├── .gitignore
├── LICENSE
├── README.md
├── __init__.py
├── assets
├── 0.png
├── 1.png
├── applications.png
├── compare-a.png
├── compare-b.png
└── compare-c.png
├── checkpoints
└── __put_instant_id_checkpoints_here__
├── download_models.py
├── examples
├── kaifu_resize.png
├── musk_resize.jpeg
├── sam_resize.png
├── schmidhuber_resize.png
└── yann-lecun_resize.jpg
├── gradio_demo
├── app.py
├── requirements.txt
└── style_template.py
├── infer.py
├── ip_adapter
├── attention_processor.py
├── resampler.py
└── utils.py
├── models
└── __put_antelopev2_here__
├── nodes.py
├── pipeline_stable_diffusion_xl_instantid.py
└── requirements.txt
/.gitignore:
--------------------------------------------------------------------------------
1 | # Byte-compiled / optimized / DLL files
2 | __pycache__/
3 | *.py[cod]
4 | *$py.class
5 |
6 | # C extensions
7 | *.so
8 |
9 | # Distribution / packaging
10 | .Python
11 | build/
12 | develop-eggs/
13 | dist/
14 | downloads/
15 | eggs/
16 | .eggs/
17 | lib/
18 | lib64/
19 | parts/
20 | sdist/
21 | var/
22 | wheels/
23 | share/python-wheels/
24 | *.egg-info/
25 | .installed.cfg
26 | *.egg
27 | MANIFEST
28 |
29 | # PyInstaller
30 | # Usually these files are written by a python script from a template
31 | # before PyInstaller builds the exe, so as to inject date/other infos into it.
32 | *.manifest
33 | *.spec
34 |
35 | # Installer logs
36 | pip-log.txt
37 | pip-delete-this-directory.txt
38 |
39 | # Unit test / coverage reports
40 | htmlcov/
41 | .tox/
42 | .nox/
43 | .coverage
44 | .coverage.*
45 | .cache
46 | nosetests.xml
47 | coverage.xml
48 | *.cover
49 | *.py,cover
50 | .hypothesis/
51 | .pytest_cache/
52 | cover/
53 |
54 | # Translations
55 | *.mo
56 | *.pot
57 |
58 | # Django stuff:
59 | *.log
60 | local_settings.py
61 | db.sqlite3
62 | db.sqlite3-journal
63 |
64 | # Flask stuff:
65 | instance/
66 | .webassets-cache
67 |
68 | # Scrapy stuff:
69 | .scrapy
70 |
71 | # Sphinx documentation
72 | docs/_build/
73 |
74 | # PyBuilder
75 | .pybuilder/
76 | target/
77 |
78 | # Jupyter Notebook
79 | .ipynb_checkpoints
80 |
81 | # IPython
82 | profile_default/
83 | ipython_config.py
84 |
85 | # pyenv
86 | # For a library or package, you might want to ignore these files since the code is
87 | # intended to run in multiple environments; otherwise, check them in:
88 | # .python-version
89 |
90 | # pipenv
91 | # According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
92 | # However, in case of collaboration, if having platform-specific dependencies or dependencies
93 | # having no cross-platform support, pipenv may install dependencies that don't work, or not
94 | # install all needed dependencies.
95 | #Pipfile.lock
96 |
97 | # poetry
98 | # Similar to Pipfile.lock, it is generally recommended to include poetry.lock in version control.
99 | # This is especially recommended for binary packages to ensure reproducibility, and is more
100 | # commonly ignored for libraries.
101 | # https://python-poetry.org/docs/basic-usage/#commit-your-poetrylock-file-to-version-control
102 | #poetry.lock
103 |
104 | # pdm
105 | # Similar to Pipfile.lock, it is generally recommended to include pdm.lock in version control.
106 | #pdm.lock
107 | # pdm stores project-wide configurations in .pdm.toml, but it is recommended to not include it
108 | # in version control.
109 | # https://pdm.fming.dev/#use-with-ide
110 | .pdm.toml
111 |
112 | # PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm
113 | __pypackages__/
114 |
115 | # Celery stuff
116 | celerybeat-schedule
117 | celerybeat.pid
118 |
119 | # SageMath parsed files
120 | *.sage.py
121 |
122 | # Environments
123 | .env
124 | .venv
125 | env/
126 | venv/
127 | ENV/
128 | env.bak/
129 | venv.bak/
130 |
131 | # Spyder project settings
132 | .spyderproject
133 | .spyproject
134 |
135 | # Rope project settings
136 | .ropeproject
137 |
138 | # mkdocs documentation
139 | /site
140 |
141 | # mypy
142 | .mypy_cache/
143 | .dmypy.json
144 | dmypy.json
145 |
146 | # Pyre type checker
147 | .pyre/
148 |
149 | # pytype static type analyzer
150 | .pytype/
151 |
152 | # Cython debug symbols
153 | cython_debug/
154 |
155 | # PyCharm
156 | # JetBrains specific template is maintained in a separate JetBrains.gitignore that can
157 | # be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore
158 | # and can be added to the global gitignore or merged into this file. For a more nuclear
159 | # option (not recommended) you can uncomment the following to ignore the entire idea folder.
160 | #.idea/
161 |
--------------------------------------------------------------------------------
/LICENSE:
--------------------------------------------------------------------------------
1 | Apache License
2 | Version 2.0, January 2004
3 | http://www.apache.org/licenses/
4 |
5 | TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
6 |
7 | 1. Definitions.
8 |
9 | "License" shall mean the terms and conditions for use, reproduction,
10 | and distribution as defined by Sections 1 through 9 of this document.
11 |
12 | "Licensor" shall mean the copyright owner or entity authorized by
13 | the copyright owner that is granting the License.
14 |
15 | "Legal Entity" shall mean the union of the acting entity and all
16 | other entities that control, are controlled by, or are under common
17 | control with that entity. For the purposes of this definition,
18 | "control" means (i) the power, direct or indirect, to cause the
19 | direction or management of such entity, whether by contract or
20 | otherwise, or (ii) ownership of fifty percent (50%) or more of the
21 | outstanding shares, or (iii) beneficial ownership of such entity.
22 |
23 | "You" (or "Your") shall mean an individual or Legal Entity
24 | exercising permissions granted by this License.
25 |
26 | "Source" form shall mean the preferred form for making modifications,
27 | including but not limited to software source code, documentation
28 | source, and configuration files.
29 |
30 | "Object" form shall mean any form resulting from mechanical
31 | transformation or translation of a Source form, including but
32 | not limited to compiled object code, generated documentation,
33 | and conversions to other media types.
34 |
35 | "Work" shall mean the work of authorship, whether in Source or
36 | Object form, made available under the License, as indicated by a
37 | copyright notice that is included in or attached to the work
38 | (an example is provided in the Appendix below).
39 |
40 | "Derivative Works" shall mean any work, whether in Source or Object
41 | form, that is based on (or derived from) the Work and for which the
42 | editorial revisions, annotations, elaborations, or other modifications
43 | represent, as a whole, an original work of authorship. For the purposes
44 | of this License, Derivative Works shall not include works that remain
45 | separable from, or merely link (or bind by name) to the interfaces of,
46 | the Work and Derivative Works thereof.
47 |
48 | "Contribution" shall mean any work of authorship, including
49 | the original version of the Work and any modifications or additions
50 | to that Work or Derivative Works thereof, that is intentionally
51 | submitted to Licensor for inclusion in the Work by the copyright owner
52 | or by an individual or Legal Entity authorized to submit on behalf of
53 | the copyright owner. For the purposes of this definition, "submitted"
54 | means any form of electronic, verbal, or written communication sent
55 | to the Licensor or its representatives, including but not limited to
56 | communication on electronic mailing lists, source code control systems,
57 | and issue tracking systems that are managed by, or on behalf of, the
58 | Licensor for the purpose of discussing and improving the Work, but
59 | excluding communication that is conspicuously marked or otherwise
60 | designated in writing by the copyright owner as "Not a Contribution."
61 |
62 | "Contributor" shall mean Licensor and any individual or Legal Entity
63 | on behalf of whom a Contribution has been received by Licensor and
64 | subsequently incorporated within the Work.
65 |
66 | 2. Grant of Copyright License. Subject to the terms and conditions of
67 | this License, each Contributor hereby grants to You a perpetual,
68 | worldwide, non-exclusive, no-charge, royalty-free, irrevocable
69 | copyright license to reproduce, prepare Derivative Works of,
70 | publicly display, publicly perform, sublicense, and distribute the
71 | Work and such Derivative Works in Source or Object form.
72 |
73 | 3. Grant of Patent License. Subject to the terms and conditions of
74 | this License, each Contributor hereby grants to You a perpetual,
75 | worldwide, non-exclusive, no-charge, royalty-free, irrevocable
76 | (except as stated in this section) patent license to make, have made,
77 | use, offer to sell, sell, import, and otherwise transfer the Work,
78 | where such license applies only to those patent claims licensable
79 | by such Contributor that are necessarily infringed by their
80 | Contribution(s) alone or by combination of their Contribution(s)
81 | with the Work to which such Contribution(s) was submitted. If You
82 | institute patent litigation against any entity (including a
83 | cross-claim or counterclaim in a lawsuit) alleging that the Work
84 | or a Contribution incorporated within the Work constitutes direct
85 | or contributory patent infringement, then any patent licenses
86 | granted to You under this License for that Work shall terminate
87 | as of the date such litigation is filed.
88 |
89 | 4. Redistribution. You may reproduce and distribute copies of the
90 | Work or Derivative Works thereof in any medium, with or without
91 | modifications, and in Source or Object form, provided that You
92 | meet the following conditions:
93 |
94 | (a) You must give any other recipients of the Work or
95 | Derivative Works a copy of this License; and
96 |
97 | (b) You must cause any modified files to carry prominent notices
98 | stating that You changed the files; and
99 |
100 | (c) You must retain, in the Source form of any Derivative Works
101 | that You distribute, all copyright, patent, trademark, and
102 | attribution notices from the Source form of the Work,
103 | excluding those notices that do not pertain to any part of
104 | the Derivative Works; and
105 |
106 | (d) If the Work includes a "NOTICE" text file as part of its
107 | distribution, then any Derivative Works that You distribute must
108 | include a readable copy of the attribution notices contained
109 | within such NOTICE file, excluding those notices that do not
110 | pertain to any part of the Derivative Works, in at least one
111 | of the following places: within a NOTICE text file distributed
112 | as part of the Derivative Works; within the Source form or
113 | documentation, if provided along with the Derivative Works; or,
114 | within a display generated by the Derivative Works, if and
115 | wherever such third-party notices normally appear. The contents
116 | of the NOTICE file are for informational purposes only and
117 | do not modify the License. You may add Your own attribution
118 | notices within Derivative Works that You distribute, alongside
119 | or as an addendum to the NOTICE text from the Work, provided
120 | that such additional attribution notices cannot be construed
121 | as modifying the License.
122 |
123 | You may add Your own copyright statement to Your modifications and
124 | may provide additional or different license terms and conditions
125 | for use, reproduction, or distribution of Your modifications, or
126 | for any such Derivative Works as a whole, provided Your use,
127 | reproduction, and distribution of the Work otherwise complies with
128 | the conditions stated in this License.
129 |
130 | 5. Submission of Contributions. Unless You explicitly state otherwise,
131 | any Contribution intentionally submitted for inclusion in the Work
132 | by You to the Licensor shall be under the terms and conditions of
133 | this License, without any additional terms or conditions.
134 | Notwithstanding the above, nothing herein shall supersede or modify
135 | the terms of any separate license agreement you may have executed
136 | with Licensor regarding such Contributions.
137 |
138 | 6. Trademarks. This License does not grant permission to use the trade
139 | names, trademarks, service marks, or product names of the Licensor,
140 | except as required for reasonable and customary use in describing the
141 | origin of the Work and reproducing the content of the NOTICE file.
142 |
143 | 7. Disclaimer of Warranty. Unless required by applicable law or
144 | agreed to in writing, Licensor provides the Work (and each
145 | Contributor provides its Contributions) on an "AS IS" BASIS,
146 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
147 | implied, including, without limitation, any warranties or conditions
148 | of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
149 | PARTICULAR PURPOSE. You are solely responsible for determining the
150 | appropriateness of using or redistributing the Work and assume any
151 | risks associated with Your exercise of permissions under this License.
152 |
153 | 8. Limitation of Liability. In no event and under no legal theory,
154 | whether in tort (including negligence), contract, or otherwise,
155 | unless required by applicable law (such as deliberate and grossly
156 | negligent acts) or agreed to in writing, shall any Contributor be
157 | liable to You for damages, including any direct, indirect, special,
158 | incidental, or consequential damages of any character arising as a
159 | result of this License or out of the use or inability to use the
160 | Work (including but not limited to damages for loss of goodwill,
161 | work stoppage, computer failure or malfunction, or any and all
162 | other commercial damages or losses), even if such Contributor
163 | has been advised of the possibility of such damages.
