init spaces

.gitattributes

@@ -1,35 +1,8 @@
-*.7z filter=lfs diff=lfs merge=lfs -text
-*.arrow filter=lfs diff=lfs merge=lfs -text
-*.bin filter=lfs diff=lfs merge=lfs -text
-*.bz2 filter=lfs diff=lfs merge=lfs -text
-*.ckpt filter=lfs diff=lfs merge=lfs -text
-*.ftz filter=lfs diff=lfs merge=lfs -text
-*.gz filter=lfs diff=lfs merge=lfs -text
-*.h5 filter=lfs diff=lfs merge=lfs -text
-*.joblib filter=lfs diff=lfs merge=lfs -text
-*.lfs.* filter=lfs diff=lfs merge=lfs -text
-*.mlmodel filter=lfs diff=lfs merge=lfs -text
-*.model filter=lfs diff=lfs merge=lfs -text
-*.msgpack filter=lfs diff=lfs merge=lfs -text
-*.npy filter=lfs diff=lfs merge=lfs -text
-*.npz filter=lfs diff=lfs merge=lfs -text
-*.onnx filter=lfs diff=lfs merge=lfs -text
-*.ot filter=lfs diff=lfs merge=lfs -text
-*.parquet filter=lfs diff=lfs merge=lfs -text
-*.pb filter=lfs diff=lfs merge=lfs -text
-*.pickle filter=lfs diff=lfs merge=lfs -text
-*.pkl filter=lfs diff=lfs merge=lfs -text
-*.pt filter=lfs diff=lfs merge=lfs -text
 *.pth filter=lfs diff=lfs merge=lfs -text
-*.rar filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
 *.safetensors filter=lfs diff=lfs merge=lfs -text
-saved_model/**/* filter=lfs diff=lfs merge=lfs -text
-*.tar.* filter=lfs diff=lfs merge=lfs -text
-*.tar filter=lfs diff=lfs merge=lfs -text
-*.tflite filter=lfs diff=lfs merge=lfs -text
-*.tgz filter=lfs diff=lfs merge=lfs -text
-*.wasm filter=lfs diff=lfs merge=lfs -text
-*.xz filter=lfs diff=lfs merge=lfs -text
-*.zip filter=lfs diff=lfs merge=lfs -text
-*.zst filter=lfs diff=lfs merge=lfs -text
-*tfevents* filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+tools/ffmpeg filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.wav filter=lfs diff=lfs merge=lfs -text
+*.MP3 filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -0,0 +1,137 @@
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+*$py.class
+
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+pip-wheel-metadata/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# PyInstaller
+#  Usually these files are written by a python script from a template
+#  before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.nox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+.hypothesis/
+.pytest_cache/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# IPython
+profile_default/
+ipython_config.py
+
+# pyenv
+.python-version
+
+# pipenv
+#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
+#   However, in case of collaboration, if having platform-specific dependencies or dependencies
+#   having no cross-platform support, pipenv may install dependencies that don’t work, or not
+#   install all needed dependencies.
+#Pipfile.lock
+
+# celery beat schedule file
+celerybeat-schedule
+
+# SageMath parsed files
+*.sage.py
+
+# Environments
+.env
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# Pyre type checker
+.pyre/
+
+*.swp
+.*.swp
+
+.DS_Store
+
+# project 
+
+.idea
+.vscode
+*.mp4
+res_path/
+tmp_path/
+flagged/

LICENSE

@@ -0,0 +1,107 @@
+Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International
+
+Creative Commons Corporation ("Creative Commons") is not a law firm and does not provide legal services or legal advice. Distribution of Creative Commons public licenses does not create a lawyer-client or other relationship. Creative Commons makes its licenses and related information available on an "as-is" basis. Creative Commons gives no warranties regarding its licenses, any material licensed under their terms and conditions, or any related information. Creative Commons disclaims all liability for damages resulting from their use to the fullest extent possible.
+
+Using Creative Commons Public Licenses
+
+Creative Commons public licenses provide a standard set of terms and conditions that creators and other rights holders may use to share original works of authorship and other material subject to copyright and certain other rights specified in the public license below. The following considerations are for informational purposes only, are not exhaustive, and do not form part of our licenses.
+
+Considerations for licensors: Our public licenses are intended for use by those authorized to give the public permission to use material in ways otherwise restricted by copyright and certain other rights. Our licenses are irrevocable. Licensors should read and understand the terms and conditions of the license they choose before applying it. Licensors should also secure all rights necessary before applying our licenses so that the public can reuse the material as expected. Licensors should clearly mark any material not subject to the license. This includes other CC-licensed material, or material used under an exception or limitation to copyright. More considerations for licensors : wiki.creativecommons.org/Considerations_for_licensors
+
+Considerations for the public: By using one of our public licenses, a licensor grants the public permission to use the licensed material under specified terms and conditions. If the licensor's permission is not necessary for any reason–for example, because of any applicable exception or limitation to copyright–then that use is not regulated by the license. Our licenses grant only permissions under copyright and certain other rights that a licensor has authority to grant. Use of the licensed material may still be restricted for other reasons, including because others have copyright or other rights in the material. A licensor may make special requests, such as asking that all changes be marked or described. Although not required by our licenses, you are encouraged to respect those requests where reasonable. More considerations for the public : wiki.creativecommons.org/Considerations_for_licensees
+
+Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International Public License
+
+By exercising the Licensed Rights (defined below), You accept and agree to be bound by the terms and conditions of this Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International Public License ("Public License"). To the extent this Public License may be interpreted as a contract, You are granted the Licensed Rights in consideration of Your acceptance of these terms and conditions, and the Licensor grants You such rights in consideration of benefits the Licensor receives from making the Licensed Material available under these terms and conditions.
+
+Section 1 – Definitions.
+
+a. Adapted Material means material subject to Copyright and Similar Rights that is derived from or based upon the Licensed Material and in which the Licensed Material is translated, altered, arranged, transformed, or otherwise modified in a manner requiring permission under the Copyright and Similar Rights held by the Licensor. For purposes of this Public License, where the Licensed Material is a musical work, performance, or sound recording, Adapted Material is always produced where the Licensed Material is synched in timed relation with a moving image.
+b. Adapter's License means the license You apply to Your Copyright and Similar Rights in Your contributions to Adapted Material in accordance with the terms and conditions of this Public License.
+c. BY-NC-SA Compatible License means a license listed at creativecommons.org/compatiblelicenses, approved by Creative Commons as essentially the equivalent of this Public License.
+d. Copyright and Similar Rights means copyright and/or similar rights closely related to copyright including, without limitation, performance, broadcast, sound recording, and Sui Generis Database Rights, without regard to how the rights are labeled or categorized. For purposes of this Public License, the rights specified in Section 2(b)(1)-(2) are not Copyright and Similar Rights.
+e. Effective Technological Measures means those measures that, in the absence of proper authority, may not be circumvented under laws fulfilling obligations under Article 11 of the WIPO Copyright Treaty adopted on December 20, 1996, and/or similar international agreements.
+f. Exceptions and Limitations means fair use, fair dealing, and/or any other exception or limitation to Copyright and Similar Rights that applies to Your use of the Licensed Material.
+g. License Elements means the license attributes listed in the name of a Creative Commons Public License. The License Elements of this Public License are Attribution, NonCommercial, and ShareAlike.
+h. Licensed Material means the artistic or literary work, database, or other material to which the Licensor applied this Public License.
+i. Licensed Rights means the rights granted to You subject to the terms and conditions of this Public License, which are limited to all Copyright and Similar Rights that apply to Your use of the Licensed Material and that the Licensor has authority to license.
+j. Licensor means the individual(s) or entity(ies) granting rights under this Public License.
+k. NonCommercial means not primarily intended for or directed towards commercial advantage or monetary compensation. For purposes of this Public License, the exchange of the Licensed Material for other material subject to Copyright and Similar Rights by digital file-sharing or similar means is NonCommercial provided there is no payment of monetary compensation in connection with the exchange.
+l. Share means to provide material to the public by any means or process that requires permission under the Licensed Rights, such as reproduction, public display, public performance, distribution, dissemination, communication, or importation, and to make material available to the public including in ways that members of the public may access the material from a place and at a time individually chosen by them.
+m. Sui Generis Database Rights means rights other than copyright resulting from Directive 96/9/EC of the European Parliament and of the Council of 11 March 1996 on the legal protection of databases, as amended and/or succeeded, as well as other essentially equivalent rights anywhere in the world.
+n. You means the individual or entity exercising the Licensed Rights under this Public License. Your has a corresponding meaning.
+Section 2 – Scope.
+
+a. License grant.
+1. Subject to the terms and conditions of this Public License, the Licensor hereby grants You a worldwide, royalty-free, non-sublicensable, non-exclusive, irrevocable license to exercise the Licensed Rights in the Licensed Material to:
+A. reproduce and Share the Licensed Material, in whole or in part, for NonCommercial purposes only; and
+B. produce, reproduce, and Share Adapted Material for NonCommercial purposes only.
+2. Exceptions and Limitations. For the avoidance of doubt, where Exceptions and Limitations apply to Your use, this Public License does not apply, and You do not need to comply with its terms and conditions.
+3. Term. The term of this Public License is specified in Section 6(a).
+4. Media and formats; technical modifications allowed. The Licensor authorizes You to exercise the Licensed Rights in all media and formats whether now known or hereafter created, and to make technical modifications necessary to do so. The Licensor waives and/or agrees not to assert any right or authority to forbid You from making technical modifications necessary to exercise the Licensed Rights, including technical modifications necessary to circumvent Effective Technological Measures. For purposes of this Public License, simply making modifications authorized by this Section 2(a)(4) never produces Adapted Material.
+5. Downstream recipients.
+A. Offer from the Licensor – Licensed Material. Every recipient of the Licensed Material automatically receives an offer from the Licensor to exercise the Licensed Rights under the terms and conditions of this Public License.
+B. Additional offer from the Licensor – Adapted Material. Every recipient of Adapted Material from You automatically receives an offer from the Licensor to exercise the Licensed Rights in the Adapted Material under the conditions of the Adapter's License You apply.
+C. No downstream restrictions. You may not offer or impose any additional or different terms or conditions on, or apply any Effective Technological Measures to, the Licensed Material if doing so restricts exercise of the Licensed Rights by any recipient of the Licensed Material.
+6. No endorsement. Nothing in this Public License constitutes or may be construed as permission to assert or imply that You are, or that Your use of the Licensed Material is, connected with, or sponsored, endorsed, or granted official status by, the Licensor or others designated to receive attribution as provided in Section 3(a)(1)(A)(i).
+b. Other rights.
+1. Moral rights, such as the right of integrity, are not licensed under this Public License, nor are publicity, privacy, and/or other similar personality rights; however, to the extent possible, the Licensor waives and/or agrees not to assert any such rights held by the Licensor to the limited extent necessary to allow You to exercise the Licensed Rights, but not otherwise.
+2. Patent and trademark rights are not licensed under this Public License.
+3. To the extent possible, the Licensor waives any right to collect royalties from You for the exercise of the Licensed Rights, whether directly or through a collecting society under any voluntary or waivable statutory or compulsory licensing scheme. In all other cases the Licensor expressly reserves any right to collect such royalties, including when the Licensed Material is used other than for NonCommercial purposes.
+Section 3 – License Conditions.
+
+Your exercise of the Licensed Rights is expressly made subject to the following conditions.
+
+a. Attribution.
+1. If You Share the Licensed Material (including in modified form), You must:
+A. retain the following if it is supplied by the Licensor with the Licensed Material:
+i. identification of the creator(s) of the Licensed Material and any others designated to receive attribution, in any reasonable manner requested by the Licensor (including by pseudonym if designated);
+ii. a copyright notice;
+iii. a notice that refers to this Public License;
+iv. a notice that refers to the disclaimer of warranties;
+v. a URI or hyperlink to the Licensed Material to the extent reasonably practicable;
+
+B. indicate if You modified the Licensed Material and retain an indication of any previous modifications; and
+C. indicate the Licensed Material is licensed under this Public License, and include the text of, or the URI or hyperlink to, this Public License.
+2. You may satisfy the conditions in Section 3(a)(1) in any reasonable manner based on the medium, means, and context in which You Share the Licensed Material. For example, it may be reasonable to satisfy the conditions by providing a URI or hyperlink to a resource that includes the required information.
+3. If requested by the Licensor, You must remove any of the information required by Section 3(a)(1)(A) to the extent reasonably practicable.
+b. ShareAlike.In addition to the conditions in Section 3(a), if You Share Adapted Material You produce, the following conditions also apply.
+1. The Adapter's License You apply must be a Creative Commons license with the same License Elements, this version or later, or a BY-NC-SA Compatible License.
+2. You must include the text of, or the URI or hyperlink to, the Adapter's License You apply. You may satisfy this condition in any reasonable manner based on the medium, means, and context in which You Share Adapted Material.
+3. You may not offer or impose any additional or different terms or conditions on, or apply any Effective Technological Measures to, Adapted Material that restrict exercise of the rights granted under the Adapter's License You apply.
+Section 4 – Sui Generis Database Rights.
+
+Where the Licensed Rights include Sui Generis Database Rights that apply to Your use of the Licensed Material:
+
+a. for the avoidance of doubt, Section 2(a)(1) grants You the right to extract, reuse, reproduce, and Share all or a substantial portion of the contents of the database for NonCommercial purposes only;
+b. if You include all or a substantial portion of the database contents in a database in which You have Sui Generis Database Rights, then the database in which You have Sui Generis Database Rights (but not its individual contents) is Adapted Material, including for purposes of Section 3(b); and
+c. You must comply with the conditions in Section 3(a) if You Share all or a substantial portion of the contents of the database.
+For the avoidance of doubt, this Section 4 supplements and does not replace Your obligations under this Public License where the Licensed Rights include other Copyright and Similar Rights.
+Section 5 – Disclaimer of Warranties and Limitation of Liability.
+
+a. Unless otherwise separately undertaken by the Licensor, to the extent possible, the Licensor offers the Licensed Material as-is and as-available, and makes no representations or warranties of any kind concerning the Licensed Material, whether express, implied, statutory, or other. This includes, without limitation, warranties of title, merchantability, fitness for a particular purpose, non-infringement, absence of latent or other defects, accuracy, or the presence or absence of errors, whether or not known or discoverable. Where disclaimers of warranties are not allowed in full or in part, this disclaimer may not apply to You.
+b. To the extent possible, in no event will the Licensor be liable to You on any legal theory (including, without limitation, negligence) or otherwise for any direct, special, indirect, incidental, consequential, punitive, exemplary, or other losses, costs, expenses, or damages arising out of this Public License or use of the Licensed Material, even if the Licensor has been advised of the possibility of such losses, costs, expenses, or damages. Where a limitation of liability is not allowed in full or in part, this limitation may not apply to You.
+c. The disclaimer of warranties and limitation of liability provided above shall be interpreted in a manner that, to the extent possible, most closely approximates an absolute disclaimer and waiver of all liability.
+Section 6 – Term and Termination.
+
+a. This Public License applies for the term of the Copyright and Similar Rights licensed here. However, if You fail to comply with this Public License, then Your rights under this Public License terminate automatically.
+b. Where Your right to use the Licensed Material has terminated under Section 6(a), it reinstates:
+1. automatically as of the date the violation is cured, provided it is cured within 30 days of Your discovery of the violation; or
+2. upon express reinstatement by the Licensor.
+For the avoidance of doubt, this Section 6(b) does not affect any right the Licensor may have to seek remedies for Your violations of this Public License.
+
+c. For the avoidance of doubt, the Licensor may also offer the Licensed Material under separate terms or conditions or stop distributing the Licensed Material at any time; however, doing so will not terminate this Public License.
+d. Sections 1, 5, 6, 7, and 8 survive termination of this Public License.
+Section 7 – Other Terms and Conditions.
+
+a. The Licensor shall not be bound by any additional or different terms or conditions communicated by You unless expressly agreed.
+b. Any arrangements, understandings, or agreements regarding the Licensed Material not stated herein are separate from and independent of the terms and conditions of this Public License.
+Section 8 – Interpretation.
+
+a. For the avoidance of doubt, this Public License does not, and shall not be interpreted to, reduce, limit, restrict, or impose conditions on any use of the Licensed Material that could lawfully be made without permission under this Public License.
+b. To the extent possible, if any provision of this Public License is deemed unenforceable, it shall be automatically reformed to the minimum extent necessary to make it enforceable. If the provision cannot be reformed, it shall be severed from this Public License without affecting the enforceability of the remaining terms and conditions.
+c. No term or condition of this Public License will be waived and no failure to comply consented to unless expressly agreed to by the Licensor.
+d. Nothing in this Public License constitutes or may be interpreted as a limitation upon, or waiver of, any privileges and immunities that apply to the Licensor or You, including from the legal processes of any jurisdiction or authority.
+Creative Commons is not a party to its public licenses. Notwithstanding, Creative Commons may elect to apply one of its public licenses to material it publishes and in those instances will be considered the "Licensor." The text of the Creative Commons public licenses is dedicated to the public domain under the CC0 Public Domain Dedication. Except for the limited purpose of indicating that material is shared under a Creative Commons public license or as otherwise permitted by the Creative Commons policies published at creativecommons.org/policies, Creative Commons does not authorize the use of the trademark "Creative Commons" or any other trademark or logo of Creative Commons without its prior written consent including, without limitation, in connection with any unauthorized modifications to any of its public licenses or any other arrangements, understandings, or agreements concerning use of licensed material. For the avoidance of doubt, this paragraph does not form part of the public licenses.
+
+Creative Commons may be contacted at creativecommons.org.

