
+
114
+ In addition the pipeline inherits the following loading methods:
115
+ - *LoRA*: [`StableDiffusionXLPipeline.load_lora_weights`]
116
+ - *Ckpt*: [`loaders.FromSingleFileMixin.from_single_file`]
117
+
118
+ as well as the following saving methods:
119
+ - *LoRA*: [`loaders.StableDiffusionXLPipeline.save_lora_weights`]
120
+
121
+ Args:
122
+ vae ([`AutoencoderKL`]):
123
+ Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
124
+ text_encoder ([`CLIPTextModel`]):
125
+ Frozen text-encoder. Stable Diffusion XL uses the text portion of
126
+ [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
127
+ the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
128
+ text_encoder_2 ([` CLIPTextModelWithProjection`]):
129
+ Second frozen text-encoder. Stable Diffusion XL uses the text and pool portion of
130
+ [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection),
131
+ specifically the
132
+ [laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k)
133
+ variant.
134
+ tokenizer (`CLIPTokenizer`):
135
+ Tokenizer of class
136
+ [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
137
+ tokenizer_2 (`CLIPTokenizer`):
138
+ Second Tokenizer of class
139
+ [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
140
+ unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
141
+ scheduler ([`SchedulerMixin`]):
142
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
143
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
144
+ force_zeros_for_empty_prompt (`bool`, *optional*, defaults to `"True"`):
145
+ Whether the negative prompt embeddings shall be forced to always be set to 0. Also see the config of
146
+ `stabilityai/stable-diffusion-xl-base-1-0`.
147
+ add_watermarker (`bool`, *optional*):
148
+ Whether to use the [invisible_watermark library](https://github.com/ShieldMnt/invisible-watermark/) to
149
+ watermark output images. If not defined, it will default to True if the package is installed, otherwise no
150
+ watermarker will be used.
151
+ """
152
+ model_cpu_offload_seq = "text_encoder->text_encoder_2->unet->vae"
153
+
154
+ def __init__(
155
+ self,
156
+ vae: AutoencoderKL,
157
+ text_encoder: CLIPTextModel,
158
+ text_encoder_2: CLIPTextModelWithProjection,
159
+ tokenizer: CLIPTokenizer,
160
+ tokenizer_2: CLIPTokenizer,
161
+ unet: UNet2DConditionModel,
162
+ scheduler: KarrasDiffusionSchedulers,
163
+ force_zeros_for_empty_prompt: bool = True,
164
+ add_watermarker: Optional[bool] = None,
165
+ ):
166
+ super().__init__()
167
+
168
+ self.register_modules(
169
+ vae=vae,
170
+ text_encoder=text_encoder,
171
+ text_encoder_2=text_encoder_2,
172
+ tokenizer=tokenizer,
173
+ tokenizer_2=tokenizer_2,
174
+ unet=unet,
175
+ scheduler=scheduler,
176
+ )
177
+ self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt)
178
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
179
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
180
+ self.default_sample_size = self.unet.config.sample_size
181
+
182
+ add_watermarker = add_watermarker if add_watermarker is not None else is_invisible_watermark_available()
183
+
184
+ if add_watermarker:
185
+ self.watermark = StableDiffusionXLWatermarker()
186
+ else:
187
+ self.watermark = None
188
+
189
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_slicing
190
+ def enable_vae_slicing(self):
191
+ r"""
192
+ Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
193
+ compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
194
+ """
195
+ self.vae.enable_slicing()
196
+
197
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_slicing
198
+ def disable_vae_slicing(self):
199
+ r"""
200
+ Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to
201
+ computing decoding in one step.
202
+ """
203
+ self.vae.disable_slicing()
204
+
205
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_tiling
206
+ def enable_vae_tiling(self):
207
+ r"""
208
+ Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
209
+ compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
210
+ processing larger images.
211
+ """
212
+ self.vae.enable_tiling()
213
+
214
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_tiling
215
+ def disable_vae_tiling(self):
216
+ r"""
217
+ Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to
218
+ computing decoding in one step.
219
+ """
220
+ self.vae.disable_tiling()
221
+
222
+ def encode_prompt(
223
+ self,
224
+ prompt: str,
225
+ prompt_2: Optional[str] = None,
226
+ device: Optional[torch.device] = None,
227
+ num_images_per_prompt: int = 1,
228
+ do_classifier_free_guidance: bool = True,
229
+ negative_prompt: Optional[str] = None,
230
+ negative_prompt_2: Optional[str] = None,
231
+ prompt_embeds: Optional[torch.FloatTensor] = None,
232
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
233
+ pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
234
+ negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
235
+ lora_scale: Optional[float] = None,
236
+ ):
237
+ r"""
238
+ Encodes the prompt into text encoder hidden states.
239
+
240
+ Args:
241
+ prompt (`str` or `List[str]`, *optional*):
242
+ prompt to be encoded
243
+ prompt_2 (`str` or `List[str]`, *optional*):
244
+ The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
245
+ used in both text-encoders
246
+ device: (`torch.device`):
247
+ torch device
248
+ num_images_per_prompt (`int`):
249
+ number of images that should be generated per prompt
250
+ do_classifier_free_guidance (`bool`):
251
+ whether to use classifier free guidance or not
252
+ negative_prompt (`str` or `List[str]`, *optional*):
253
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
254
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
255
+ less than `1`).
256
+ negative_prompt_2 (`str` or `List[str]`, *optional*):
257
+ The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
258
+ `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders
259
+ prompt_embeds (`torch.FloatTensor`, *optional*):
260
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
261
+ provided, text embeddings will be generated from `prompt` input argument.
262
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
263
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
264
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
265
+ argument.
266
+ pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
267
+ Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
268
+ If not provided, pooled text embeddings will be generated from `prompt` input argument.
269
+ negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
270
+ Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
271
+ weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
272
+ input argument.
273
+ lora_scale (`float`, *optional*):
274
+ A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
275
+ """
276
+ device = device or self._execution_device
277
+
278
+ # set lora scale so that monkey patched LoRA
279
+ # function of text encoder can correctly access it
280
+ if lora_scale is not None and isinstance(self, LoraLoaderMixin):
281
+ self._lora_scale = lora_scale
282
+
283
+ # dynamically adjust the LoRA scale
284
+ adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
285
+ adjust_lora_scale_text_encoder(self.text_encoder_2, lora_scale)
286
+
287
+ if prompt is not None and isinstance(prompt, str):
288
+ batch_size = 1
289
+ elif prompt is not None and isinstance(prompt, list):
290
+ batch_size = len(prompt)
291
+ else:
292
+ batch_size = prompt_embeds.shape[0]
293
+
294
+ # Define tokenizers and text encoders
295
+ tokenizers = [self.tokenizer, self.tokenizer_2] if self.tokenizer is not None else [self.tokenizer_2]
296
+ text_encoders = (
297
+ [self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2]
298
+ )
299
+
300
+ if prompt_embeds is None:
301
+ prompt_2 = prompt_2 or prompt
302
+ # textual inversion: procecss multi-vector tokens if necessary
303
+ prompt_embeds_list = []
304
+ prompts = [prompt, prompt_2]
305
+ for prompt, tokenizer, text_encoder in zip(prompts, tokenizers, text_encoders):
306
+ if isinstance(self, TextualInversionLoaderMixin):
307
+ prompt = self.maybe_convert_prompt(prompt, tokenizer)
308
+
309
+ text_inputs = tokenizer(
310
+ prompt,
311
+ padding="max_length",
312
+ max_length=tokenizer.model_max_length,
313
+ truncation=True,
314
+ return_tensors="pt",
315
+ )
316
+
317
+ text_input_ids = text_inputs.input_ids
318
+ untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
319
+
320
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
321
+ text_input_ids, untruncated_ids
322
+ ):
323
+ removed_text = tokenizer.batch_decode(untruncated_ids[:, tokenizer.model_max_length - 1 : -1])
324
+ logger.warning(
325
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
326
+ f" {tokenizer.model_max_length} tokens: {removed_text}"
327
+ )
328
+
329
+ prompt_embeds = text_encoder(
330
+ text_input_ids.to(device),
331
+ output_hidden_states=True,
332
+ )
333
+
334
+ # We are only ALWAYS interested in the pooled output of the final text encoder
335
+ pooled_prompt_embeds = prompt_embeds[0]
336
+ prompt_embeds = prompt_embeds.hidden_states[-2]
337
+
338
+ prompt_embeds_list.append(prompt_embeds)
339
+
340
+ prompt_embeds = torch.concat(prompt_embeds_list, dim=-1)
341
+
342
+ # get unconditional embeddings for classifier free guidance
343
+ zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt
344
+ if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt:
345
+ negative_prompt_embeds = torch.zeros_like(prompt_embeds)
346
+ negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds)
347
+ elif do_classifier_free_guidance and negative_prompt_embeds is None:
348
+ negative_prompt = negative_prompt or ""
349
+ negative_prompt_2 = negative_prompt_2 or negative_prompt
350
+
351
+ uncond_tokens: List[str]
352
+ if prompt is not None and type(prompt) is not type(negative_prompt):
353
+ raise TypeError(
354
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
355
+ f" {type(prompt)}."
356
+ )
357
+ elif isinstance(negative_prompt, str):
358
+ uncond_tokens = [negative_prompt, negative_prompt_2]
359
+ elif batch_size != len(negative_prompt):
360
+ raise ValueError(
361
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
362
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
363
+ " the batch size of `prompt`."
364
+ )
365
+ else:
366
+ uncond_tokens = [negative_prompt, negative_prompt_2]
367
+
368
+ negative_prompt_embeds_list = []
369
+ for negative_prompt, tokenizer, text_encoder in zip(uncond_tokens, tokenizers, text_encoders):
370
+ if isinstance(self, TextualInversionLoaderMixin):
371
+ negative_prompt = self.maybe_convert_prompt(negative_prompt, tokenizer)
372
+
373
+ max_length = prompt_embeds.shape[1]
374
+ uncond_input = tokenizer(
375
+ negative_prompt,
376
+ padding="max_length",
377
+ max_length=max_length,
378
+ truncation=True,
379
+ return_tensors="pt",
380
+ )
381
+
382
+ negative_prompt_embeds = text_encoder(
383
+ uncond_input.input_ids.to(device),
384
+ output_hidden_states=True,
385
+ )
386
+ # We are only ALWAYS interested in the pooled output of the final text encoder
387
+ negative_pooled_prompt_embeds = negative_prompt_embeds[0]
388
+ negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2]
389
+
390
+ negative_prompt_embeds_list.append(negative_prompt_embeds)
391
+
392
+ negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1)
393
+
394
+ prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device)
395
+ bs_embed, seq_len, _ = prompt_embeds.shape
396
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
397
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
398
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
399
+
400
+ if do_classifier_free_guidance:
401
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
402
+ seq_len = negative_prompt_embeds.shape[1]
403
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device)
404
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
405
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
406
+
407
+ pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
408
+ bs_embed * num_images_per_prompt, -1
409
+ )
410
+ if do_classifier_free_guidance:
411
+ negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
412
+ bs_embed * num_images_per_prompt, -1
413
+ )
414
+
415
+ return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds
416
+
417
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
418
+ def prepare_extra_step_kwargs(self, generator, eta):
419
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
420
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
421
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
422
+ # and should be between [0, 1]
423
+
424
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
425
+ extra_step_kwargs = {}
426
+ if accepts_eta:
427
+ extra_step_kwargs["eta"] = eta
428
+
429
+ # check if the scheduler accepts generator
430
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
431
+ if accepts_generator:
432
+ extra_step_kwargs["generator"] = generator
433
+ return extra_step_kwargs
434
+
435
+ def check_inputs(
436
+ self,
437
+ prompt,
438
+ prompt_2,
439
+ height,
440
+ width,
441
+ callback_steps,
442
+ negative_prompt=None,
443
+ negative_prompt_2=None,
444
+ prompt_embeds=None,
445
+ negative_prompt_embeds=None,
446
+ pooled_prompt_embeds=None,
447
+ negative_pooled_prompt_embeds=None,
448
+ num_images_per_prompt=None,
449
+ ):
450
+ if height % 8 != 0 or width % 8 != 0:
451
+ raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
452
+
453
+ if (callback_steps is None) or (
454
+ callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
455
+ ):
456
+ raise ValueError(
457
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
458
+ f" {type(callback_steps)}."
459
+ )
460
+
461
+ if prompt is not None and prompt_embeds is not None:
462
+ raise ValueError(
463
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
464
+ " only forward one of the two."
465
+ )
466
+ elif prompt_2 is not None and prompt_embeds is not None:
467
+ raise ValueError(
468
+ f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
469
+ " only forward one of the two."
470
+ )
471
+ elif prompt is None and prompt_embeds is None:
472
+ raise ValueError(
473
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
474
+ )
475
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
476
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
477
+ elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)):
478
+ raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}")
479
+
480
+ if negative_prompt is not None and negative_prompt_embeds is not None:
481
+ raise ValueError(
482
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
483
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
484
+ )
485
+ elif negative_prompt_2 is not None and negative_prompt_embeds is not None:
486
+ raise ValueError(
487
+ f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:"
488
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
489
+ )
490
+
491
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
492
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
493
+ raise ValueError(
494
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
495
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
496
+ f" {negative_prompt_embeds.shape}."
497
+ )
498
+
499
+ if prompt_embeds is not None and pooled_prompt_embeds is None:
500
+ raise ValueError(
501
+ "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`."
502
+ )
503
+
504
+ if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None:
505
+ raise ValueError(
506
+ "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`."
507
+ )
508
+
509
+ # DemoFusion specific checks
510
+ if max(height, width) % 1024 != 0:
511
+ raise ValueError(f"the larger one of `height` and `width` has to be divisible by 1024 but are {height} and {width}.")
512
+
513
+ if num_images_per_prompt != 1:
514
+ warnings.warn("num_images_per_prompt != 1 is not supported by DemoFusion and will be ignored.")
515
+ num_images_per_prompt = 1
516
+
517
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
518
+ def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
519
+ shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
520
+ if isinstance(generator, list) and len(generator) != batch_size:
521
+ raise ValueError(
522
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
523
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
524
+ )
525
+
526
+ if latents is None:
527
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
528
+ else:
529
+ latents = latents.to(device)
530
+
531
+ # scale the initial noise by the standard deviation required by the scheduler
532
+ latents = latents * self.scheduler.init_noise_sigma
533
+ return latents
534
+
535
+ def _get_add_time_ids(self, original_size, crops_coords_top_left, target_size, dtype):
536
+ add_time_ids = list(original_size + crops_coords_top_left + target_size)
537
+
538
+ passed_add_embed_dim = (
539
+ self.unet.config.addition_time_embed_dim * len(add_time_ids) + self.text_encoder_2.config.projection_dim
540
+ )
541
+ expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features
542
+
543
+ if expected_add_embed_dim != passed_add_embed_dim:
544
+ raise ValueError(
545
+ 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`."