164 |
165 | 9. Accepting Warranty or Additional Liability. While redistributing
166 | the Work or Derivative Works thereof, You may choose to offer,
167 | and charge a fee for, acceptance of support, warranty, indemnity,
168 | or other liability obligations and/or rights consistent with this
169 | License. However, in accepting such obligations, You may act only
170 | on Your own behalf and on Your sole responsibility, not on behalf
171 | of any other Contributor, and only if You agree to indemnify,
172 | defend, and hold each Contributor harmless for any liability
173 | incurred by, or claims asserted against, such Contributor by reason
174 | of your accepting any such warranty or additional liability.
175 |
176 | END OF TERMS AND CONDITIONS
177 |
178 | APPENDIX: How to apply the Apache License to your work.
179 |
180 | To apply the Apache License to your work, attach the following
181 | boilerplate notice, with the fields enclosed by brackets "[]"
182 | replaced with your own identifying information. (Don't include
183 | the brackets!) The text should be enclosed in the appropriate
184 | comment syntax for the file format. We also recommend that a
185 | file or class name and description of purpose be included on the
186 | same "printed page" as the copyright notice for easier
187 | identification within third-party archives.
188 |
189 | Copyright [yyyy] [name of copyright owner]
190 |
191 | Licensed under the Apache License, Version 2.0 (the "License");
192 | you may not use this file except in compliance with the License.
193 | You may obtain a copy of the License at
194 |
195 | http://www.apache.org/licenses/LICENSE-2.0
196 |
197 | Unless required by applicable law or agreed to in writing, software
198 | distributed under the License is distributed on an "AS IS" BASIS,
199 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
200 | See the License for the specific language governing permissions and
201 | limitations under the License.
202 |
--------------------------------------------------------------------------------
/README.md:
--------------------------------------------------------------------------------
1 | # ComfyUI InstantID
2 |
3 | This is a thin wrapper custom node for [Instant ID](https://github.com/InstantID/InstantID). It's providing basic testing interface for playing around with Instant ID functions. Forgive me for not implementing stepping progress indicator.
4 |
5 | It's not following ComfyUI module design nicely, but I just want to set it up for quick testing. Hope IPAdapterPlus will do better integrating to ComfyUI ecosystems...
6 |
7 | A little more explanation: Yes, I know it's great to break down nodes; but it's diffuser based implementation and its inputs / outputs are not compatible with existing other nodes. Even if you break down nodes, those nodes are just connecting each others within the group. Let's wait for better IPAdapterPlus implementation instead of introducing yet another bunch of fancy nodes just for one purpose.
8 |
9 | ## Install
10 |
11 | Just as other custom nodes:
12 | ```
13 | cd ComfyUI/custom_nodes/
14 | git clone https://github.com/huxiuhan/ComfyUI-InstantID.git
15 | pip install -r requirements.txt
16 | ```
17 |
18 | ## Download Models
19 |
20 | ## Download
21 |
22 | You can directly download the model from [Huggingface](https://huggingface.co/InstantX/InstantID).
23 | You also can download the model in python script:
24 |
25 | ```python
26 | from huggingface_hub import hf_hub_download
27 | hf_hub_download(repo_id="InstantX/InstantID", filename="ControlNetModel/config.json", local_dir="./checkpoints")
28 | hf_hub_download(repo_id="InstantX/InstantID", filename="ControlNetModel/diffusion_pytorch_model.safetensors", local_dir="./checkpoints")
29 | hf_hub_download(repo_id="InstantX/InstantID", filename="ip-adapter.bin", local_dir="./checkpoints")
30 | ```
31 |
32 | If you cannot access to Huggingface, you can use [hf-mirror](https://hf-mirror.com/) to download models.
33 | ```python
34 | export HF_ENDPOINT=https://hf-mirror.com
35 | huggingface-cli download --resume-download InstantX/InstantID --local-dir checkpoints
36 | ```
37 |
38 | For face encoder, you need to manually download via this [URL](https://github.com/deepinsight/insightface/issues/1896#issuecomment-1023867304) to `models/antelopev2` as the default link is invalid. Once you have prepared all models, the folder tree should be like:
39 |
40 | ```
41 | .
42 | ├── models
43 | ├── checkpoints
44 | ├── ip_adapter
45 | ├── pipeline_stable_diffusion_xl_instantid.py
46 | └── README.md
47 | ```
48 |
49 | ## Usage
50 |
51 | Choose a SDXL base ckpt. You can also try SDXL Turbo with 4 steps, very efficient for fast testing.
52 |
53 | First time loading usually takes more than 60s, but the node will try its best to cache models.
54 |
55 | 
56 |
57 | 
58 |
59 | ## Original Project
60 | 
61 | 
62 | 
63 | 
64 |
65 | **InstantID : Zero-shot Identity-Preserving Generation in Seconds**
66 |
67 | InstantID is a new state-of-the-art tuning-free method to achieve ID-Preserving generation with only single image, supporting various downstream tasks.
68 |
69 | 
70 |
71 | ## Release
72 | - [2024/1/22] 🔥 We release the [pre-trained checkpoints](https://huggingface.co/InstantX/InstantID), [inference code](https://github.com/InstantID/InstantID/blob/main/infer.py) and [gradio demo](https://huggingface.co/spaces/InstantX/InstantID)!
73 | - [2024/1/15] 🔥 We release the technical report.
74 | - [2023/12/11] 🔥 We launch the project page.
75 |
76 | ## Demos
77 |
78 | ### Stylized Synthesis
79 |
80 |
81 | 
82 |
83 |
84 | ### Comparison with Previous Works
85 |
86 |
87 | 
88 |
89 |
90 | Comparison with existing tuning-free state-of-the-art techniques. InstantID achieves better fidelity and retain good text editability (faces and styles blend better).
91 |
92 |
93 | 
94 |
95 |
96 | Comparison with pre-trained character LoRAs. We don't need multiple images and still can achieve competitive results as LoRAs without any training.
97 |
98 |
99 | 
100 |
101 |
102 | Comparison with InsightFace Swapper (also known as ROOP or Refactor). However, in non-realistic style, our work is more flexible on the integration of face and background.
103 |
104 | ## Usage Tips
105 | - For higher similarity, increase the weight of controlnet_conditioning_scale (IdentityNet) and ip_adapter_scale (Adapter).
106 | - For over-saturation, decrease the ip_adapter_scale. If not work, decrease controlnet_conditioning_scale.
107 | - For higher text control ability, decrease ip_adapter_scale.
108 | - For specific styles, choose corresponding base model makes differences.
109 |
110 | ## Acknowledgements
111 | - Our work is highly inspired by [IP-Adapter](https://github.com/tencent-ailab/IP-Adapter) and [ControlNet](https://github.com/lllyasviel/ControlNet). Thanks for their great works!
112 | - Thanks to the HuggingFace team for their generous GPU support!
113 |
114 | ## Disclaimer
115 | This project is released under [Apache License](https://github.com/InstantID/InstantID?tab=Apache-2.0-1-ov-file#readme) and aims to positively impact the field of AI-driven image generation. Users are granted the freedom to create images using this tool, but they are obligated to comply with local laws and utilize it responsibly. The developers will not assume any responsibility for potential misuse by users.
116 |
117 | ## Cite
118 | If you find InstantID useful for your research and applications, please cite us using this BibTeX:
119 |
120 | ```bibtex
121 | @article{wang2024instantid,
122 | title={InstantID: Zero-shot Identity-Preserving Generation in Seconds},
123 | author={Wang, Qixun and Bai, Xu and Wang, Haofan and Qin, Zekui and Chen, Anthony},
124 | journal={arXiv preprint arXiv:2401.07519},
125 | year={2024}
126 | }
127 |
--------------------------------------------------------------------------------
/__init__.py:
--------------------------------------------------------------------------------
1 | import os
2 | from .nodes import InstantIDSampler
3 |
4 | NODE_CLASS_MAPPINGS = {
5 | "Instant ID Sampler": InstantIDSampler,
6 | }
7 | NODE_DISPLAY_NAME_MAPPINGS = {
8 | "Instant ID Sampler": "Instant ID Sampler For SDXL",
9 | }
10 |
--------------------------------------------------------------------------------
/assets/0.png:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/huxiuhan/ComfyUI-InstantID/c725820fcae81cfb7bddd6e80d5d19df634eb050/assets/0.png
--------------------------------------------------------------------------------
/assets/1.png:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/huxiuhan/ComfyUI-InstantID/c725820fcae81cfb7bddd6e80d5d19df634eb050/assets/1.png
--------------------------------------------------------------------------------
/assets/applications.png:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/huxiuhan/ComfyUI-InstantID/c725820fcae81cfb7bddd6e80d5d19df634eb050/assets/applications.png
--------------------------------------------------------------------------------
/assets/compare-a.png:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/huxiuhan/ComfyUI-InstantID/c725820fcae81cfb7bddd6e80d5d19df634eb050/assets/compare-a.png
--------------------------------------------------------------------------------
/assets/compare-b.png:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/huxiuhan/ComfyUI-InstantID/c725820fcae81cfb7bddd6e80d5d19df634eb050/assets/compare-b.png
--------------------------------------------------------------------------------
/assets/compare-c.png:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/huxiuhan/ComfyUI-InstantID/c725820fcae81cfb7bddd6e80d5d19df634eb050/assets/compare-c.png
--------------------------------------------------------------------------------
/checkpoints/__put_instant_id_checkpoints_here__:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/huxiuhan/ComfyUI-InstantID/c725820fcae81cfb7bddd6e80d5d19df634eb050/checkpoints/__put_instant_id_checkpoints_here__
--------------------------------------------------------------------------------
/download_models.py:
--------------------------------------------------------------------------------
1 | from huggingface_hub import hf_hub_download
2 | hf_hub_download(repo_id="InstantX/InstantID", filename="ControlNetModel/config.json", local_dir="./checkpoints")
3 | hf_hub_download(repo_id="InstantX/InstantID", filename="ControlNetModel/diffusion_pytorch_model.safetensors", local_dir="./checkpoints")
4 | hf_hub_download(repo_id="InstantX/InstantID", filename="ip-adapter.bin", local_dir="./checkpoints")
--------------------------------------------------------------------------------
/examples/kaifu_resize.png:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/huxiuhan/ComfyUI-InstantID/c725820fcae81cfb7bddd6e80d5d19df634eb050/examples/kaifu_resize.png
--------------------------------------------------------------------------------
/examples/musk_resize.jpeg:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/huxiuhan/ComfyUI-InstantID/c725820fcae81cfb7bddd6e80d5d19df634eb050/examples/musk_resize.jpeg
--------------------------------------------------------------------------------
/examples/sam_resize.png:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/huxiuhan/ComfyUI-InstantID/c725820fcae81cfb7bddd6e80d5d19df634eb050/examples/sam_resize.png
--------------------------------------------------------------------------------
/examples/schmidhuber_resize.png:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/huxiuhan/ComfyUI-InstantID/c725820fcae81cfb7bddd6e80d5d19df634eb050/examples/schmidhuber_resize.png