app.py

@@ -0,0 +1,125 @@
+import gradio as gr
+import os
+import numpy as np
+from pydub import AudioSegment
+import hashlib
+from sonic import Sonic
+
+cmd = 'python3 -m pip install "huggingface_hub[cli]"; \
+huggingface-cli download LeonJoe13/Sonic --local-dir  checkpoints; \
+huggingface-cli download stabilityai/stable-video-diffusion-img2vid-xt --local-dir  checkpoints/stable-video-diffusion-img2vid-xt; \
+huggingface-cli download openai/whisper-tiny --local-dir checkpoints/whisper-tiny'
+os.system(cmd)
+
+
+pipe = Sonic(0)
+
+def get_md5(content):
+    md5hash = hashlib.md5(content)
+    md5 = md5hash.hexdigest()
+    return md5
+
+def get_video_res(img_path, audio_path, res_video_path, dynamic_scale=1.0):
+
+    expand_ratio = 0.5
+    min_resolution = 512
+    inference_steps = 25
+
+    face_info = pipe.preprocess(img_path, expand_ratio=expand_ratio)
+    print(face_info)
+    if face_info['face_num'] > 0:
+        crop_image_path = img_path + '.crop.png'
+        pipe.crop_image(img_path, crop_image_path, face_info['crop_bbox'])
+        img_path = crop_image_path
+        os.makedirs(os.path.dirname(res_video_path), exist_ok=True)
+        pipe.process(img_path, audio_path, res_video_path, min_resolution=min_resolution, inference_steps=inference_steps, dynamic_scale=dynamic_scale)
+    else:
+        return -1
+tmp_path = './tmp_path/'
+res_path = './res_path/'
+os.makedirs(tmp_path,exist_ok=1)
+os.makedirs(res_path,exist_ok=1)
+
+def process_sonic(image,audio,s0):
+    img_md5= get_md5(np.array(image))
+    audio_md5 = get_md5(audio[1])
+    print(img_md5,audio_md5)
+    sampling_rate, arr = audio[:2]
+    if len(arr.shape)==1:
+        arr = arr[:,None]
+    audio = AudioSegment(
+        arr.tobytes(),
+        frame_rate=sampling_rate,
+        sample_width=arr.dtype.itemsize,
+        channels=arr.shape[1]
+    )
+    audio = audio.set_frame_rate(sampling_rate)
+    image_path = os.path.abspath(tmp_path+'{0}.png'.format(img_md5))
+    audio_path = os.path.abspath(tmp_path+'{0}.wav'.format(audio_md5))
+    if not os.path.exists(image_path):
+        image.save(image_path)
+    if not os.path.exists(audio_path):
+        audio.export(audio_path, format="wav")
+    res_video_path = os.path.abspath(res_path+f'{img_md5}_{audio_md5}_{s0}.mp4')
+    if os.path.exists(res_video_path):
+        return res_video_path
+    else:
+        get_video_res(image_path, audio_path, res_video_path,s0)
+    return res_video_path
+    
+inputs = [
+    gr.Image(type='pil',label="Upload Image"),
+    gr.Audio(label="Upload Audio"),
+    gr.Slider(0.5, 2.0, value=1.0, step=0.1, label="Dynamic scale", info="Increase/decrease to obtain more/less movements"),
+]
+outputs = gr.Video(label="output.mp4")
+
+
+html_description = """
+<div style="display: flex; justify-content: center; align-items: center;">
+  <a href="https://github.com/jixiaozhong/Sonic.git" style="margin: 0 2px;">
+    <img src='https://img.shields.io/badge/GitHub-Repo-blue?style=flat&logo=GitHub' alt='GitHub'>
+  </a>
+  <a href="https://arxiv.org/pdf/2411.16331" style="margin: 0 2px;">
+    <img src='https://img.shields.io/badge/arXiv-2411.16331-red?style=flat&logo=arXiv&logoColor=red' alt='arxiv'>
+  </a>
+  <a href='https://jixiaozhong.github.io/Sonic/' style="margin: 0 2px;">
+    <img src='https://img.shields.io/badge/Webpage-Project-silver?style=flat&logo=&logoColor=orange' alt='webpage'>
+  </a>
+  <a href="https://github.com/jixiaozhong/Sonic/blob/main/LICENSE" style="margin: 0 2px;">
+    <img src='https://img.shields.io/badge/License-CC BY--NC--SA--4.0-lightgreen?style=flat&logo=Lisence' alt='License'>
+  </a>
+</div>
+
+The demo can only be used for <b>Non-commercial Use</b>.
+<br>If you like our work, please star <a href='https://jixiaozhong.github.io/Sonic/' style="margin: 0 2px;">Sonic</a>.
+<br>Note: Audio longer than 10s will be truncated due to computing resources.
+"""
+TAIL = """
+<div style="display: flex; justify-content: center; align-items: center;">
+<a href="https://clustrmaps.com/site/1c38t"  title="ClustrMaps"><img src="//www.clustrmaps.com/map_v2.png?d=BI2nzSldyixPC88l8Kev4wjjqsU4IOk7gcvpOijolGI&cl=ffffff" /></a>
+</div>
+"""
+
+def get_example():
+    return [
+        ["examples/image/female_diaosu.png", "examples/wav/sing_female_rap_10s.MP3", 1.0],
+        ["examples/image/hair.png", "examples/wav/sing_female_10s.wav", 1.0],
+        ["examples/image/anime1.png", "examples/wav/talk_female_english_10s.MP3", 1.0],
+        ["examples/image/leonnado.jpg", "examples/wav/talk_male_law_10s.wav", 1.0],
+        
+    ]
+
+with gr.Blocks(title="Sonic") as demo:
+    gr.Interface(fn=process_sonic, inputs=inputs, outputs=outputs, title="Sonic: Shifting Focus to Global Audio Perception in Portrait Animation", description=html_description)
+    gr.Examples(
+        examples=get_example(),
+        fn=process_sonic,
+        inputs=inputs,
+        outputs=outputs,
+        cache_examples=False,)
+    gr.Markdown(TAIL)
+    
+demo.launch(server_name='0.0.0.0', server_port=8081, share=True, enable_queue=True)
+
+

config/inference/sonic.yaml

@@ -0,0 +1,28 @@
+pretrained_model_name_or_path: "checkpoints/stable-video-diffusion-img2vid-xt"
+unet_checkpoint_path: "checkpoints/Sonic/unet.pth"
+audio2token_checkpoint_path: "checkpoints/Sonic/audio2token.pth"
+audio2bucket_checkpoint_path: "checkpoints/Sonic/audio2bucket.pth"
+
+weight_dtype: 'fp16'  # [fp16, fp32]
+
+num_inference_steps: 25
+n_sample_frames: 25
+fps: 12.5
+decode_chunk_size: 8
+motion_bucket_scale: 1.0
+image_size: 512
+area: 1.1
+frame_num: 10000
+step: 2
+overlap: 0
+shift_offset: 7
+min_appearance_guidance_scale: 2.0
+max_appearance_guidance_scale: 2.0
+audio_guidance_scale: 7.5
+i2i_noise_strength: 1.0
+ip_audio_scale: 1.0
+noise_aug_strength: 0.00
+
+use_interframe: True
+
+seed: 72589

examples/wav/sing_female_10s.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:104729e65b4824df3caa05786508fddea5ccd4dfbe6c66a01dd311f82e725428
+size 640078

examples/wav/sing_female_rap_10s.MP3

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:44333e198b646dfb5389513961609515558276daefc9114336dc3ee5d75ce902
+size 160749

examples/wav/talk_female_english_10s.MP3

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7a6fac3e0fa99dab99a2049d3658f31651ec687ba670f47e6e57d8632ac30a29
+size 160749

examples/wav/talk_male_law_10s.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:686fb83475d540edd6c2899e01d57119cefb8fb441734ed10cc55819fc83d3f0
+size 1764078

requirements.txt

@@ -0,0 +1,14 @@
+diffusers==0.29.0
+torch==2.2.1
+torchaudio==2.2.1
+torchvision==0.17.1
+transformers==4.43.2
+imageio==2.31.1
+imageio-ffmpeg==0.5.1
+gradio==3.50.0
+omegaconf==2.3.0
+tqdm==4.65.2
+librosa==0.10.2.post1
+einops==0.7.0
+pydub==0.25.1
+gradio==4.3.0