546
+ )
547
+
548
+ add_time_ids = torch.tensor([add_time_ids], dtype=dtype)
549
+ return add_time_ids
550
+
551
+ def get_views(self, height, width, window_size=128, stride=64, random_jitter=False):
552
+ # Here, we define the mappings F_i (see Eq. 7 in the MultiDiffusion paper https://arxiv.org/abs/2302.08113)
553
+ # if panorama's height/width < window_size, num_blocks of height/width should return 1
554
+ height //= self.vae_scale_factor
555
+ width //= self.vae_scale_factor
556
+ num_blocks_height = int((height - window_size) / stride - 1e-6) + 2 if height > window_size else 1
557
+ num_blocks_width = int((width - window_size) / stride - 1e-6) + 2 if width > window_size else 1
558
+ total_num_blocks = int(num_blocks_height * num_blocks_width)
559
+ views = []
560
+ for i in range(total_num_blocks):
561
+ h_start = int((i // num_blocks_width) * stride)
562
+ h_end = h_start + window_size
563
+ w_start = int((i % num_blocks_width) * stride)
564
+ w_end = w_start + window_size
565
+
566
+ if h_end > height:
567
+ h_start = int(h_start + height - h_end)
568
+ h_end = int(height)
569
+ if w_end > width:
570
+ w_start = int(w_start + width - w_end)
571
+ w_end = int(width)
572
+ if h_start < 0:
573
+ h_end = int(h_end - h_start)
574
+ h_start = 0
575
+ if w_start < 0:
576
+ w_end = int(w_end - w_start)
577
+ w_start = 0
578
+
579
+ if random_jitter:
580
+ jitter_range = (window_size - stride) // 4
581
+ w_jitter = 0
582
+ h_jitter = 0
583
+ if (w_start != 0) and (w_end != width):
584
+ w_jitter = random.randint(-jitter_range, jitter_range)
585
+ elif (w_start == 0) and (w_end != width):
586
+ w_jitter = random.randint(-jitter_range, 0)
587
+ elif (w_start != 0) and (w_end == width):
588
+ w_jitter = random.randint(0, jitter_range)
589
+ if (h_start != 0) and (h_end != height):
590
+ h_jitter = random.randint(-jitter_range, jitter_range)
591
+ elif (h_start == 0) and (h_end != height):
592
+ h_jitter = random.randint(-jitter_range, 0)
593
+ elif (h_start != 0) and (h_end == height):
594
+ h_jitter = random.randint(0, jitter_range)
595
+ h_start += (h_jitter + jitter_range)
596
+ h_end += (h_jitter + jitter_range)
597
+ w_start += (w_jitter + jitter_range)
598
+ w_end += (w_jitter + jitter_range)
599
+
600
+ views.append((h_start, h_end, w_start, w_end))
601
+ return views
602
+
603
+ def tiled_decode(self, latents, current_height, current_width):
604
+ sample_size = self.unet.config.sample_size
605
+ core_size = self.unet.config.sample_size // 4
606
+ core_stride = core_size
607
+ pad_size = self.unet.config.sample_size // 8 * 3
608
+ decoder_view_batch_size = 1
609
+
610
+ if self.lowvram:
611
+ core_stride = core_size // 2
612
+ pad_size = core_size
613
+
614
+ views = self.get_views(current_height, current_width, stride=core_stride, window_size=core_size)
615
+ views_batch = [views[i : i + decoder_view_batch_size] for i in range(0, len(views), decoder_view_batch_size)]
616
+ latents_ = F.pad(latents, (pad_size, pad_size, pad_size, pad_size), 'constant', 0)
617
+ image = torch.zeros(latents.size(0), 3, current_height, current_width).to(latents.device)
618
+ count = torch.zeros_like(image).to(latents.device)
619
+ # get the latents corresponding to the current view coordinates
620
+ with self.progress_bar(total=len(views_batch)) as progress_bar:
621
+ for j, batch_view in enumerate(views_batch):
622
+ vb_size = len(batch_view)
623
+ latents_for_view = torch.cat(
624
+ [
625
+ latents_[:, :, h_start:h_end+pad_size*2, w_start:w_end+pad_size*2]
626
+ for h_start, h_end, w_start, w_end in batch_view
627
+ ]
628
+ ).to(self.vae.device)
629
+ image_patch = self.vae.decode(latents_for_view / self.vae.config.scaling_factor, return_dict=False)[0]
630
+ h_start, h_end, w_start, w_end = views[j]
631
+ 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
632
+ 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
633
+ 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)
634
+ count[:, :, h_start:h_end, w_start:w_end] += 1
635
+ progress_bar.update()
636
+ image = image / count
637
+
638
+ return image
639
+
640
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_upscale.StableDiffusionUpscalePipeline.upcast_vae
641
+ def upcast_vae(self):
642
+ dtype = self.vae.dtype
643
+ self.vae.to(dtype=torch.float32)
644
+ use_torch_2_0_or_xformers = isinstance(
645
+ self.vae.decoder.mid_block.attentions[0].processor,
646
+ (
647
+ AttnProcessor2_0,
648
+ XFormersAttnProcessor,
649
+ LoRAXFormersAttnProcessor,
650
+ LoRAAttnProcessor2_0,
651
+ ),
652
+ )
653
+ # if xformers or torch_2_0 is used attention block does not need
654
+ # to be in float32 which can save lots of memory
655
+ if use_torch_2_0_or_xformers:
656
+ self.vae.post_quant_conv.to(dtype)
657
+ self.vae.decoder.conv_in.to(dtype)
658
+ self.vae.decoder.mid_block.to(dtype)
659
+
660
+ @torch.no_grad()
661
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
662
+ def __call__(
663
+ self,
664
+ prompt: Union[str, List[str]] = None,
665
+ prompt_2: Optional[Union[str, List[str]]] = None,
666
+ height: Optional[int] = None,
667
+ width: Optional[int] = None,
668
+ num_inference_steps: int = 50,
669
+ denoising_end: Optional[float] = None,
670
+ guidance_scale: float = 5.0,
671
+ negative_prompt: Optional[Union[str, List[str]]] = None,
672
+ negative_prompt_2: Optional[Union[str, List[str]]] = None,
673
+ num_images_per_prompt: Optional[int] = 1,
674
+ eta: float = 0.0,
675
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
676
+ latents: Optional[torch.FloatTensor] = None,
677
+ prompt_embeds: Optional[torch.FloatTensor] = None,
678
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
679
+ pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
680
+ negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
681
+ output_type: Optional[str] = "pil",
682
+ return_dict: bool = False,
683
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
684
+ callback_steps: int = 1,
685
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
686
+ guidance_rescale: float = 0.0,
687
+ original_size: Optional[Tuple[int, int]] = None,
688
+ crops_coords_top_left: Tuple[int, int] = (0, 0),
689
+ target_size: Optional[Tuple[int, int]] = None,
690
+ negative_original_size: Optional[Tuple[int, int]] = None,
691
+ negative_crops_coords_top_left: Tuple[int, int] = (0, 0),
692
+ negative_target_size: Optional[Tuple[int, int]] = None,
693
+ ################### DemoFusion specific parameters ####################
694
+ image_lr: Optional[torch.FloatTensor] = None,
695
+ view_batch_size: int = 16,
696
+ multi_decoder: bool = True,
697
+ stride: Optional[int] = 64,
698
+ cosine_scale_1: Optional[float] = 3.,
699
+ cosine_scale_2: Optional[float] = 1.,
700
+ cosine_scale_3: Optional[float] = 1.,
701
+ sigma: Optional[float] = 1.0,
702
+ show_image: bool = False,
703
+ lowvram: bool = False,
704
+ ):
705
+ r"""
706
+ Function invoked when calling the pipeline for generation.
707
+
708
+ Args:
709
+ prompt (`str` or `List[str]`, *optional*):
710
+ The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
711
+ instead.
712
+ prompt_2 (`str` or `List[str]`, *optional*):
713
+ The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
714
+ used in both text-encoders
715
+ height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
716
+ The height in pixels of the generated image. This is set to 1024 by default for the best results.
717
+ Anything below 512 pixels won't work well for
718
+ [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0)
719
+ and checkpoints that are not specifically fine-tuned on low resolutions.
720
+ width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
721
+ The width in pixels of the generated image. This is set to 1024 by default for the best results.
722
+ Anything below 512 pixels won't work well for
723
+ [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0)
724
+ and checkpoints that are not specifically fine-tuned on low resolutions.
725
+ num_inference_steps (`int`, *optional*, defaults to 50):
726
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
727
+ expense of slower inference.
728
+ denoising_end (`float`, *optional*):
729
+ When specified, determines the fraction (between 0.0 and 1.0) of the total denoising process to be
730
+ completed before it is intentionally prematurely terminated. As a result, the returned sample will
731
+ still retain a substantial amount of noise as determined by the discrete timesteps selected by the
732
+ scheduler. The denoising_end parameter should ideally be utilized when this pipeline forms a part of a
733
+ "Mixture of Denoisers" multi-pipeline setup, as elaborated in [**Refining the Image
734
+ Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output)
735
+ guidance_scale (`float`, *optional*, defaults to 5.0):
736
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
737
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
738
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
739
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
740
+ usually at the expense of lower image quality.
741
+ negative_prompt (`str` or `List[str]`, *optional*):
742
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
743
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
744
+ less than `1`).
745
+ negative_prompt_2 (`str` or `List[str]`, *optional*):
746
+ The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
747
+ `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders
748
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
749
+ The number of images to generate per prompt.
750
+ eta (`float`, *optional*, defaults to 0.0):
751
+ Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
752
+ [`schedulers.DDIMScheduler`], will be ignored for others.
753
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
754
+ One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
755
+ to make generation deterministic.
756
+ latents (`torch.FloatTensor`, *optional*):
757
+ Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
758
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
759
+ tensor will ge generated by sampling using the supplied random `generator`.
760
+ prompt_embeds (`torch.FloatTensor`, *optional*):
761
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
762
+ provided, text embeddings will be generated from `prompt` input argument.
763
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
764
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
765
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
766
+ argument.
767
+ pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
768
+ Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
769
+ If not provided, pooled text embeddings will be generated from `prompt` input argument.
770
+ negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
771
+ Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
772
+ weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
773
+ input argument.
774
+ output_type (`str`, *optional*, defaults to `"pil"`):
775
+ The output format of the generate image. Choose between
776
+ [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
777
+ return_dict (`bool`, *optional*, defaults to `True`):
778
+ Whether or not to return a [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] instead
779
+ of a plain tuple.
780
+ callback (`Callable`, *optional*):
781
+ A function that will be called every `callback_steps` steps during inference. The function will be
782
+ called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
783
+ callback_steps (`int`, *optional*, defaults to 1):
784
+ The frequency at which the `callback` function will be called. If not specified, the callback will be
785
+ called at every step.
786
+ cross_attention_kwargs (`dict`, *optional*):
787
+ A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
788
+ `self.processor` in
789
+ [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
790
+ guidance_rescale (`float`, *optional*, defaults to 0.7):
791
+ Guidance rescale factor proposed by [Common Diffusion Noise Schedules and Sample Steps are
792
+ Flawed](https://arxiv.org/pdf/2305.08891.pdf) `guidance_scale` is defined as `φ` in equation 16. of
793
+ [Common Diffusion Noise Schedules and Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf).
794
+ Guidance rescale factor should fix overexposure when using zero terminal SNR.
795
+ original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
796
+ If `original_size` is not the same as `target_size` the image will appear to be down- or upsampled.
797
+ `original_size` defaults to `(width, height)` if not specified. Part of SDXL's micro-conditioning as
798
+ explained in section 2.2 of
799
+ [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
800
+ crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
801
+ `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position
802
+ `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting
803
+ `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of
804
+ [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
805
+ target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
806
+ For most cases, `target_size` should be set to the desired height and width of the generated image. If
807
+ not specified it will default to `(width, height)`. Part of SDXL's micro-conditioning as explained in
808
+ section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
809
+ negative_original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
810
+ To negatively condition the generation process based on a specific image resolution. Part of SDXL's
811
+ micro-conditioning as explained in section 2.2 of
812
+ [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
813
+ information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
814
+ negative_crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
815
+ To negatively condition the generation process based on a specific crop coordinates. Part of SDXL's
816
+ micro-conditioning as explained in section 2.2 of
817
+ [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
818
+ information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
819
+ negative_target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
820
+ To negatively condition the generation process based on a target image resolution. It should be as same
821
+ as the `target_size` for most cases. Part of SDXL's micro-conditioning as explained in section 2.2 of
822
+ [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
823
+ information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
824
+ ################### DemoFusion specific parameters ####################
825
+ image_lr (`torch.FloatTensor`, *optional*, , defaults to None):
826
+ Low-resolution image input for upscaling. If provided, DemoFusion will encode it as the initial latent representation.
827
+ view_batch_size (`int`, defaults to 16):
828
+ The batch size for multiple denoising paths. Typically, a larger batch size can result in higher
829
+ efficiency but comes with increased GPU memory requirements.
830
+ multi_decoder (`bool`, defaults to True):
831
+ Determine whether to use a tiled decoder. Generally, when the resolution exceeds 3072x3072,
832
+ a tiled decoder becomes necessary.
833
+ stride (`int`, defaults to 64):
834
+ The stride of moving local patches. A smaller stride is better for alleviating seam issues,
835
+ but it also introduces additional computational overhead and inference time.
836
+ cosine_scale_1 (`float`, defaults to 3):
837
+ Control the strength of skip-residual. For specific impacts, please refer to Appendix C
838
+ in the DemoFusion paper.
839
+ cosine_scale_2 (`float`, defaults to 1):
840
+ Control the strength of dilated sampling. For specific impacts, please refer to Appendix C
841
+ in the DemoFusion paper.
842
+ cosine_scale_3 (`float`, defaults to 1):
843
+ Control the strength of the gaussion filter. For specific impacts, please refer to Appendix C
844
+ in the DemoFusion paper.
845
+ sigma (`float`, defaults to 1):
846
+ The standard value of the gaussian filter.
847
+ show_image (`bool`, defaults to False):
848
+ Determine whether to show intermediate results during generation.
849
+ lowvram (`bool`, defaults to False):
850
+ Try to fit in 8 Gb of VRAM, with xformers installed.
851
+
852
+ Examples:
853
+
854
+ Returns:
855
+ a `list` with the generated images at each phase.
856
+ """
857
+
858
+ # 0. Default height and width to unet
859
+ height = height or self.default_sample_size * self.vae_scale_factor
860
+ width = width or self.default_sample_size * self.vae_scale_factor
861
+
862
+ x1_size = self.default_sample_size * self.vae_scale_factor
863
+
864
+ height_scale = height / x1_size
865
+ width_scale = width / x1_size
866
+ scale_num = int(max(height_scale, width_scale))
867
+ aspect_ratio = min(height_scale, width_scale) / max(height_scale, width_scale)
868
+
869
+ original_size = original_size or (height, width)
870
+ target_size = target_size or (height, width)
871
+
872
+ # 1. Check inputs. Raise error if not correct
873
+ self.check_inputs(
874
+ prompt,
875
+ prompt_2,
876
+ height,
877
+ width,
878
+ callback_steps,
879
+ negative_prompt,
880
+ negative_prompt_2,
881
+ prompt_embeds,
882
+ negative_prompt_embeds,
883
+ pooled_prompt_embeds,
884
+ negative_pooled_prompt_embeds,
885
+ num_images_per_prompt,
886
+ )
887
+
888
+ # 2. Define call parameters
889
+ if prompt is not None and isinstance(prompt, str):
890
+ batch_size = 1
891
+ elif prompt is not None and isinstance(prompt, list):
892
+ batch_size = len(prompt)
893
+ else:
894
+ batch_size = prompt_embeds.shape[0]
895
+
896
+ device = self._execution_device
897
+ self.lowvram = lowvram
898
+ if self.lowvram:
899
+ self.vae.cpu()
900
+ self.unet.cpu()
901
+ self.text_encoder.to(device)
902
+ self.text_encoder_2.to(device)
903
+ image_lr.cpu()
904
+
905
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
906
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
907
+ # corresponds to doing no classifier free guidance.