--------------------------------------------------------------------------------
/examples/yann-lecun_resize.jpg:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/huxiuhan/ComfyUI-InstantID/c725820fcae81cfb7bddd6e80d5d19df634eb050/examples/yann-lecun_resize.jpg
--------------------------------------------------------------------------------
/gradio_demo/app.py:
--------------------------------------------------------------------------------
1 | import os
2 | import cv2
3 | import math
4 | import torch
5 | import random
6 | import numpy as np
7 |
8 | import PIL
9 | from PIL import Image
10 |
11 | import diffusers
12 | from diffusers.utils import load_image
13 | from diffusers.models import ControlNetModel
14 |
15 | import insightface
16 | from insightface.app import FaceAnalysis
17 |
18 | from style_template import styles
19 | from pipeline_stable_diffusion_xl_instantid import StableDiffusionXLInstantIDPipeline
20 |
21 | import gradio as gr
22 |
23 | # global variable
24 | MAX_SEED = np.iinfo(np.int32).max
25 | device = "cuda" if torch.cuda.is_available() else "cpu"
26 | STYLE_NAMES = list(styles.keys())
27 | DEFAULT_STYLE_NAME = "Watercolor"
28 |
29 | # Load face encoder
30 | app = FaceAnalysis(name='antelopev2', root='./', providers=['CUDAExecutionProvider', 'CPUExecutionProvider'])
31 | app.prepare(ctx_id=0, det_size=(640, 640))
32 |
33 | # Path to InstantID models
34 | face_adapter = f'./checkpoints/ip-adapter.bin'
35 | controlnet_path = f'./checkpoints/ControlNetModel'
36 |
37 | # Load pipeline
38 | controlnet = ControlNetModel.from_pretrained(controlnet_path, torch_dtype=torch.float16)
39 |
40 | base_model_path = 'wangqixun/YamerMIX_v8'
41 |
42 | pipe = StableDiffusionXLInstantIDPipeline.from_pretrained(
43 | base_model_path,
44 | controlnet=controlnet,
45 | torch_dtype=torch.float16,
46 | safety_checker=None,
47 | feature_extractor=None,
48 | )
49 | pipe.cuda()
50 | pipe.load_ip_adapter_instantid(face_adapter)
51 |
52 | def randomize_seed_fn(seed: int, randomize_seed: bool) -> int:
53 | if randomize_seed:
54 | seed = random.randint(0, MAX_SEED)
55 | return seed
56 |
57 | def swap_to_gallery(images):
58 | return gr.update(value=images, visible=True), gr.update(visible=True), gr.update(visible=False)
59 |
60 | def upload_example_to_gallery(images, prompt, style, negative_prompt):
61 | return gr.update(value=images, visible=True), gr.update(visible=True), gr.update(visible=False)
62 |
63 | def remove_back_to_files():
64 | return gr.update(visible=False), gr.update(visible=False), gr.update(visible=True)
65 |
66 | def remove_tips():
67 | return gr.update(visible=False)
68 |
69 | def get_example():
70 | case = [
71 | [
72 | ['./examples/yann-lecun_resize.jpg'],
73 | "a man",
74 | "Snow",
75 | "(lowres, low quality, worst quality:1.2), (text:1.2), watermark, (frame:1.2), deformed, ugly, deformed eyes, blur, out of focus, blurry, deformed cat, deformed, photo, anthropomorphic cat, monochrome, photo, pet collar, gun, weapon, blue, 3d, drones, drone, buildings in background, green",
76 | ],
77 | [
78 | ['./examples/musk_resize.jpeg'],
79 | "a man",
80 | "Mars",
81 | "(lowres, low quality, worst quality:1.2), (text:1.2), watermark, (frame:1.2), deformed, ugly, deformed eyes, blur, out of focus, blurry, deformed cat, deformed, photo, anthropomorphic cat, monochrome, photo, pet collar, gun, weapon, blue, 3d, drones, drone, buildings in background, green",
82 | ],
83 | [
84 | ['./examples/sam_resize.png'],
85 | "a man",
86 | "Jungle",
87 | "(lowres, low quality, worst quality:1.2), (text:1.2), watermark, (frame:1.2), deformed, ugly, deformed eyes, blur, out of focus, blurry, deformed cat, deformed, photo, anthropomorphic cat, monochrome, photo, pet collar, gun, weapon, blue, 3d, drones, drone, buildings in background, gree",
88 | ],
89 | [
90 | ['./examples/schmidhuber_resize.png'],
91 | "a man",
92 | "Neon",
93 | "(lowres, low quality, worst quality:1.2), (text:1.2), watermark, (frame:1.2), deformed, ugly, deformed eyes, blur, out of focus, blurry, deformed cat, deformed, photo, anthropomorphic cat, monochrome, photo, pet collar, gun, weapon, blue, 3d, drones, drone, buildings in background, green",
94 | ],
95 | [
96 | ['./examples/kaifu_resize.png'],
97 | "a man",
98 | "Vibrant Color",
99 | "(lowres, low quality, worst quality:1.2), (text:1.2), watermark, (frame:1.2), deformed, ugly, deformed eyes, blur, out of focus, blurry, deformed cat, deformed, photo, anthropomorphic cat, monochrome, photo, pet collar, gun, weapon, blue, 3d, drones, drone, buildings in background, green",
100 | ],
101 | ]
102 | return case
103 |
104 | def convert_from_cv2_to_image(img: np.ndarray) -> Image:
105 | return Image.fromarray(cv2.cvtColor(img, cv2.COLOR_BGR2RGB))
106 |
107 | def convert_from_image_to_cv2(img: Image) -> np.ndarray:
108 | return cv2.cvtColor(np.array(img), cv2.COLOR_RGB2BGR)
109 |
110 | def draw_kps(image_pil, kps, color_list=[(255,0,0), (0,255,0), (0,0,255), (255,255,0), (255,0,255)]):
111 | stickwidth = 4
112 | limbSeq = np.array([[0, 2], [1, 2], [3, 2], [4, 2]])
113 | kps = np.array(kps)
114 |
115 | w, h = image_pil.size
116 | out_img = np.zeros([h, w, 3])
117 |
118 | for i in range(len(limbSeq)):
119 | index = limbSeq[i]
120 | color = color_list[index[0]]
121 |
122 | x = kps[index][:, 0]
123 | y = kps[index][:, 1]
124 | length = ((x[0] - x[1]) ** 2 + (y[0] - y[1]) ** 2) ** 0.5
125 | angle = math.degrees(math.atan2(y[0] - y[1], x[0] - x[1]))
126 | polygon = cv2.ellipse2Poly((int(np.mean(x)), int(np.mean(y))), (int(length / 2), stickwidth), int(angle), 0, 360, 1)
127 | out_img = cv2.fillConvexPoly(out_img.copy(), polygon, color)
128 | out_img = (out_img * 0.6).astype(np.uint8)
129 |
130 | for idx_kp, kp in enumerate(kps):
131 | color = color_list[idx_kp]
132 | x, y = kp
133 | out_img = cv2.circle(out_img.copy(), (int(x), int(y)), 10, color, -1)
134 |
135 | out_img_pil = Image.fromarray(out_img.astype(np.uint8))
136 | return out_img_pil
137 |
138 | def resize_img(input_image, max_side=1280, min_side=1024, size=None,
139 | pad_to_max_side=False, mode=PIL.Image.BILINEAR, base_pixel_number=64):
140 |
141 | w, h = input_image.size
142 | if size is not None:
143 | w_resize_new, h_resize_new = size
144 | else:
145 | ratio = min_side / min(h, w)
146 | w, h = round(ratio*w), round(ratio*h)
147 | ratio = max_side / max(h, w)
148 | input_image = input_image.resize([round(ratio*w), round(ratio*h)], mode)
149 | w_resize_new = (round(ratio * w) // base_pixel_number) * base_pixel_number
150 | h_resize_new = (round(ratio * h) // base_pixel_number) * base_pixel_number
151 | input_image = input_image.resize([w_resize_new, h_resize_new], mode)
152 |
153 | if pad_to_max_side:
154 | res = np.ones([max_side, max_side, 3], dtype=np.uint8) * 255
155 | offset_x = (max_side - w_resize_new) // 2
156 | offset_y = (max_side - h_resize_new) // 2
157 | res[offset_y:offset_y+h_resize_new, offset_x:offset_x+w_resize_new] = np.array(input_image)
158 | input_image = Image.fromarray(res)
159 | return input_image
160 |
161 | def apply_style(style_name: str, positive: str, negative: str = "") -> tuple[str, str]:
162 | p, n = styles.get(style_name, styles[DEFAULT_STYLE_NAME])
163 | return p.replace("{prompt}", positive), n + ' ' + negative
164 |
165 | def generate_image(face_image, pose_image, prompt, negative_prompt, style_name, num_steps, identitynet_strength_ratio, adapter_strength_ratio, guidance_scale, seed, progress=gr.Progress(track_tqdm=True)):
166 |
167 | if face_image is None:
168 | raise gr.Error(f"Cannot find any input face image! Please upload the face image")
169 |
170 | if prompt is None:
171 | prompt = "a person"
172 |
173 | # apply the style template
174 | prompt, negative_prompt = apply_style(style_name, prompt, negative_prompt)
175 |
176 | face_image = load_image(face_image[0])
177 | face_image = resize_img(face_image)
178 | face_image_cv2 = convert_from_image_to_cv2(face_image)
179 | height, width, _ = face_image_cv2.shape
180 |
181 | # Extract face features
182 | face_info = app.get(face_image_cv2)
183 |
184 | if len(face_info) == 0:
185 | raise gr.Error(f"Cannot find any face in the image! Please upload another person image")
186 |
187 | face_info = face_info[-1]
188 | face_emb = face_info['embedding']
189 | face_kps = draw_kps(convert_from_cv2_to_image(face_image_cv2), face_info['kps'])
190 |
191 | if pose_image is not None:
192 | pose_image = load_image(pose_image[0])
193 | pose_image = resize_img(pose_image)
194 | pose_image_cv2 = convert_from_image_to_cv2(pose_image)
195 |
196 | face_info = app.get(pose_image_cv2)
197 |
198 | if len(face_info) == 0:
199 | raise gr.Error(f"Cannot find any face in the reference image! Please upload another person image")
200 |
201 | face_info = face_info[-1]
202 | face_kps = draw_kps(pose_image, face_info['kps'])
203 |
204 | width, height = face_kps.size
205 |
206 | generator = torch.Generator(device=device).manual_seed(seed)
207 |
208 | print("Start inference...")
209 | print(f"[Debug] Prompt: {prompt}, \n[Debug] Neg Prompt: {negative_prompt}")
210 |
211 | pipe.set_ip_adapter_scale(adapter_strength_ratio)
212 | images = pipe(
213 | prompt=prompt,
214 | negative_prompt=negative_prompt,
215 | image_embeds=face_emb,
216 | image=face_kps,
217 | controlnet_conditioning_scale=float(identitynet_strength_ratio),
218 | num_inference_steps=num_steps,
219 | guidance_scale=guidance_scale,
220 | height=height,
221 | width=width,
222 | generator=generator
223 | ).images
224 |
225 | return images, gr.update(visible=True)
226 |
227 | ### Description
228 | title = r"""
229 | InstantID: Zero-shot Identity-Preserving Generation in Seconds
230 | """
231 |
232 | description = r"""
233 | Official 🤗 Gradio demo for InstantID: Zero-shot Identity-Preserving Generation in Seconds.
234 |
235 | How to use:
236 | 1. Upload a person image. For multiple person images, we will only detect the biggest face. Make sure face is not too small and not significantly blocked or blurred.
237 | 2. (Optionally) upload another person image as reference pose. If not uploaded, we will use the first person image to extract landmarks. If you use a cropped face at step1, it is recommeneded to upload it to extract a new pose.
238 | 3. Enter a text prompt as done in normal text-to-image models.
239 | 4. Click the Submit button to start customizing.
240 | 5. Share your customizd photo with your friends, enjoy😊!
241 | """
242 |
243 | article = r"""
244 | ---
245 | 📝 **Citation**
246 |
247 | If our work is helpful for your research or applications, please cite us via:
248 | ```bibtex
249 | @article{wang2024instantid,
250 | title={InstantID: Zero-shot Identity-Preserving Generation in Seconds},
251 | author={Wang, Qixun and Bai, Xu and Wang, Haofan and Qin, Zekui and Chen, Anthony},
252 | journal={arXiv preprint arXiv:2401.07519},
253 | year={2024}
254 | }
255 | ```
256 | 📧 **Contact**
257 |
258 | If you have any questions, please feel free to open an issue or directly reach us out at haofanwang.ai@gmail.com.
259 | """
260 |
261 | tips = r"""
262 | ### Usage tips of InstantID
263 | 1. If you're not satisfied with the similarity, try to increase the weight of "IdentityNet Strength" and "Adapter Strength".
264 | 2. If you feel that the saturation is too high, first decrease the Adapter strength. If it is still too high, then decrease the IdentityNet strength.
265 | 3. If you find that text control is not as expected, decrease Adapter strength.
266 | 4. If you find that realistic style is not good enough, go for our Github repo and use a more realistic base model.