sonic.py

@@ -0,0 +1,330 @@
+import os
+import torch
+import torch.utils.checkpoint
+from PIL import Image
+import numpy as np
+from omegaconf import OmegaConf
+from tqdm import tqdm
+import cv2
+
+from diffusers import AutoencoderKLTemporalDecoder
+from diffusers.schedulers import EulerDiscreteScheduler
+from transformers import WhisperModel, CLIPVisionModelWithProjection, AutoFeatureExtractor
+
+from src.utils.util import save_videos_grid, seed_everything
+from src.dataset.test_preprocess import process_bbox, image_audio_to_tensor
+from src.models.base.unet_spatio_temporal_condition import UNetSpatioTemporalConditionModel, add_ip_adapters
+from src.pipelines.pipeline_sonic import SonicPipeline
+from src.models.audio_adapter.audio_proj import AudioProjModel
+from src.models.audio_adapter.audio_to_bucket import Audio2bucketModel
+from src.utils.RIFE.RIFE_HDv3 import RIFEModel
+from src.dataset.face_align.align import AlignImage
+
+
+BASE_DIR = os.path.dirname(os.path.abspath(__file__))
+
+def test(
+    pipe,
+    config,
+    wav_enc,
+    audio_pe,
+    audio2bucket,
+    image_encoder,
+    width,
+    height,
+    batch
+):
+    for k, v in batch.items():
+        if isinstance(v, torch.Tensor):
+            batch[k] = v.unsqueeze(0).to(pipe.device).float()
+    ref_img = batch['ref_img']
+    clip_img = batch['clip_images']
+    face_mask = batch['face_mask']
+    image_embeds = image_encoder(
+        clip_img
+            ).image_embeds
+
+    audio_feature = batch['audio_feature']
+    audio_len = batch['audio_len']
+    step = int(config.step)
+
+    window = 3000
+    audio_prompts = []
+    last_audio_prompts = []
+    for i in range(0, audio_feature.shape[-1], window):
+        audio_prompt = wav_enc.encoder(audio_feature[:,:,i:i+window], output_hidden_states=True).hidden_states
+        last_audio_prompt = wav_enc.encoder(audio_feature[:,:,i:i+window]).last_hidden_state
+        last_audio_prompt = last_audio_prompt.unsqueeze(-2)
+        audio_prompt = torch.stack(audio_prompt, dim=2)
+        audio_prompts.append(audio_prompt)
+        last_audio_prompts.append(last_audio_prompt)
+
+    audio_prompts = torch.cat(audio_prompts, dim=1)
+    audio_prompts = audio_prompts[:,:audio_len*2]
+    audio_prompts = torch.cat([torch.zeros_like(audio_prompts[:,:4]), audio_prompts, torch.zeros_like(audio_prompts[:,:6])], 1)
+
+    last_audio_prompts = torch.cat(last_audio_prompts, dim=1)
+    last_audio_prompts = last_audio_prompts[:,:audio_len*2]
+    last_audio_prompts = torch.cat([torch.zeros_like(last_audio_prompts[:,:24]), last_audio_prompts, torch.zeros_like(last_audio_prompts[:,:26])], 1)
+
+
+    ref_tensor_list = []
+    audio_tensor_list = []
+    uncond_audio_tensor_list = []
+    motion_buckets = []
+    for i in tqdm(range(audio_len//step)):
+
+
+        audio_clip = audio_prompts[:,i*2*step:i*2*step+10].unsqueeze(0)
+        audio_clip_for_bucket = last_audio_prompts[:,i*2*step:i*2*step+50].unsqueeze(0)
+        motion_bucket = audio2bucket(audio_clip_for_bucket, image_embeds)
+        motion_bucket = motion_bucket * 16 + 16
+        motion_buckets.append(motion_bucket[0])
+
+        cond_audio_clip = audio_pe(audio_clip).squeeze(0)
+        uncond_audio_clip = audio_pe(torch.zeros_like(audio_clip)).squeeze(0)
+
+        ref_tensor_list.append(ref_img[0])
+        audio_tensor_list.append(cond_audio_clip[0])
+        uncond_audio_tensor_list.append(uncond_audio_clip[0])
+
+    video = pipe(
+        ref_img,
+        clip_img,
+        face_mask,
+        audio_tensor_list,
+        uncond_audio_tensor_list,
+        motion_buckets,
+        height=height,
+        width=width,
+        num_frames=len(audio_tensor_list),
+        decode_chunk_size=config.decode_chunk_size,
+        motion_bucket_scale=config.motion_bucket_scale,
+        fps=config.fps,
+        noise_aug_strength=config.noise_aug_strength,
+        min_guidance_scale1=config.min_appearance_guidance_scale, # 1.0,
+        max_guidance_scale1=config.max_appearance_guidance_scale,
+        min_guidance_scale2=config.audio_guidance_scale, # 1.0,
+        max_guidance_scale2=config.audio_guidance_scale,
+        overlap=config.overlap,
+        shift_offset=config.shift_offset,
+        frames_per_batch=config.n_sample_frames,
+        num_inference_steps=config.num_inference_steps,
+        i2i_noise_strength=config.i2i_noise_strength
+    ).frames
+
+
+    # Concat it with pose tensor
+    # pose_tensor = torch.stack(pose_tensor_list,1).unsqueeze(0)
+    video = (video*0.5 + 0.5).clamp(0, 1)
+    video = torch.cat([video.to(pipe.device)], dim=0).cpu()
+
+    return video
+
+
+class Sonic():
+    config_file = os.path.join(BASE_DIR, 'config/inference/sonic.yaml')
+    config = OmegaConf.load(config_file)
+
+    def __init__(self, 
+                 device_id=0,
+                 enable_interpolate_frame=True,
+                 ):
+        
+        config = self.config
+        config.use_interframe = enable_interpolate_frame
+
+        device = 'cuda:{}'.format(device_id) if device_id > -1 else 'cpu'
+
+        config.pretrained_model_name_or_path = os.path.join(BASE_DIR, config.pretrained_model_name_or_path)
+
+        vae = AutoencoderKLTemporalDecoder.from_pretrained(
+            config.pretrained_model_name_or_path, 
+            subfolder="vae",
+            variant="fp16")
+        
+        val_noise_scheduler = EulerDiscreteScheduler.from_pretrained(
+            config.pretrained_model_name_or_path, 
+            subfolder="scheduler")
+        
+        image_encoder = CLIPVisionModelWithProjection.from_pretrained(
+            config.pretrained_model_name_or_path, 
+            subfolder="image_encoder",
+            variant="fp16")
+        unet = UNetSpatioTemporalConditionModel.from_pretrained(
+            config.pretrained_model_name_or_path,
+            subfolder="unet",
+            variant="fp16")
+        add_ip_adapters(unet, [32], [config.ip_audio_scale])
+        
+        audio2token = AudioProjModel(seq_len=10, blocks=5, channels=384, intermediate_dim=1024, output_dim=1024, context_tokens=32).to(device)
+        audio2bucket = Audio2bucketModel(seq_len=50, blocks=1, channels=384, clip_channels=1024, intermediate_dim=1024, output_dim=1, context_tokens=2).to(device)
+
+        unet_checkpoint_path = os.path.join(BASE_DIR, config.unet_checkpoint_path)
+        audio2token_checkpoint_path = os.path.join(BASE_DIR, config.audio2token_checkpoint_path)
+        audio2bucket_checkpoint_path = os.path.join(BASE_DIR, config.audio2bucket_checkpoint_path)
+
+        unet.load_state_dict(
+            torch.load(unet_checkpoint_path, map_location="cpu"),
+            strict=True,
+        )
+        
+        audio2token.load_state_dict(
+            torch.load(audio2token_checkpoint_path, map_location="cpu"),
+            strict=True,
+        )
+
+        audio2bucket.load_state_dict(
+            torch.load(audio2bucket_checkpoint_path, map_location="cpu"),
+            strict=True,
+        )
+        
+
+        if config.weight_dtype == "fp16":
+            weight_dtype = torch.float16
+        elif config.weight_dtype == "fp32":
+            weight_dtype = torch.float32
+        elif config.weight_dtype == "bf16":
+            weight_dtype = torch.bfloat16
+        else:
+            raise ValueError(
+                f"Do not support weight dtype: {config.weight_dtype} during training"
+            )
+
+        whisper = WhisperModel.from_pretrained(os.path.join(BASE_DIR, 'checkpoints/whisper-tiny/')).to(device).eval()
+        
+        whisper.requires_grad_(False)
+
+        self.feature_extractor = AutoFeatureExtractor.from_pretrained(os.path.join(BASE_DIR, 'checkpoints/whisper-tiny/'))
+
+        det_path = os.path.join(BASE_DIR, os.path.join(BASE_DIR, 'checkpoints/yoloface_v5m.pt'))
+        self.face_det = AlignImage(device, det_path=det_path)
+        if config.use_interframe:
+            rife = RIFEModel(device=device)
+            rife.load_model(os.path.join(BASE_DIR, 'checkpoints', 'RIFE/'))
+            self.rife = rife
+
+
+        image_encoder.to(weight_dtype)
+        vae.to(weight_dtype)
+        unet.to(weight_dtype)
+
+        pipe = SonicPipeline(
+            unet=unet,
+            image_encoder=image_encoder,
+            vae=vae,
+            scheduler=val_noise_scheduler,
+        )
+        pipe = pipe.to(device=device, dtype=weight_dtype)
+
+
+        self.pipe = pipe
+        self.whisper = whisper
+        self.audio2token = audio2token
+        self.audio2bucket = audio2bucket
+        self.image_encoder = image_encoder
+        self.device = device
+
+        print('init done')
+
+
+    def preprocess(self,
+              image_path, expand_ratio=1.0):
+        face_image = cv2.imread(image_path)
+        h, w = face_image.shape[:2]
+        _, _, bboxes = self.face_det(face_image, maxface=True)
+        face_num = len(bboxes)
+        bbox = []
+        if face_num > 0:
+            x1, y1, ww, hh = bboxes[0]
+            x2, y2 = x1 + ww, y1 + hh
+            bbox = x1, y1, x2, y2
+            bbox_s = process_bbox(bbox, expand_radio=expand_ratio, height=h, width=w)
+
+        return {
+            'face_num': face_num,
+            'crop_bbox': bbox_s,
+        }
+    
+    def crop_image(self,
+                   input_image_path,
+                   output_image_path,
+                   crop_bbox):
+        face_image = cv2.imread(input_image_path)
+        crop_image = face_image[crop_bbox[1]:crop_bbox[3], crop_bbox[0]:crop_bbox[2]]
+        cv2.imwrite(output_image_path, crop_image)
+
+    @torch.no_grad()
+    def process(self,
+                image_path,
+                audio_path,
+                output_path,
+                min_resolution=512,
+                inference_steps=25,
+                dynamic_scale=1.0,
+                keep_resolution=False,
+                seed=None):
+        
+        config = self.config
+        device = self.device
+        pipe = self.pipe
+        whisper = self.whisper
+        audio2token = self.audio2token
+        audio2bucket = self.audio2bucket
+        image_encoder = self.image_encoder
+
+        # specific parameters
+        if seed:
+            config.seed = seed
+
+        config.num_inference_steps = inference_steps
+
+        config.motion_bucket_scale = dynamic_scale
+
+        seed_everything(config.seed)
+
+        video_path = output_path.replace('.mp4', '_noaudio.mp4')
+        audio_video_path = output_path
+
+        imSrc_ = Image.open(image_path).convert('RGB')
+        raw_w, raw_h = imSrc_.size
+
+        test_data = image_audio_to_tensor(self.face_det, self.feature_extractor, image_path, audio_path, limit=config.frame_num, image_size=min_resolution, area=config.area)
+        if test_data is None:
+            return -1
+        height, width = test_data['ref_img'].shape[-2:]
+        if keep_resolution:
+            resolution = f'{raw_w//2*2}x{raw_h//2*2}'
+        else:
+            resolution = f'{width}x{height}'
+
+        video = test(
+            pipe,
+            config,
+            wav_enc=whisper,
+            audio_pe=audio2token,
+            audio2bucket=audio2bucket,
+            image_encoder=image_encoder,
+            width=width,
+            height=height,
+            batch=test_data,
+            )
+
+        if config.use_interframe:
+            rife = self.rife
+            out = video.to(device)
+            results = []
+            video_len = out.shape[2]
+            for idx in tqdm(range(video_len-1), ncols=0):
+                I1 = out[:, :, idx]
+                I2 = out[:, :, idx+1]
+                middle = rife.inference(I1, I2).clamp(0, 1).detach()
+                results.append(out[:, :, idx])
+                results.append(middle)
+            results.append(out[:, :, video_len-1])
+            video = torch.stack(results, 2).cpu()
+        
+        save_videos_grid(video, video_path, n_rows=video.shape[0], fps=config.fps * 2 if config.use_interframe else config.fps)
+        os.system(f"ffmpeg -i '{video_path}'  -i '{audio_path}' -s {resolution} -vcodec libx264 -acodec aac -crf 18 -shortest '{audio_video_path}' -y; rm '{video_path}'")
+        return 0
+        

src/dataset/face_align/align.py

@@ -0,0 +1,36 @@
+import os
+import sys
+BASE_DIR = os.path.dirname(os.path.abspath(__file__))
+sys.path.append(BASE_DIR)
+import torch
+from src.dataset.face_align.yoloface import YoloFace
+
+class AlignImage(object):
+    def __init__(self, device='cuda', det_path='checkpoints/yoloface_v5m.pt'):
+        self.facedet = YoloFace(pt_path=det_path, confThreshold=0.5, nmsThreshold=0.45, device=device)
+
+    @torch.no_grad()
+    def __call__(self, im, maxface=False):
+        bboxes, kpss, scores = self.facedet.detect(im)
+        face_num = bboxes.shape[0]
+
+        five_pts_list = []
+        scores_list = []
+        bboxes_list = []
+        for i in range(face_num):
+            five_pts_list.append(kpss[i].reshape(5,2))
+            scores_list.append(scores[i])
+            bboxes_list.append(bboxes[i])
+
+        if maxface and face_num>1:
+            max_idx = 0
+            max_area = (bboxes[0, 2])*(bboxes[0, 3])
+            for i in range(1, face_num):
+                area = (bboxes[i,2])*(bboxes[i,3])
+                if area>max_area:
+                    max_idx = i
+            five_pts_list = [five_pts_list[max_idx]]
+            scores_list = [scores_list[max_idx]]
+            bboxes_list = [bboxes_list[max_idx]]
+
+        return five_pts_list, scores_list, bboxes_list