908
+ do_classifier_free_guidance = guidance_scale > 1.0
909
+
910
+ # 3. Encode input prompt
911
+ text_encoder_lora_scale = (
912
+ cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None
913
+ )
914
+ (
915
+ prompt_embeds,
916
+ negative_prompt_embeds,
917
+ pooled_prompt_embeds,
918
+ negative_pooled_prompt_embeds,
919
+ ) = self.encode_prompt(
920
+ prompt=prompt,
921
+ prompt_2=prompt_2,
922
+ device=device,
923
+ num_images_per_prompt=num_images_per_prompt,
924
+ do_classifier_free_guidance=do_classifier_free_guidance,
925
+ negative_prompt=negative_prompt,
926
+ negative_prompt_2=negative_prompt_2,
927
+ prompt_embeds=prompt_embeds,
928
+ negative_prompt_embeds=negative_prompt_embeds,
929
+ pooled_prompt_embeds=pooled_prompt_embeds,
930
+ negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
931
+ lora_scale=text_encoder_lora_scale,
932
+ )
933
+
934
+ # 4. Prepare timesteps
935
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
936
+
937
+ timesteps = self.scheduler.timesteps
938
+
939
+ # 5. Prepare latent variables
940
+ num_channels_latents = self.unet.config.in_channels
941
+ latents = self.prepare_latents(
942
+ batch_size * num_images_per_prompt,
943
+ num_channels_latents,
944
+ height // scale_num,
945
+ width // scale_num,
946
+ prompt_embeds.dtype,
947
+ device,
948
+ generator,
949
+ latents,
950
+ )
951
+
952
+ # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
953
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
954
+
955
+ # 7. Prepare added time ids & embeddings
956
+ add_text_embeds = pooled_prompt_embeds
957
+ add_time_ids = self._get_add_time_ids(
958
+ original_size, crops_coords_top_left, target_size, dtype=prompt_embeds.dtype
959
+ )
960
+ if negative_original_size is not None and negative_target_size is not None:
961
+ negative_add_time_ids = self._get_add_time_ids(
962
+ negative_original_size,
963
+ negative_crops_coords_top_left,
964
+ negative_target_size,
965
+ dtype=prompt_embeds.dtype,
966
+ )
967
+ else:
968
+ negative_add_time_ids = add_time_ids
969
+
970
+ if do_classifier_free_guidance:
971
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
972
+ add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0)
973
+ add_time_ids = torch.cat([negative_add_time_ids, add_time_ids], dim=0)
974
+ del negative_prompt_embeds, negative_pooled_prompt_embeds, negative_add_time_ids
975
+
976
+ prompt_embeds = prompt_embeds.to(device)
977
+ add_text_embeds = add_text_embeds.to(device)
978
+ add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1)
979
+
980
+ # 8. Denoising loop
981
+ num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
982
+
983
+ # 7.1 Apply denoising_end
984
+ if denoising_end is not None and isinstance(denoising_end, float) and denoising_end > 0 and denoising_end < 1:
985
+ discrete_timestep_cutoff = int(
986
+ round(
987
+ self.scheduler.config.num_train_timesteps
988
+ - (denoising_end * self.scheduler.config.num_train_timesteps)
989
+ )
990
+ )
991
+ num_inference_steps = len(list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps)))
992
+ timesteps = timesteps[:num_inference_steps]
993
+
994
+ output_images = []
995
+
996
+ ###################################################### Phase Initialization ########################################################
997
+
998
+ if self.lowvram:
999
+ self.text_encoder.cpu()
1000
+ self.text_encoder_2.cpu()
1001
+
1002
+ if image_lr == None:
1003
+ print("### Phase 1 Denoising ###")
1004
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
1005
+ for i, t in enumerate(timesteps):
1006
+
1007
+ if self.lowvram:
1008
+ self.vae.cpu()
1009
+ self.unet.to(device)
1010
+
1011
+ latents_for_view = latents
1012
+
1013
+ # expand the latents if we are doing classifier free guidance
1014
+ latent_model_input = (
1015
+ latents.repeat_interleave(2, dim=0)
1016
+ if do_classifier_free_guidance
1017
+ else latents
1018
+ )
1019
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
1020
+
1021
+ # predict the noise residual
1022
+ added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids}
1023
+ noise_pred = self.unet(
1024
+ latent_model_input,
1025
+ t,
1026
+ encoder_hidden_states=prompt_embeds,
1027
+ cross_attention_kwargs=cross_attention_kwargs,
1028
+ added_cond_kwargs=added_cond_kwargs,
1029
+ return_dict=False,
1030
+ )[0]
1031
+
1032
+ # perform guidance
1033
+ if do_classifier_free_guidance:
1034
+ noise_pred_uncond, noise_pred_text = noise_pred[::2], noise_pred[1::2]
1035
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
1036
+
1037
+ if do_classifier_free_guidance and guidance_rescale > 0.0:
1038
+ # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
1039
+ noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=guidance_rescale)
1040
+
1041
+ # compute the previous noisy sample x_t -> x_t-1
1042
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
1043
+
1044
+ # call the callback, if provided
1045
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
1046
+ progress_bar.update()
1047
+ if callback is not None and i % callback_steps == 0:
1048
+ step_idx = i // getattr(self.scheduler, "order", 1)
1049
+ callback(step_idx, t, latents)
1050
+ del latents_for_view, latent_model_input, noise_pred, noise_pred_text, noise_pred_uncond
1051
+ else:
1052
+ print("### Encoding Real Image ###")
1053
+ latents = self.vae.encode(image_lr)
1054
+ latents = latents.latent_dist.sample() * self.vae.config.scaling_factor
1055
+
1056
+ anchor_mean = latents.mean()
1057
+ anchor_std = latents.std()
1058
+ if self.lowvram:
1059
+ latents = latents.cpu()
1060
+ torch.cuda.empty_cache()
1061
+ if not output_type == "latent":
1062
+ # make sure the VAE is in float32 mode, as it overflows in float16
1063
+ needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
1064
+
1065
+ if self.lowvram:
1066
+ needs_upcasting = False # use madebyollin/sdxl-vae-fp16-fix in lowvram mode!
1067
+ self.unet.cpu()
1068
+ self.vae.to(device)
1069
+
1070
+ if needs_upcasting:
1071
+ self.upcast_vae()
1072
+ latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
1073
+ if self.lowvram and multi_decoder:
1074
+ current_width_height = self.unet.config.sample_size * self.vae_scale_factor
1075
+ image = self.tiled_decode(latents, current_width_height, current_width_height)
1076
+ else:
1077
+ image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
1078
+ # cast back to fp16 if needed
1079
+ if needs_upcasting:
1080
+ self.vae.to(dtype=torch.float16)
1081
+
1082
+ image = self.image_processor.postprocess(image, output_type=output_type)
1083
+ if show_image:
1084
+ plt.figure(figsize=(10, 10))
1085
+ plt.imshow(image[0])
1086
+ plt.axis('off') # Turn off axis numbers and ticks
1087
+ plt.show()
1088
+ output_images.append(image[0])
1089
+
1090
+ ####################################################### Phase Upscaling #####################################################
1091
+ if image_lr == None:
1092
+ starting_scale = 2
1093
+ else:
1094
+ starting_scale = 1
1095
+ for current_scale_num in range(starting_scale, scale_num + 1):
1096
+ if self.lowvram:
1097
+ latents = latents.to(device)
1098
+ self.unet.to(device)
1099
+ torch.cuda.empty_cache()
1100
+ print("### Phase {} Denoising ###".format(current_scale_num))
1101
+ current_height = self.unet.config.sample_size * self.vae_scale_factor * current_scale_num
1102
+ current_width = self.unet.config.sample_size * self.vae_scale_factor * current_scale_num
1103
+ if height > width:
1104
+ current_width = int(current_width * aspect_ratio)
1105
+ else:
1106
+ current_height = int(current_height * aspect_ratio)
1107
+
1108
+ latents = F.interpolate(latents.to(device), size=(int(current_height / self.vae_scale_factor), int(current_width / self.vae_scale_factor)), mode='bicubic')
1109
+
1110
+ noise_latents = []
1111
+ noise = torch.randn_like(latents)
1112
+ for timestep in timesteps:
1113
+ noise_latent = self.scheduler.add_noise(latents, noise, timestep.unsqueeze(0))
1114
+ noise_latents.append(noise_latent)
1115
+ latents = noise_latents[0]
1116
+
1117
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
1118
+ for i, t in enumerate(timesteps):
1119
+ count = torch.zeros_like(latents)
1120
+ value = torch.zeros_like(latents)
1121
+ cosine_factor = 0.5 * (1 + torch.cos(torch.pi * (self.scheduler.config.num_train_timesteps - t) / self.scheduler.config.num_train_timesteps)).cpu()
1122
+
1123
+ c1 = cosine_factor ** cosine_scale_1
1124
+ latents = latents * (1 - c1) + noise_latents[i] * c1
1125
+
1126
+ ############################################# MultiDiffusion #############################################
1127
+
1128
+ views = self.get_views(current_height, current_width, stride=stride, window_size=self.unet.config.sample_size, random_jitter=True)
1129
+ views_batch = [views[i : i + view_batch_size] for i in range(0, len(views), view_batch_size)]
1130
+
1131
+ jitter_range = (self.unet.config.sample_size - stride) // 4
1132
+ latents_ = F.pad(latents, (jitter_range, jitter_range, jitter_range, jitter_range), 'constant', 0)
1133
+
1134
+ count_local = torch.zeros_like(latents_)
1135
+ value_local = torch.zeros_like(latents_)
1136
+
1137
+ for j, batch_view in enumerate(views_batch):
1138
+ vb_size = len(batch_view)
1139
+
1140
+ # get the latents corresponding to the current view coordinates
1141
+ latents_for_view = torch.cat(
1142
+ [
1143
+ latents_[:, :, h_start:h_end, w_start:w_end]
1144
+ for h_start, h_end, w_start, w_end in batch_view
1145
+ ]
1146
+ )
1147
+
1148
+ # expand the latents if we are doing classifier free guidance
1149
+ latent_model_input = latents_for_view
1150
+ latent_model_input = (
1151
+ latent_model_input.repeat_interleave(2, dim=0)
1152
+ if do_classifier_free_guidance
1153
+ else latent_model_input
1154
+ )
1155
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
1156
+
1157
+ prompt_embeds_input = torch.cat([prompt_embeds] * vb_size)
1158
+ add_text_embeds_input = torch.cat([add_text_embeds] * vb_size)
1159
+ add_time_ids_input = []
1160
+ for h_start, h_end, w_start, w_end in batch_view:
1161
+ add_time_ids_ = add_time_ids.clone()
1162
+ add_time_ids_[:, 2] = h_start * self.vae_scale_factor
1163
+ add_time_ids_[:, 3] = w_start * self.vae_scale_factor
1164
+ add_time_ids_input.append(add_time_ids_)
1165
+ add_time_ids_input = torch.cat(add_time_ids_input)
1166
+
1167
+ # predict the noise residual
1168
+ added_cond_kwargs = {"text_embeds": add_text_embeds_input, "time_ids": add_time_ids_input}
1169
+ noise_pred = self.unet(
1170
+ latent_model_input,
1171
+ t,
1172
+ encoder_hidden_states=prompt_embeds_input,
1173
+ cross_attention_kwargs=cross_attention_kwargs,
1174
+ added_cond_kwargs=added_cond_kwargs,
1175
+ return_dict=False,
1176
+ )[0]
1177
+
1178
+ if do_classifier_free_guidance:
1179
+ noise_pred_uncond, noise_pred_text = noise_pred[::2], noise_pred[1::2]
1180
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
1181
+
1182
+ if do_classifier_free_guidance and guidance_rescale > 0.0:
1183
+ # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
1184
+ noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=guidance_rescale)
1185
+
1186
+ # compute the previous noisy sample x_t -> x_t-1
1187
+ self.scheduler._init_step_index(t)
1188
+ latents_denoised_batch = self.scheduler.step(
1189
+ noise_pred, t, latents_for_view, **extra_step_kwargs, return_dict=False)[0]
1190
+
1191
+ # extract value from batch
1192
+ for latents_view_denoised, (h_start, h_end, w_start, w_end) in zip(
1193
+ latents_denoised_batch.chunk(vb_size), batch_view
1194
+ ):
1195
+ value_local[:, :, h_start:h_end, w_start:w_end] += latents_view_denoised
1196
+ count_local[:, :, h_start:h_end, w_start:w_end] += 1
1197
+
1198
+ value_local = value_local[: ,:, jitter_range: jitter_range + current_height // self.vae_scale_factor, jitter_range: jitter_range + current_width // self.vae_scale_factor]
1199
+ count_local = count_local[: ,:, jitter_range: jitter_range + current_height // self.vae_scale_factor, jitter_range: jitter_range + current_width // self.vae_scale_factor]
1200
+
1201
+ c2 = cosine_factor ** cosine_scale_2
1202
+
1203
+ value += value_local / count_local * (1 - c2)
1204
+ count += torch.ones_like(value_local) * (1 - c2)
1205
+
1206
+ ############################################# Dilated Sampling #############################################
1207
+
1208
+ views = [[h, w] for h in range(current_scale_num) for w in range(current_scale_num)]
1209
+ views_batch = [views[i : i + view_batch_size] for i in range(0, len(views), view_batch_size)]
1210
+
1211
+ h_pad = (current_scale_num - (latents.size(2) % current_scale_num)) % current_scale_num
1212
+ w_pad = (current_scale_num - (latents.size(3) % current_scale_num)) % current_scale_num
1213
+ latents_ = F.pad(latents, (w_pad, 0, h_pad, 0), 'constant', 0)
1214
+
1215
+ count_global = torch.zeros_like(latents_)
1216
+ value_global = torch.zeros_like(latents_)
1217
+
1218
+ c3 = 0.99 * cosine_factor ** cosine_scale_3 + 1e-2
1219
+ std_, mean_ = latents_.std(), latents_.mean()
1220
+ latents_gaussian = gaussian_filter(latents_, kernel_size=(2*current_scale_num-1), sigma=sigma*c3)
1221
+ latents_gaussian = (latents_gaussian - latents_gaussian.mean()) / latents_gaussian.std() * std_ + mean_
1222
+
1223
+ for j, batch_view in enumerate(views_batch):
1224
+ latents_for_view = torch.cat(
1225
+ [
1226
+ latents_[:, :, h::current_scale_num, w::current_scale_num]
1227
+ for h, w in batch_view
1228
+ ]
1229
+ )
1230
+ latents_for_view_gaussian = torch.cat(
1231
+ [
1232
+ latents_gaussian[:, :, h::current_scale_num, w::current_scale_num]
1233
+ for h, w in batch_view
1234
+ ]
1235
+ )
1236
+
1237
+ vb_size = latents_for_view.size(0)
1238
+
1239
+ # expand the latents if we are doing classifier free guidance
1240
+ latent_model_input = latents_for_view_gaussian
1241
+ latent_model_input = (
1242
+ latent_model_input.repeat_interleave(2, dim=0)
1243
+ if do_classifier_free_guidance
1244
+ else latent_model_input
1245
+ )
1246
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
1247
+
1248
+ prompt_embeds_input = torch.cat([prompt_embeds] * vb_size)
1249
+ add_text_embeds_input = torch.cat([add_text_embeds] * vb_size)
1250
+ add_time_ids_input = torch.cat([add_time_ids] * vb_size)
1251
+
1252
+ # predict the noise residual
1253
+ added_cond_kwargs = {"text_embeds": add_text_embeds_input, "time_ids": add_time_ids_input}
1254
+ noise_pred = self.unet(
1255
+ latent_model_input,
1256
+ t,
1257
+ encoder_hidden_states=prompt_embeds_input,
1258
+ cross_attention_kwargs=cross_attention_kwargs,
1259
+ added_cond_kwargs=added_cond_kwargs,
1260
+ return_dict=False,
1261
+ )[0]
1262
+
1263
+ if do_classifier_free_guidance:
1264
+ noise_pred_uncond, noise_pred_text = noise_pred[::2], noise_pred[1::2]
1265
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
1266
+
1267
+ if do_classifier_free_guidance and guidance_rescale > 0.0:
1268
+ # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
1269
+ noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=guidance_rescale)
1270
+
1271
+ # compute the previous noisy sample x_t -> x_t-1
1272
+ self.scheduler._init_step_index(t)
1273
+ latents_denoised_batch = self.scheduler.step(
1274
+ noise_pred, t, latents_for_view, **extra_step_kwargs, return_dict=False)[0]
1275
+
1276
+ # extract value from batch
1277
+ for latents_view_denoised, (h, w) in zip(
1278
+ latents_denoised_batch.chunk(vb_size), batch_view
1279
+ ):
1280
+ value_global[:, :, h::current_scale_num, w::current_scale_num] += latents_view_denoised
1281
+ count_global[:, :, h::current_scale_num, w::current_scale_num] += 1
1282
+
1283
+ c2 = cosine_factor ** cosine_scale_2
1284
+
1285
+ value_global = value_global[: ,:, h_pad:, w_pad:]
1286
+
1287
+ value += value_global * c2
1288
+ count += torch.ones_like(value_global) * c2
1289
+
1290
+ ###########################################################
1291
+
1292
+ latents = torch.where(count > 0, value / count, value)
1293
+
1294
+ # call the callback, if provided
1295
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
1296
+ progress_bar.update()
1297
+ if callback is not None and i % callback_steps == 0:
1298
+ step_idx = i // getattr(self.scheduler, "order", 1)
1299
+ callback(step_idx, t, latents)
1300
+
1301
+ #########################################################################################################################################
1302
+
1303
+ latents = (latents - latents.mean()) / latents.std() * anchor_std + anchor_mean
1304
+ if self.lowvram:
1305
+ latents = latents.cpu()
1306
+ torch.cuda.empty_cache()
1307
+ if not output_type == "latent":
1308
+ # make sure the VAE is in float32 mode, as it overflows in float16
1309
+ needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
1310
+
1311
+ if self.lowvram:
1312
+ needs_upcasting = False # use madebyollin/sdxl-vae-fp16-fix in lowvram mode!
1313
+ self.unet.cpu()
1314
+ self.vae.to(device)
1315
+
1316
+ if needs_upcasting:
1317
+ self.upcast_vae()
1318
+ latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
1319
+
1320
+ print("### Phase {} Decoding ###".format(current_scale_num))
1321
+ if multi_decoder:
1322
+ image = self.tiled_decode(latents, current_height, current_width)
1323
+ else:
1324
+ image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
1325
+
1326
+ # cast back to fp16 if needed
1327
+ if needs_upcasting:
1328
+ self.vae.to(dtype=torch.float16)
1329
+ else:
1330
+ image = latents
1331
+
1332
+ if not output_type == "latent":
1333
+ image = self.image_processor.postprocess(image, output_type=output_type)
1334
+ if show_image:
1335
+ plt.figure(figsize=(10, 10))
1336
+ plt.imshow(image[0])
1337
+ plt.axis('off') # Turn off axis numbers and ticks
1338
+ plt.show()
1339
+ output_images.append(image[0])
1340
+
1341
+ # Offload all models
1342
+ self.maybe_free_model_hooks()
1343
+
1344
+ return output_images
1345
+
1346
+ # Overrride to properly handle the loading and unloading of the additional text encoder.