267 | """
268 |
269 | css = '''
270 | .gradio-container {width: 85% !important}
271 | '''
272 | with gr.Blocks(css=css) as demo:
273 |
274 | # description
275 | gr.Markdown(title)
276 | gr.Markdown(description)
277 |
278 | with gr.Row():
279 | with gr.Column():
280 |
281 | # upload face image
282 | face_files = gr.Files(
283 | label="Upload a photo of your face",
284 | file_types=["image"]
285 | )
286 | uploaded_faces = gr.Gallery(label="Your images", visible=False, columns=1, rows=1, height=512)
287 | with gr.Column(visible=False) as clear_button_face:
288 | remove_and_reupload_faces = gr.ClearButton(value="Remove and upload new ones", components=face_files, size="sm")
289 |
290 | # optional: upload a reference pose image
291 | pose_files = gr.Files(
292 | label="Upload a reference pose image (optional)",
293 | file_types=["image"]
294 | )
295 | uploaded_poses = gr.Gallery(label="Your images", visible=False, columns=1, rows=1, height=512)
296 | with gr.Column(visible=False) as clear_button_pose:
297 | remove_and_reupload_poses = gr.ClearButton(value="Remove and upload new ones", components=pose_files, size="sm")
298 |
299 | # prompt
300 | prompt = gr.Textbox(label="Prompt",
301 | info="Give simple prompt is enough to achieve good face fedility",
302 | placeholder="A photo of a person",
303 | value="")
304 |
305 | submit = gr.Button("Submit", variant="primary")
306 |
307 | style = gr.Dropdown(label="Style template", choices=STYLE_NAMES, value=DEFAULT_STYLE_NAME)
308 |
309 | # strength
310 | identitynet_strength_ratio = gr.Slider(
311 | label="IdentityNet strength (for fedility)",
312 | minimum=0,
313 | maximum=1.5,
314 | step=0.05,
315 | value=0.80,
316 | )
317 | adapter_strength_ratio = gr.Slider(
318 | label="Image adapter strength (for detail)",
319 | minimum=0,
320 | maximum=1.5,
321 | step=0.05,
322 | value=0.80,
323 | )
324 |
325 | with gr.Accordion(open=False, label="Advanced Options"):
326 | negative_prompt = gr.Textbox(
327 | label="Negative Prompt",
328 | placeholder="low quality",
329 | value="(lowres, low quality, worst quality:1.2), (text:1.2), watermark, (frame:1.2), deformed, ugly, deformed eyes, blur, out of focus, blurry, deformed cat, deformed, photo, anthropomorphic cat, monochrome, pet collar, gun, weapon, blue, 3d, drones, drone, buildings in background, green",
330 | )
331 | num_steps = gr.Slider(
332 | label="Number of sample steps",
333 | minimum=20,
334 | maximum=100,
335 | step=1,
336 | value=30,
337 | )
338 | guidance_scale = gr.Slider(
339 | label="Guidance scale",
340 | minimum=0.1,
341 | maximum=10.0,
342 | step=0.1,
343 | value=5,
344 | )
345 | seed = gr.Slider(
346 | label="Seed",
347 | minimum=0,
348 | maximum=MAX_SEED,
349 | step=1,
350 | value=42,
351 | )
352 | randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
353 |
354 | with gr.Column():
355 | gallery = gr.Gallery(label="Generated Images")
356 | usage_tips = gr.Markdown(label="Usage tips of InstantID", value=tips ,visible=False)
357 |
358 | face_files.upload(fn=swap_to_gallery, inputs=face_files, outputs=[uploaded_faces, clear_button_face, face_files])
359 | pose_files.upload(fn=swap_to_gallery, inputs=pose_files, outputs=[uploaded_poses, clear_button_pose, pose_files])
360 |
361 | remove_and_reupload_faces.click(fn=remove_back_to_files, outputs=[uploaded_faces, clear_button_face, face_files])
362 | remove_and_reupload_poses.click(fn=remove_back_to_files, outputs=[uploaded_poses, clear_button_pose, pose_files])
363 |
364 | submit.click(
365 | fn=remove_tips,
366 | outputs=usage_tips,
367 | ).then(
368 | fn=randomize_seed_fn,
369 | inputs=[seed, randomize_seed],
370 | outputs=seed,
371 | queue=False,
372 | api_name=False,
373 | ).then(
374 | fn=generate_image,
375 | inputs=[face_files, pose_files, prompt, negative_prompt, style, num_steps, identitynet_strength_ratio, adapter_strength_ratio, guidance_scale, seed],
376 | outputs=[gallery, usage_tips]
377 | )
378 |
379 | gr.Examples(
380 | examples=get_example(),
381 | inputs=[face_files, prompt, style, negative_prompt],
382 | run_on_click=True,
383 | fn=upload_example_to_gallery,
384 | outputs=[uploaded_faces, clear_button_face, face_files],
385 | )
386 |
387 | gr.Markdown(article)
388 |
389 | demo.launch()
--------------------------------------------------------------------------------
/gradio_demo/requirements.txt:
--------------------------------------------------------------------------------
1 | diffusers==0.25.0
2 | torch==2.0.0
3 | torchvision==0.15.1
4 | transformers==4.36.2
5 | accelerate
6 | safetensors
7 | einops
8 | onnxruntime-gpu
9 | spaces==0.19.4
10 | omegaconf
11 | peft
12 | huggingface-hub==0.20.2
13 | opencv-python
14 | insightface
--------------------------------------------------------------------------------
/gradio_demo/style_template.py:
--------------------------------------------------------------------------------
1 | style_list = [
2 | {
3 | "name": "(No style)",
4 | "prompt": "{prompt}",
5 | "negative_prompt": "",
6 | },
7 | {
8 | "name": "Watercolor",
9 | "prompt": "watercolor painting, {prompt}. vibrant, beautiful, painterly, detailed, textural, artistic",
10 | "negative_prompt": "(lowres, low quality, worst quality:1.2), (text:1.2), watermark, anime, photorealistic, 35mm film, deformed, glitch, low contrast, noisy",
11 | },
12 | {
13 | "name": "Film Noir",
14 | "prompt": "film noir style, ink sketch|vector, {prompt} highly detailed, sharp focus, ultra sharpness, monochrome, high contrast, dramatic shadows, 1940s style, mysterious, cinematic",
15 | "negative_prompt": "(lowres, low quality, worst quality:1.2), (text:1.2), watermark, (frame:1.2), deformed, ugly, deformed eyes, blur, out of focus, blurry, deformed cat, deformed, photo, anthropomorphic cat, monochrome, photo, pet collar, gun, weapon, blue, 3d, drones, drone, buildings in background, green",
16 | },
17 | {
18 | "name": "Neon",
19 | "prompt": "masterpiece painting, buildings in the backdrop, kaleidoscope, lilac orange blue cream fuchsia bright vivid gradient colors, the scene is cinematic, {prompt}, emotional realism, double exposure, watercolor ink pencil, graded wash, color layering, magic realism, figurative painting, intricate motifs, organic tracery, polished",
20 | "negative_prompt": "(lowres, low quality, worst quality:1.2), (text:1.2), watermark, (frame:1.2), deformed, ugly, deformed eyes, blur, out of focus, blurry, deformed cat, deformed, photo, anthropomorphic cat, monochrome, photo, pet collar, gun, weapon, blue, 3d, drones, drone, buildings in background, green",
21 | },
22 | {
23 | "name": "Jungle",
24 | "prompt": 'waist-up "{prompt} in a Jungle" by Syd Mead, tangerine cold color palette, muted colors, detailed, 8k,photo r3al,dripping paint,3d toon style,3d style,Movie Still',
25 | "negative_prompt": "(lowres, low quality, worst quality:1.2), (text:1.2), watermark, (frame:1.2), deformed, ugly, deformed eyes, blur, out of focus, blurry, deformed cat, deformed, photo, anthropomorphic cat, monochrome, photo, pet collar, gun, weapon, blue, 3d, drones, drone, buildings in background, green",
26 | },
27 | {
28 | "name": "Mars",
29 | "prompt": "{prompt}, Post-apocalyptic. Mars Colony, Scavengers roam the wastelands searching for valuable resources, rovers, bright morning sunlight shining, (detailed) (intricate) (8k) (HDR) (cinematic lighting) (sharp focus)",
30 | "negative_prompt": "(lowres, low quality, worst quality:1.2), (text:1.2), watermark, (frame:1.2), deformed, ugly, deformed eyes, blur, out of focus, blurry, deformed cat, deformed, photo, anthropomorphic cat, monochrome, photo, pet collar, gun, weapon, blue, 3d, drones, drone, buildings in background, green",
31 | },
32 | {
33 | "name": "Vibrant Color",
34 | "prompt": "vibrant colorful, ink sketch|vector|2d colors, at nightfall, sharp focus, {prompt}, highly detailed, sharp focus, the clouds,colorful,ultra sharpness",
35 | "negative_prompt": "(lowres, low quality, worst quality:1.2), (text:1.2), watermark, (frame:1.2), deformed, ugly, deformed eyes, blur, out of focus, blurry, deformed cat, deformed, photo, anthropomorphic cat, monochrome, photo, pet collar, gun, weapon, blue, 3d, drones, drone, buildings in background, green",
36 | },
37 | {
38 | "name": "Snow",
39 | "prompt": "cinema 4d render, {prompt}, high contrast, vibrant and saturated, sico style, surrounded by magical glow,floating ice shards, snow crystals, cold, windy background, frozen natural landscape in background cinematic atmosphere,highly detailed, sharp focus, intricate design, 3d, unreal engine, octane render, CG best quality, highres, photorealistic, dramatic lighting, artstation, concept art, cinematic, epic Steven Spielberg movie still, sharp focus, smoke, sparks, art by pascal blanche and greg rutkowski and repin, trending on artstation, hyperrealism painting, matte painting, 4k resolution",
40 | "negative_prompt": "(lowres, low quality, worst quality:1.2), (text:1.2), watermark, (frame:1.2), deformed, ugly, deformed eyes, blur, out of focus, blurry, deformed cat, deformed, photo, anthropomorphic cat, monochrome, photo, pet collar, gun, weapon, blue, 3d, drones, drone, buildings in background, green",
41 | },
42 | {
43 | "name": "Line art",
44 | "prompt": "line art drawing {prompt} . professional, sleek, modern, minimalist, graphic, line art, vector graphics",
45 | "negative_prompt": "anime, photorealistic, 35mm film, deformed, glitch, blurry, noisy, off-center, deformed, cross-eyed, closed eyes, bad anatomy, ugly, disfigured, mutated, realism, realistic, impressionism, expressionism, oil, acrylic",
46 | },
47 | ]
48 |
49 | styles = {k["name"]: (k["prompt"], k["negative_prompt"]) for k in style_list}
--------------------------------------------------------------------------------
/infer.py:
--------------------------------------------------------------------------------
1 | import cv2
2 | import torch
3 | import numpy as np
4 | from PIL import Image
5 |
6 | from diffusers.utils import load_image
7 | from diffusers.models import ControlNetModel
8 | from diffusers import StableDiffusionXLPipeline
9 |
10 | from insightface.app import FaceAnalysis
11 | from .pipeline_stable_diffusion_xl_instantid import StableDiffusionXLInstantIDPipeline, draw_kps
12 |
13 | def resize_img(input_image, max_side=1280, min_side=1024, size=None,
14 | pad_to_max_side=False, mode=Image.BILINEAR, base_pixel_number=64):
15 |
16 | w, h = input_image.size
17 | if size is not None:
18 | w_resize_new, h_resize_new = size
19 | else:
20 | ratio = min_side / min(h, w)
21 | w, h = round(ratio*w), round(ratio*h)
22 | ratio = max_side / max(h, w)
23 | input_image = input_image.resize([round(ratio*w), round(ratio*h)], mode)
24 | w_resize_new = (round(ratio * w) // base_pixel_number) * base_pixel_number
25 | h_resize_new = (round(ratio * h) // base_pixel_number) * base_pixel_number
26 | input_image = input_image.resize([w_resize_new, h_resize_new], mode)
27 |
28 | if pad_to_max_side:
29 | res = np.ones([max_side, max_side, 3], dtype=np.uint8) * 255
30 | offset_x = (max_side - w_resize_new) // 2
31 | offset_y = (max_side - h_resize_new) // 2
32 | res[offset_y:offset_y+h_resize_new, offset_x:offset_x+w_resize_new] = np.array(input_image)
33 | input_image = Image.fromarray(res)
34 | return input_image
35 |
36 |
37 | if __name__ == "__main__":
38 |
39 | app = FaceAnalysis(name='antelopev2', root='./', providers=['CUDAExecutionProvider', 'CPUExecutionProvider'])
40 | app.prepare(ctx_id=0, det_size=(640, 640))
41 |
42 | # Path to InstantID models
43 | face_adapter = f'./checkpoints/ip-adapter.bin'
44 | controlnet_path = f'./checkpoints/ControlNetModel'
45 |
46 | # Load pipeline
47 | controlnet = ControlNetModel.from_pretrained(controlnet_path, torch_dtype=torch.float16, use_safetensors=True)
48 |
49 | base_model_path = '/nieta_fs/ops/models/checkpoint/sd_xl_base_1.0.safetensors'
50 | ckpt_cache_path = './tmp/tmpckpt.safetensors'
51 | StableDiffusionXLPipeline.from_single_file(
52 | pretrained_model_link_or_path=base_model_path,
53 | torch_dtype=torch.float16,
54 | cache_dir='./tmp',
55 | ).save_pretrained(ckpt_cache_path, safe_serialization=True)
56 |
57 |
58 | pipe = StableDiffusionXLInstantIDPipeline.from_pretrained(
59 | ckpt_cache_path,
60 | controlnet=controlnet,
61 | torch_dtype=torch.float16,
62 | )
63 | pipe.cuda()
64 | pipe.load_ip_adapter_instantid(face_adapter)
65 |
66 | prompt = "anime film screenshot of a man, masterpiece, best quality"
67 | n_prompt = "(lowres, low quality, worst quality:1.2), (text:1.2), watermark, painting, drawing, illustration, glitch, deformed, mutated, cross-eyed, ugly, disfigured (lowres, low quality, worst quality:1.2), (text:1.2), watermark, painting, drawing, illustration, glitch,deformed, mutated, cross-eyed, ugly, disfigured"
68 |
69 | face_image = load_image("./examples/yann-lecun_resize.jpg")
70 | face_image = resize_img(face_image)
71 |
72 | face_info = app.get(cv2.cvtColor(np.array(face_image), cv2.COLOR_RGB2BGR))
73 | face_info = sorted(face_info, key=lambda x:(x['bbox'][2]-x['bbox'][0])*x['bbox'][3]-x['bbox'][1])[-1] # only use the maximum face
74 | face_emb = face_info['embedding']
75 | face_kps = draw_kps(face_image, face_info['kps'])
76 |
77 | pipe.set_ip_adapter_scale(0.8)
78 | image = pipe(
79 | prompt=prompt,
80 | negative_prompt=n_prompt,
81 | image_embeds=face_emb,
82 | image=face_kps,
83 | controlnet_conditioning_scale=0.8,
84 | num_inference_steps=30,
85 | guidance_scale=5,
86 | ).images[0]
87 |
88 | image.save('result.jpg')
--------------------------------------------------------------------------------
/ip_adapter/attention_processor.py:
--------------------------------------------------------------------------------
1 | # modified from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py
2 | import torch
3 | import torch.nn as nn
4 | import torch.nn.functional as F
5 |
6 | try:
7 | import xformers
8 | import xformers.ops
9 | xformers_available = True
10 | except Exception as e:
11 | xformers_available = False
12 |
13 |
14 |
15 | class RegionControler(object):
16 | def __init__(self) -> None:
17 | self.prompt_image_conditioning = []
18 | region_control = RegionControler()
19 |
20 |
21 | class AttnProcessor(nn.Module):
22 | r"""
23 | Default processor for performing attention-related computations.