src/dataset/face_align/yoloface.py

@@ -0,0 +1,310 @@
+# -*- coding: UTF-8 -*-
+import os
+import cv2
+import numpy as np
+import torch
+import torchvision
+
+
+def xyxy2xywh(x):
+    # Convert nx4 boxes from [x1, y1, x2, y2] to [x, y, w, h] where xy1=top-left, xy2=bottom-right
+    y = x.clone() if isinstance(x, torch.Tensor) else np.copy(x)
+    y[:, 0] = (x[:, 0] + x[:, 2]) / 2  # x center
+    y[:, 1] = (x[:, 1] + x[:, 3]) / 2  # y center
+    y[:, 2] = x[:, 2] - x[:, 0]  # width
+    y[:, 3] = x[:, 3] - x[:, 1]  # height
+    return y
+
+
+def xywh2xyxy(x):
+    # Convert nx4 boxes from [x, y, w, h] to [x1, y1, x2, y2] where xy1=top-left, xy2=bottom-right
+    y = x.clone() if isinstance(x, torch.Tensor) else np.copy(x)
+    y[:, 0] = x[:, 0] - x[:, 2] / 2  # top left x
+    y[:, 1] = x[:, 1] - x[:, 3] / 2  # top left y
+    y[:, 2] = x[:, 0] + x[:, 2] / 2  # bottom right x
+    y[:, 3] = x[:, 1] + x[:, 3] / 2  # bottom right y
+    return y
+
+
+def box_iou(box1, box2):
+    # https://github.com/pytorch/vision/blob/master/torchvision/ops/boxes.py
+    """
+    Return intersection-over-union (Jaccard index) of boxes.
+    Both sets of boxes are expected to be in (x1, y1, x2, y2) format.
+    Arguments:
+        box1 (Tensor[N, 4])
+        box2 (Tensor[M, 4])
+    Returns:
+        iou (Tensor[N, M]): the NxM matrix containing the pairwise
+            IoU values for every element in boxes1 and boxes2
+    """
+
+    def box_area(box):
+        # box = 4xn
+        return (box[2] - box[0]) * (box[3] - box[1])
+
+    area1 = box_area(box1.T)
+    area2 = box_area(box2.T)
+
+    # inter(N,M) = (rb(N,M,2) - lt(N,M,2)).clamp(0).prod(2)
+    inter = (torch.min(box1[:, None, 2:], box2[:, 2:]) -
+             torch.max(box1[:, None, :2], box2[:, :2])).clamp(0).prod(2)
+    # iou = inter / (area1 + area2 - inter)
+    return inter / (area1[:, None] + area2 - inter)
+
+
+def scale_coords(img1_shape, coords, img0_shape, ratio_pad=None):
+    # Rescale coords (xyxy) from img1_shape to img0_shape
+    if ratio_pad is None:  # calculate from img0_shape
+        gain = min(img1_shape[0] / img0_shape[0], img1_shape[1] / img0_shape[1])  # gain  = old / new
+        pad = (img1_shape[1] - img0_shape[1] * gain) / 2, (img1_shape[0] - img0_shape[0] * gain) / 2  # wh padding
+    else:
+        gain = ratio_pad[0][0]
+        pad = ratio_pad[1]
+
+    coords[:, [0, 2]] -= pad[0]  # x padding
+    coords[:, [1, 3]] -= pad[1]  # y padding
+    coords[:, :4] /= gain
+    clip_coords(coords, img0_shape)
+    return coords
+
+
+def clip_coords(boxes, img_shape):
+    # Clip bounding xyxy bounding boxes to image shape (height, width)
+    boxes[:, 0].clamp_(0, img_shape[1])  # x1
+    boxes[:, 1].clamp_(0, img_shape[0])  # y1
+    boxes[:, 2].clamp_(0, img_shape[1])  # x2
+    boxes[:, 3].clamp_(0, img_shape[0])  # y2
+
+
+def scale_coords_landmarks(img1_shape, coords, img0_shape, ratio_pad=None):
+    # Rescale coords (xyxy) from img1_shape to img0_shape
+    if ratio_pad is None:  # calculate from img0_shape
+        gain = min(img1_shape[0] / img0_shape[0], img1_shape[1] / img0_shape[1])  # gain  = old / new
+        pad = (img1_shape[1] - img0_shape[1] * gain) / 2, (img1_shape[0] - img0_shape[0] * gain) / 2  # wh padding
+    else:
+        gain = ratio_pad[0][0]
+        pad = ratio_pad[1]
+
+    coords[:, [0, 2, 4, 6, 8]] -= pad[0]  # x padding
+    coords[:, [1, 3, 5, 7, 9]] -= pad[1]  # y padding
+    coords[:, :10] /= gain
+    #clip_coords(coords, img0_shape)
+    coords[:, 0].clamp_(0, img0_shape[1])  # x1
+    coords[:, 1].clamp_(0, img0_shape[0])  # y1
+    coords[:, 2].clamp_(0, img0_shape[1])  # x2
+    coords[:, 3].clamp_(0, img0_shape[0])  # y2
+    coords[:, 4].clamp_(0, img0_shape[1])  # x3
+    coords[:, 5].clamp_(0, img0_shape[0])  # y3
+    coords[:, 6].clamp_(0, img0_shape[1])  # x4
+    coords[:, 7].clamp_(0, img0_shape[0])  # y4
+    coords[:, 8].clamp_(0, img0_shape[1])  # x5
+    coords[:, 9].clamp_(0, img0_shape[0])  # y5
+    return coords
+
+
+def show_results(img, xywh, conf, landmarks, class_num):
+    h,w,c = img.shape
+    tl = 1 or round(0.002 * (h + w) / 2) + 1  # line/font thickness
+    x1 = int(xywh[0] * w - 0.5 * xywh[2] * w)
+    y1 = int(xywh[1] * h - 0.5 * xywh[3] * h)
+    x2 = int(xywh[0] * w + 0.5 * xywh[2] * w)
+    y2 = int(xywh[1] * h + 0.5 * xywh[3] * h)
+    cv2.rectangle(img, (x1,y1), (x2, y2), (0,255,0), thickness=tl, lineType=cv2.LINE_AA)
+
+    clors = [(255,0,0),(0,255,0),(0,0,255),(255,255,0),(0,255,255)]
+
+    for i in range(5):
+        point_x = int(landmarks[2 * i] * w)
+        point_y = int(landmarks[2 * i + 1] * h)
+        cv2.circle(img, (point_x, point_y), tl+1, clors[i], -1)
+
+    tf = max(tl - 1, 1)  # font thickness
+    label = str(conf)[:5]
+    cv2.putText(img, label, (x1, y1 - 2), 0, tl / 3, [225, 255, 255], thickness=tf, lineType=cv2.LINE_AA)
+    return img
+
+
+def make_divisible(x, divisor):
+    # Returns x evenly divisible by divisor
+    return (x // divisor) * divisor
+
+
+def non_max_suppression_face(prediction, conf_thres=0.5, iou_thres=0.45, classes=None, agnostic=False, labels=()):
+    """Performs Non-Maximum Suppression (NMS) on inference results
+    Returns:
+         detections with shape: nx6 (x1, y1, x2, y2, conf, cls)
+    """
+
+    nc = prediction.shape[2] - 15  # number of classes
+    xc = prediction[..., 4] > conf_thres  # candidates
+
+    # Settings
+    min_wh, max_wh = 2, 4096  # (pixels) minimum and maximum box width and height
+    # time_limit = 10.0  # seconds to quit after
+    redundant = True  # require redundant detections
+    multi_label = nc > 1  # multiple labels per box (adds 0.5ms/img)
+    merge = False  # use merge-NMS
+
+    # t = time.time()
+    output = [torch.zeros((0, 16), device=prediction.device)] * prediction.shape[0]
+    for xi, x in enumerate(prediction):  # image index, image inference
+        # Apply constraints
+        # x[((x[..., 2:4] < min_wh) | (x[..., 2:4] > max_wh)).any(1), 4] = 0  # width-height
+        x = x[xc[xi]]  # confidence
+
+        # Cat apriori labels if autolabelling
+        if labels and len(labels[xi]):
+            l = labels[xi]
+            v = torch.zeros((len(l), nc + 15), device=x.device)
+            v[:, :4] = l[:, 1:5]  # box
+            v[:, 4] = 1.0  # conf
+            v[range(len(l)), l[:, 0].long() + 15] = 1.0  # cls
+            x = torch.cat((x, v), 0)
+
+        # If none remain process next image
+        if not x.shape[0]:
+            continue
+
+        # Compute conf
+        x[:, 15:] *= x[:, 4:5]  # conf = obj_conf * cls_conf
+
+        # Box (center x, center y, width, height) to (x1, y1, x2, y2)
+        box = xywh2xyxy(x[:, :4])
+
+        # Detections matrix nx6 (xyxy, conf, landmarks, cls)
+        if multi_label:
+            i, j = (x[:, 15:] > conf_thres).nonzero(as_tuple=False).T
+            x = torch.cat((box[i], x[i, j + 15, None], x[i, 5:15] ,j[:, None].float()), 1)
+        else:  # best class only
+            conf, j = x[:, 15:].max(1, keepdim=True)
+            x = torch.cat((box, conf, x[:, 5:15], j.float()), 1)[conf.view(-1) > conf_thres]
+
+        # Filter by class
+        if classes is not None:
+            x = x[(x[:, 5:6] == torch.tensor(classes, device=x.device)).any(1)]
+
+        # If none remain process next image
+        n = x.shape[0]  # number of boxes
+        if not n:
+            continue
+
+        # Batched NMS
+        c = x[:, 15:16] * (0 if agnostic else max_wh)  # classes
+        boxes, scores = x[:, :4] + c, x[:, 4]  # boxes (offset by class), scores
+        i = torchvision.ops.nms(boxes, scores, iou_thres)  # NMS
+        #if i.shape[0] > max_det:  # limit detections
+        #    i = i[:max_det]
+        if merge and (1 < n < 3E3):  # Merge NMS (boxes merged using weighted mean)
+            # update boxes as boxes(i,4) = weights(i,n) * boxes(n,4)
+            iou = box_iou(boxes[i], boxes) > iou_thres  # iou matrix
+            weights = iou * scores[None]  # box weights
+            x[i, :4] = torch.mm(weights, x[:, :4]).float() / weights.sum(1, keepdim=True)  # merged boxes
+            if redundant:
+                i = i[iou.sum(1) > 1]  # require redundancy
+
+        output[xi] = x[i]
+        # if (time.time() - t) > time_limit:
+        #     break  # time limit exceeded
+
+    return output
+
+
+class YoloFace():
+    def __init__(self, pt_path='checkpoints/yolov5m-face.pt', confThreshold=0.5, nmsThreshold=0.45, device='cuda'):
+        assert os.path.exists(pt_path)
+
+        self.inpSize = 416
+        self.conf_thres = confThreshold
+        self.iou_thres = nmsThreshold
+        self.test_device = torch.device(device if torch.cuda.is_available() else "cpu")
+        self.model = torch.jit.load(pt_path).to(self.test_device)
+        self.last_w = 416
+        self.last_h = 416
+        self.grids = None
+
+    @torch.no_grad()
+    def detect(self, srcimg):
+        # t0=time.time()
+
+        h0, w0 = srcimg.shape[:2]  # orig hw
+        r = self.inpSize / min(h0, w0)  # resize image to img_size
+        h1 = int(h0*r+31)//32*32
+        w1 = int(w0*r+31)//32*32
+
+        img = cv2.resize(srcimg, (w1,h1), interpolation=cv2.INTER_LINEAR)
+
+        # Convert
+        img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB) # BGR to RGB
+
+        # Run inference
+        img = torch.from_numpy(img).to(self.test_device).permute(2,0,1)
+        img = img.float()/255  # uint8 to fp16/32  0-1
+        if img.ndimension() == 3:
+            img = img.unsqueeze(0)
+
+        # Inference
+        if h1 != self.last_h or w1 != self.last_w or self.grids is None:
+            grids = []
+            for scale in [8,16,32]:
+                ny = h1//scale
+                nx = w1//scale
+                yv, xv = torch.meshgrid([torch.arange(ny), torch.arange(nx)])
+                grid = torch.stack((xv, yv), 2).view((1,1,ny, nx, 2)).float()
+                grids.append(grid.to(self.test_device))
+            self.grids = grids
+            self.last_w = w1
+            self.last_h = h1
+
+        pred = self.model(img, self.grids).cpu()
+
+        # Apply NMS
+        det = non_max_suppression_face(pred, self.conf_thres, self.iou_thres)[0]
+        # Process detections
+        # det = pred[0]
+        bboxes = np.zeros((det.shape[0], 4))
+        kpss = np.zeros((det.shape[0], 5, 2))
+        scores = np.zeros((det.shape[0]))
+        # gn = torch.tensor([w0, h0, w0, h0]).to(pred)  # normalization gain whwh
+        # gn_lks = torch.tensor([w0, h0, w0, h0, w0, h0, w0, h0, w0, h0]).to(pred)  # normalization gain landmarks
+        det = det.cpu().numpy()
+
+        for j in range(det.shape[0]):
+            # xywh = (xyxy2xywh(det[j, :4].view(1, 4)) / gn).view(4).cpu().numpy()
+            bboxes[j, 0] = det[j, 0] * w0/w1
+            bboxes[j, 1] = det[j, 1] * h0/h1
+            bboxes[j, 2] = det[j, 2] * w0/w1 - bboxes[j, 0]
+            bboxes[j, 3] = det[j, 3] * h0/h1 - bboxes[j, 1]
+            scores[j] = det[j, 4]
+            # landmarks = (det[j, 5:15].view(1, 10) / gn_lks).view(5,2).cpu().numpy()
+            kpss[j, :, :] = det[j, 5:15].reshape(5, 2) * np.array([[w0/w1,h0/h1]])
+                # class_num = det[j, 15].cpu().numpy()
+                # orgimg = show_results(orgimg, xywh, conf, landmarks, class_num)
+        return bboxes, kpss, scores
+
+
+
+if __name__ == '__main__':
+    import time
+
+    imgpath = 'test.png'
+
+    yoloface = YoloFace(pt_path='../checkpoints/yoloface_v5m.pt')
+    srcimg = cv2.imread(imgpath)
+
+    #warpup
+    bboxes, kpss, scores = yoloface.detect(srcimg)
+    bboxes, kpss, scores = yoloface.detect(srcimg)
+    bboxes, kpss, scores = yoloface.detect(srcimg)
+
+    t1 = time.time()
+    for _ in range(10):
+        bboxes, kpss, scores = yoloface.detect(srcimg)
+    t2 = time.time()
+    print('total time: {} ms'.format((t2 - t1) * 1000))
+    for i in range(bboxes.shape[0]):
+        xmin, ymin, xamx, ymax = int(bboxes[i, 0]), int(bboxes[i, 1]), int(bboxes[i, 0] + bboxes[i, 2]), int(bboxes[i, 1] + bboxes[i, 3])
+        cv2.rectangle(srcimg, (xmin, ymin), (xamx, ymax), (0, 0, 255), thickness=2)
+        for j in range(5):
+            cv2.circle(srcimg, (int(kpss[i, j, 0]), int(kpss[i, j, 1])), 1, (0, 255, 0), thickness=5)
+    cv2.imwrite('test_yoloface.jpg', srcimg)