1347
+ def load_lora_weights(self, pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]], **kwargs):
1348
+ # We could have accessed the unet config from `lora_state_dict()` too. We pass
1349
+ # it here explicitly to be able to tell that it's coming from an SDXL
1350
+ # pipeline.
1351
+
1352
+ # Remove any existing hooks.
1353
+ if is_accelerate_available() and is_accelerate_version(">=", "0.17.0.dev0"):
1354
+ from accelerate.hooks import AlignDevicesHook, CpuOffload, remove_hook_from_module
1355
+ else:
1356
+ raise ImportError("Offloading requires `accelerate v0.17.0` or higher.")
1357
+
1358
+ is_model_cpu_offload = False
1359
+ is_sequential_cpu_offload = False
1360
+ recursive = False
1361
+ for _, component in self.components.items():
1362
+ if isinstance(component, torch.nn.Module):
1363
+ if hasattr(component, "_hf_hook"):
1364
+ is_model_cpu_offload = isinstance(getattr(component, "_hf_hook"), CpuOffload)
1365
+ is_sequential_cpu_offload = isinstance(getattr(component, "_hf_hook"), AlignDevicesHook)
1366
+ logger.info(
1367
+ "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."
1368
+ )
1369
+ recursive = is_sequential_cpu_offload
1370
+ remove_hook_from_module(component, recurse=recursive)
1371
+ state_dict, network_alphas = self.lora_state_dict(
1372
+ pretrained_model_name_or_path_or_dict,
1373
+ unet_config=self.unet.config,
1374
+ **kwargs,
1375
+ )
1376
+ self.load_lora_into_unet(state_dict, network_alphas=network_alphas, unet=self.unet)
1377
+
1378
+ text_encoder_state_dict = {k: v for k, v in state_dict.items() if "text_encoder." in k}
1379
+ if len(text_encoder_state_dict) > 0:
1380
+ self.load_lora_into_text_encoder(
1381
+ text_encoder_state_dict,
1382
+ network_alphas=network_alphas,
1383
+ text_encoder=self.text_encoder,
1384
+ prefix="text_encoder",
1385
+ lora_scale=self.lora_scale,
1386
+ )
1387
+
1388
+ text_encoder_2_state_dict = {k: v for k, v in state_dict.items() if "text_encoder_2." in k}
1389
+ if len(text_encoder_2_state_dict) > 0:
1390
+ self.load_lora_into_text_encoder(
1391
+ text_encoder_2_state_dict,
1392
+ network_alphas=network_alphas,
1393
+ text_encoder=self.text_encoder_2,
1394
+ prefix="text_encoder_2",
1395
+ lora_scale=self.lora_scale,
1396
+ )
1397
+
1398
+ # Offload back.
1399
+ if is_model_cpu_offload:
1400
+ self.enable_model_cpu_offload()
1401
+ elif is_sequential_cpu_offload:
1402
+ self.enable_sequential_cpu_offload()
1403
+
1404
+ @classmethod
1405
+ def save_lora_weights(
1406
+ self,
1407
+ save_directory: Union[str, os.PathLike],
1408
+ unet_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None,
1409
+ text_encoder_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None,
1410
+ text_encoder_2_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None,
1411
+ is_main_process: bool = True,
1412
+ weight_name: str = None,
1413
+ save_function: Callable = None,
1414
+ safe_serialization: bool = True,
1415
+ ):
1416
+ state_dict = {}
1417
+
1418
+ def pack_weights(layers, prefix):
1419
+ layers_weights = layers.state_dict() if isinstance(layers, torch.nn.Module) else layers
1420
+ layers_state_dict = {f"{prefix}.{module_name}": param for module_name, param in layers_weights.items()}
1421
+ return layers_state_dict
1422
+
1423
+ if not (unet_lora_layers or text_encoder_lora_layers or text_encoder_2_lora_layers):
1424
+ raise ValueError(
1425
+ "You must pass at least one of `unet_lora_layers`, `text_encoder_lora_layers` or `text_encoder_2_lora_layers`."
1426
+ )
1427
+
1428
+ if unet_lora_layers:
1429
+ state_dict.update(pack_weights(unet_lora_layers, "unet"))
1430
+
1431
+ if text_encoder_lora_layers and text_encoder_2_lora_layers:
1432
+ state_dict.update(pack_weights(text_encoder_lora_layers, "text_encoder"))
1433
+ state_dict.update(pack_weights(text_encoder_2_lora_layers, "text_encoder_2"))
1434
+
1435
+ self.write_lora_layers(
1436
+ state_dict=state_dict,
1437
+ save_directory=save_directory,
1438
+ is_main_process=is_main_process,
1439
+ weight_name=weight_name,
1440
+ save_function=save_function,
1441
+ safe_serialization=safe_serialization,
1442
+ )
1443
+
1444
+ def _remove_text_encoder_monkey_patch(self):
1445
+ self._remove_text_encoder_monkey_patch_classmethod(self.text_encoder)
1446
+ self._remove_text_encoder_monkey_patch_classmethod(self.text_encoder_2)
pipeline_demofusion_sdxl_controlnet.py ADDED
@@ -0,0 +1,1795 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright 2023 The HuggingFace Team. All rights reserved.
2
+ #
3
+ # Licensed under the Apache License, Version 2.0 (the "License");
4
+ # you may not use this file except in compliance with the License.
5
+ # You may obtain a copy of the License at
6
+ #
7
+ # http://www.apache.org/licenses/LICENSE-2.0
8
+ #
9
+ # Unless required by applicable law or agreed to in writing, software
10
+ # distributed under the License is distributed on an "AS IS" BASIS,
11
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12
+ # See the License for the specific language governing permissions and
13
+ # limitations under the License.
14
+
15
+
16
+ import inspect
17
+ import os
18
+ from typing import Any, Callable, Dict, List, Optional, Tuple, Union
19
+ import matplotlib.pyplot as plt
20
+
21
+ import numpy as np
22
+ import PIL.Image
23
+ import torch
24
+ import torch.nn.functional as F
25
+ import random
26
+ import warnings
27
+ from transformers import CLIPTextModel, CLIPTextModelWithProjection, CLIPTokenizer
28
+
29
+ from diffusers.utils.import_utils import is_invisible_watermark_available
30
+
31
+ from diffusers.image_processor import PipelineImageInput, VaeImageProcessor
32
+ from diffusers.loaders import FromSingleFileMixin, LoraLoaderMixin, TextualInversionLoaderMixin
33
+ from diffusers.models import AutoencoderKL, ControlNetModel, UNet2DConditionModel
34
+ from diffusers.models.attention_processor import (
35
+ AttnProcessor2_0,
36
+ LoRAAttnProcessor2_0,
37
+ LoRAXFormersAttnProcessor,
38
+ XFormersAttnProcessor,
39
+ )
40
+ from diffusers.models.lora import adjust_lora_scale_text_encoder
41
+ from diffusers.schedulers import KarrasDiffusionSchedulers
42
+ from diffusers.utils import (
43
+ is_accelerate_available,
44
+ is_accelerate_version,
45
+ logging,
46
+ replace_example_docstring,
47
+ )
48
+ from diffusers.utils.torch_utils import is_compiled_module, randn_tensor
49
+ from diffusers.pipelines.pipeline_utils import DiffusionPipeline
50
+ from diffusers.pipelines.stable_diffusion_xl import StableDiffusionXLPipelineOutput
51
+
52
+
53
+ if is_invisible_watermark_available():
54
+ from diffusers.pipelines.stable_diffusion_xl.watermark import StableDiffusionXLWatermarker
55
+
56
+ from diffusers.pipelines.controlnet.multicontrolnet import MultiControlNetModel
57
+
58
+
59
+ logger = logging.get_logger(__name__) # pylint: disable=invalid-name
60
+
61
+
62
+ EXAMPLE_DOC_STRING = """
63
+ Examples:
64
+ """
65
+
66
+ def gaussian_kernel(kernel_size=3, sigma=1.0, channels=3):
67
+ x_coord = torch.arange(kernel_size)
68
+ gaussian_1d = torch.exp(-(x_coord - (kernel_size - 1) / 2) ** 2 / (2 * sigma ** 2))
69
+ gaussian_1d = gaussian_1d / gaussian_1d.sum()
70
+ gaussian_2d = gaussian_1d[:, None] * gaussian_1d[None, :]
71
+ kernel = gaussian_2d[None, None, :, :].repeat(channels, 1, 1, 1)
72
+
73
+ return kernel
74
+
75
+ def gaussian_filter(latents, kernel_size=3, sigma=1.0):
76
+ channels = latents.shape[1]
77
+ kernel = gaussian_kernel(kernel_size, sigma, channels).to(latents.device, latents.dtype)
78
+ blurred_latents = F.conv2d(latents, kernel, padding=kernel_size//2, groups=channels)
79
+
80
+ return blurred_latents
81
+
82
+ class DemoFusionSDXLControlNetPipeline(
83
+ DiffusionPipeline, TextualInversionLoaderMixin, LoraLoaderMixin, FromSingleFileMixin
84
+ ):
85
+ r"""
86
+ Pipeline for text-to-image generation using Stable Diffusion XL with ControlNet guidance.
87
+
88
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
89
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
90
+
91
+ The pipeline also inherits the following loading methods:
92
+ - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
93
+ - [`loaders.LoraLoaderMixin.load_lora_weights`] for loading LoRA weights
94
+ - [`loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files
95
+
96
+ Args:
97
+ vae ([`AutoencoderKL`]):
98
+ Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
99
+ text_encoder ([`~transformers.CLIPTextModel`]):
100
+ Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
101
+ text_encoder_2 ([`~transformers.CLIPTextModelWithProjection`]):
102
+ Second frozen text-encoder
103
+ ([laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k)).
104
+ tokenizer ([`~transformers.CLIPTokenizer`]):
105
+ A `CLIPTokenizer` to tokenize text.
106
+ tokenizer_2 ([`~transformers.CLIPTokenizer`]):
107
+ A `CLIPTokenizer` to tokenize text.
108
+ unet ([`UNet2DConditionModel`]):
109
+ A `UNet2DConditionModel` to denoise the encoded image latents.
110
+ controlnet ([`ControlNetModel`] or `List[ControlNetModel]`):
111
+ Provides additional conditioning to the `unet` during the denoising process. If you set multiple
112
+ ControlNets as a list, the outputs from each ControlNet are added together to create one combined
113
+ additional conditioning.
114
+ scheduler ([`SchedulerMixin`]):
115
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
116
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
117
+ force_zeros_for_empty_prompt (`bool`, *optional*, defaults to `"True"`):
118
+ Whether the negative prompt embeddings should always be set to 0. Also see the config of
119
+ `stabilityai/stable-diffusion-xl-base-1-0`.
120
+ add_watermarker (`bool`, *optional*):
121
+ Whether to use the [invisible_watermark](https://github.com/ShieldMnt/invisible-watermark/) library to
122
+ watermark output images. If not defined, it defaults to `True` if the package is installed; otherwise no
123
+ watermarker is used.
124
+ """
125
+ model_cpu_offload_seq = (
126
+ "text_encoder->text_encoder_2->unet->vae" # leave controlnet out on purpose because it iterates with unet
127
+ )
128
+
129
+ def __init__(
130
+ self,
131
+ vae: AutoencoderKL,
132
+ text_encoder: CLIPTextModel,
133
+ text_encoder_2: CLIPTextModelWithProjection,
134
+ tokenizer: CLIPTokenizer,
135
+ tokenizer_2: CLIPTokenizer,
136
+ unet: UNet2DConditionModel,
137
+ controlnet: Union[ControlNetModel, List[ControlNetModel], Tuple[ControlNetModel], MultiControlNetModel],
138
+ scheduler: KarrasDiffusionSchedulers,
139
+ force_zeros_for_empty_prompt: bool = True,
140
+ add_watermarker: Optional[bool] = None,
141
+ ):
142
+ super().__init__()
143
+
144
+ if isinstance(controlnet, (list, tuple)):
145
+ controlnet = MultiControlNetModel(controlnet)
146
+
147
+ self.register_modules(
148
+ vae=vae,
149
+ text_encoder=text_encoder,
150
+ text_encoder_2=text_encoder_2,
151
+ tokenizer=tokenizer,
152
+ tokenizer_2=tokenizer_2,
153
+ unet=unet,
154
+ controlnet=controlnet,
155
+ scheduler=scheduler,
156
+ )
157
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
158
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True)
159
+ self.control_image_processor = VaeImageProcessor(
160
+ vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True, do_normalize=False
161
+ )
162
+ add_watermarker = add_watermarker if add_watermarker is not None else is_invisible_watermark_available()
163
+
164
+ if add_watermarker:
165
+ self.watermark = StableDiffusionXLWatermarker()
166
+ else:
167
+ self.watermark = None
168
+
169
+ self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt)
170
+
171
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_slicing
172
+ def enable_vae_slicing(self):
173
+ r"""
174
+ Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
175
+ compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
176
+ """
177
+ self.vae.enable_slicing()
178
+
179
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_slicing
180
+ def disable_vae_slicing(self):
181
+ r"""
182
+ Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to
183
+ computing decoding in one step.
184
+ """
185
+ self.vae.disable_slicing()
186
+
187
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_tiling
188
+ def enable_vae_tiling(self):
189
+ r"""
190
+ Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
191
+ compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
192
+ processing larger images.
193
+ """
194
+ self.vae.enable_tiling()
195
+
196
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_tiling
197
+ def disable_vae_tiling(self):
198
+ r"""
199
+ Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to
200
+ computing decoding in one step.
201
+ """
202
+ self.vae.disable_tiling()
203
+
204
+ # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.encode_prompt
205
+ def encode_prompt(
206
+ self,
207
+ prompt: str,
208
+ prompt_2: Optional[str] = None,
209
+ device: Optional[torch.device] = None,
210
+ num_images_per_prompt: int = 1,
211
+ do_classifier_free_guidance: bool = True,
212
+ negative_prompt: Optional[str] = None,
213
+ negative_prompt_2: Optional[str] = None,
214
+ prompt_embeds: Optional[torch.FloatTensor] = None,
215
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
216
+ pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
217
+ negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
218
+ lora_scale: Optional[float] = None,
219
+ ):
220
+ r"""
221
+ Encodes the prompt into text encoder hidden states.
222
+
223
+ Args:
224
+ prompt (`str` or `List[str]`, *optional*):
225
+ prompt to be encoded
226
+ prompt_2 (`str` or `List[str]`, *optional*):
227
+ The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
228
+ used in both text-encoders
229
+ device: (`torch.device`):
230
+ torch device
231
+ num_images_per_prompt (`int`):
232
+ number of images that should be generated per prompt
233
+ do_classifier_free_guidance (`bool`):
234
+ whether to use classifier free guidance or not
235
+ negative_prompt (`str` or `List[str]`, *optional*):
236
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
237
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
238
+ less than `1`).
239
+ negative_prompt_2 (`str` or `List[str]`, *optional*):
240
+ The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
241
+ `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders
242
+ prompt_embeds (`torch.FloatTensor`, *optional*):
243
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
244
+ provided, text embeddings will be generated from `prompt` input argument.
245
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
246
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
247
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
248
+ argument.
249
+ pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
250
+ Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
251
+ If not provided, pooled text embeddings will be generated from `prompt` input argument.
252
+ negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
253
+ Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
254
+ weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
255
+ input argument.