24 | """
25 | def __init__(
26 | self,
27 | hidden_size=None,
28 | cross_attention_dim=None,
29 | ):
30 | super().__init__()
31 |
32 | def __call__(
33 | self,
34 | attn,
35 | hidden_states,
36 | encoder_hidden_states=None,
37 | attention_mask=None,
38 | temb=None,
39 | ):
40 | residual = hidden_states
41 |
42 | if attn.spatial_norm is not None:
43 | hidden_states = attn.spatial_norm(hidden_states, temb)
44 |
45 | input_ndim = hidden_states.ndim
46 |
47 | if input_ndim == 4:
48 | batch_size, channel, height, width = hidden_states.shape
49 | hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
50 |
51 | batch_size, sequence_length, _ = (
52 | hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
53 | )
54 | attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
55 |
56 | if attn.group_norm is not None:
57 | hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
58 |
59 | query = attn.to_q(hidden_states)
60 |
61 | if encoder_hidden_states is None:
62 | encoder_hidden_states = hidden_states
63 | elif attn.norm_cross:
64 | encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
65 |
66 | key = attn.to_k(encoder_hidden_states)
67 | value = attn.to_v(encoder_hidden_states)
68 |
69 | query = attn.head_to_batch_dim(query)
70 | key = attn.head_to_batch_dim(key)
71 | value = attn.head_to_batch_dim(value)
72 |
73 | attention_probs = attn.get_attention_scores(query, key, attention_mask)
74 | hidden_states = torch.bmm(attention_probs, value)
75 | hidden_states = attn.batch_to_head_dim(hidden_states)
76 |
77 | # linear proj
78 | hidden_states = attn.to_out[0](hidden_states)
79 | # dropout
80 | hidden_states = attn.to_out[1](hidden_states)
81 |
82 | if input_ndim == 4:
83 | hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
84 |
85 | if attn.residual_connection:
86 | hidden_states = hidden_states + residual
87 |
88 | hidden_states = hidden_states / attn.rescale_output_factor
89 |
90 | return hidden_states
91 |
92 |
93 | class IPAttnProcessor(nn.Module):
94 | r"""
95 | Attention processor for IP-Adapater.
96 | Args:
97 | hidden_size (`int`):
98 | The hidden size of the attention layer.
99 | cross_attention_dim (`int`):
100 | The number of channels in the `encoder_hidden_states`.
101 | scale (`float`, defaults to 1.0):
102 | the weight scale of image prompt.
103 | num_tokens (`int`, defaults to 4 when do ip_adapter_plus it should be 16):
104 | The context length of the image features.
105 | """
106 |
107 | def __init__(self, hidden_size, cross_attention_dim=None, scale=1.0, num_tokens=4):
108 | super().__init__()
109 |
110 | self.hidden_size = hidden_size
111 | self.cross_attention_dim = cross_attention_dim
112 | self.scale = scale
113 | self.num_tokens = num_tokens
114 |
115 | self.to_k_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
116 | self.to_v_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
117 |
118 | def __call__(
119 | self,
120 | attn,
121 | hidden_states,
122 | encoder_hidden_states=None,
123 | attention_mask=None,
124 | temb=None,
125 | ):
126 | residual = hidden_states
127 |
128 | if attn.spatial_norm is not None:
129 | hidden_states = attn.spatial_norm(hidden_states, temb)
130 |
131 | input_ndim = hidden_states.ndim
132 |
133 | if input_ndim == 4:
134 | batch_size, channel, height, width = hidden_states.shape
135 | hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
136 |
137 | batch_size, sequence_length, _ = (
138 | hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
139 | )
140 | attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
141 |
142 | if attn.group_norm is not None:
143 | hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
144 |
145 | query = attn.to_q(hidden_states)
146 |
147 | if encoder_hidden_states is None:
148 | encoder_hidden_states = hidden_states
149 | else:
150 | # get encoder_hidden_states, ip_hidden_states
151 | end_pos = encoder_hidden_states.shape[1] - self.num_tokens
152 | encoder_hidden_states, ip_hidden_states = encoder_hidden_states[:, :end_pos, :], encoder_hidden_states[:, end_pos:, :]
153 | if attn.norm_cross:
154 | encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
155 |
156 | key = attn.to_k(encoder_hidden_states)
157 | value = attn.to_v(encoder_hidden_states)
158 |
159 | query = attn.head_to_batch_dim(query)
160 | key = attn.head_to_batch_dim(key)
161 | value = attn.head_to_batch_dim(value)
162 |
163 | if xformers_available:
164 | hidden_states = self._memory_efficient_attention_xformers(query, key, value, attention_mask)
165 | else:
166 | attention_probs = attn.get_attention_scores(query, key, attention_mask)
167 | hidden_states = torch.bmm(attention_probs, value)
168 | hidden_states = attn.batch_to_head_dim(hidden_states)
169 |
170 | # for ip-adapter
171 | ip_key = self.to_k_ip(ip_hidden_states)
172 | ip_value = self.to_v_ip(ip_hidden_states)
173 |
174 | ip_key = attn.head_to_batch_dim(ip_key)
175 | ip_value = attn.head_to_batch_dim(ip_value)
176 |
177 | if xformers_available:
178 | ip_hidden_states = self._memory_efficient_attention_xformers(query, ip_key, ip_value, None)
179 | else:
180 | ip_attention_probs = attn.get_attention_scores(query, ip_key, None)
181 | ip_hidden_states = torch.bmm(ip_attention_probs, ip_value)
182 | ip_hidden_states = attn.batch_to_head_dim(ip_hidden_states)
183 |
184 | # region control
185 | if len(region_control.prompt_image_conditioning) == 1:
186 | region_mask = region_control.prompt_image_conditioning[0].get('region_mask', None)
187 | if region_mask is not None:
188 | h, w = region_mask.shape[:2]
189 | ratio = (h * w / query.shape[1]) ** 0.5
190 | mask = F.interpolate(region_mask[None, None], scale_factor=1/ratio, mode='nearest').reshape([1, -1, 1])
191 | else:
192 | mask = torch.ones_like(ip_hidden_states)
193 | ip_hidden_states = ip_hidden_states * mask
194 |
195 | hidden_states = hidden_states + self.scale * ip_hidden_states
196 |
197 | # linear proj
198 | hidden_states = attn.to_out[0](hidden_states)
199 | # dropout
200 | hidden_states = attn.to_out[1](hidden_states)
201 |
202 | if input_ndim == 4:
203 | hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
204 |
205 | if attn.residual_connection:
206 | hidden_states = hidden_states + residual
207 |
208 | hidden_states = hidden_states / attn.rescale_output_factor
209 |
210 | return hidden_states
211 |
212 |
213 | def _memory_efficient_attention_xformers(self, query, key, value, attention_mask):
214 | # TODO attention_mask
215 | query = query.contiguous()
216 | key = key.contiguous()
217 | value = value.contiguous()
218 | hidden_states = xformers.ops.memory_efficient_attention(query, key, value, attn_bias=attention_mask)
219 | # hidden_states = self.reshape_batch_dim_to_heads(hidden_states)
220 | return hidden_states
221 |
222 |
223 | class AttnProcessor2_0(torch.nn.Module):
224 | r"""
225 | Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0).
226 | """
227 | def __init__(
228 | self,
229 | hidden_size=None,
230 | cross_attention_dim=None,
231 | ):
232 | super().__init__()
233 | if not hasattr(F, "scaled_dot_product_attention"):
234 | raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
235 |
236 | def __call__(
237 | self,
238 | attn,
239 | hidden_states,
240 | encoder_hidden_states=None,
241 | attention_mask=None,
242 | temb=None,
243 | ):
244 | residual = hidden_states
245 |
246 | if attn.spatial_norm is not None:
247 | hidden_states = attn.spatial_norm(hidden_states, temb)
248 |
249 | input_ndim = hidden_states.ndim
250 |
251 | if input_ndim == 4:
252 | batch_size, channel, height, width = hidden_states.shape
253 | hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
254 |
255 | batch_size, sequence_length, _ = (
256 | hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
257 | )
258 |
259 | if attention_mask is not None:
260 | attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
261 | # scaled_dot_product_attention expects attention_mask shape to be
262 | # (batch, heads, source_length, target_length)
263 | attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
264 |
265 | if attn.group_norm is not None:
266 | hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
267 |
268 | query = attn.to_q(hidden_states)
269 |
270 | if encoder_hidden_states is None:
271 | encoder_hidden_states = hidden_states
272 | elif attn.norm_cross:
273 | encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
274 |
275 | key = attn.to_k(encoder_hidden_states)
276 | value = attn.to_v(encoder_hidden_states)
277 |
278 | inner_dim = key.shape[-1]
279 | head_dim = inner_dim // attn.heads
280 |
281 | query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
282 |
283 | key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
284 | value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
285 |
286 | # the output of sdp = (batch, num_heads, seq_len, head_dim)
287 | # TODO: add support for attn.scale when we move to Torch 2.1
288 | hidden_states = F.scaled_dot_product_attention(
289 | query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
290 | )
291 |
292 | hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
293 | hidden_states = hidden_states.to(query.dtype)
294 |
295 | # linear proj
296 | hidden_states = attn.to_out[0](hidden_states)
297 | # dropout
298 | hidden_states = attn.to_out[1](hidden_states)
299 |
300 | if input_ndim == 4:
301 | hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
302 |
303 | if attn.residual_connection:
304 | hidden_states = hidden_states + residual
305 |
306 | hidden_states = hidden_states / attn.rescale_output_factor
307 |
308 | return hidden_states
--------------------------------------------------------------------------------
/ip_adapter/resampler.py:
--------------------------------------------------------------------------------
1 | # modified from https://github.com/mlfoundations/open_flamingo/blob/main/open_flamingo/src/helpers.py
2 | import math
3 |
4 | import torch
5 | import torch.nn as nn
6 |
7 |
8 | # FFN
9 | def FeedForward(dim, mult=4):
10 | inner_dim = int(dim * mult)
11 | return nn.Sequential(
12 | nn.LayerNorm(dim),
13 | nn.Linear(dim, inner_dim, bias=False),
14 | nn.GELU(),
15 | nn.Linear(inner_dim, dim, bias=False),
16 | )
17 |
18 |
19 | def reshape_tensor(x, heads):
20 | bs, length, width = x.shape
21 | #(bs, length, width) --> (bs, length, n_heads, dim_per_head)
22 | x = x.view(bs, length, heads, -1)
23 | # (bs, length, n_heads, dim_per_head) --> (bs, n_heads, length, dim_per_head)
24 | x = x.transpose(1, 2)
25 | # (bs, n_heads, length, dim_per_head) --> (bs*n_heads, length, dim_per_head)
26 | x = x.reshape(bs, heads, length, -1)
27 | return x
28 |
29 |
30 | class PerceiverAttention(nn.Module):
31 | def __init__(self, *, dim, dim_head=64, heads=8):
32 | super().__init__()
33 | self.scale = dim_head**-0.5
34 | self.dim_head = dim_head
35 | self.heads = heads
36 | inner_dim = dim_head * heads
37 |
38 | self.norm1 = nn.LayerNorm(dim)
39 | self.norm2 = nn.LayerNorm(dim)
40 |
41 | self.to_q = nn.Linear(dim, inner_dim, bias=False)
42 | self.to_kv = nn.Linear(dim, inner_dim * 2, bias=False)
43 | self.to_out = nn.Linear(inner_dim, dim, bias=False)
44 |
45 |
46 | def forward(self, x, latents):
47 | """
48 | Args:
49 | x (torch.Tensor): image features
50 | shape (b, n1, D)
51 | latent (torch.Tensor): latent features
52 | shape (b, n2, D)
53 | """
54 | x = self.norm1(x)
55 | latents = self.norm2(latents)
56 |
57 | b, l, _ = latents.shape
58 |
59 | q = self.to_q(latents)
60 | kv_input = torch.cat((x, latents), dim=-2)
61 | k, v = self.to_kv(kv_input).chunk(2, dim=-1)
62 |
63 | q = reshape_tensor(q, self.heads)
64 | k = reshape_tensor(k, self.heads)
65 | v = reshape_tensor(v, self.heads)
66 |
67 | # attention
68 | scale = 1 / math.sqrt(math.sqrt(self.dim_head))
69 | weight = (q * scale) @ (k * scale).transpose(-2, -1) # More stable with f16 than dividing afterwards
70 | weight = torch.softmax(weight.float(), dim=-1).type(weight.dtype)
71 | out = weight @ v
72 |
73 | out = out.permute(0, 2, 1, 3).reshape(b, l, -1)
74 |
75 | return self.to_out(out)
76 |
77 |
78 | class Resampler(nn.Module):