src/dataset/test_preprocess.py

@@ -0,0 +1,155 @@
+import os
+import numpy as np
+from PIL import Image
+import torch
+import torchvision.transforms as transforms
+from transformers import CLIPImageProcessor
+import librosa
+
+
+def process_bbox(bbox, expand_radio, height, width):
+    """
+    raw_vid_path:
+    bbox: format: x1, y1, x2, y2
+    radio: expand radio against bbox size
+    height,width: source image height and width
+    """
+
+    def expand(bbox, ratio, height, width):
+        
+        bbox_h = bbox[3] - bbox[1]
+        bbox_w = bbox[2] - bbox[0]
+        
+        expand_x1 = max(bbox[0] - ratio * bbox_w, 0)
+        expand_y1 = max(bbox[1] - ratio * bbox_h, 0)
+        expand_x2 = min(bbox[2] + ratio * bbox_w, width)
+        expand_y2 = min(bbox[3] + ratio * bbox_h, height)
+
+        return [expand_x1,expand_y1,expand_x2,expand_y2]
+
+    def to_square(bbox_src, bbox_expend, height, width):
+
+        h = bbox_expend[3] - bbox_expend[1]
+        w = bbox_expend[2] - bbox_expend[0]
+        c_h = (bbox_expend[1] + bbox_expend[3]) / 2
+        c_w = (bbox_expend[0] + bbox_expend[2]) / 2
+
+        c = min(h, w) / 2
+
+        c_src_h = (bbox_src[1] + bbox_src[3]) / 2
+        c_src_w = (bbox_src[0] + bbox_src[2]) / 2
+
+        s_h, s_w = 0, 0
+        if w < h:
+            d = abs((h - w) / 2)
+            s_h = min(d, abs(c_src_h-c_h))
+            s_h = s_h if  c_src_h > c_h else s_h * (-1)
+        else:
+            d = abs((h - w) / 2)
+            s_w = min(d, abs(c_src_w-c_w))
+            s_w = s_w if  c_src_w > c_w else s_w * (-1)
+
+
+        c_h = (bbox_expend[1] + bbox_expend[3]) / 2 + s_h
+        c_w = (bbox_expend[0] + bbox_expend[2]) / 2 + s_w
+
+        square_x1 = c_w - c
+        square_y1 = c_h - c
+        square_x2 = c_w + c
+        square_y2 = c_h + c
+
+        x1, y1, x2, y2 = square_x1, square_y1, square_x2, square_y2
+        ww = x2 - x1
+        hh = y2 - y1
+        cc_x = (x1 + x2)/2
+        cc_y = (y1 + y2)/2
+        # 1:1
+        ww = hh = min(ww, hh)
+        x1, x2 = round(cc_x - ww/2), round(cc_x + ww/2)
+        y1, y2 = round(cc_y - hh/2), round(cc_y + hh/2) 
+
+        return [round(x1), round(y1), round(x2), round(y2)]
+
+
+    bbox_expend = expand(bbox, expand_radio, height=height, width=width)
+    processed_bbox = to_square(bbox, bbox_expend, height=height, width=width)
+
+    return processed_bbox
+
+
+def get_audio_feature(audio_path, feature_extractor):
+    audio_input, sampling_rate = librosa.load(audio_path, sr=16000)
+    assert sampling_rate == 16000
+
+    audio_features = []
+    window = 750*640
+    for i in range(0, len(audio_input), window):
+        audio_feature = feature_extractor(audio_input[i:i+window], 
+                                        sampling_rate=sampling_rate, 
+                                        return_tensors="pt", 
+                                        ).input_features
+        audio_features.append(audio_feature)
+    audio_features = torch.cat(audio_features, dim=-1)
+    return audio_features, len(audio_input) // 640
+
+def image_audio_to_tensor(align_instance, feature_extractor, image_path, audio_path, limit=100, image_size=512, area=1.25):
+    
+    clip_processor = CLIPImageProcessor()
+    
+    to_tensor = transforms.Compose([
+            transforms.ToTensor(),
+            transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
+        ])
+    mask_to_tensor = transforms.Compose([
+            transforms.ToTensor(),
+        ])
+    
+
+    imSrc_ = Image.open(image_path).convert('RGB')
+    w, h = imSrc_.size
+    
+    _, _, bboxes_list = align_instance(np.array(imSrc_)[:,:,[2,1,0]], maxface=True)
+
+    if len(bboxes_list) == 0:
+        return None
+    bboxSrc = bboxes_list[0]
+
+    x1, y1, ww, hh = bboxSrc
+    x2, y2 = x1 + ww, y1 + hh
+
+    mask_img = np.zeros_like(np.array(imSrc_))
+    ww, hh = (x2-x1) * area, (y2-y1) * area
+    center = [(x2+x1)//2, (y2+y1)//2]
+    x1 = max(center[0] - ww//2, 0)
+    y1 = max(center[1] - hh//2, 0)
+    x2 = min(center[0] + ww//2, w)
+    y2 = min(center[1] + hh//2, h)
+    mask_img[int(y1):int(y2), int(x1):int(x2)] = 255
+    mask_img = Image.fromarray(mask_img)
+    
+    w, h = imSrc_.size
+    scale = image_size / min(w, h)
+    new_w = round(w * scale / 64) * 64
+    new_h = round(h * scale / 64) * 64
+    if new_h != h or new_w != w:
+        imSrc = imSrc_.resize((new_w, new_h), Image.LANCZOS)
+        mask_img = mask_img.resize((new_w, new_h), Image.LANCZOS)
+    else:
+        imSrc = imSrc_
+
+    clip_image = clip_processor(
+            images=imSrc.resize((224, 224), Image.LANCZOS), return_tensors="pt"
+        ).pixel_values[0]
+    audio_input, audio_len = get_audio_feature(audio_path, feature_extractor)
+
+    audio_len = min(limit, audio_len)
+
+    sample = dict(
+                face_mask=mask_to_tensor(mask_img),
+                ref_img=to_tensor(imSrc),
+                clip_images=clip_image,
+                audio_feature=audio_input[0],
+                audio_len=audio_len
+            )
+
+    return sample

src/models/audio_adapter/audio_proj.py

@@ -0,0 +1,124 @@
+"""
+This module provides the implementation of an Audio Projection Model, which is designed for
+audio processing tasks. The model takes audio embeddings as input and outputs context tokens
+that can be used for various downstream applications, such as audio analysis or synthesis.
+
+The AudioProjModel class is based on the ModelMixin class from the diffusers library, which
+provides a foundation for building custom models. This implementation includes multiple linear
+layers with ReLU activation functions and a LayerNorm for normalization.
+
+Key Features:
+- Audio embedding input with flexible sequence length and block structure.
+- Multiple linear layers for feature transformation.
+- ReLU activation for non-linear transformation.
+- LayerNorm for stabilizing and speeding up training.
+- Rearrangement of input embeddings to match the model's expected input shape.
+- Customizable number of blocks, channels, and context tokens for adaptability.
+
+The module is structured to be easily integrated into larger systems or used as a standalone
+component for audio feature extraction and processing.
+
+Classes:
+- AudioProjModel: A class representing the audio projection model with configurable parameters.
+
+Functions:
+- (none)
+
+Dependencies:
+- torch: For tensor operations and neural network components.
+- diffusers: For the ModelMixin base class.
+- einops: For tensor rearrangement operations.
+
+"""
+
+import torch
+from diffusers import ModelMixin
+from einops import rearrange
+from torch import nn
+
+
+class AudioProjModel(ModelMixin):
+    """Audio Projection Model
+
+    This class defines an audio projection model that takes audio embeddings as input
+    and produces context tokens as output. The model is based on the ModelMixin class
+    and consists of multiple linear layers and activation functions. It can be used
+    for various audio processing tasks.
+
+    Attributes:
+        seq_len (int): The length of the audio sequence.
+        blocks (int): The number of blocks in the audio projection model.
+        channels (int): The number of channels in the audio projection model.
+        intermediate_dim (int): The intermediate dimension of the model.
+        context_tokens (int): The number of context tokens in the output.
+        output_dim (int): The output dimension of the context tokens.
+
+    Methods:
+        __init__(self, seq_len=5, blocks=12, channels=768, intermediate_dim=512, context_tokens=32, output_dim=768):
+            Initializes the AudioProjModel with the given parameters.
+        forward(self, audio_embeds):
+            Defines the forward pass for the AudioProjModel.
+            Parameters:
+            audio_embeds (torch.Tensor): The input audio embeddings with shape (batch_size, video_length, blocks, channels).
+            Returns:
+            context_tokens (torch.Tensor): The output context tokens with shape (batch_size, video_length, context_tokens, output_dim).
+
+    """
+
+    def __init__(
+        self,
+        seq_len=5,
+        blocks=12,  # add a new parameter blocks
+        channels=768,  # add a new parameter channels
+        intermediate_dim=512,
+        output_dim=768,
+        context_tokens=32,
+    ):
+        super().__init__()
+
+        self.seq_len = seq_len
+        self.blocks = blocks
+        self.channels = channels
+        self.input_dim = (
+            seq_len * blocks * channels
+        )  # update input_dim to be the product of blocks and channels.
+        self.intermediate_dim = intermediate_dim
+        self.context_tokens = context_tokens
+        self.output_dim = output_dim
+
+        # define multiple linear layers
+        self.proj1 = nn.Linear(self.input_dim, intermediate_dim)
+        self.proj2 = nn.Linear(intermediate_dim, intermediate_dim)
+        self.proj3 = nn.Linear(intermediate_dim, context_tokens * output_dim)
+
+        self.norm = nn.LayerNorm(output_dim)
+
+    def forward(self, audio_embeds):
+        """
+        Defines the forward pass for the AudioProjModel.
+
+        Parameters:
+            audio_embeds (torch.Tensor): The input audio embeddings with shape (batch_size, video_length, blocks, channels).
+
+        Returns:
+            context_tokens (torch.Tensor): The output context tokens with shape (batch_size, video_length, context_tokens, output_dim).
+        """
+        # merge
+        video_length = audio_embeds.shape[1]
+        audio_embeds = rearrange(audio_embeds, "bz f w b c -> (bz f) w b c")
+        batch_size, window_size, blocks, channels = audio_embeds.shape
+        audio_embeds = audio_embeds.view(batch_size, window_size * blocks * channels)
+
+        audio_embeds = torch.relu(self.proj1(audio_embeds))
+        audio_embeds = torch.relu(self.proj2(audio_embeds))
+
+        context_tokens = self.proj3(audio_embeds).reshape(
+            batch_size, self.context_tokens, self.output_dim
+        )
+
+        context_tokens = self.norm(context_tokens)
+        context_tokens = rearrange(
+            context_tokens, "(bz f) m c -> bz f m c", f=video_length
+        )
+
+        return context_tokens