256
+ lora_scale (`float`, *optional*):
257
+ A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
258
+ """
259
+ device = device or self._execution_device
260
+
261
+ # set lora scale so that monkey patched LoRA
262
+ # function of text encoder can correctly access it
263
+ if lora_scale is not None and isinstance(self, LoraLoaderMixin):
264
+ self._lora_scale = lora_scale
265
+
266
+ # dynamically adjust the LoRA scale
267
+ adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
268
+ adjust_lora_scale_text_encoder(self.text_encoder_2, lora_scale)
269
+
270
+ if prompt is not None and isinstance(prompt, str):
271
+ batch_size = 1
272
+ elif prompt is not None and isinstance(prompt, list):
273
+ batch_size = len(prompt)
274
+ else:
275
+ batch_size = prompt_embeds.shape[0]
276
+
277
+ # Define tokenizers and text encoders
278
+ tokenizers = [self.tokenizer, self.tokenizer_2] if self.tokenizer is not None else [self.tokenizer_2]
279
+ text_encoders = (
280
+ [self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2]
281
+ )
282
+
283
+ if prompt_embeds is None:
284
+ prompt_2 = prompt_2 or prompt
285
+ # textual inversion: procecss multi-vector tokens if necessary
286
+ prompt_embeds_list = []
287
+ prompts = [prompt, prompt_2]
288
+ for prompt, tokenizer, text_encoder in zip(prompts, tokenizers, text_encoders):
289
+ if isinstance(self, TextualInversionLoaderMixin):
290
+ prompt = self.maybe_convert_prompt(prompt, tokenizer)
291
+
292
+ text_inputs = tokenizer(
293
+ prompt,
294
+ padding="max_length",
295
+ max_length=tokenizer.model_max_length,
296
+ truncation=True,
297
+ return_tensors="pt",
298
+ )
299
+
300
+ text_input_ids = text_inputs.input_ids
301
+ untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
302
+
303
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
304
+ text_input_ids, untruncated_ids
305
+ ):
306
+ removed_text = tokenizer.batch_decode(untruncated_ids[:, tokenizer.model_max_length - 1 : -1])
307
+ logger.warning(
308
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
309
+ f" {tokenizer.model_max_length} tokens: {removed_text}"
310
+ )
311
+
312
+ prompt_embeds = text_encoder(
313
+ text_input_ids.to(device),
314
+ output_hidden_states=True,
315
+ )
316
+
317
+ # We are only ALWAYS interested in the pooled output of the final text encoder
318
+ pooled_prompt_embeds = prompt_embeds[0]
319
+ prompt_embeds = prompt_embeds.hidden_states[-2]
320
+
321
+ prompt_embeds_list.append(prompt_embeds)
322
+
323
+ prompt_embeds = torch.concat(prompt_embeds_list, dim=-1)
324
+
325
+ # get unconditional embeddings for classifier free guidance
326
+ zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt
327
+ if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt:
328
+ negative_prompt_embeds = torch.zeros_like(prompt_embeds)
329
+ negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds)
330
+ elif do_classifier_free_guidance and negative_prompt_embeds is None:
331
+ negative_prompt = negative_prompt or ""
332
+ negative_prompt_2 = negative_prompt_2 or negative_prompt
333
+
334
+ uncond_tokens: List[str]
335
+ if prompt is not None and type(prompt) is not type(negative_prompt):
336
+ raise TypeError(
337
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
338
+ f" {type(prompt)}."
339
+ )
340
+ elif isinstance(negative_prompt, str):
341
+ uncond_tokens = [negative_prompt, negative_prompt_2]
342
+ elif batch_size != len(negative_prompt):
343
+ raise ValueError(
344
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
345
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
346
+ " the batch size of `prompt`."
347
+ )
348
+ else:
349
+ uncond_tokens = [negative_prompt, negative_prompt_2]
350
+
351
+ negative_prompt_embeds_list = []
352
+ for negative_prompt, tokenizer, text_encoder in zip(uncond_tokens, tokenizers, text_encoders):
353
+ if isinstance(self, TextualInversionLoaderMixin):
354
+ negative_prompt = self.maybe_convert_prompt(negative_prompt, tokenizer)
355
+
356
+ max_length = prompt_embeds.shape[1]
357
+ uncond_input = tokenizer(
358
+ negative_prompt,
359
+ padding="max_length",
360
+ max_length=max_length,
361
+ truncation=True,
362
+ return_tensors="pt",
363
+ )
364
+
365
+ negative_prompt_embeds = text_encoder(
366
+ uncond_input.input_ids.to(device),
367
+ output_hidden_states=True,
368
+ )
369
+ # We are only ALWAYS interested in the pooled output of the final text encoder
370
+ negative_pooled_prompt_embeds = negative_prompt_embeds[0]
371
+ negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2]
372
+
373
+ negative_prompt_embeds_list.append(negative_prompt_embeds)
374
+
375
+ negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1)
376
+
377
+ prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device)
378
+ bs_embed, seq_len, _ = prompt_embeds.shape
379
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
380
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
381
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
382
+
383
+ if do_classifier_free_guidance:
384
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
385
+ seq_len = negative_prompt_embeds.shape[1]
386
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device)
387
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
388
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
389
+
390
+ pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
391
+ bs_embed * num_images_per_prompt, -1
392
+ )
393
+ if do_classifier_free_guidance:
394
+ negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
395
+ bs_embed * num_images_per_prompt, -1
396
+ )
397
+
398
+ return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds
399
+
400
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
401
+ def prepare_extra_step_kwargs(self, generator, eta):
402
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
403
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
404
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
405
+ # and should be between [0, 1]
406
+
407
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
408
+ extra_step_kwargs = {}
409
+ if accepts_eta:
410
+ extra_step_kwargs["eta"] = eta
411
+
412
+ # check if the scheduler accepts generator
413
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
414
+ if accepts_generator:
415
+ extra_step_kwargs["generator"] = generator
416
+ return extra_step_kwargs
417
+
418
+ def check_inputs(
419
+ self,
420
+ prompt,
421
+ prompt_2,
422
+ image,
423
+ callback_steps,
424
+ negative_prompt=None,
425
+ negative_prompt_2=None,
426
+ prompt_embeds=None,
427
+ negative_prompt_embeds=None,
428
+ pooled_prompt_embeds=None,
429
+ negative_pooled_prompt_embeds=None,
430
+ controlnet_conditioning_scale=1.0,
431
+ control_guidance_start=0.0,
432
+ control_guidance_end=1.0,
433
+ ):
434
+ if (callback_steps is None) or (
435
+ callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
436
+ ):
437
+ raise ValueError(
438
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
439
+ f" {type(callback_steps)}."
440
+ )
441
+
442
+ if prompt is not None and prompt_embeds is not None:
443
+ raise ValueError(
444
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
445
+ " only forward one of the two."
446
+ )
447
+ elif prompt_2 is not None and prompt_embeds is not None:
448
+ raise ValueError(
449
+ f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
450
+ " only forward one of the two."
451
+ )
452
+ elif prompt is None and prompt_embeds is None:
453
+ raise ValueError(
454
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
455
+ )
456
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
457
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
458
+ elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)):
459
+ raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}")
460
+
461
+ if negative_prompt is not None and negative_prompt_embeds is not None:
462
+ raise ValueError(
463
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
464
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
465
+ )
466
+ elif negative_prompt_2 is not None and negative_prompt_embeds is not None:
467
+ raise ValueError(
468
+ f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:"
469
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
470
+ )
471
+
472
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
473
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
474
+ raise ValueError(
475
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
476
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
477
+ f" {negative_prompt_embeds.shape}."
478
+ )
479
+
480
+ if prompt_embeds is not None and pooled_prompt_embeds is None:
481
+ raise ValueError(
482
+ "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`."
483
+ )
484
+
485
+ if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None:
486
+ raise ValueError(
487
+ "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`."
488
+ )
489
+
490
+ # `prompt` needs more sophisticated handling when there are multiple
491
+ # conditionings.
492
+ if isinstance(self.controlnet, MultiControlNetModel):
493
+ if isinstance(prompt, list):
494
+ logger.warning(
495
+ f"You have {len(self.controlnet.nets)} ControlNets and you have passed {len(prompt)}"
496
+ " prompts. The conditionings will be fixed across the prompts."
497
+ )
498
+
499
+ # Check `image`
500
+ is_compiled = hasattr(F, "scaled_dot_product_attention") and isinstance(
501
+ self.controlnet, torch._dynamo.eval_frame.OptimizedModule
502
+ )
503
+ if (
504
+ isinstance(self.controlnet, ControlNetModel)
505
+ or is_compiled
506
+ and isinstance(self.controlnet._orig_mod, ControlNetModel)
507
+ ):
508
+ self.check_image(image, prompt, prompt_embeds)
509
+ elif (
510
+ isinstance(self.controlnet, MultiControlNetModel)
511
+ or is_compiled
512
+ and isinstance(self.controlnet._orig_mod, MultiControlNetModel)
513
+ ):
514
+ if not isinstance(image, list):
515
+ raise TypeError("For multiple controlnets: `image` must be type `list`")
516
+
517
+ # When `image` is a nested list:
518
+ # (e.g. [[canny_image_1, pose_image_1], [canny_image_2, pose_image_2]])
519
+ elif any(isinstance(i, list) for i in image):
520
+ raise ValueError("A single batch of multiple conditionings are supported at the moment.")
521
+ elif len(image) != len(self.controlnet.nets):
522
+ raise ValueError(
523
+ 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."
524
+ )
525
+
526
+ for image_ in image:
527
+ self.check_image(image_, prompt, prompt_embeds)
528
+ else:
529
+ assert False
530
+
531
+ # Check `controlnet_conditioning_scale`
532
+ if (
533
+ isinstance(self.controlnet, ControlNetModel)
534
+ or is_compiled
535
+ and isinstance(self.controlnet._orig_mod, ControlNetModel)
536
+ ):
537
+ if not isinstance(controlnet_conditioning_scale, float):
538
+ raise TypeError("For single controlnet: `controlnet_conditioning_scale` must be type `float`.")
539
+ elif (
540
+ isinstance(self.controlnet, MultiControlNetModel)
541
+ or is_compiled
542
+ and isinstance(self.controlnet._orig_mod, MultiControlNetModel)
543
+ ):
544
+ if isinstance(controlnet_conditioning_scale, list):
545
+ if any(isinstance(i, list) for i in controlnet_conditioning_scale):
546
+ raise ValueError("A single batch of multiple conditionings are supported at the moment.")
547
+ elif isinstance(controlnet_conditioning_scale, list) and len(controlnet_conditioning_scale) != len(
548
+ self.controlnet.nets
549
+ ):
550
+ raise ValueError(
551
+ "For multiple controlnets: When `controlnet_conditioning_scale` is specified as `list`, it must have"
552
+ " the same length as the number of controlnets"
553
+ )
554
+ else:
555
+ assert False
556
+
557
+ if not isinstance(control_guidance_start, (tuple, list)):
558
+ control_guidance_start = [control_guidance_start]
559
+
560
+ if not isinstance(control_guidance_end, (tuple, list)):
561
+ control_guidance_end = [control_guidance_end]
562
+
563
+ if len(control_guidance_start) != len(control_guidance_end):
564
+ raise ValueError(
565
+ 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."
566
+ )
567
+
568
+ if isinstance(self.controlnet, MultiControlNetModel):
569
+ if len(control_guidance_start) != len(self.controlnet.nets):
570
+ raise ValueError(
571
+ 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)}."
572
+ )
573
+
574
+ for start, end in zip(control_guidance_start, control_guidance_end):
575
+ if start >= end:
576
+ raise ValueError(
577
+ f"control guidance start: {start} cannot be larger or equal to control guidance end: {end}."
578
+ )
579
+ if start < 0.0:
580
+ raise ValueError(f"control guidance start: {start} can't be smaller than 0.")
581
+ if end > 1.0:
582
+ raise ValueError(f"control guidance end: {end} can't be larger than 1.0.")
583
+
584
+ # Copied from diffusers.pipelines.controlnet.pipeline_controlnet.StableDiffusionControlNetPipeline.check_image
585
+ def check_image(self, image, prompt, prompt_embeds):
586
+ image_is_pil = isinstance(image, PIL.Image.Image)
587
+ image_is_tensor = isinstance(image, torch.Tensor)
588
+ image_is_np = isinstance(image, np.ndarray)
589
+ image_is_pil_list = isinstance(image, list) and isinstance(image[0], PIL.Image.Image)
590
+ image_is_tensor_list = isinstance(image, list) and isinstance(image[0], torch.Tensor)
591
+ image_is_np_list = isinstance(image, list) and isinstance(image[0], np.ndarray)
592
+
593
+ if (
594
+ not image_is_pil
595
+ and not image_is_tensor
596
+ and not image_is_np
597
+ and not image_is_pil_list
598
+ and not image_is_tensor_list
599
+ and not image_is_np_list
600
+ ):
601
+ raise TypeError(
602
+ 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)}"
603
+ )
604
+
605
+ if image_is_pil:
606
+ image_batch_size = 1
607
+ else:
608
+ image_batch_size = len(image)
609
+
610
+ if prompt is not None and isinstance(prompt, str):
611
+ prompt_batch_size = 1
612
+ elif prompt is not None and isinstance(prompt, list):
613
+ prompt_batch_size = len(prompt)
614
+ elif prompt_embeds is not None:
615
+ prompt_batch_size = prompt_embeds.shape[0]
616
+
617
+ if image_batch_size != 1 and image_batch_size != prompt_batch_size:
618
+ raise ValueError(
619
+ 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}"
620
+ )
621
+
622
+ # Copied from diffusers.pipelines.controlnet.pipeline_controlnet.StableDiffusionControlNetPipeline.prepare_image
623
+ def prepare_image(
624
+ self,
625
+ image,
626
+ width,
627
+ height,
628
+ batch_size,
629
+ num_images_per_prompt,
630
+ device,
631
+ dtype,
632
+ do_classifier_free_guidance=False,
633
+ guess_mode=False,
634
+ ):
635
+ image = self.control_image_processor.preprocess(image, height=height, width=width).to(dtype=torch.float32)
636
+ image_batch_size = image.shape[0]
637
+
638
+ if image_batch_size == 1:
639
+ repeat_by = batch_size
640
+ else:
641
+ # image batch size is the same as prompt batch size
642
+ repeat_by = num_images_per_prompt
643
+
644
+ image = image.repeat_interleave(repeat_by, dim=0)
645
+
646
+ image = image.to(device=device, dtype=dtype)
647
+
648
+ if do_classifier_free_guidance and not guess_mode:
649
+ image = torch.cat([image] * 2)
650
+
651
+ return image
652
+
653
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
654
+ def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
655
+ shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
656
+ if isinstance(generator, list) and len(generator) != batch_size:
657
+ raise ValueError(
658
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
659
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
660
+ )
661
+
662
+ if latents is None:
663
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
664
+ else:
665
+ latents = latents.to(device)
666
+
667
+ # scale the initial noise by the standard deviation required by the scheduler
668
+ latents = latents * self.scheduler.init_noise_sigma
669
+ return latents
670
+
671
+ # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline._get_add_time_ids
672
+ def _get_add_time_ids(self, original_size, crops_coords_top_left, target_size, dtype):
673
+ add_time_ids = list(original_size + crops_coords_top_left + target_size)
674
+
675
+ passed_add_embed_dim = (
676
+ self.unet.config.addition_time_embed_dim * len(add_time_ids) + self.text_encoder_2.config.projection_dim
677
+ )
678
+ expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features
679
+
680
+ if expected_add_embed_dim != passed_add_embed_dim:
681
+ raise ValueError(
682
+ 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`."