79 | def __init__(
80 | self,
81 | dim=1024,
82 | depth=8,
83 | dim_head=64,
84 | heads=16,
85 | num_queries=8,
86 | embedding_dim=768,
87 | output_dim=1024,
88 | ff_mult=4,
89 | ):
90 | super().__init__()
91 |
92 | self.latents = nn.Parameter(torch.randn(1, num_queries, dim) / dim**0.5)
93 |
94 | self.proj_in = nn.Linear(embedding_dim, dim)
95 |
96 | self.proj_out = nn.Linear(dim, output_dim)
97 | self.norm_out = nn.LayerNorm(output_dim)
98 |
99 | self.layers = nn.ModuleList([])
100 | for _ in range(depth):
101 | self.layers.append(
102 | nn.ModuleList(
103 | [
104 | PerceiverAttention(dim=dim, dim_head=dim_head, heads=heads),
105 | FeedForward(dim=dim, mult=ff_mult),
106 | ]
107 | )
108 | )
109 |
110 | def forward(self, x):
111 |
112 | latents = self.latents.repeat(x.size(0), 1, 1)
113 |
114 | x = self.proj_in(x)
115 |
116 | for attn, ff in self.layers:
117 | latents = attn(x, latents) + latents
118 | latents = ff(latents) + latents
119 |
120 | latents = self.proj_out(latents)
121 | return self.norm_out(latents)
--------------------------------------------------------------------------------
/ip_adapter/utils.py:
--------------------------------------------------------------------------------
1 | import torch.nn.functional as F
2 |
3 |
4 | def is_torch2_available():
5 | return hasattr(F, "scaled_dot_product_attention")
6 |
--------------------------------------------------------------------------------
/models/__put_antelopev2_here__:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/huxiuhan/ComfyUI-InstantID/c725820fcae81cfb7bddd6e80d5d19df634eb050/models/__put_antelopev2_here__
--------------------------------------------------------------------------------
/nodes.py:
--------------------------------------------------------------------------------
1 | import copy
2 | import os
3 | import torch
4 | from safetensors.torch import load_file
5 | from torchvision.transforms import ToTensor
6 | from comfy.model_management import get_torch_device
7 | from .pipeline_stable_diffusion_xl_instantid import StableDiffusionXLInstantIDPipeline, draw_kps
8 | from .infer import resize_img
9 | import folder_paths
10 | import cv2
11 | import numpy as np
12 | from PIL import Image
13 |
14 | from diffusers.utils import load_image
15 | from diffusers.models import ControlNetModel
16 | from diffusers import StableDiffusionXLPipeline
17 |
18 | from insightface.app import FaceAnalysis
19 |
20 | CUSTOM_NODE_CWD = os.path.dirname(os.path.realpath(__file__))
21 |
22 | class InstantIDSampler:
23 | def __init__(self):
24 | self.torch_device = get_torch_device()
25 | self.tmp_dir = folder_paths.get_temp_directory()
26 | self.face_app = None
27 | self.controlnet = None
28 | self.previous_ckpt = None
29 | self.pipe = None
30 |
31 | @classmethod
32 | def INPUT_TYPES(s):
33 | return {
34 | "required": {
35 | "ckpt_name": (folder_paths.get_filename_list("checkpoints"), ),
36 | "positive": ("STRING", {"multiline": True}),
37 | "negative": ("STRING", {"multiline": True}),
38 | "steps": ("INT", {"default": 30, "min": 1, "max": 10000}),
39 | "cfg": ("FLOAT", {"default": 5, "min": 0.0, "max": 100.0}),
40 | "strength": ("FLOAT", {"default": 0.8, "min": 0.0, "max": 1.0}),
41 | "seed": ("INT", {"default": 0, "min": 0, "max": 0xffffffffffffffff}),
42 | "image" : ("IMAGE", )
43 | }
44 | }
45 |
46 | RETURN_TYPES = ("IMAGE",)
47 |
48 | FUNCTION = "sample"
49 |
50 | CATEGORY = "Instant ID"
51 |
52 |
53 | def sample(self, ckpt_name, positive, negative, steps, cfg, strength, seed, image):
54 |
55 | print("Instant ID Current Working Dir:", CUSTOM_NODE_CWD)
56 |
57 | if self.face_app is None:
58 | app = FaceAnalysis(name='antelopev2', root=CUSTOM_NODE_CWD, providers=['CUDAExecutionProvider', 'CPUExecutionProvider'])
59 | app.prepare(ctx_id=0, det_size=(640, 640))
60 | self.face_app = app
61 | print("Instant ID Insightface App Loaded.")
62 |
63 | # Path to InstantID models
64 | face_adapter = os.path.join(CUSTOM_NODE_CWD, f'./checkpoints/ip-adapter.bin')
65 | controlnet_path = os.path.join(CUSTOM_NODE_CWD, f'./checkpoints/ControlNetModel')
66 |
67 | # Load pipeline
68 | if self.controlnet is None:
69 | self.controlnet = ControlNetModel.from_pretrained(controlnet_path, torch_dtype=torch.float16, use_safetensors=True)
70 | print("Instant ID Controlnet Loaded.")
71 |
72 | # prepare ckpt for diffusers
73 |
74 | if self.previous_ckpt != ckpt_name:
75 | ckpt_cache_path = os.path.join(self.tmp_dir, ckpt_name)
76 | StableDiffusionXLPipeline.from_single_file(
77 | pretrained_model_link_or_path=folder_paths.get_full_path("checkpoints", ckpt_name),
78 | torch_dtype=torch.float16,
79 | cache_dir=self.tmp_dir,
80 | ).save_pretrained(ckpt_cache_path, safe_serialization=True)
81 |
82 | self.previous_ckpt = ckpt_name
83 | self.pipe = StableDiffusionXLInstantIDPipeline.from_pretrained(
84 | ckpt_cache_path,
85 | controlnet=self.controlnet,
86 | torch_dtype=torch.float16,
87 | )
88 | self.pipe.cuda()
89 | self.pipe.load_ip_adapter_instantid(face_adapter)
90 | print("Instant ID Ckpt Reloaded.")
91 |
92 |
93 | prompt = positive
94 | n_prompt = negative
95 |
96 | face_image = Image.fromarray(np.clip(255. * image[0].cpu().numpy(), 0, 255).astype(np.uint8))
97 | face_image = resize_img(face_image)
98 |
99 | face_info = self.face_app.get(cv2.cvtColor(np.array(face_image), cv2.COLOR_RGB2BGR))
100 | if len(face_info) < 1:
101 | raise ValueError("Cannot find any face.")
102 | face_info = sorted(face_info, key=lambda x:(x['bbox'][2]-x['bbox'][0])*x['bbox'][3]-x['bbox'][1])[-1] # only use the maximum face
103 | face_emb = face_info['embedding']
104 | face_kps = draw_kps(face_image, face_info['kps'])
105 | print("Instant ID Face Info Updated.")
106 |
107 |
108 | self.pipe.set_ip_adapter_scale(strength)
109 |
110 | g_cpu = torch.Generator()
111 | g_cpu.manual_seed(seed)
112 | result = self.pipe(
113 | prompt=prompt,
114 | negative_prompt=n_prompt,
115 | image_embeds=face_emb,
116 | image=face_kps,
117 | controlnet_conditioning_scale=strength,
118 | num_inference_steps=steps,
119 | guidance_scale=cfg,
120 | generator=g_cpu
121 | ).images
122 |
123 | def convert_images_to_tensors(images: list[Image.Image]):
124 | return torch.stack([np.transpose(ToTensor()(image), (1, 2, 0)) for image in images])
125 |
126 | return (convert_images_to_tensors(result),)
127 |
--------------------------------------------------------------------------------
/pipeline_stable_diffusion_xl_instantid.py:
--------------------------------------------------------------------------------
1 | # Copyright 2024 The InstantX 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 | from typing import Any, Callable, Dict, List, Optional, Tuple, Union
17 |
18 | import cv2
19 | import math
20 |
21 | import numpy as np
22 | import PIL.Image
23 | import torch
24 | import torch.nn.functional as F
25 |
26 | from diffusers.image_processor import PipelineImageInput
27 |
28 | from diffusers.models import ControlNetModel
29 |
30 | from diffusers.utils import (
31 | deprecate,
32 | logging,
33 | replace_example_docstring,
34 | )
35 | from diffusers.utils.torch_utils import is_compiled_module, is_torch_version
36 | from diffusers.pipelines.stable_diffusion_xl import StableDiffusionXLPipelineOutput
37 |
38 | from diffusers import StableDiffusionXLControlNetPipeline
39 | from diffusers.pipelines.controlnet.multicontrolnet import MultiControlNetModel
40 | from diffusers.utils.import_utils import is_xformers_available
41 |
42 | from .ip_adapter.resampler import Resampler
43 | from .ip_adapter.utils import is_torch2_available
44 |
45 | from .ip_adapter.attention_processor import AttnProcessor, IPAttnProcessor
46 |
47 | logger = logging.get_logger(__name__) # pylint: disable=invalid-name
48 |
49 |
50 | EXAMPLE_DOC_STRING = """
51 | Examples:
52 | ```py
53 | >>> # !pip install opencv-python transformers accelerate insightface
54 | >>> import diffusers
55 | >>> from diffusers.utils import load_image
56 | >>> from diffusers.models import ControlNetModel
57 |
58 | >>> import cv2
59 | >>> import torch
60 | >>> import numpy as np
61 | >>> from PIL import Image
62 |
63 | >>> from insightface.app import FaceAnalysis
64 | >>> from pipeline_stable_diffusion_xl_instantid import StableDiffusionXLInstantIDPipeline, draw_kps
65 |
66 | >>> # download 'antelopev2' under ./models
67 | >>> app = FaceAnalysis(name='antelopev2', root='./', providers=['CUDAExecutionProvider', 'CPUExecutionProvider'])
68 | >>> app.prepare(ctx_id=0, det_size=(640, 640))
69 |
70 | >>> # download models under ./checkpoints
71 | >>> face_adapter = f'./checkpoints/ip-adapter.bin'
72 | >>> controlnet_path = f'./checkpoints/ControlNetModel'
73 |
74 | >>> # load IdentityNet
75 | >>> controlnet = ControlNetModel.from_pretrained(controlnet_path, torch_dtype=torch.float16)
76 |
77 | >>> pipe = StableDiffusionXLInstantIDPipeline.from_pretrained(
78 | ... "stabilityai/stable-diffusion-xl-base-1.0", controlnet=controlnet, torch_dtype=torch.float16
79 | ... )
80 | >>> pipe.cuda()
81 |
82 | >>> # load adapter
83 | >>> pipe.load_ip_adapter_instantid(face_adapter)
84 |
85 | >>> prompt = "analog film photo of a man. faded film, desaturated, 35mm photo, grainy, vignette, vintage, Kodachrome, Lomography, stained, highly detailed, found footage, masterpiece, best quality"
86 | >>> negative_prompt = "(lowres, low quality, worst quality:1.2), (text:1.2), watermark, painting, drawing, illustration, glitch, deformed, mutated, cross-eyed, ugly, disfigured (lowres, low quality, worst quality:1.2), (text:1.2), watermark, painting, drawing, illustration, glitch,deformed, mutated, cross-eyed, ugly, disfigured"
87 |
88 | >>> # load an image
89 | >>> image = load_image("your-example.jpg")
90 |
91 | >>> face_info = app.get(cv2.cvtColor(np.array(face_image), cv2.COLOR_RGB2BGR))[-1]
92 | >>> face_emb = face_info['embedding']
93 | >>> face_kps = draw_kps(face_image, face_info['kps'])
94 |
95 | >>> pipe.set_ip_adapter_scale(0.8)
96 |
97 | >>> # generate image
98 | >>> image = pipe(
99 | ... prompt, image_embeds=face_emb, image=face_kps, controlnet_conditioning_scale=0.8
100 | ... ).images[0]
101 | ```
102 | """
103 |
104 | def draw_kps(image_pil, kps, color_list=[(255,0,0), (0,255,0), (0,0,255), (255,255,0), (255,0,255)]):
105 |
106 | stickwidth = 4
107 | limbSeq = np.array([[0, 2], [1, 2], [3, 2], [4, 2]])
108 | kps = np.array(kps)
109 |
110 | w, h = image_pil.size
111 | out_img = np.zeros([h, w, 3])
112 |
113 | for i in range(len(limbSeq)):
114 | index = limbSeq[i]
115 | color = color_list[index[0]]
116 |
117 | x = kps[index][:, 0]
118 | y = kps[index][:, 1]
119 | length = ((x[0] - x[1]) ** 2 + (y[0] - y[1]) ** 2) ** 0.5
120 | angle = math.degrees(math.atan2(y[0] - y[1], x[0] - x[1]))
121 | polygon = cv2.ellipse2Poly((int(np.mean(x)), int(np.mean(y))), (int(length / 2), stickwidth), int(angle), 0, 360, 1)
122 | out_img = cv2.fillConvexPoly(out_img.copy(), polygon, color)
123 | out_img = (out_img * 0.6).astype(np.uint8)
124 |
125 | for idx_kp, kp in enumerate(kps):
126 | color = color_list[idx_kp]
127 | x, y = kp
128 | out_img = cv2.circle(out_img.copy(), (int(x), int(y)), 10, color, -1)
129 |
130 | out_img_pil = PIL.Image.fromarray(out_img.astype(np.uint8))
131 | return out_img_pil
132 |
133 | class StableDiffusionXLInstantIDPipeline(StableDiffusionXLControlNetPipeline):
134 |
135 | def cuda(self, dtype=torch.float16, use_xformers=False):
136 | self.to('cuda', dtype)
137 |
138 | if hasattr(self, 'image_proj_model'):
139 | self.image_proj_model.to(self.unet.device).to(self.unet.dtype)
140 |
141 | if use_xformers:
142 | if is_xformers_available():
143 | import xformers
144 | from packaging import version
145 |
146 | xformers_version = version.parse(xformers.__version__)
147 | if xformers_version == version.parse("0.0.16"):
148 | logger.warn(
149 | "xFormers 0.0.16 cannot be used for training in some GPUs. If you observe problems during training, please update xFormers to at least 0.0.17. See https://huggingface.co/docs/diffusers/main/en/optimization/xformers for more details."