src/models/audio_adapter/audio_to_bucket.py

@@ -0,0 +1,127 @@
+"""
+This module provides the implementation of an Audio Projection Model, which is designed for
+audio processing tasks. The model takes audio embeddings as input and outputs context tokens
+that can be used for various downstream applications, such as audio analysis or synthesis.
+
+The AudioProjModel class is based on the ModelMixin class from the diffusers library, which
+provides a foundation for building custom models. This implementation includes multiple linear
+layers with ReLU activation functions and a LayerNorm for normalization.
+
+Key Features:
+- Audio embedding input with flexible sequence length and block structure.
+- Multiple linear layers for feature transformation.
+- ReLU activation for non-linear transformation.
+- LayerNorm for stabilizing and speeding up training.
+- Rearrangement of input embeddings to match the model's expected input shape.
+- Customizable number of blocks, channels, and context tokens for adaptability.
+
+The module is structured to be easily integrated into larger systems or used as a standalone
+component for audio feature extraction and processing.
+
+Classes:
+- AudioProjModel: A class representing the audio projection model with configurable parameters.
+
+Functions:
+- (none)
+
+Dependencies:
+- torch: For tensor operations and neural network components.
+- diffusers: For the ModelMixin base class.
+- einops: For tensor rearrangement operations.
+
+"""
+
+import torch
+from diffusers import ModelMixin
+from einops import rearrange
+from torch import nn
+
+
+class Audio2bucketModel(ModelMixin):
+    """Audio Projection Model
+
+    This class defines an audio projection model that takes audio embeddings as input
+    and produces context tokens as output. The model is based on the ModelMixin class
+    and consists of multiple linear layers and activation functions. It can be used
+    for various audio processing tasks.
+
+    Attributes:
+        seq_len (int): The length of the audio sequence.
+        blocks (int): The number of blocks in the audio projection model.
+        channels (int): The number of channels in the audio projection model.
+        intermediate_dim (int): The intermediate dimension of the model.
+        context_tokens (int): The number of context tokens in the output.
+        output_dim (int): The output dimension of the context tokens.
+
+    Methods:
+        __init__(self, seq_len=5, blocks=12, channels=768, intermediate_dim=512, context_tokens=32, output_dim=768):
+            Initializes the AudioProjModel with the given parameters.
+        forward(self, audio_embeds):
+            Defines the forward pass for the AudioProjModel.
+            Parameters:
+            audio_embeds (torch.Tensor): The input audio embeddings with shape (batch_size, video_length, blocks, channels).
+            Returns:
+            context_tokens (torch.Tensor): The output context tokens with shape (batch_size, video_length, context_tokens, output_dim).
+
+    """
+
+    def __init__(
+        self,
+        seq_len=5,
+        blocks=12,  # add a new parameter blocks
+        channels=768,  # add a new parameter channels
+        clip_channels=768,  # add a new parameter channels
+        intermediate_dim=512,
+        output_dim=768,
+        context_tokens=32,
+    ):
+        super().__init__()
+
+        self.seq_len = seq_len
+        self.blocks = blocks
+        self.channels = channels
+        self.input_dim = (
+            seq_len * blocks * channels + clip_channels
+        )  # update input_dim to be the product of blocks and channels.
+        self.intermediate_dim = intermediate_dim
+        self.context_tokens = context_tokens
+        self.output_dim = output_dim
+
+        # define multiple linear layers
+        self.proj1 = nn.Linear(self.input_dim, intermediate_dim)
+        self.proj2 = nn.Linear(intermediate_dim, intermediate_dim)
+        self.proj3 = nn.Linear(intermediate_dim, context_tokens * output_dim)
+        self.act = nn.SiLU()
+
+        # self.norm = nn.LayerNorm(output_dim)
+
+    def forward(self, audio_embeds, clip_embeds):
+        """
+        Defines the forward pass for the AudioProjModel.
+
+        Parameters:
+            audio_embeds (torch.Tensor): The input audio embeddings with shape (batch_size, video_length, blocks, channels).
+
+        Returns:
+            context_tokens (torch.Tensor): The output context tokens with shape (batch_size, video_length, context_tokens, output_dim).
+        """
+        # merge
+        video_length = audio_embeds.shape[1]
+        audio_embeds = rearrange(audio_embeds, "bz f w b c -> (bz f) w b c")
+        batch_size, window_size, blocks, channels = audio_embeds.shape
+        audio_embeds = audio_embeds.view(batch_size, window_size * blocks * channels)
+        audio_embeds = torch.cat([audio_embeds, clip_embeds], dim=-1)
+
+        audio_embeds = self.act(self.proj1(audio_embeds))
+        audio_embeds = self.act(self.proj2(audio_embeds))
+
+        context_tokens = self.proj3(audio_embeds).reshape(
+            batch_size, self.context_tokens, self.output_dim
+        )
+
+        # context_tokens = self.norm(context_tokens)
+        context_tokens = rearrange(
+            context_tokens, "(bz f) m c -> bz f m c", f=video_length
+        )
+
+        return context_tokens