683
+ )
684
+
685
+ add_time_ids = torch.tensor([add_time_ids], dtype=dtype)
686
+ return add_time_ids
687
+
688
+ def get_views(self, height, width, window_size=128, stride=64, random_jitter=False):
689
+ # Here, we define the mappings F_i (see Eq. 7 in the MultiDiffusion paper https://arxiv.org/abs/2302.08113)
690
+ # if panorama's height/width < window_size, num_blocks of height/width should return 1
691
+ height //= self.vae_scale_factor
692
+ width //= self.vae_scale_factor
693
+ num_blocks_height = int((height - window_size) / stride - 1e-6) + 2 if height > window_size else 1
694
+ num_blocks_width = int((width - window_size) / stride - 1e-6) + 2 if width > window_size else 1
695
+ total_num_blocks = int(num_blocks_height * num_blocks_width)
696
+ views = []
697
+ for i in range(total_num_blocks):
698
+ h_start = int((i // num_blocks_width) * stride)
699
+ h_end = h_start + window_size
700
+ w_start = int((i % num_blocks_width) * stride)
701
+ w_end = w_start + window_size
702
+
703
+ if h_end > height:
704
+ h_start = int(h_start + height - h_end)
705
+ h_end = int(height)
706
+ if w_end > width:
707
+ w_start = int(w_start + width - w_end)
708
+ w_end = int(width)
709
+ if h_start < 0:
710
+ h_end = int(h_end - h_start)
711
+ h_start = 0
712
+ if w_start < 0:
713
+ w_end = int(w_end - w_start)
714
+ w_start = 0
715
+
716
+ if random_jitter:
717
+ jitter_range = (window_size - stride) // 4
718
+ w_jitter = 0
719
+ h_jitter = 0
720
+ if (w_start != 0) and (w_end != width):
721
+ w_jitter = random.randint(-jitter_range, jitter_range)
722
+ elif (w_start == 0) and (w_end != width):
723
+ w_jitter = random.randint(-jitter_range, 0)
724
+ elif (w_start != 0) and (w_end == width):
725
+ w_jitter = random.randint(0, jitter_range)
726
+ if (h_start != 0) and (h_end != height):
727
+ h_jitter = random.randint(-jitter_range, jitter_range)
728
+ elif (h_start == 0) and (h_end != height):
729
+ h_jitter = random.randint(-jitter_range, 0)
730
+ elif (h_start != 0) and (h_end == height):
731
+ h_jitter = random.randint(0, jitter_range)
732
+ h_start += (h_jitter + jitter_range)
733
+ h_end += (h_jitter + jitter_range)
734
+ w_start += (w_jitter + jitter_range)
735
+ w_end += (w_jitter + jitter_range)
736
+
737
+ views.append((h_start, h_end, w_start, w_end))
738
+ return views
739
+
740
+ def tiled_decode(self, latents, current_height, current_width):
741
+ sample_size = self.unet.config.sample_size
742
+ core_size = self.unet.config.sample_size // 4
743
+ core_stride = core_size
744
+ pad_size = self.unet.config.sample_size // 8 * 3
745
+ decoder_view_batch_size = 1
746
+
747
+ if self.lowvram:
748
+ core_stride = core_size // 2
749
+ pad_size = core_size
750
+
751
+ views = self.get_views(current_height, current_width, stride=core_stride, window_size=core_size)
752
+ views_batch = [views[i : i + decoder_view_batch_size] for i in range(0, len(views), decoder_view_batch_size)]
753
+ latents_ = F.pad(latents, (pad_size, pad_size, pad_size, pad_size), 'constant', 0)
754
+ image = torch.zeros(latents.size(0), 3, current_height, current_width).to(latents.device)
755
+ count = torch.zeros_like(image).to(latents.device)
756
+ # get the latents corresponding to the current view coordinates
757
+ with self.progress_bar(total=len(views_batch)) as progress_bar:
758
+ for j, batch_view in enumerate(views_batch):
759
+ vb_size = len(batch_view)
760
+ latents_for_view = torch.cat(
761
+ [
762
+ latents_[:, :, h_start:h_end+pad_size*2, w_start:w_end+pad_size*2]
763
+ for h_start, h_end, w_start, w_end in batch_view
764
+ ]
765
+ ).to(self.vae.device)
766
+ image_patch = self.vae.decode(latents_for_view / self.vae.config.scaling_factor, return_dict=False)[0]
767
+ h_start, h_end, w_start, w_end = views[j]
768
+ 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
769
+ 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
770
+ 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)
771
+ count[:, :, h_start:h_end, w_start:w_end] += 1
772
+ progress_bar.update()
773
+ image = image / count
774
+
775
+ return image
776
+
777
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_upscale.StableDiffusionUpscalePipeline.upcast_vae
778
+ def upcast_vae(self):
779
+ dtype = self.vae.dtype
780
+ self.vae.to(dtype=torch.float32)
781
+ use_torch_2_0_or_xformers = isinstance(
782
+ self.vae.decoder.mid_block.attentions[0].processor,
783
+ (
784
+ AttnProcessor2_0,
785
+ XFormersAttnProcessor,
786
+ LoRAXFormersAttnProcessor,
787
+ LoRAAttnProcessor2_0,
788
+ ),
789
+ )
790
+ # if xformers or torch_2_0 is used attention block does not need
791
+ # to be in float32 which can save lots of memory
792
+ if use_torch_2_0_or_xformers:
793
+ self.vae.post_quant_conv.to(dtype)
794
+ self.vae.decoder.conv_in.to(dtype)
795
+ self.vae.decoder.mid_block.to(dtype)
796
+
797
+ @torch.no_grad()
798
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
799
+ def __call__(
800
+ self,
801
+ prompt: Union[str, List[str]] = None,
802
+ prompt_2: Optional[Union[str, List[str]]] = None,
803
+ condition_image: PipelineImageInput = None,
804
+ height: Optional[int] = None,
805
+ width: Optional[int] = None,
806
+ num_inference_steps: int = 50,
807
+ guidance_scale: float = 5.0,
808
+ negative_prompt: Optional[Union[str, List[str]]] = None,
809
+ negative_prompt_2: Optional[Union[str, List[str]]] = None,
810
+ num_images_per_prompt: Optional[int] = 1,
811
+ eta: float = 0.0,
812
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
813
+ latents: Optional[torch.FloatTensor] = None,
814
+ prompt_embeds: Optional[torch.FloatTensor] = None,
815
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
816
+ pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
817
+ negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
818
+ output_type: Optional[str] = "pil",
819
+ return_dict: bool = True,
820
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
821
+ callback_steps: int = 1,
822
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
823
+ controlnet_conditioning_scale: Union[float, List[float]] = 1.0,
824
+ guess_mode: bool = False,
825
+ control_guidance_start: Union[float, List[float]] = 0.0,
826
+ control_guidance_end: Union[float, List[float]] = 1.0,
827
+ original_size: Tuple[int, int] = None,
828
+ crops_coords_top_left: Tuple[int, int] = (0, 0),
829
+ target_size: Tuple[int, int] = None,
830
+ negative_original_size: Optional[Tuple[int, int]] = None,
831
+ negative_crops_coords_top_left: Tuple[int, int] = (0, 0),
832
+ negative_target_size: Optional[Tuple[int, int]] = None,
833
+ ################### DemoFusion specific parameters ####################
834
+ image_lr: Optional[torch.FloatTensor] = None,
835
+ view_batch_size: int = 16,
836
+ multi_decoder: bool = True,
837
+ stride: Optional[int] = 64,
838
+ cosine_scale_1: Optional[float] = 3.,
839
+ cosine_scale_2: Optional[float] = 1.,
840
+ cosine_scale_3: Optional[float] = 1.,
841
+ sigma: Optional[float] = 1.0,
842
+ show_image: bool = False,
843
+ lowvram: bool = False,
844
+ ):
845
+ r"""
846
+ The call function to the pipeline for generation.
847
+
848
+ Args:
849
+ prompt (`str` or `List[str]`, *optional*):
850
+ The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
851
+ prompt_2 (`str` or `List[str]`, *optional*):
852
+ The prompt or prompts to be sent to `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
853
+ used in both text-encoders.
854
+ image (`torch.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.FloatTensor]`, `List[PIL.Image.Image]`, `List[np.ndarray]`,:
855
+ `List[List[torch.FloatTensor]]`, `List[List[np.ndarray]]` or `List[List[PIL.Image.Image]]`):
856
+ The ControlNet input condition to provide guidance to the `unet` for generation. If the type is
857
+ specified as `torch.FloatTensor`, it is passed to ControlNet as is. `PIL.Image.Image` can also be
858
+ accepted as an image. The dimensions of the output image defaults to `image`'s dimensions. If height
859
+ and/or width are passed, `image` is resized accordingly. If multiple ControlNets are specified in
860
+ `init`, images must be passed as a list such that each element of the list can be correctly batched for
861
+ input to a single ControlNet.
862
+ height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
863
+ The height in pixels of the generated image. Anything below 512 pixels won't work well for
864
+ [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0)
865
+ and checkpoints that are not specifically fine-tuned on low resolutions.
866
+ width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
867
+ The width in pixels of the generated image. Anything below 512 pixels won't work well for
868
+ [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0)
869
+ and checkpoints that are not specifically fine-tuned on low resolutions.
870
+ num_inference_steps (`int`, *optional*, defaults to 50):
871
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
872
+ expense of slower inference.
873
+ guidance_scale (`float`, *optional*, defaults to 5.0):
874
+ A higher guidance scale value encourages the model to generate images closely linked to the text
875
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
876
+ negative_prompt (`str` or `List[str]`, *optional*):
877
+ The prompt or prompts to guide what to not include in image generation. If not defined, you need to
878
+ pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
879
+ negative_prompt_2 (`str` or `List[str]`, *optional*):
880
+ The prompt or prompts to guide what to not include in image generation. This is sent to `tokenizer_2`
881
+ and `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders.
882
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
883
+ The number of images to generate per prompt.
884
+ eta (`float`, *optional*, defaults to 0.0):
885
+ Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
886
+ to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
887
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
888
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
889
+ generation deterministic.
890
+ latents (`torch.FloatTensor`, *optional*):
891
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
892
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
893
+ tensor is generated by sampling using the supplied random `generator`.
894
+ prompt_embeds (`torch.FloatTensor`, *optional*):
895
+ Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
896
+ provided, text embeddings are generated from the `prompt` input argument.
897
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
898
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
899
+ not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
900
+ pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
901
+ Pre-generated pooled text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
902
+ not provided, pooled text embeddings are generated from `prompt` input argument.
903
+ negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
904
+ Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs (prompt
905
+ weighting). If not provided, pooled `negative_prompt_embeds` are generated from `negative_prompt` input
906
+ argument.
907
+ output_type (`str`, *optional*, defaults to `"pil"`):
908
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
909
+ return_dict (`bool`, *optional*, defaults to `True`):
910
+ Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
911
+ plain tuple.
912
+ callback (`Callable`, *optional*):
913
+ A function that calls every `callback_steps` steps during inference. The function is called with the
914
+ following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
915
+ callback_steps (`int`, *optional*, defaults to 1):
916
+ The frequency at which the `callback` function is called. If not specified, the callback is called at
917
+ every step.
918
+ cross_attention_kwargs (`dict`, *optional*):
919
+ A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
920
+ [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
921
+ controlnet_conditioning_scale (`float` or `List[float]`, *optional*, defaults to 1.0):
922
+ The outputs of the ControlNet are multiplied by `controlnet_conditioning_scale` before they are added
923
+ to the residual in the original `unet`. If multiple ControlNets are specified in `init`, you can set
924
+ the corresponding scale as a list.
925
+ guess_mode (`bool`, *optional*, defaults to `False`):
926
+ The ControlNet encoder tries to recognize the content of the input image even if you remove all
927
+ prompts. A `guidance_scale` value between 3.0 and 5.0 is recommended.
928
+ control_guidance_start (`float` or `List[float]`, *optional*, defaults to 0.0):
929
+ The percentage of total steps at which the ControlNet starts applying.
930
+ control_guidance_end (`float` or `List[float]`, *optional*, defaults to 1.0):
931
+ The percentage of total steps at which the ControlNet stops applying.
932
+ original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
933
+ If `original_size` is not the same as `target_size` the image will appear to be down- or upsampled.
934
+ `original_size` defaults to `(width, height)` if not specified. Part of SDXL's micro-conditioning as
935
+ explained in section 2.2 of
936
+ [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
937
+ crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
938
+ `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position
939
+ `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting
940
+ `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of
941
+ [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
942
+ target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
943
+ For most cases, `target_size` should be set to the desired height and width of the generated image. If
944
+ not specified it will default to `(width, height)`. Part of SDXL's micro-conditioning as explained in
945
+ section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
946
+ negative_original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
947
+ To negatively condition the generation process based on a specific image resolution. Part of SDXL's
948
+ micro-conditioning as explained in section 2.2 of
949
+ [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
950
+ information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
951
+ negative_crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
952
+ To negatively condition the generation process based on a specific crop coordinates. Part of SDXL's
953
+ micro-conditioning as explained in section 2.2 of
954
+ [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
955
+ information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
956
+ negative_target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
957
+ To negatively condition the generation process based on a target image resolution. It should be as same
958
+ as the `target_size` for most cases. Part of SDXL's micro-conditioning as explained in section 2.2 of
959
+ [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
960
+ information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
961
+ ################### DemoFusion specific parameters ####################
962
+ image_lr (`torch.FloatTensor`, *optional*, , defaults to None):
963
+ Low-resolution image input for upscaling. If provided, DemoFusion will encode it as the initial latent representation.
964
+ view_batch_size (`int`, defaults to 16):
965
+ The batch size for multiple denoising paths. Typically, a larger batch size can result in higher
966
+ efficiency but comes with increased GPU memory requirements.
967
+ multi_decoder (`bool`, defaults to True):
968
+ Determine whether to use a tiled decoder. Generally, when the resolution exceeds 3072x3072,
969
+ a tiled decoder becomes necessary.
970
+ stride (`int`, defaults to 64):
971
+ The stride of moving local patches. A smaller stride is better for alleviating seam issues,
972
+ but it also introduces additional computational overhead and inference time.
973
+ cosine_scale_1 (`float`, defaults to 3):
974
+ Control the strength of skip-residual. For specific impacts, please refer to Appendix C
975
+ in the DemoFusion paper.
976
+ cosine_scale_2 (`float`, defaults to 1):
977
+ Control the strength of dilated sampling. For specific impacts, please refer to Appendix C
978
+ in the DemoFusion paper.
979
+ cosine_scale_3 (`float`, defaults to 1):
980
+ Control the strength of the gaussion filter. For specific impacts, please refer to Appendix C
981
+ in the DemoFusion paper.
982
+ sigma (`float`, defaults to 1):
983
+ The standard value of the gaussian filter.
984
+ show_image (`bool`, defaults to False):
985
+ Determine whether to show intermediate results during generation.
986
+ lowvram (`bool`, defaults to False):
987
+ Try to fit in 8 Gb of VRAM, with xformers installed.
988
+ Examples:
989
+
990
+ Returns:
991
+ [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
992
+ If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned,
993
+ otherwise a `tuple` is returned containing the output images.
994
+ """
995
+ controlnet = self.controlnet._orig_mod if is_compiled_module(self.controlnet) else self.controlnet
996
+
997
+ # align format for control guidance
998
+ if not isinstance(control_guidance_start, list) and isinstance(control_guidance_end, list):
999
+ control_guidance_start = len(control_guidance_end) * [control_guidance_start]
1000
+ elif not isinstance(control_guidance_end, list) and isinstance(control_guidance_start, list):
1001
+ control_guidance_end = len(control_guidance_start) * [control_guidance_end]
1002
+ elif not isinstance(control_guidance_start, list) and not isinstance(control_guidance_end, list):
1003
+ mult = len(controlnet.nets) if isinstance(controlnet, MultiControlNetModel) else 1
1004
+ control_guidance_start, control_guidance_end = mult * [control_guidance_start], mult * [
1005
+ control_guidance_end
1006
+ ]
1007
+
1008
+ # 0. Default height and width to unet
1009
+ height = height or self.unet.config.sample_size * self.vae_scale_factor
1010
+ width = width or self.unet.config.sample_size * self.vae_scale_factor
1011
+
1012
+ x1_size = self.unet.config.sample_size * self.vae_scale_factor
1013
+
1014
+ height_scale = height / x1_size
1015
+ width_scale = width / x1_size
1016
+ scale_num = int(max(height_scale, width_scale))
1017
+ aspect_ratio = min(height_scale, width_scale) / max(height_scale, width_scale)
1018
+
1019
+ original_size = original_size or (height, width)
1020
+ target_size = target_size or (height, width)
1021
+
1022
+ # 1. Check inputs. Raise error if not correct
1023
+ self.check_inputs(
1024
+ prompt,
1025
+ prompt_2,
1026
+ condition_image,
1027
+ callback_steps,
1028
+ negative_prompt,
1029
+ negative_prompt_2,
1030
+ prompt_embeds,
1031
+ negative_prompt_embeds,
1032
+ pooled_prompt_embeds,
1033
+ negative_pooled_prompt_embeds,
1034
+ controlnet_conditioning_scale,
1035
+ control_guidance_start,
1036
+ control_guidance_end,
1037
+ )
1038
+
1039
+ # 2. Define call parameters
1040
+ if prompt is not None and isinstance(prompt, str):
1041
+ batch_size = 1
1042
+ elif prompt is not None and isinstance(prompt, list):
1043
+ batch_size = len(prompt)
1044
+ else:
1045
+ batch_size = prompt_embeds.shape[0]
1046
+
1047
+ device = self._execution_device
1048
+ self.lowvram = lowvram
1049
+ if self.lowvram:
1050
+ self.vae.cpu()
1051
+ self.unet.cpu()
1052
+ self.text_encoder.to(device)
1053
+ self.text_encoder_2.to(device)
1054
+ image_lr.cpu()
1055
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
1056
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
1057
+ # corresponds to doing no classifier free guidance.