150 | )
151 | self.enable_xformers_memory_efficient_attention()
152 | else:
153 | raise ValueError("xformers is not available. Make sure it is installed correctly")
154 |
155 | def load_ip_adapter_instantid(self, model_ckpt, image_emb_dim=512, num_tokens=16, scale=0.5):
156 | self.set_image_proj_model(model_ckpt, image_emb_dim, num_tokens)
157 | self.set_ip_adapter(model_ckpt, num_tokens, scale)
158 |
159 | def set_image_proj_model(self, model_ckpt, image_emb_dim=512, num_tokens=16):
160 |
161 | image_proj_model = Resampler(
162 | dim=1280,
163 | depth=4,
164 | dim_head=64,
165 | heads=20,
166 | num_queries=num_tokens,
167 | embedding_dim=image_emb_dim,
168 | output_dim=self.unet.config.cross_attention_dim,
169 | ff_mult=4,
170 | )
171 |
172 | image_proj_model.eval()
173 |
174 | self.image_proj_model = image_proj_model.to(self.device, dtype=self.dtype)
175 | state_dict = torch.load(model_ckpt, map_location="cpu")
176 | if 'image_proj' in state_dict:
177 | state_dict = state_dict["image_proj"]
178 | self.image_proj_model.load_state_dict(state_dict)
179 |
180 | self.image_proj_model_in_features = image_emb_dim
181 |
182 | def set_ip_adapter(self, model_ckpt, num_tokens, scale):
183 |
184 | unet = self.unet
185 | attn_procs = {}
186 | for name in unet.attn_processors.keys():
187 | cross_attention_dim = None if name.endswith("attn1.processor") else unet.config.cross_attention_dim
188 | if name.startswith("mid_block"):
189 | hidden_size = unet.config.block_out_channels[-1]
190 | elif name.startswith("up_blocks"):
191 | block_id = int(name[len("up_blocks.")])
192 | hidden_size = list(reversed(unet.config.block_out_channels))[block_id]
193 | elif name.startswith("down_blocks"):
194 | block_id = int(name[len("down_blocks.")])
195 | hidden_size = unet.config.block_out_channels[block_id]
196 | if cross_attention_dim is None:
197 | attn_procs[name] = AttnProcessor().to(unet.device, dtype=unet.dtype)
198 | else:
199 | attn_procs[name] = IPAttnProcessor(hidden_size=hidden_size,
200 | cross_attention_dim=cross_attention_dim,
201 | scale=scale,
202 | num_tokens=num_tokens).to(unet.device, dtype=unet.dtype)
203 | unet.set_attn_processor(attn_procs)
204 |
205 | state_dict = torch.load(model_ckpt, map_location="cpu")
206 | ip_layers = torch.nn.ModuleList(self.unet.attn_processors.values())
207 | if 'ip_adapter' in state_dict:
208 | state_dict = state_dict['ip_adapter']
209 | ip_layers.load_state_dict(state_dict)
210 |
211 | def set_ip_adapter_scale(self, scale):
212 | unet = getattr(self, self.unet_name) if not hasattr(self, "unet") else self.unet
213 | for attn_processor in unet.attn_processors.values():
214 | if isinstance(attn_processor, IPAttnProcessor):
215 | attn_processor.scale = scale
216 |
217 | def _encode_prompt_image_emb(self, prompt_image_emb, device, dtype, do_classifier_free_guidance):
218 |
219 | if isinstance(prompt_image_emb, torch.Tensor):
220 | prompt_image_emb = prompt_image_emb.clone().detach()
221 | else:
222 | prompt_image_emb = torch.tensor(prompt_image_emb)
223 |
224 | prompt_image_emb = prompt_image_emb.to(device=device, dtype=dtype)
225 | prompt_image_emb = prompt_image_emb.reshape([1, -1, self.image_proj_model_in_features])
226 |
227 | if do_classifier_free_guidance:
228 | prompt_image_emb = torch.cat([torch.zeros_like(prompt_image_emb), prompt_image_emb], dim=0)
229 | else:
230 | prompt_image_emb = torch.cat([prompt_image_emb], dim=0)
231 |
232 | prompt_image_emb = self.image_proj_model(prompt_image_emb)
233 | return prompt_image_emb
234 |
235 | @torch.no_grad()
236 | @replace_example_docstring(EXAMPLE_DOC_STRING)
237 | def __call__(
238 | self,
239 | prompt: Union[str, List[str]] = None,
240 | prompt_2: Optional[Union[str, List[str]]] = None,
241 | image: PipelineImageInput = None,
242 | height: Optional[int] = None,
243 | width: Optional[int] = None,
244 | num_inference_steps: int = 50,
245 | guidance_scale: float = 5.0,
246 | negative_prompt: Optional[Union[str, List[str]]] = None,
247 | negative_prompt_2: Optional[Union[str, List[str]]] = None,
248 | num_images_per_prompt: Optional[int] = 1,
249 | eta: float = 0.0,
250 | generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
251 | latents: Optional[torch.FloatTensor] = None,
252 | prompt_embeds: Optional[torch.FloatTensor] = None,
253 | negative_prompt_embeds: Optional[torch.FloatTensor] = None,
254 | pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
255 | negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
256 | image_embeds: Optional[torch.FloatTensor] = None,
257 | output_type: Optional[str] = "pil",
258 | return_dict: bool = True,
259 | cross_attention_kwargs: Optional[Dict[str, Any]] = None,
260 | controlnet_conditioning_scale: Union[float, List[float]] = 1.0,
261 | guess_mode: bool = False,
262 | control_guidance_start: Union[float, List[float]] = 0.0,
263 | control_guidance_end: Union[float, List[float]] = 1.0,
264 | original_size: Tuple[int, int] = None,
265 | crops_coords_top_left: Tuple[int, int] = (0, 0),
266 | target_size: Tuple[int, int] = None,
267 | negative_original_size: Optional[Tuple[int, int]] = None,
268 | negative_crops_coords_top_left: Tuple[int, int] = (0, 0),
269 | negative_target_size: Optional[Tuple[int, int]] = None,
270 | clip_skip: Optional[int] = None,
271 | callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
272 | callback_on_step_end_tensor_inputs: List[str] = ["latents"],
273 | **kwargs,
274 | ):
275 | r"""
276 | The call function to the pipeline for generation.
277 |
278 | Args:
279 | prompt (`str` or `List[str]`, *optional*):
280 | The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
281 | prompt_2 (`str` or `List[str]`, *optional*):
282 | The prompt or prompts to be sent to `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
283 | used in both text-encoders.
284 | image (`torch.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.FloatTensor]`, `List[PIL.Image.Image]`, `List[np.ndarray]`,:
285 | `List[List[torch.FloatTensor]]`, `List[List[np.ndarray]]` or `List[List[PIL.Image.Image]]`):
286 | The ControlNet input condition to provide guidance to the `unet` for generation. If the type is
287 | specified as `torch.FloatTensor`, it is passed to ControlNet as is. `PIL.Image.Image` can also be
288 | accepted as an image. The dimensions of the output image defaults to `image`'s dimensions. If height
289 | and/or width are passed, `image` is resized accordingly. If multiple ControlNets are specified in
290 | `init`, images must be passed as a list such that each element of the list can be correctly batched for
291 | input to a single ControlNet.
292 | height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
293 | The height in pixels of the generated image. Anything below 512 pixels won't work well for
294 | [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0)
295 | and checkpoints that are not specifically fine-tuned on low resolutions.
296 | width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
297 | The width in pixels of the generated image. Anything below 512 pixels won't work well for
298 | [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0)
299 | and checkpoints that are not specifically fine-tuned on low resolutions.
300 | num_inference_steps (`int`, *optional*, defaults to 50):
301 | The number of denoising steps. More denoising steps usually lead to a higher quality image at the
302 | expense of slower inference.
303 | guidance_scale (`float`, *optional*, defaults to 5.0):
304 | A higher guidance scale value encourages the model to generate images closely linked to the text
305 | `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
306 | negative_prompt (`str` or `List[str]`, *optional*):
307 | The prompt or prompts to guide what to not include in image generation. If not defined, you need to
308 | pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
309 | negative_prompt_2 (`str` or `List[str]`, *optional*):
310 | The prompt or prompts to guide what to not include in image generation. This is sent to `tokenizer_2`
311 | and `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders.
312 | num_images_per_prompt (`int`, *optional*, defaults to 1):
313 | The number of images to generate per prompt.
314 | eta (`float`, *optional*, defaults to 0.0):
315 | Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
316 | to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
317 | generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
318 | A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
319 | generation deterministic.
320 | latents (`torch.FloatTensor`, *optional*):
321 | Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
322 | generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
323 | tensor is generated by sampling using the supplied random `generator`.
324 | prompt_embeds (`torch.FloatTensor`, *optional*):
325 | Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
326 | provided, text embeddings are generated from the `prompt` input argument.
327 | negative_prompt_embeds (`torch.FloatTensor`, *optional*):
328 | Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
329 | not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
330 | pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
331 | Pre-generated pooled text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
332 | not provided, pooled text embeddings are generated from `prompt` input argument.
333 | negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
334 | Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs (prompt
335 | weighting). If not provided, pooled `negative_prompt_embeds` are generated from `negative_prompt` input
336 | argument.
337 | image_embeds (`torch.FloatTensor`, *optional*):
338 | Pre-generated image embeddings.
339 | output_type (`str`, *optional*, defaults to `"pil"`):
340 | The output format of the generated image. Choose between `PIL.Image` or `np.array`.
341 | return_dict (`bool`, *optional*, defaults to `True`):
342 | Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
343 | plain tuple.
344 | cross_attention_kwargs (`dict`, *optional*):
345 | A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
346 | [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
347 | controlnet_conditioning_scale (`float` or `List[float]`, *optional*, defaults to 1.0):
348 | The outputs of the ControlNet are multiplied by `controlnet_conditioning_scale` before they are added
349 | to the residual in the original `unet`. If multiple ControlNets are specified in `init`, you can set
350 | the corresponding scale as a list.
351 | guess_mode (`bool`, *optional*, defaults to `False`):
352 | The ControlNet encoder tries to recognize the content of the input image even if you remove all
353 | prompts. A `guidance_scale` value between 3.0 and 5.0 is recommended.
354 | control_guidance_start (`float` or `List[float]`, *optional*, defaults to 0.0):
355 | The percentage of total steps at which the ControlNet starts applying.
356 | control_guidance_end (`float` or `List[float]`, *optional*, defaults to 1.0):
357 | The percentage of total steps at which the ControlNet stops applying.
358 | original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
359 | If `original_size` is not the same as `target_size` the image will appear to be down- or upsampled.
360 | `original_size` defaults to `(height, width)` if not specified. Part of SDXL's micro-conditioning as
361 | explained in section 2.2 of
362 | [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
363 | crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
364 | `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position
365 | `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting
366 | `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of
367 | [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
368 | target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
369 | For most cases, `target_size` should be set to the desired height and width of the generated image. If
370 | not specified it will default to `(height, width)`. Part of SDXL's micro-conditioning as explained in
371 | section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
372 | negative_original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
373 | To negatively condition the generation process based on a specific image resolution. Part of SDXL's
374 | micro-conditioning as explained in section 2.2 of
375 | [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
376 | information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
377 | negative_crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
378 | To negatively condition the generation process based on a specific crop coordinates. Part of SDXL's
379 | micro-conditioning as explained in section 2.2 of
380 | [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
381 | information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
382 | negative_target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
383 | To negatively condition the generation process based on a target image resolution. It should be as same
384 | as the `target_size` for most cases. Part of SDXL's micro-conditioning as explained in section 2.2 of
385 | [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
386 | information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
387 | clip_skip (`int`, *optional*):
388 | Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
389 | the output of the pre-final layer will be used for computing the prompt embeddings.
390 | callback_on_step_end (`Callable`, *optional*):
391 | A function that calls at the end of each denoising steps during the inference. The function is called
392 | with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
393 | callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
394 | `callback_on_step_end_tensor_inputs`.
395 | callback_on_step_end_tensor_inputs (`List`, *optional*):
396 | The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
397 | will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
398 | `._callback_tensor_inputs` attribute of your pipeine class.
399 |
400 | Examples:
401 |
402 | Returns:
403 | [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
404 | If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned,
405 | otherwise a `tuple` is returned containing the output images.