src/models/base/unet_spatio_temporal_condition.py

@@ -0,0 +1,600 @@
+from dataclasses import dataclass
+from typing import Dict, Optional, Tuple, Union, Any
+
+import torch
+import torch.nn as nn
+
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.loaders import UNet2DConditionLoadersMixin
+from diffusers.utils import BaseOutput, logging
+# from diffusers.models.attention_processor import CROSS_ATTENTION_PROCESSORS, AttentionProcessor, AttnProcessor
+
+from diffusers.models.embeddings import TimestepEmbedding, Timesteps
+from diffusers.models.modeling_utils import ModelMixin
+from .unet_3d_blocks import UNetMidBlockSpatioTemporal, get_down_block, get_up_block
+from .attention_processor import CROSS_ATTENTION_PROCESSORS, AttentionProcessor, AttnProcessor, AttnProcessor2_0, IPAdapterAttnProcessor, IPAdapterAttnProcessor2_0
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+@dataclass
+class UNetSpatioTemporalConditionOutput(BaseOutput):
+    """
+    The output of [`UNetSpatioTemporalConditionModel`].
+
+    Args:
+        sample (`torch.Tensor` of shape `(batch_size, num_frames, num_channels, height, width)`):
+            The hidden states output conditioned on `encoder_hidden_states` input. Output of last layer of model.
+    """
+
+    sample: torch.Tensor = None
+
+
+class UNetSpatioTemporalConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin):
+    r"""
+    A conditional Spatio-Temporal UNet model that takes a noisy video frames, conditional state, and a timestep and
+    returns a sample shaped output.
+
+    This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented
+    for all models (such as downloading or saving).
+
+    Parameters:
+        sample_size (`int` or `Tuple[int, int]`, *optional*, defaults to `None`):
+            Height and width of input/output sample.
+        in_channels (`int`, *optional*, defaults to 8): Number of channels in the input sample.
+        out_channels (`int`, *optional*, defaults to 4): Number of channels in the output.
+        down_block_types (`Tuple[str]`, *optional*, defaults to `("CrossAttnDownBlockSpatioTemporal", "CrossAttnDownBlockSpatioTemporal", "CrossAttnDownBlockSpatioTemporal", "DownBlockSpatioTemporal")`):
+            The tuple of downsample blocks to use.
+        up_block_types (`Tuple[str]`, *optional*, defaults to `("UpBlockSpatioTemporal", "CrossAttnUpBlockSpatioTemporal", "CrossAttnUpBlockSpatioTemporal", "CrossAttnUpBlockSpatioTemporal")`):
+            The tuple of upsample blocks to use.
+        block_out_channels (`Tuple[int]`, *optional*, defaults to `(320, 640, 1280, 1280)`):
+            The tuple of output channels for each block.
+        addition_time_embed_dim: (`int`, defaults to 256):
+            Dimension to to encode the additional time ids.
+        projection_class_embeddings_input_dim (`int`, defaults to 768):
+            The dimension of the projection of encoded `added_time_ids`.
+        layers_per_block (`int`, *optional*, defaults to 2): The number of layers per block.
+        cross_attention_dim (`int` or `Tuple[int]`, *optional*, defaults to 1280):
+            The dimension of the cross attention features.
+        transformer_layers_per_block (`int`, `Tuple[int]`, or `Tuple[Tuple]` , *optional*, defaults to 1):
+            The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`]. Only relevant for
+            [`~models.unets.unet_3d_blocks.CrossAttnDownBlockSpatioTemporal`],
+            [`~models.unets.unet_3d_blocks.CrossAttnUpBlockSpatioTemporal`],
+            [`~models.unets.unet_3d_blocks.UNetMidBlockSpatioTemporal`].
+        num_attention_heads (`int`, `Tuple[int]`, defaults to `(5, 10, 10, 20)`):
+            The number of attention heads.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+    """
+
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        sample_size: Optional[int] = None,
+        in_channels: int = 8,
+        out_channels: int = 4,
+        down_block_types: Tuple[str] = (
+            "CrossAttnDownBlockSpatioTemporal",
+            "CrossAttnDownBlockSpatioTemporal",
+            "CrossAttnDownBlockSpatioTemporal",
+            "DownBlockSpatioTemporal",
+        ),
+        up_block_types: Tuple[str] = (
+            "UpBlockSpatioTemporal",
+            "CrossAttnUpBlockSpatioTemporal",
+            "CrossAttnUpBlockSpatioTemporal",
+            "CrossAttnUpBlockSpatioTemporal",
+        ),
+        block_out_channels: Tuple[int] = (320, 640, 1280, 1280),
+        addition_time_embed_dim: int = 256,
+        projection_class_embeddings_input_dim: int = 768,
+        layers_per_block: Union[int, Tuple[int]] = 2,
+        cross_attention_dim: Union[int, Tuple[int]] = 1024,
+        transformer_layers_per_block: Union[int, Tuple[int], Tuple[Tuple]] = 1,
+        num_attention_heads: Union[int, Tuple[int]] = (5, 10, 20, 20),
+        num_frames: int = 25,
+    ):
+        super().__init__()
+
+        self.sample_size = sample_size
+
+        # Check inputs
+        if len(down_block_types) != len(up_block_types):
+            raise ValueError(
+                f"Must provide the same number of `down_block_types` as `up_block_types`. `down_block_types`: {down_block_types}. `up_block_types`: {up_block_types}."
+            )
+
+        if len(block_out_channels) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `block_out_channels` as `down_block_types`. `block_out_channels`: {block_out_channels}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(num_attention_heads, int) and len(num_attention_heads) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `num_attention_heads` as `down_block_types`. `num_attention_heads`: {num_attention_heads}. `down_block_types`: {down_block_types}."
+            )
+
+        if isinstance(cross_attention_dim, list) and len(cross_attention_dim) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `cross_attention_dim` as `down_block_types`. `cross_attention_dim`: {cross_attention_dim}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(layers_per_block, int) and len(layers_per_block) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `layers_per_block` as `down_block_types`. `layers_per_block`: {layers_per_block}. `down_block_types`: {down_block_types}."
+            )
+
+        # input
+        self.conv_in = nn.Conv2d(
+            in_channels,
+            block_out_channels[0],
+            kernel_size=3,
+            padding=1,
+        )
+
+        # time
+        time_embed_dim = block_out_channels[0] * 4
+
+        self.time_proj = Timesteps(block_out_channels[0], True, downscale_freq_shift=0)
+        timestep_input_dim = block_out_channels[0]
+
+        self.time_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+
+        self.add_time_proj = Timesteps(addition_time_embed_dim, True, downscale_freq_shift=0)
+        self.add_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim)
+
+        self.down_blocks = nn.ModuleList([])
+        self.up_blocks = nn.ModuleList([])
+
+        if isinstance(num_attention_heads, int):
+            num_attention_heads = (num_attention_heads,) * len(down_block_types)
+
+        if isinstance(cross_attention_dim, int):
+            cross_attention_dim = (cross_attention_dim,) * len(down_block_types)
+
+        if isinstance(layers_per_block, int):
+            layers_per_block = [layers_per_block] * len(down_block_types)
+
+        if isinstance(transformer_layers_per_block, int):
+            transformer_layers_per_block = [transformer_layers_per_block] * len(down_block_types)
+
+        blocks_time_embed_dim = time_embed_dim
+
+        # down
+        output_channel = block_out_channels[0]
+        for i, down_block_type in enumerate(down_block_types):
+            input_channel = output_channel
+            output_channel = block_out_channels[i]
+            is_final_block = i == len(block_out_channels) - 1
+
+            down_block = get_down_block(
+                down_block_type,
+                num_layers=layers_per_block[i],
+                transformer_layers_per_block=transformer_layers_per_block[i],
+                in_channels=input_channel,
+                out_channels=output_channel,
+                temb_channels=blocks_time_embed_dim,
+                add_downsample=not is_final_block,
+                resnet_eps=1e-5,
+                cross_attention_dim=cross_attention_dim[i],
+                num_attention_heads=num_attention_heads[i],
+                resnet_act_fn="silu",
+            )
+            self.down_blocks.append(down_block)
+
+        # mid
+        self.mid_block = UNetMidBlockSpatioTemporal(
+            block_out_channels[-1],
+            temb_channels=blocks_time_embed_dim,
+            transformer_layers_per_block=transformer_layers_per_block[-1],
+            cross_attention_dim=cross_attention_dim[-1],
+            num_attention_heads=num_attention_heads[-1],
+        )
+
+        # count how many layers upsample the images
+        self.num_upsamplers = 0
+
+        # up
+        reversed_block_out_channels = list(reversed(block_out_channels))
+        reversed_num_attention_heads = list(reversed(num_attention_heads))
+        reversed_layers_per_block = list(reversed(layers_per_block))
+        reversed_cross_attention_dim = list(reversed(cross_attention_dim))
+        reversed_transformer_layers_per_block = list(reversed(transformer_layers_per_block))
+
+        output_channel = reversed_block_out_channels[0]
+        for i, up_block_type in enumerate(up_block_types):
+            is_final_block = i == len(block_out_channels) - 1
+
+            prev_output_channel = output_channel
+            output_channel = reversed_block_out_channels[i]
+            input_channel = reversed_block_out_channels[min(i + 1, len(block_out_channels) - 1)]
+
+            # add upsample block for all BUT final layer
+            if not is_final_block:
+                add_upsample = True
+                self.num_upsamplers += 1
+            else:
+                add_upsample = False
+
+            up_block = get_up_block(
+                up_block_type,
+                num_layers=reversed_layers_per_block[i] + 1,
+                transformer_layers_per_block=reversed_transformer_layers_per_block[i],
+                in_channels=input_channel,
+                out_channels=output_channel,
+                prev_output_channel=prev_output_channel,
+                temb_channels=blocks_time_embed_dim,
+                add_upsample=add_upsample,
+                resnet_eps=1e-5,
+                resolution_idx=i,
+                cross_attention_dim=reversed_cross_attention_dim[i],
+                num_attention_heads=reversed_num_attention_heads[i],
+                resnet_act_fn="silu",
+            )
+            self.up_blocks.append(up_block)
+            prev_output_channel = output_channel
+
+        # out
+        self.conv_norm_out = nn.GroupNorm(num_channels=block_out_channels[0], num_groups=32, eps=1e-5)
+        self.conv_act = nn.SiLU()
+
+        self.conv_out = nn.Conv2d(
+            block_out_channels[0],
+            out_channels,
+            kernel_size=3,
+            padding=1,
+        )
+
+    @property
+    def attn_processors(self) -> Dict[str, AttentionProcessor]:
+        r"""
+        Returns:
+            `dict` of attention processors: A dictionary containing all attention processors used in the model with
+            indexed by its weight name.
+        """
+        # set recursively
+        processors = {}
+
+        def fn_recursive_add_processors(
+            name: str,
+            module: torch.nn.Module,
+            processors: Dict[str, AttentionProcessor],
+        ):
+            if hasattr(module, "get_processor"):
+                processors[f"{name}.processor"] = module.get_processor()
+
+            for sub_name, child in module.named_children():
+                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
+
+            return processors
+
+        for name, module in self.named_children():
+            fn_recursive_add_processors(name, module, processors)
+
+        return processors
+
+    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
+        r"""
+        Sets the attention processor to use to compute attention.
+
+        Parameters:
+            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
+                The instantiated processor class or a dictionary of processor classes that will be set as the processor
+                for **all** `Attention` layers.
+
+                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
+                processor. This is strongly recommended when setting trainable attention processors.
+
+        """
+        count = len(self.attn_processors.keys())
+
+        if isinstance(processor, dict) and len(processor) != count:
+            raise ValueError(
+                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
+                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
+            )
+
+        def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
+            if hasattr(module, "set_processor"):
+                if not isinstance(processor, dict):
+                    module.set_processor(processor)
+                else:
+                    module.set_processor(processor.pop(f"{name}.processor"))
+
+            for sub_name, child in module.named_children():
+                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
+
+        for name, module in self.named_children():
+            fn_recursive_attn_processor(name, module, processor)
+
+    def set_default_attn_processor(self):
+        """
+        Disables custom attention processors and sets the default attention implementation.
+        """
+        if all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+            processor = AttnProcessor()
+        else:
+            raise ValueError(
+                f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
+            )
+
+        self.set_attn_processor(processor)
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if hasattr(module, "gradient_checkpointing"):
+            module.gradient_checkpointing = value
+
+    # Copied from diffusers.models.unets.unet_3d_condition.UNet3DConditionModel.enable_forward_chunking
+    def enable_forward_chunking(self, chunk_size: Optional[int] = None, dim: int = 0) -> None:
+        """
+        Sets the attention processor to use [feed forward
+        chunking](https://huggingface.co/blog/reformer#2-chunked-feed-forward-layers).
+
+        Parameters:
+            chunk_size (`int`, *optional*):
+                The chunk size of the feed-forward layers. If not specified, will run feed-forward layer individually
+                over each tensor of dim=`dim`.
+            dim (`int`, *optional*, defaults to `0`):
+                The dimension over which the feed-forward computation should be chunked. Choose between dim=0 (batch)
+                or dim=1 (sequence length).
+        """
+        if dim not in [0, 1]:
+            raise ValueError(f"Make sure to set `dim` to either 0 or 1, not {dim}")
+
+        # By default chunk size is 1
+        chunk_size = chunk_size or 1
+
+        def fn_recursive_feed_forward(module: torch.nn.Module, chunk_size: int, dim: int):
+            if hasattr(module, "set_chunk_feed_forward"):
+                module.set_chunk_feed_forward(chunk_size=chunk_size, dim=dim)
+
+            for child in module.children():
+                fn_recursive_feed_forward(child, chunk_size, dim)
+
+        for module in self.children():
+            fn_recursive_feed_forward(module, chunk_size, dim)
+
+    def forward(
+        self,
+        sample: torch.Tensor,
+        timestep: Union[torch.Tensor, float, int],
+        encoder_hidden_states: torch.Tensor,
+        added_time_ids: torch.Tensor,
+        spatial_condition: Optional[torch.Tensor] = None,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        return_dict: bool = True,
+    ) -> Union[UNetSpatioTemporalConditionOutput, Tuple]:
+        r"""
+        The [`UNetSpatioTemporalConditionModel`] forward method.
+
+        Args:
+            sample (`torch.Tensor`):
+                The noisy input tensor with the following shape `(batch, num_frames, channel, height, width)`.
+            timestep (`torch.Tensor` or `float` or `int`): The number of timesteps to denoise an input.
+            encoder_hidden_states (`torch.Tensor`):
+                The encoder hidden states with shape `(batch*num_frames, sequence_length, cross_attention_dim)`.
+            added_time_ids: (`torch.Tensor`):
+                The additional time ids with shape `(batch, num_additional_ids)`. These are encoded with sinusoidal
+                embeddings and added to the time embeddings.
+            spatial_condition (`torch.Tensor`, *optional*, defaults to `None`):
+                The spatial_condition embedding with shape `(batch, num_frames, channel_in(320), height, width)`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~models.unet_slatio_temporal.UNetSpatioTemporalConditionOutput`] instead
+                of a plain tuple.
+        Returns:
+            [`~models.unet_slatio_temporal.UNetSpatioTemporalConditionOutput`] or `tuple`:
+                If `return_dict` is True, an [`~models.unet_slatio_temporal.UNetSpatioTemporalConditionOutput`] is
+                returned, otherwise a `tuple` is returned where the first element is the sample tensor.
+        """
+        # 1. time
+        timesteps = timestep
+        if not torch.is_tensor(timesteps):
+            # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
+            # This would be a good case for the `match` statement (Python 3.10+)
+            is_mps = sample.device.type == "mps"
+            if isinstance(timestep, float):
+                dtype = torch.float32 if is_mps else torch.float64
+            else:
+                dtype = torch.int32 if is_mps else torch.int64
+            timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
+        elif len(timesteps.shape) == 0:
+            timesteps = timesteps[None].to(sample.device)
+
+        # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+        batch_size, num_frames = sample.shape[:2]
+        timesteps = timesteps.expand(batch_size)
+
+        t_emb = self.time_proj(timesteps)
+
+        # `Timesteps` does not contain any weights and will always return f32 tensors
+        # but time_embedding might actually be running in fp16. so we need to cast here.
+        # there might be better ways to encapsulate this.
+        t_emb = t_emb.to(dtype=sample.dtype)
+
+        emb = self.time_embedding(t_emb)
+
+        time_embeds = self.add_time_proj(added_time_ids.flatten())
+        # import ipdb
+        # ipdb.set_trace()
+        time_embeds = time_embeds.reshape((batch_size, -1))
+        time_embeds = time_embeds.to(emb.dtype)
+        aug_emb = self.add_embedding(time_embeds)
+        emb = emb + aug_emb
+
+        # Flatten the batch and frames dimensions
+        # sample: [batch, frames, channels, height, width] -> [batch * frames, channels, height, width]
+        sample = sample.flatten(0, 1)
+        # Repeat the embeddings num_video_frames times
+        # emb: [batch, channels] -> [batch * frames, channels]
+        emb = emb.repeat_interleave(num_frames, dim=0)
+        # encoder_hidden_states: [batch, 1, channels] -> [batch * frames, 1, channels]
+        
+        ### 20240731 process encoder_hidden_states ###
+        if isinstance(encoder_hidden_states, tuple):
+            # ip_hidden_states is a list
+            encoder_hidden_states, ip_hidden_states = encoder_hidden_states
+            if encoder_hidden_states.shape[0]==batch_size:
+                encoder_hidden_states = encoder_hidden_states.repeat_interleave(num_frames, dim=0)
+            encoder_hidden_states = (encoder_hidden_states, ip_hidden_states)
+        elif encoder_hidden_states.shape[0]==batch_size:
+            ### if framewised feature is not provided, repeat_interleave
+            encoder_hidden_states = encoder_hidden_states.repeat_interleave(num_frames, dim=0)
+            
+
+        # 2. pre-process
+        sample = self.conv_in(sample)
+        
+        ### 20240731 add spatial_condition here ###
+        if spatial_condition is not None:
+            sample = sample + spatial_condition.flatten(0,1)
+
+        image_only_indicator = torch.zeros(batch_size, num_frames, dtype=sample.dtype, device=sample.device)
+
+        down_block_res_samples = (sample,)
+        for downsample_block in self.down_blocks:
+            if hasattr(downsample_block, "has_cross_attention") and downsample_block.has_cross_attention:
+                sample, res_samples = downsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    encoder_hidden_states=encoder_hidden_states,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    image_only_indicator=image_only_indicator,
+                )
+            else:
+                sample, res_samples = downsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    image_only_indicator=image_only_indicator,
+                )
+
+            down_block_res_samples += res_samples
+
+        # 4. mid
+        sample = self.mid_block(
+            hidden_states=sample,
+            temb=emb,
+            encoder_hidden_states=encoder_hidden_states,
+            cross_attention_kwargs=cross_attention_kwargs,
+            image_only_indicator=image_only_indicator,
+        )
+
+        # 5. up
+        for i, upsample_block in enumerate(self.up_blocks):
+            res_samples = down_block_res_samples[-len(upsample_block.resnets) :]
+            down_block_res_samples = down_block_res_samples[: -len(upsample_block.resnets)]
+
+            if hasattr(upsample_block, "has_cross_attention") and upsample_block.has_cross_attention:
+                sample = upsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    res_hidden_states_tuple=res_samples,
+                    encoder_hidden_states=encoder_hidden_states,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    image_only_indicator=image_only_indicator,
+                )
+            else:
+                sample = upsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    res_hidden_states_tuple=res_samples,
+                    image_only_indicator=image_only_indicator,
+                )
+
+        # 6. post-process
+        sample = self.conv_norm_out(sample)
+        sample = self.conv_act(sample)
+        sample = self.conv_out(sample)
+
+        # 7. Reshape back to original shape
+        sample = sample.reshape(batch_size, num_frames, *sample.shape[1:])
+
+        if not return_dict:
+            return (sample,)
+
+        return UNetSpatioTemporalConditionOutput(sample=sample)
+
+
+
+def add_ip_adapters(unet, num_adapter_embeds=[32,], scale=[1.0,]):
+    
+    assert len(num_adapter_embeds)==len(scale)
+    
+    
+    # init adapter modules
+    attn_procs = {}
+    unet_sd = unet.state_dict()
+    for name in unet.attn_processors.keys():
+        cross_attention_dim = None if name.endswith("attn1.processor") else unet.config.cross_attention_dim
+        if name.startswith("mid_block"):
+            hidden_size = unet.config.block_out_channels[-1]
+        elif name.startswith("up_blocks"):
+            block_id = int(name[len("up_blocks.")])
+            hidden_size = list(reversed(unet.config.block_out_channels))[block_id]
+        elif name.startswith("down_blocks"):
+            block_id = int(name[len("down_blocks.")])
+            hidden_size = unet.config.block_out_channels[block_id]
+        # if cross_attention_dim is None or "temporal_transformer_blocks" in name:
+        if cross_attention_dim is None:
+            attn_processor_class = (
+                    AttnProcessor2_0 if hasattr(torch.nn.functional, "scaled_dot_product_attention") else AttnProcessor
+                )
+            attn_procs[name] = attn_processor_class()
+        else:
+            attn_processor_class = (
+                    IPAdapterAttnProcessor2_0 if hasattr(torch.nn.functional, "scaled_dot_product_attention") else IPAdapterAttnProcessor
+                )
+            
+            attn_procs[name] = attn_processor_class(
+                        hidden_size=hidden_size,
+                        cross_attention_dim=cross_attention_dim,
+                        num_tokens=num_adapter_embeds,
+                        scale=scale
+                    ).to(device=unet.device, dtype=unet.dtype)
+
+            layer_name = name.split(".processor")[0]
+            weights = {}
+         
+            for i in range(len(num_adapter_embeds)):
+                weights.update({f"to_k_ip.{i}.weight": unet_sd[layer_name + ".to_k.weight"]})
+                weights.update({f"to_v_ip.{i}.weight": unet_sd[layer_name + ".to_v.weight"]})
+                    
+    
+            attn_procs[name].load_state_dict(weights)
+
+    unet.set_attn_processor(attn_procs)
+
+    adapter_modules = torch.nn.ModuleList([m for m in unet.attn_processors.values() if isinstance(m, IPAdapterAttnProcessor) or isinstance(m, IPAdapterAttnProcessor2_0)])
+    return adapter_modules
+
+
+def load_adapter_states(adapter_modules, state_dict_list):
+    assert len(state_dict_list)>0
+    
+    merged_stete_dict = {}
+    for state_dict in state_dict_list:
+        for k, v in state_dict.items():
+            if k in merged_stete_dict.keys():
+                k_split = k.split('.')
+                adapter_idx = int(k_split[2])
+                adapter_idx += 1
+                k_split[2] = str(adapter_idx)
+                new_k = '.'.join(k_split)
+                while(new_k in merged_stete_dict.keys()):
+                    adapter_idx += 1
+                    k_split[2] = str(adapter_idx)
+                    new_k = '.'.join(k_split)
+                merged_stete_dict[new_k] = v    
+            else:
+                merged_stete_dict[k] = v
+                
+    info = adapter_modules.load_state_dict(merged_stete_dict, strict=True)
+    return info
+            
+        
+        
+            
+        
+    
+        
+    