1058
+ do_classifier_free_guidance = guidance_scale > 1.0
1059
+
1060
+ if isinstance(controlnet, MultiControlNetModel) and isinstance(controlnet_conditioning_scale, float):
1061
+ controlnet_conditioning_scale = [controlnet_conditioning_scale] * len(controlnet.nets)
1062
+
1063
+ global_pool_conditions = (
1064
+ controlnet.config.global_pool_conditions
1065
+ if isinstance(controlnet, ControlNetModel)
1066
+ else controlnet.nets[0].config.global_pool_conditions
1067
+ )
1068
+ guess_mode = guess_mode or global_pool_conditions
1069
+
1070
+ # 3. Encode input prompt
1071
+ text_encoder_lora_scale = (
1072
+ cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None
1073
+ )
1074
+ (
1075
+ prompt_embeds,
1076
+ negative_prompt_embeds,
1077
+ pooled_prompt_embeds,
1078
+ negative_pooled_prompt_embeds,
1079
+ ) = self.encode_prompt(
1080
+ prompt,
1081
+ prompt_2,
1082
+ device,
1083
+ num_images_per_prompt,
1084
+ do_classifier_free_guidance,
1085
+ negative_prompt,
1086
+ negative_prompt_2,
1087
+ prompt_embeds=prompt_embeds,
1088
+ negative_prompt_embeds=negative_prompt_embeds,
1089
+ pooled_prompt_embeds=pooled_prompt_embeds,
1090
+ negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
1091
+ lora_scale=text_encoder_lora_scale,
1092
+ )
1093
+
1094
+ # 4. Prepare image
1095
+ if isinstance(controlnet, ControlNetModel):
1096
+ condition_image = self.prepare_image(
1097
+ image=condition_image,
1098
+ width=width // scale_num,
1099
+ height=height // scale_num,
1100
+ batch_size=batch_size * num_images_per_prompt,
1101
+ num_images_per_prompt=num_images_per_prompt,
1102
+ device=device,
1103
+ dtype=controlnet.dtype,
1104
+ do_classifier_free_guidance=do_classifier_free_guidance,
1105
+ guess_mode=guess_mode,
1106
+ )
1107
+ # height, width = condition_image.shape[-2:]
1108
+ # condition_image.shape ([2, 3, 1024, 1024])
1109
+ elif isinstance(controlnet, MultiControlNetModel):
1110
+ condition_images = []
1111
+
1112
+ for image_ in condition_image:
1113
+ image_ = self.prepare_image(
1114
+ image=image_,
1115
+ width=width // scale_num,
1116
+ height=height // scale_num,
1117
+ batch_size=batch_size * num_images_per_prompt,
1118
+ num_images_per_prompt=num_images_per_prompt,
1119
+ device=device,
1120
+ dtype=controlnet.dtype,
1121
+ do_classifier_free_guidance=do_classifier_free_guidance,
1122
+ guess_mode=guess_mode,
1123
+ )
1124
+
1125
+ condition_images.append(image_)
1126
+
1127
+ condition_image = condition_images
1128
+ # height, width = condition_image[0].shape[-2:]
1129
+ else:
1130
+ assert False
1131
+
1132
+ # 5. Prepare timesteps
1133
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
1134
+ timesteps = self.scheduler.timesteps
1135
+
1136
+ # 6. Prepare latent variables
1137
+ num_channels_latents = self.unet.config.in_channels
1138
+ latents = self.prepare_latents(
1139
+ batch_size * num_images_per_prompt,
1140
+ num_channels_latents,
1141
+ height // scale_num,
1142
+ width // scale_num,
1143
+ prompt_embeds.dtype,
1144
+ device,
1145
+ generator,
1146
+ latents,
1147
+ )
1148
+
1149
+ # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
1150
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
1151
+
1152
+ # 7.1 Create tensor stating which controlnets to keep
1153
+ controlnet_keep = []
1154
+ for i in range(len(timesteps)):
1155
+ keeps = [
1156
+ 1.0 - float(i / len(timesteps) < s or (i + 1) / len(timesteps) > e)
1157
+ for s, e in zip(control_guidance_start, control_guidance_end)
1158
+ ]
1159
+ controlnet_keep.append(keeps[0] if isinstance(controlnet, ControlNetModel) else keeps)
1160
+
1161
+ # 7.2 Prepare added time ids & embeddings
1162
+ if isinstance(condition_image, list):
1163
+ original_size = original_size or condition_image[0].shape[-2:]
1164
+ else:
1165
+ original_size = original_size or condition_image.shape[-2:]
1166
+ target_size = target_size or (height, width)
1167
+
1168
+ add_text_embeds = pooled_prompt_embeds
1169
+ add_time_ids = self._get_add_time_ids(
1170
+ original_size, crops_coords_top_left, target_size, dtype=prompt_embeds.dtype
1171
+ )
1172
+
1173
+ if negative_original_size is not None and negative_target_size is not None:
1174
+ negative_add_time_ids = self._get_add_time_ids(
1175
+ negative_original_size,
1176
+ negative_crops_coords_top_left,
1177
+ negative_target_size,
1178
+ dtype=prompt_embeds.dtype,
1179
+ )
1180
+ else:
1181
+ negative_add_time_ids = add_time_ids
1182
+
1183
+ if do_classifier_free_guidance:
1184
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
1185
+ add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0)
1186
+ add_time_ids = torch.cat([negative_add_time_ids, add_time_ids], dim=0)
1187
+
1188
+ prompt_embeds = prompt_embeds.to(device)
1189
+ add_text_embeds = add_text_embeds.to(device)
1190
+ add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1)
1191
+
1192
+
1193
+
1194
+ # 8. Denoising loop
1195
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
1196
+
1197
+ output_images = []
1198
+
1199
+ ###################################################### Phase Initialization ########################################################
1200
+
1201
+ if self.lowvram:
1202
+ self.text_encoder.cpu()
1203
+ self.text_encoder_2.cpu()
1204
+
1205
+ if image_lr == None:
1206
+ print("### Phase 1 Denoising ###")
1207
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
1208
+ for i, t in enumerate(timesteps):
1209
+
1210
+ if self.lowvram:
1211
+ self.vae.cpu()
1212
+ self.unet.to(device)
1213
+
1214
+ latents_for_view = latents
1215
+
1216
+ # expand the latents if we are doing classifier free guidance
1217
+ latent_model_input = (
1218
+ latents.repeat_interleave(2, dim=0)
1219
+ if do_classifier_free_guidance
1220
+ else latents
1221
+ )
1222
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
1223
+
1224
+
1225
+ added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids}
1226
+
1227
+ # controlnet(s) inference
1228
+ if guess_mode and do_classifier_free_guidance:
1229
+ # Infer ControlNet only for the conditional batch.
1230
+ control_model_input = latents
1231
+ control_model_input = self.scheduler.scale_model_input(control_model_input, t)
1232
+ controlnet_prompt_embeds = prompt_embeds.chunk(2)[1]
1233
+ controlnet_added_cond_kwargs = {
1234
+ "text_embeds": add_text_embeds.chunk(2)[1],
1235
+ "time_ids": add_time_ids.chunk(2)[1],
1236
+ }
1237
+ else:
1238
+ control_model_input = latent_model_input
1239
+ controlnet_prompt_embeds = prompt_embeds
1240
+ controlnet_added_cond_kwargs = added_cond_kwargs
1241
+
1242
+ if isinstance(controlnet_keep[i], list):
1243
+ cond_scale = [c * s for c, s in zip(controlnet_conditioning_scale, controlnet_keep[i])]
1244
+ else:
1245
+ controlnet_cond_scale = controlnet_conditioning_scale
1246
+ if isinstance(controlnet_cond_scale, list):
1247
+ controlnet_cond_scale = controlnet_cond_scale[0]
1248
+ cond_scale = controlnet_cond_scale * controlnet_keep[i]
1249
+
1250
+ # print(condition_image.shape, control_model_input.shape, controlnet_prompt_embeds.shape, t, cond_scale, guess_mode)
1251
+ # print(controlnet_added_cond_kwargs["text_embeds"].shape, controlnet_added_cond_kwargs["time_ids"].shape)
1252
+ down_block_res_samples, mid_block_res_sample = self.controlnet(
1253
+ control_model_input,
1254
+ t,
1255
+ encoder_hidden_states=controlnet_prompt_embeds,
1256
+ controlnet_cond=condition_image,
1257
+ conditioning_scale=cond_scale,
1258
+ guess_mode=guess_mode,
1259
+ added_cond_kwargs=controlnet_added_cond_kwargs,
1260
+ return_dict=False,
1261
+ )
1262
+
1263
+ if guess_mode and do_classifier_free_guidance:
1264
+ # Infered ControlNet only for the conditional batch.
1265
+ # To apply the output of ControlNet to both the unconditional and conditional batches,
1266
+ # add 0 to the unconditional batch to keep it unchanged.
1267
+ down_block_res_samples = [torch.cat([torch.zeros_like(d), d]) for d in down_block_res_samples]
1268
+ mid_block_res_sample = torch.cat([torch.zeros_like(mid_block_res_sample), mid_block_res_sample])
1269
+
1270
+ # predict the noise residual
1271
+ noise_pred = self.unet(
1272
+ latent_model_input,
1273
+ t,
1274
+ encoder_hidden_states=prompt_embeds,
1275
+ cross_attention_kwargs=cross_attention_kwargs,
1276
+ down_block_additional_residuals=down_block_res_samples,
1277
+ mid_block_additional_residual=mid_block_res_sample,
1278
+ added_cond_kwargs=added_cond_kwargs,
1279
+ return_dict=False,
1280
+ )[0]
1281
+
1282
+ # perform guidance
1283
+ if do_classifier_free_guidance:
1284
+ noise_pred_uncond, noise_pred_text = noise_pred[::2], noise_pred[1::2]
1285
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
1286
+
1287
+ # compute the previous noisy sample x_t -> x_t-1
1288
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
1289
+
1290
+ # call the callback, if provided
1291
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
1292
+ progress_bar.update()
1293
+ if callback is not None and i % callback_steps == 0:
1294
+ step_idx = i // getattr(self.scheduler, "order", 1)
1295
+ callback(step_idx, t, latents)
1296
+ else:
1297
+ print("### Encoding Real Image ###")
1298
+ latents = self.vae.encode(image_lr)
1299
+ latents = latents.latent_dist.sample() * self.vae.config.scaling_factor
1300
+
1301
+ anchor_mean = latents.mean()
1302
+ anchor_std = latents.std()
1303
+ if self.lowvram:
1304
+ latents = latents.cpu()
1305
+ torch.cuda.empty_cache()
1306
+ if not output_type == "latent":
1307
+ # make sure the VAE is in float32 mode, as it overflows in float16
1308
+ needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
1309
+
1310
+ if self.lowvram:
1311
+ needs_upcasting = False # use madebyollin/sdxl-vae-fp16-fix in lowvram mode!
1312
+ self.unet.cpu()
1313
+ self.vae.to(device)
1314
+
1315
+ if needs_upcasting:
1316
+ self.upcast_vae()
1317
+ latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
1318
+ if self.lowvram and multi_decoder:
1319
+ current_width_height = self.unet.config.sample_size * self.vae_scale_factor
1320
+ image = self.tiled_decode(latents, current_width_height, current_width_height)
1321
+ else:
1322
+ image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
1323
+ # cast back to fp16 if needed
1324
+ if needs_upcasting:
1325
+ self.vae.to(dtype=torch.float16)
1326
+
1327
+ image = self.image_processor.postprocess(image, output_type=output_type)
1328
+ if show_image:
1329
+ plt.figure(figsize=(10, 10))
1330
+ plt.imshow(image[0])
1331
+ plt.axis('off') # Turn off axis numbers and ticks
1332
+ plt.show()
1333
+ output_images.append(image[0])
1334
+
1335
+ ####################################################### Phase Upscaling #####################################################
1336
+ if image_lr == None:
1337
+ starting_scale = 2
1338
+ else:
1339
+ starting_scale = 1
1340
+ for current_scale_num in range(starting_scale, scale_num + 1):
1341
+ if self.lowvram:
1342
+ latents = latents.to(device)
1343
+ self.unet.to(device)
1344
+ torch.cuda.empty_cache()
1345
+ print("### Phase {} Denoising ###".format(current_scale_num))
1346
+ current_height = self.unet.config.sample_size * self.vae_scale_factor * current_scale_num
1347
+ current_width = self.unet.config.sample_size * self.vae_scale_factor * current_scale_num
1348
+ if height > width:
1349
+ current_width = int(current_width * aspect_ratio)
1350
+ else:
1351
+ current_height = int(current_height * aspect_ratio)
1352
+
1353
+ latents = F.interpolate(latents, size=(int(current_height / self.vae_scale_factor), int(current_width / self.vae_scale_factor)), mode='bicubic')
1354
+ condition_image = F.interpolate(condition_image, size=(current_height, current_width), mode='bicubic')
1355
+
1356
+ noise_latents = []
1357
+ noise = torch.randn_like(latents)
1358
+ for timestep in timesteps:
1359
+ noise_latent = self.scheduler.add_noise(latents, noise, timestep.unsqueeze(0))
1360
+ noise_latents.append(noise_latent)
1361
+ latents = noise_latents[0]
1362
+
1363
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
1364
+ for i, t in enumerate(timesteps):
1365
+ count = torch.zeros_like(latents)
1366
+ value = torch.zeros_like(latents)
1367
+ cosine_factor = 0.5 * (1 + torch.cos(torch.pi * (self.scheduler.config.num_train_timesteps - t) / self.scheduler.config.num_train_timesteps)).cpu()
1368
+
1369
+ c1 = cosine_factor ** cosine_scale_1
1370
+ latents = latents * (1 - c1) + noise_latents[i] * c1
1371
+
1372
+ ############################################# MultiDiffusion #############################################
1373
+
1374
+ views = self.get_views(current_height, current_width, stride=stride, window_size=self.unet.config.sample_size, random_jitter=True)
1375
+ views_batch = [views[i : i + view_batch_size] for i in range(0, len(views), view_batch_size)]
1376
+
1377
+ jitter_range = (self.unet.config.sample_size - stride) // 4
1378
+ latents_ = F.pad(latents, (jitter_range, jitter_range, jitter_range, jitter_range), 'constant', 0)
1379
+ 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)
1380
+
1381
+ count_local = torch.zeros_like(latents_)
1382
+ value_local = torch.zeros_like(latents_)
1383
+
1384
+ for j, batch_view in enumerate(views_batch):
1385
+ vb_size = len(batch_view)
1386
+
1387
+ # get the latents corresponding to the current view coordinates
1388
+ latents_for_view = torch.cat(
1389
+ [
1390
+ latents_[:, :, h_start:h_end, w_start:w_end]
1391
+ for h_start, h_end, w_start, w_end in batch_view
1392
+ ]
1393
+ )
1394
+ condition_image_for_view = torch.cat(
1395
+ [
1396
+ 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]
1397
+ for h_start, h_end, w_start, w_end in batch_view
1398
+ ]
1399
+ )
1400
+
1401
+ # expand the latents if we are doing classifier free guidance
1402
+ latent_model_input = latents_for_view
1403
+ latent_model_input = (
1404
+ latent_model_input.repeat_interleave(2, dim=0)
1405
+ if do_classifier_free_guidance
1406
+ else latent_model_input
1407
+ )
1408
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
1409
+
1410
+ condition_image_input = condition_image_for_view
1411
+ condition_image_input = (
1412
+ condition_image_input.repeat_interleave(2, dim=0)
1413
+ if do_classifier_free_guidance
1414
+ else condition_image_input
1415
+ )
1416
+
1417
+ prompt_embeds_input = torch.cat([prompt_embeds] * vb_size)
1418
+ add_text_embeds_input = torch.cat([add_text_embeds] * vb_size)
1419
+ add_time_ids_input = []
1420
+ for h_start, h_end, w_start, w_end in batch_view:
1421
+ add_time_ids_ = add_time_ids.clone()
1422
+ add_time_ids_[:, 2] = h_start * self.vae_scale_factor
1423
+ add_time_ids_[:, 3] = w_start * self.vae_scale_factor
1424
+ add_time_ids_input.append(add_time_ids_)
1425
+ add_time_ids_input = torch.cat(add_time_ids_input)
1426
+
1427
+ added_cond_kwargs = {"text_embeds": add_text_embeds_input, "time_ids": add_time_ids_input}
1428
+
1429
+ # controlnet(s) inference
1430
+ if guess_mode and do_classifier_free_guidance:
1431
+ # Infer ControlNet only for the conditional batch.
1432
+ control_model_input = latent_model_input
1433
+ control_model_input = self.scheduler.scale_model_input(control_model_input, t)
1434
+ controlnet_prompt_embeds = prompt_embeds_input.chunk(2)[1]
1435
+ controlnet_added_cond_kwargs = {
1436
+ "text_embeds": add_text_embeds_input.chunk(2)[1],
1437
+ "time_ids": add_time_ids_input.chunk(2)[1],
1438
+ }
1439
+ else:
1440
+ control_model_input = latent_model_input
1441
+ controlnet_prompt_embeds = prompt_embeds_input
1442
+ controlnet_added_cond_kwargs = added_cond_kwargs
1443
+
1444
+ if isinstance(controlnet_keep[i], list):
1445
+ cond_scale = [c * s for c, s in zip(controlnet_conditioning_scale, controlnet_keep[i])]
1446
+ else:
1447
+ controlnet_cond_scale = controlnet_conditioning_scale
1448
+ if isinstance(controlnet_cond_scale, list):
1449
+ controlnet_cond_scale = controlnet_cond_scale[0]
1450
+ cond_scale = controlnet_cond_scale * controlnet_keep[i]
1451
+
1452
+ down_block_res_samples, mid_block_res_sample = self.controlnet(
1453
+ control_model_input,
1454
+ t,
1455
+ encoder_hidden_states=controlnet_prompt_embeds,
1456
+ controlnet_cond=condition_image_input,
1457
+ conditioning_scale=cond_scale,
1458
+ guess_mode=guess_mode,
1459
+ added_cond_kwargs=controlnet_added_cond_kwargs,
1460
+ return_dict=False,
1461
+ )
1462
+
1463
+ if guess_mode and do_classifier_free_guidance:
1464
+ # Infered ControlNet only for the conditional batch.