406 | """
407 |
408 | callback = kwargs.pop("callback", None)
409 | callback_steps = kwargs.pop("callback_steps", None)
410 |
411 | if callback is not None:
412 | deprecate(
413 | "callback",
414 | "1.0.0",
415 | "Passing `callback` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`",
416 | )
417 | if callback_steps is not None:
418 | deprecate(
419 | "callback_steps",
420 | "1.0.0",
421 | "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`",
422 | )
423 |
424 | controlnet = self.controlnet._orig_mod if is_compiled_module(self.controlnet) else self.controlnet
425 |
426 | # align format for control guidance
427 | if not isinstance(control_guidance_start, list) and isinstance(control_guidance_end, list):
428 | control_guidance_start = len(control_guidance_end) * [control_guidance_start]
429 | elif not isinstance(control_guidance_end, list) and isinstance(control_guidance_start, list):
430 | control_guidance_end = len(control_guidance_start) * [control_guidance_end]
431 | elif not isinstance(control_guidance_start, list) and not isinstance(control_guidance_end, list):
432 | mult = len(controlnet.nets) if isinstance(controlnet, MultiControlNetModel) else 1
433 | control_guidance_start, control_guidance_end = (
434 | mult * [control_guidance_start],
435 | mult * [control_guidance_end],
436 | )
437 |
438 | # 1. Check inputs. Raise error if not correct
439 | self.check_inputs(
440 | prompt,
441 | prompt_2,
442 | image,
443 | callback_steps,
444 | negative_prompt,
445 | negative_prompt_2,
446 | prompt_embeds,
447 | negative_prompt_embeds,
448 | pooled_prompt_embeds,
449 | negative_pooled_prompt_embeds,
450 | controlnet_conditioning_scale,
451 | control_guidance_start,
452 | control_guidance_end,
453 | callback_on_step_end_tensor_inputs,
454 | )
455 |
456 | self._guidance_scale = guidance_scale
457 | self._clip_skip = clip_skip
458 | self._cross_attention_kwargs = cross_attention_kwargs
459 |
460 | # 2. Define call parameters
461 | if prompt is not None and isinstance(prompt, str):
462 | batch_size = 1
463 | elif prompt is not None and isinstance(prompt, list):
464 | batch_size = len(prompt)
465 | else:
466 | batch_size = prompt_embeds.shape[0]
467 |
468 | device = self._execution_device
469 |
470 | if isinstance(controlnet, MultiControlNetModel) and isinstance(controlnet_conditioning_scale, float):
471 | controlnet_conditioning_scale = [controlnet_conditioning_scale] * len(controlnet.nets)
472 |
473 | global_pool_conditions = (
474 | controlnet.config.global_pool_conditions
475 | if isinstance(controlnet, ControlNetModel)
476 | else controlnet.nets[0].config.global_pool_conditions
477 | )
478 | guess_mode = guess_mode or global_pool_conditions
479 |
480 | # 3.1 Encode input prompt
481 | text_encoder_lora_scale = (
482 | self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None
483 | )
484 | (
485 | prompt_embeds,
486 | negative_prompt_embeds,
487 | pooled_prompt_embeds,
488 | negative_pooled_prompt_embeds,
489 | ) = self.encode_prompt(
490 | prompt,
491 | prompt_2,
492 | device,
493 | num_images_per_prompt,
494 | self.do_classifier_free_guidance,
495 | negative_prompt,
496 | negative_prompt_2,
497 | prompt_embeds=prompt_embeds,
498 | negative_prompt_embeds=negative_prompt_embeds,
499 | pooled_prompt_embeds=pooled_prompt_embeds,
500 | negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
501 | lora_scale=text_encoder_lora_scale,
502 | clip_skip=self.clip_skip,
503 | )
504 |
505 | # 3.2 Encode image prompt
506 | prompt_image_emb = self._encode_prompt_image_emb(image_embeds,
507 | device,
508 | self.unet.dtype,
509 | self.do_classifier_free_guidance)
510 |
511 | # 4. Prepare image
512 | if isinstance(controlnet, ControlNetModel):
513 | image = self.prepare_image(
514 | image=image,
515 | width=width,
516 | height=height,
517 | batch_size=batch_size * num_images_per_prompt,
518 | num_images_per_prompt=num_images_per_prompt,
519 | device=device,
520 | dtype=controlnet.dtype,
521 | do_classifier_free_guidance=self.do_classifier_free_guidance,
522 | guess_mode=guess_mode,
523 | )
524 | height, width = image.shape[-2:]
525 | elif isinstance(controlnet, MultiControlNetModel):
526 | images = []
527 |
528 | for image_ in image:
529 | image_ = self.prepare_image(
530 | image=image_,
531 | width=width,
532 | height=height,
533 | batch_size=batch_size * num_images_per_prompt,
534 | num_images_per_prompt=num_images_per_prompt,
535 | device=device,
536 | dtype=controlnet.dtype,
537 | do_classifier_free_guidance=self.do_classifier_free_guidance,
538 | guess_mode=guess_mode,
539 | )
540 |
541 | images.append(image_)
542 |
543 | image = images
544 | height, width = image[0].shape[-2:]
545 | else:
546 | assert False
547 |
548 | # 5. Prepare timesteps
549 | self.scheduler.set_timesteps(num_inference_steps, device=device)
550 | timesteps = self.scheduler.timesteps
551 | self._num_timesteps = len(timesteps)
552 |
553 | # 6. Prepare latent variables
554 | num_channels_latents = self.unet.config.in_channels
555 | latents = self.prepare_latents(
556 | batch_size * num_images_per_prompt,
557 | num_channels_latents,
558 | height,
559 | width,
560 | prompt_embeds.dtype,
561 | device,
562 | generator,
563 | latents,
564 | )
565 |
566 | # 6.5 Optionally get Guidance Scale Embedding
567 | timestep_cond = None
568 | if self.unet.config.time_cond_proj_dim is not None:
569 | guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt)
570 | timestep_cond = self.get_guidance_scale_embedding(
571 | guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim
572 | ).to(device=device, dtype=latents.dtype)
573 |
574 | # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
575 | extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
576 |
577 | # 7.1 Create tensor stating which controlnets to keep
578 | controlnet_keep = []
579 | for i in range(len(timesteps)):
580 | keeps = [
581 | 1.0 - float(i / len(timesteps) < s or (i + 1) / len(timesteps) > e)
582 | for s, e in zip(control_guidance_start, control_guidance_end)
583 | ]
584 | controlnet_keep.append(keeps[0] if isinstance(controlnet, ControlNetModel) else keeps)
585 |
586 | # 7.2 Prepare added time ids & embeddings
587 | if isinstance(image, list):
588 | original_size = original_size or image[0].shape[-2:]
589 | else:
590 | original_size = original_size or image.shape[-2:]
591 | target_size = target_size or (height, width)
592 |
593 | add_text_embeds = pooled_prompt_embeds
594 | if self.text_encoder_2 is None:
595 | text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1])
596 | else:
597 | text_encoder_projection_dim = self.text_encoder_2.config.projection_dim
598 |
599 | add_time_ids = self._get_add_time_ids(
600 | original_size,
601 | crops_coords_top_left,
602 | target_size,
603 | dtype=prompt_embeds.dtype,
604 | text_encoder_projection_dim=text_encoder_projection_dim,
605 | )
606 |
607 | if negative_original_size is not None and negative_target_size is not None:
608 | negative_add_time_ids = self._get_add_time_ids(
609 | negative_original_size,
610 | negative_crops_coords_top_left,
611 | negative_target_size,
612 | dtype=prompt_embeds.dtype,
613 | text_encoder_projection_dim=text_encoder_projection_dim,
614 | )
615 | else:
616 | negative_add_time_ids = add_time_ids
617 |
618 | if self.do_classifier_free_guidance:
619 | prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
620 | add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0)
621 | add_time_ids = torch.cat([negative_add_time_ids, add_time_ids], dim=0)
622 |
623 | prompt_embeds = prompt_embeds.to(device)
624 | add_text_embeds = add_text_embeds.to(device)
625 | add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1)
626 | encoder_hidden_states = torch.cat([prompt_embeds, prompt_image_emb], dim=1)
627 |
628 | # 8. Denoising loop
629 | num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
630 | is_unet_compiled = is_compiled_module(self.unet)
631 | is_controlnet_compiled = is_compiled_module(self.controlnet)
632 | is_torch_higher_equal_2_1 = is_torch_version(">=", "2.1")
633 |
634 | with self.progress_bar(total=num_inference_steps) as progress_bar:
635 | for i, t in enumerate(timesteps):
636 | # Relevant thread:
637 | # https://dev-discuss.pytorch.org/t/cudagraphs-in-pytorch-2-0/1428
638 | if (is_unet_compiled and is_controlnet_compiled) and is_torch_higher_equal_2_1:
639 | torch._inductor.cudagraph_mark_step_begin()
640 | # expand the latents if we are doing classifier free guidance
641 | latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
642 | latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
643 |
644 | added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids}
645 |
646 | # controlnet(s) inference
647 | if guess_mode and self.do_classifier_free_guidance:
648 | # Infer ControlNet only for the conditional batch.
649 | control_model_input = latents
650 | control_model_input = self.scheduler.scale_model_input(control_model_input, t)
651 | controlnet_prompt_embeds = prompt_embeds.chunk(2)[1]
652 | controlnet_added_cond_kwargs = {
653 | "text_embeds": add_text_embeds.chunk(2)[1],
654 | "time_ids": add_time_ids.chunk(2)[1],
655 | }
656 | else:
657 | control_model_input = latent_model_input
658 | controlnet_prompt_embeds = prompt_embeds
659 | controlnet_added_cond_kwargs = added_cond_kwargs
660 |
661 | if isinstance(controlnet_keep[i], list):
662 | cond_scale = [c * s for c, s in zip(controlnet_conditioning_scale, controlnet_keep[i])]
663 | else:
664 | controlnet_cond_scale = controlnet_conditioning_scale
665 | if isinstance(controlnet_cond_scale, list):
666 | controlnet_cond_scale = controlnet_cond_scale[0]
667 | cond_scale = controlnet_cond_scale * controlnet_keep[i]
668 |
669 | down_block_res_samples, mid_block_res_sample = self.controlnet(
670 | control_model_input,
671 | t,
672 | encoder_hidden_states=prompt_image_emb,
673 | controlnet_cond=image,
674 | conditioning_scale=cond_scale,
675 | guess_mode=guess_mode,
676 | added_cond_kwargs=controlnet_added_cond_kwargs,
677 | return_dict=False,
678 | )
679 |
680 | if guess_mode and self.do_classifier_free_guidance:
681 | # Infered ControlNet only for the conditional batch.
682 | # To apply the output of ControlNet to both the unconditional and conditional batches,
683 | # add 0 to the unconditional batch to keep it unchanged.
684 | down_block_res_samples = [torch.cat([torch.zeros_like(d), d]) for d in down_block_res_samples]
685 | mid_block_res_sample = torch.cat([torch.zeros_like(mid_block_res_sample), mid_block_res_sample])
686 |
687 | # predict the noise residual
688 | noise_pred = self.unet(
689 | latent_model_input,
690 | t,
691 | encoder_hidden_states=encoder_hidden_states,
692 | timestep_cond=timestep_cond,
693 | cross_attention_kwargs=self.cross_attention_kwargs,
694 | down_block_additional_residuals=down_block_res_samples,
695 | mid_block_additional_residual=mid_block_res_sample,
696 | added_cond_kwargs=added_cond_kwargs,
697 | return_dict=False,
698 | )[0]
699 |
700 | # perform guidance
701 | if self.do_classifier_free_guidance:
702 | noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
703 | noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
704 |
705 | # compute the previous noisy sample x_t -> x_t-1
706 | latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
707 |
708 | if callback_on_step_end is not None:
709 | callback_kwargs = {}
710 | for k in callback_on_step_end_tensor_inputs:
711 | callback_kwargs[k] = locals()[k]
712 | callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
713 |
714 | latents = callback_outputs.pop("latents", latents)
715 | prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
716 | negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
717 |
718 | # call the callback, if provided
719 | if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
720 | progress_bar.update()
721 | if callback is not None and i % callback_steps == 0:
722 | step_idx = i // getattr(self.scheduler, "order", 1)
723 | callback(step_idx, t, latents)
724 |
725 | if not output_type == "latent":
726 | # make sure the VAE is in float32 mode, as it overflows in float16
727 | needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
728 | if needs_upcasting:
729 | self.upcast_vae()
730 | latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
731 |
732 | image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
733 |
734 | # cast back to fp16 if needed
735 | if needs_upcasting:
736 | self.vae.to(dtype=torch.float16)
737 | else:
738 | image = latents
739 |
740 | if not output_type == "latent":
741 | # apply watermark if available
742 | if self.watermark is not None:
743 | image = self.watermark.apply_watermark(image)
744 |
745 | image = self.image_processor.postprocess(image, output_type=output_type)
746 |
747 | # Offload all models
748 | self.maybe_free_model_hooks()
749 |
750 | if not return_dict:
751 | return (image,)
752 |
753 | return StableDiffusionXLPipelineOutput(images=image)
754 |
--------------------------------------------------------------------------------
/requirements.txt:
--------------------------------------------------------------------------------
1 | opencv-python
2 | transformers
3 | accelerate
4 | insightface
5 | safetensors
6 | diffusers[torch]
7 | omegaconf
8 | ip_adapter
--------------------------------------------------------------------------------