1465
+ # To apply the output of ControlNet to both the unconditional and conditional batches,
1466
+ # add 0 to the unconditional batch to keep it unchanged.
1467
+ down_block_res_samples = [torch.cat([torch.zeros_like(d), d]) for d in down_block_res_samples]
1468
+ mid_block_res_sample = torch.cat([torch.zeros_like(mid_block_res_sample), mid_block_res_sample])
1469
+
1470
+ # predict the noise residual
1471
+ noise_pred = self.unet(
1472
+ latent_model_input,
1473
+ t,
1474
+ encoder_hidden_states=prompt_embeds_input,
1475
+ cross_attention_kwargs=cross_attention_kwargs,
1476
+ down_block_additional_residuals=down_block_res_samples,
1477
+ mid_block_additional_residual=mid_block_res_sample,
1478
+ added_cond_kwargs=added_cond_kwargs,
1479
+ return_dict=False,
1480
+ )[0]
1481
+
1482
+ if do_classifier_free_guidance:
1483
+ noise_pred_uncond, noise_pred_text = noise_pred[::2], noise_pred[1::2]
1484
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) * 1
1485
+
1486
+ # compute the previous noisy sample x_t -> x_t-1
1487
+ self.scheduler._init_step_index(t)
1488
+ latents_denoised_batch = self.scheduler.step(
1489
+ noise_pred, t, latents_for_view, **extra_step_kwargs, return_dict=False)[0]
1490
+
1491
+ # extract value from batch
1492
+ for latents_view_denoised, (h_start, h_end, w_start, w_end) in zip(
1493
+ latents_denoised_batch.chunk(vb_size), batch_view
1494
+ ):
1495
+ value_local[:, :, h_start:h_end, w_start:w_end] += latents_view_denoised
1496
+ count_local[:, :, h_start:h_end, w_start:w_end] += 1
1497
+
1498
+ value_local = value_local[: ,:, jitter_range: jitter_range + current_height // self.vae_scale_factor, jitter_range: jitter_range + current_width // self.vae_scale_factor]
1499
+ count_local = count_local[: ,:, jitter_range: jitter_range + current_height // self.vae_scale_factor, jitter_range: jitter_range + current_width // self.vae_scale_factor]
1500
+
1501
+ c2 = cosine_factor ** cosine_scale_2
1502
+
1503
+ value += value_local / count_local * (1 - c2)
1504
+ count += torch.ones_like(value_local) * (1 - c2)
1505
+
1506
+ ############################################# Dilated Sampling #############################################
1507
+
1508
+ h_pad = (current_scale_num - (latents.size(2) % current_scale_num)) % current_scale_num
1509
+ w_pad = (current_scale_num - (latents.size(3) % current_scale_num)) % current_scale_num
1510
+ latents_ = F.pad(latents, (w_pad, 0, h_pad, 0), 'constant', 0)
1511
+
1512
+ count_global = torch.zeros_like(latents_)
1513
+ value_global = torch.zeros_like(latents_)
1514
+
1515
+ c3 = 0.99 * cosine_factor ** cosine_scale_3 + 1e-2
1516
+ std_, mean_ = latents_.std(), latents_.mean()
1517
+ latents_gaussian = gaussian_filter(latents_, kernel_size=(2*current_scale_num-1), sigma=sigma*c3)
1518
+ latents_gaussian = (latents_gaussian - latents_gaussian.mean()) / latents_gaussian.std() * std_ + mean_
1519
+
1520
+ latents_for_view = []
1521
+ for h in range(current_scale_num):
1522
+ for w in range(current_scale_num):
1523
+ latents_for_view.append(latents_[:, :, h::current_scale_num, w::current_scale_num])
1524
+ latents_for_view = torch.cat(latents_for_view)
1525
+
1526
+ latents_for_view_gaussian = []
1527
+ for h in range(current_scale_num):
1528
+ for w in range(current_scale_num):
1529
+ latents_for_view_gaussian.append(latents_gaussian[:, :, h::current_scale_num, w::current_scale_num])
1530
+ latents_for_view_gaussian = torch.cat(latents_for_view_gaussian)
1531
+
1532
+ condition_image_for_view = []
1533
+ for h in range(current_scale_num):
1534
+ for w in range(current_scale_num):
1535
+ 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)
1536
+ 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])
1537
+ condition_image_for_view = torch.cat(condition_image_for_view)
1538
+
1539
+ vb_size = latents_for_view.size(0)
1540
+
1541
+ # expand the latents if we are doing classifier free guidance
1542
+ latent_model_input = latents_for_view_gaussian
1543
+ latent_model_input = (
1544
+ latent_model_input.repeat_interleave(2, dim=0)
1545
+ if do_classifier_free_guidance
1546
+ else latent_model_input
1547
+ )
1548
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
1549
+
1550
+ condition_image_input = condition_image_for_view
1551
+ condition_image_input = (
1552
+ condition_image_input.repeat_interleave(2, dim=0)
1553
+ if do_classifier_free_guidance
1554
+ else condition_image_input
1555
+ )
1556
+
1557
+ prompt_embeds_input = torch.cat([prompt_embeds] * vb_size)
1558
+ add_text_embeds_input = torch.cat([add_text_embeds] * vb_size)
1559
+ add_time_ids_input = torch.cat([add_time_ids] * vb_size)
1560
+
1561
+ added_cond_kwargs = {"text_embeds": add_text_embeds_input, "time_ids": add_time_ids_input}
1562
+
1563
+ # controlnet(s) inference
1564
+ if guess_mode and do_classifier_free_guidance:
1565
+ # Infer ControlNet only for the conditional batch.
1566
+ control_model_input = latent_model_input
1567
+ control_model_input = self.scheduler.scale_model_input(control_model_input, t)
1568
+ controlnet_prompt_embeds = prompt_embeds_input.chunk(2)[1]
1569
+ controlnet_added_cond_kwargs = {
1570
+ "text_embeds": add_text_embeds_input.chunk(2)[1],
1571
+ "time_ids": add_time_ids_input.chunk(2)[1],
1572
+ }
1573
+ else:
1574
+ control_model_input = latent_model_input
1575
+ controlnet_prompt_embeds = prompt_embeds_input
1576
+ controlnet_added_cond_kwargs = added_cond_kwargs
1577
+
1578
+ if isinstance(controlnet_keep[i], list):
1579
+ cond_scale = [c * s for c, s in zip(controlnet_conditioning_scale, controlnet_keep[i])]
1580
+ else:
1581
+ controlnet_cond_scale = controlnet_conditioning_scale
1582
+ if isinstance(controlnet_cond_scale, list):
1583
+ controlnet_cond_scale = controlnet_cond_scale[0]
1584
+ cond_scale = controlnet_cond_scale * controlnet_keep[i]
1585
+
1586
+ down_block_res_samples, mid_block_res_sample = self.controlnet(
1587
+ control_model_input,
1588
+ t,
1589
+ encoder_hidden_states=controlnet_prompt_embeds,
1590
+ controlnet_cond=condition_image_input,
1591
+ conditioning_scale=cond_scale,
1592
+ guess_mode=guess_mode,
1593
+ added_cond_kwargs=controlnet_added_cond_kwargs,
1594
+ return_dict=False,
1595
+ )
1596
+
1597
+ if guess_mode and do_classifier_free_guidance:
1598
+ # Infered ControlNet only for the conditional batch.
1599
+ # To apply the output of ControlNet to both the unconditional and conditional batches,
1600
+ # add 0 to the unconditional batch to keep it unchanged.
1601
+ down_block_res_samples = [torch.cat([torch.zeros_like(d), d]) for d in down_block_res_samples]
1602
+ mid_block_res_sample = torch.cat([torch.zeros_like(mid_block_res_sample), mid_block_res_sample])
1603
+
1604
+ # predict the noise residual
1605
+ noise_pred = self.unet(
1606
+ latent_model_input,
1607
+ t,
1608
+ encoder_hidden_states=prompt_embeds_input,
1609
+ cross_attention_kwargs=cross_attention_kwargs,
1610
+ down_block_additional_residuals=down_block_res_samples,
1611
+ mid_block_additional_residual=mid_block_res_sample,
1612
+ added_cond_kwargs=added_cond_kwargs,
1613
+ return_dict=False,
1614
+ )[0]
1615
+
1616
+ if do_classifier_free_guidance:
1617
+ noise_pred_uncond, noise_pred_text = noise_pred[::2], noise_pred[1::2]
1618
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
1619
+
1620
+ # extract value from batch
1621
+ for h in range(current_scale_num):
1622
+ for w in range(current_scale_num):
1623
+ noise_pred_ = noise_pred.chunk(vb_size)[h*current_scale_num+w]
1624
+ value_global[:, :, h::current_scale_num, w::current_scale_num] += noise_pred_
1625
+ count_global[:, :, h::current_scale_num, w::current_scale_num] += 1
1626
+
1627
+ # compute the previous noisy sample x_t -> x_t-1
1628
+ self.scheduler._init_step_index(t)
1629
+ value_global = self.scheduler.step(
1630
+ value_global, t, latents_, **extra_step_kwargs, return_dict=False)[0]
1631
+
1632
+ c2 = cosine_factor ** cosine_scale_2
1633
+
1634
+ value_global = value_global[: ,:, h_pad:, w_pad:]
1635
+
1636
+ value += value_global * c2
1637
+ count += torch.ones_like(value_global) * c2
1638
+
1639
+ ###########################################################
1640
+
1641
+ latents = torch.where(count > 0, value / count, value)
1642
+
1643
+ # call the callback, if provided
1644
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
1645
+ progress_bar.update()
1646
+ if callback is not None and i % callback_steps == 0:
1647
+ step_idx = i // getattr(self.scheduler, "order", 1)
1648
+ callback(step_idx, t, latents)
1649
+
1650
+ #########################################################################################################################################
1651
+
1652
+ latents = (latents - latents.mean()) / latents.std() * anchor_std + anchor_mean
1653
+ if self.lowvram:
1654
+ latents = latents.cpu()
1655
+ torch.cuda.empty_cache()
1656
+ if not output_type == "latent":
1657
+ # make sure the VAE is in float32 mode, as it overflows in float16
1658
+ needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
1659
+
1660
+ if self.lowvram:
1661
+ needs_upcasting = False # use madebyollin/sdxl-vae-fp16-fix in lowvram mode!
1662
+ self.unet.cpu()
1663
+ self.vae.to(device)
1664
+
1665
+ if needs_upcasting:
1666
+ self.upcast_vae()
1667
+ latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
1668
+
1669
+ print("### Phase {} Decoding ###".format(current_scale_num))
1670
+ if multi_decoder:
1671
+ image = self.tiled_decode(latents, current_height, current_width)
1672
+ else:
1673
+ image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
1674
+
1675
+ # cast back to fp16 if needed
1676
+ if needs_upcasting:
1677
+ self.vae.to(dtype=torch.float16)
1678
+ else:
1679
+ image = latents
1680
+
1681
+ if not output_type == "latent":
1682
+ image = self.image_processor.postprocess(image, output_type=output_type)
1683
+ if show_image:
1684
+ plt.figure(figsize=(10, 10))
1685
+ plt.imshow(image[0])
1686
+ plt.axis('off') # Turn off axis numbers and ticks
1687
+ plt.show()
1688
+ output_images.append(image[0])
1689
+
1690
+ # Offload all models
1691
+ self.maybe_free_model_hooks()
1692
+
1693
+ return output_images
1694
+
1695
+ # Overrride to properly handle the loading and unloading of the additional text encoder.
1696
+ def load_lora_weights(self, pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]], **kwargs):
1697
+ # We could have accessed the unet config from `lora_state_dict()` too. We pass
1698
+ # it here explicitly to be able to tell that it's coming from an SDXL
1699
+ # pipeline.
1700
+
1701
+ # Remove any existing hooks.
1702
+ if is_accelerate_available() and is_accelerate_version(">=", "0.17.0.dev0"):
1703
+ from accelerate.hooks import AlignDevicesHook, CpuOffload, remove_hook_from_module
1704
+ else:
1705
+ raise ImportError("Offloading requires `accelerate v0.17.0` or higher.")
1706
+
1707
+ is_model_cpu_offload = False
1708
+ is_sequential_cpu_offload = False
1709
+ recursive = False
1710
+ for _, component in self.components.items():
1711
+ if isinstance(component, torch.nn.Module):
1712
+ if hasattr(component, "_hf_hook"):
1713
+ is_model_cpu_offload = isinstance(getattr(component, "_hf_hook"), CpuOffload)
1714
+ is_sequential_cpu_offload = isinstance(getattr(component, "_hf_hook"), AlignDevicesHook)
1715
+ logger.info(
1716
+ "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."
1717
+ )
1718
+ recursive = is_sequential_cpu_offload
1719
+ remove_hook_from_module(component, recurse=recursive)
1720
+ state_dict, network_alphas = self.lora_state_dict(
1721
+ pretrained_model_name_or_path_or_dict,
1722
+ unet_config=self.unet.config,
1723
+ **kwargs,
1724
+ )
1725
+ self.load_lora_into_unet(state_dict, network_alphas=network_alphas, unet=self.unet)
1726
+
1727
+ text_encoder_state_dict = {k: v for k, v in state_dict.items() if "text_encoder." in k}
1728
+ if len(text_encoder_state_dict) > 0:
1729
+ self.load_lora_into_text_encoder(
1730
+ text_encoder_state_dict,
1731
+ network_alphas=network_alphas,
1732
+ text_encoder=self.text_encoder,
1733
+ prefix="text_encoder",
1734
+ lora_scale=self.lora_scale,
1735
+ )
1736
+
1737
+ text_encoder_2_state_dict = {k: v for k, v in state_dict.items() if "text_encoder_2." in k}
1738
+ if len(text_encoder_2_state_dict) > 0:
1739
+ self.load_lora_into_text_encoder(
1740
+ text_encoder_2_state_dict,
1741
+ network_alphas=network_alphas,
1742
+ text_encoder=self.text_encoder_2,
1743
+ prefix="text_encoder_2",
1744
+ lora_scale=self.lora_scale,
1745
+ )
1746
+
1747
+ # Offload back.
1748
+ if is_model_cpu_offload:
1749
+ self.enable_model_cpu_offload()
1750
+ elif is_sequential_cpu_offload:
1751
+ self.enable_sequential_cpu_offload()
1752
+
1753
+ @classmethod
1754
+ def save_lora_weights(
1755
+ self,
1756
+ save_directory: Union[str, os.PathLike],
1757
+ unet_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None,
1758
+ text_encoder_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None,
1759
+ text_encoder_2_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None,
1760
+ is_main_process: bool = True,
1761
+ weight_name: str = None,
1762
+ save_function: Callable = None,
1763
+ safe_serialization: bool = True,
1764
+ ):
1765
+ state_dict = {}
1766
+
1767
+ def pack_weights(layers, prefix):
1768
+ layers_weights = layers.state_dict() if isinstance(layers, torch.nn.Module) else layers
1769
+ layers_state_dict = {f"{prefix}.{module_name}": param for module_name, param in layers_weights.items()}
1770
+ return layers_state_dict
1771
+
1772
+ if not (unet_lora_layers or text_encoder_lora_layers or text_encoder_2_lora_layers):
1773
+ raise ValueError(
1774
+ "You must pass at least one of `unet_lora_layers`, `text_encoder_lora_layers` or `text_encoder_2_lora_layers`."
1775
+ )
1776
+
1777
+ if unet_lora_layers:
1778
+ state_dict.update(pack_weights(unet_lora_layers, "unet"))
1779
+
1780
+ if text_encoder_lora_layers and text_encoder_2_lora_layers:
1781
+ state_dict.update(pack_weights(text_encoder_lora_layers, "text_encoder"))
1782
+ state_dict.update(pack_weights(text_encoder_2_lora_layers, "text_encoder_2"))
1783
+
1784
+ self.write_lora_layers(
1785
+ state_dict=state_dict,
1786
+ save_directory=save_directory,
1787
+ is_main_process=is_main_process,
1788
+ weight_name=weight_name,
1789
+ save_function=save_function,
1790
+ safe_serialization=safe_serialization,
1791
+ )
1792
+
1793
+ def _remove_text_encoder_monkey_patch(self):
1794
+ self._remove_text_encoder_monkey_patch_classmethod(self.text_encoder)
1795
+ self._remove_text_encoder_monkey_patch_classmethod(self.text_encoder_2)
requirements.txt ADDED
@@ -0,0 +1,11 @@
 
 
 
 
 
 
 
 
 
 
 
 
1
+ diffusers~=0.21.4
2
+ torch~=2.1.0
3
+ scipy~=1.11.3
4
+ omegaconf~=2.3.0
5
+ accelerate~=0.23.0
6
+ transformers~=4.34.0
7
+ tqdm
8
+ einops
9
+ matplotlib
10
+ gradio
11
+ gradio_imageslider