src/pipelines/pipeline_sonic.py

@@ -0,0 +1,632 @@
+import inspect
+from dataclasses import dataclass
+from typing import Callable, Dict, List, Optional, Union
+
+import numpy as np
+import PIL.Image
+import torch
+from transformers import CLIPVisionModelWithProjection
+
+from diffusers.image_processor import VaeImageProcessor
+from diffusers.utils import BaseOutput, logging
+from diffusers.utils.torch_utils import randn_tensor, is_compiled_module
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline
+from diffusers import (
+    AutoencoderKLTemporalDecoder,
+    EulerDiscreteScheduler,
+)
+
+from src.models.base.unet_spatio_temporal_condition import UNetSpatioTemporalConditionModel
+
+logger = logging.get_logger(__name__)
+
+
+@dataclass
+class Pose2VideoSVDPipelineOutput(BaseOutput):
+    r"""
+    Output class for zero-shot text-to-video pipeline.
+
+    Args:
+        frames (`[List[PIL.Image.Image]`, `np.ndarray`]):
+            List of denoised PIL images of length `batch_size` or NumPy array of shape `(batch_size, height, width,
+            num_channels)`.
+    """
+
+    frames: Union[List[PIL.Image.Image], np.ndarray]
+
+
+class SonicPipeline(DiffusionPipeline):
+    r"""
+    Pipeline to generate video from an input image using Stable Video Diffusion.
+
+    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+    implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+    Args:
+        vae ([`AutoencoderKL`]):
+            Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
+        image_encoder ([`~transformers.CLIPVisionModelWithProjection`]):
+            Frozen CLIP image-encoder ([laion/CLIP-ViT-H-14-laion2B-s32B-b79K](https://huggingface.co/laion/CLIP-ViT-H-14-laion2B-s32B-b79K)).
+        unet ([`UNetSpatioTemporalConditionModel`]):
+            A `UNetSpatioTemporalConditionModel` to denoise the encoded image latents.
+        scheduler ([`EulerDiscreteScheduler`]):
+            A scheduler to be used in combination with `unet` to denoise the encoded image latents.
+        feature_extractor ([`~transformers.CLIPImageProcessor`]):
+            A `CLIPImageProcessor` to extract features from generated images.
+    """
+
+    model_cpu_offload_seq = "image_encoder->unet->vae"
+    _callback_tensor_inputs = ["latents"]
+
+    def __init__(
+        self,
+        vae: AutoencoderKLTemporalDecoder,
+        image_encoder: CLIPVisionModelWithProjection,
+        unet: UNetSpatioTemporalConditionModel,
+        scheduler: EulerDiscreteScheduler,
+    ):
+        super().__init__()
+        self.register_modules(
+            vae=vae,
+            image_encoder=image_encoder,
+            unet=unet,
+            scheduler=scheduler,
+        )
+        
+        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+
+        self.image_processor = VaeImageProcessor(
+            vae_scale_factor=self.vae_scale_factor,
+            do_convert_rgb=True)
+        
+        self.pose_image_processor = VaeImageProcessor(
+            vae_scale_factor=self.vae_scale_factor,
+            do_convert_rgb=True,
+            do_normalize=False,
+        )
+
+
+    def _clip_encode_image(self, image, audio_prompts, uncond_audio_prompts, num_frames, device, num_videos_per_prompt, do_classifier_free_guidance, frames_per_batch):
+        dtype = next(self.image_encoder.parameters()).dtype
+
+        image = image.to(device=device, dtype=dtype)
+        image_embeddings = self.image_encoder(image).image_embeds
+        image_embeddings = image_embeddings.unsqueeze(1)
+
+        # duplicate image embeddings for each generation per prompt, using mps friendly method
+        bs_embed, seq_len, _ = image_embeddings.shape
+        image_embeddings = image_embeddings.repeat(1, num_videos_per_prompt, 1)
+        image_embeddings = image_embeddings.view(bs_embed * num_videos_per_prompt, seq_len, -1)
+        
+        image_embeddings = image_embeddings.unsqueeze(1).repeat((1, num_frames, 1, 1))
+        
+        if do_classifier_free_guidance:
+            negative_image_embeddings = torch.zeros_like(image_embeddings)
+
+            
+            audio_prompts = torch.stack(audio_prompts, dim=0).to(device=device, dtype=dtype)
+            audio_prompts = audio_prompts.unsqueeze(0)
+            image_embeddings = torch.cat([negative_image_embeddings, image_embeddings, image_embeddings])
+
+
+            uncond_audio_prompts = torch.stack(uncond_audio_prompts, dim=0).to(device=device, dtype=dtype)
+            uncond_audio_prompts = uncond_audio_prompts.unsqueeze(0)
+
+
+            # For classifier free guidance, we need to do two forward passes.
+            # Here we concatenate the unconditional and text embeddings into a single batch
+            # to avoid doing two forward passes
+            audio_prompts = torch.cat([uncond_audio_prompts, uncond_audio_prompts, audio_prompts])
+
+        return image_embeddings, audio_prompts
+
+    def _encode_vae_image(
+        self,
+        image: torch.Tensor,
+        device,
+        num_videos_per_prompt,
+        do_classifier_free_guidance,
+    ):
+        image = image.to(device=device)
+        image_latents = self.vae.encode(image).latent_dist.mode()
+        
+        if do_classifier_free_guidance:
+            negative_image_latents = torch.zeros_like(image_latents)
+
+            # For classifier free guidance, we need to do two forward passes.
+            # Here we concatenate the unconditional and text embeddings into a single batch
+            # to avoid doing two forward passes
+            image_latents = torch.cat([negative_image_latents, image_latents, image_latents])
+
+        # duplicate image_latents for each generation per prompt, using mps friendly method
+        image_latents = image_latents.repeat(num_videos_per_prompt, 1, 1, 1)
+
+        return image_latents
+    
+    def _get_add_time_ids(
+        self,
+        fps,
+        motion_bucket_id,
+        noise_aug_strength,
+        dtype,
+        batch_size,
+        num_videos_per_prompt,
+        do_classifier_free_guidance,
+    ):
+        add_time_ids = [fps, motion_bucket_id, noise_aug_strength]
+
+        passed_add_embed_dim = self.unet.config.addition_time_embed_dim * len(add_time_ids)
+        expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features
+
+        if expected_add_embed_dim != passed_add_embed_dim:
+            raise ValueError(
+                f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`."
+            )
+
+        add_time_ids = torch.tensor([add_time_ids], dtype=dtype)
+        add_time_ids = add_time_ids.repeat(batch_size * num_videos_per_prompt, 1)
+
+        if do_classifier_free_guidance:
+            add_time_ids = torch.cat([add_time_ids, add_time_ids, add_time_ids])
+
+        return add_time_ids
+    
+    def decode_latents(self, latents, num_frames, decode_chunk_size=14):
+        # [batch, frames, channels, height, width] -> [batch*frames, channels, height, width]
+        latents = latents.flatten(0, 1)
+
+        latents = 1 / self.vae.config.scaling_factor * latents
+
+        forward_vae_fn = self.vae._orig_mod.forward if is_compiled_module(self.vae) else self.vae.forward
+        accepts_num_frames = "num_frames" in set(inspect.signature(forward_vae_fn).parameters.keys())
+
+        # decode decode_chunk_size frames at a time to avoid OOM
+        frames = []
+        for i in range(0, latents.shape[0], decode_chunk_size):
+            num_frames_in = latents[i : i + decode_chunk_size].shape[0]
+            decode_kwargs = {}
+            if accepts_num_frames:
+                # we only pass num_frames_in if it's expected
+                decode_kwargs["num_frames"] = num_frames_in
+
+            frame = self.vae.decode(latents[i : i + decode_chunk_size], **decode_kwargs).sample
+            frames.append(frame.cpu())
+        frames = torch.cat(frames, dim=0)
+
+        # [batch*frames, channels, height, width] -> [batch, channels, frames, height, width]
+        frames = frames.reshape(-1, num_frames, *frames.shape[1:]).permute(0, 2, 1, 3, 4)
+
+        # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+        frames = frames.float()
+        return frames
+
+    def check_inputs(self, image, height, width):
+        if (
+            not isinstance(image, torch.Tensor)
+            and not isinstance(image, PIL.Image.Image)
+            and not isinstance(image, list)
+        ):
+            raise ValueError(
+                "`image` has to be of type `torch.FloatTensor` or `PIL.Image.Image` or `List[PIL.Image.Image]` but is"
+                f" {type(image)}"
+            )
+
+        if height % 8 != 0 or width % 8 != 0:
+            raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+    def prepare_latents(
+        self,
+        batch_size,
+        num_frames,
+        num_channels_latents,
+        height,
+        width,
+        dtype,
+        device,
+        generator,
+        latents=None,
+        ref_image_latents=None,
+        timestep=None
+    ):
+        shape = (
+            batch_size,
+            num_frames,
+            num_channels_latents // 2,
+            height // self.vae_scale_factor,
+            width // self.vae_scale_factor,
+        )
+        if isinstance(generator, list) and len(generator) != batch_size:
+            raise ValueError(
+                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+            )
+
+        if latents is None:
+            noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+        else:
+            noise = latents.to(device)
+
+        # scale the initial noise by the standard deviation required by the scheduler
+        if timestep is not None:
+            init_latents = ref_image_latents.unsqueeze(1)
+            latents = self.scheduler.add_noise(init_latents, noise, timestep)
+        else:
+            latents = noise * self.scheduler.init_noise_sigma
+        return latents
+    
+    def get_timesteps(self, num_inference_steps, strength, device):
+        # get the original timestep using init_timestep
+        init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
+
+        t_start = max(num_inference_steps - init_timestep, 0)
+        timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :]
+
+        return timesteps, num_inference_steps - t_start
+
+    @property
+    def guidance_scale1(self):
+        return self._guidance_scale1
+    
+    @property
+    def guidance_scale2(self):
+        return self._guidance_scale2
+
+    # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+    # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+    # corresponds to doing no classifier free guidance.
+    @property
+    def do_classifier_free_guidance(self):
+        return True
+
+    @property
+    def num_timesteps(self):
+        return self._num_timesteps
+
+    @torch.no_grad()
+    def __call__(
+        self,
+        ref_image: Union[PIL.Image.Image, List[PIL.Image.Image], torch.FloatTensor],
+        clip_image: Union[PIL.Image.Image, List[PIL.Image.Image], torch.FloatTensor],
+        face_mask: Union[PIL.Image.Image, List[PIL.Image.Image], torch.FloatTensor],
+        audio_prompts: Union[PIL.Image.Image, List[PIL.Image.Image], torch.FloatTensor],
+        uncond_audio_prompts: Union[PIL.Image.Image, List[PIL.Image.Image], torch.FloatTensor],
+        motion_buckets: Union[PIL.Image.Image, List[PIL.Image.Image], torch.FloatTensor],
+        height: int = 576,
+        width: int = 1024,
+        num_frames: Optional[int] = None,
+        num_inference_steps: int = 25,
+        min_guidance_scale1=1.0, # 1.0,
+        max_guidance_scale1=3.0,
+        min_guidance_scale2=1.0, # 1.0,
+        max_guidance_scale2=3.0,
+        fps: int = 7,
+        motion_bucket_scale=1.0,
+        noise_aug_strength: int = 0.02,
+        decode_chunk_size: Optional[int] = None,
+        num_videos_per_prompt: Optional[int] = 1,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        output_type: Optional[str] = "pil",
+        callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+        return_dict: bool = True,
+        overlap=7,
+        shift_offset=3,
+        frames_per_batch=14,
+        i2i_noise_strength=1.0,
+    ):
+        r"""
+        The call function to the pipeline for generation.
+
+        Args:
+            image (`PIL.Image.Image` or `List[PIL.Image.Image]` or `torch.FloatTensor`):
+                Image or images to guide image generation. If you provide a tensor, it needs to be compatible with
+                [`CLIPImageProcessor`](https://huggingface.co/lambdalabs/sd-image-variations-diffusers/blob/main/feature_extractor/preprocessor_config.json).
+            height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+                The height in pixels of the generated image.
+            width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+                The width in pixels of the generated image.
+            num_frames (`int`, *optional*):
+                The number of video frames to generate. Defaults to 14 for `stable-video-diffusion-img2vid` and to 25 for `stable-video-diffusion-img2vid-xt`
+            num_inference_steps (`int`, *optional*, defaults to 25):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference. This parameter is modulated by `strength`.
+            min_guidance_scale (`float`, *optional*, defaults to 1.0):
+                The minimum guidance scale. Used for the classifier free guidance with first frame.
+            max_guidance_scale (`float`, *optional*, defaults to 3.0):
+                The maximum guidance scale. Used for the classifier free guidance with last frame.
+            fps (`int`, *optional*, defaults to 7):
+                Frames per second. The rate at which the generated images shall be exported to a video after generation.
+                Note that Stable Diffusion Video's UNet was micro-conditioned on fps-1 during training.
+            motion_bucket_id (`int`, *optional*, defaults to 127):
+                The motion bucket ID. Used as conditioning for the generation. The higher the number the more motion will be in the video.
+            noise_aug_strength (`int`, *optional*, defaults to 0.02):
+                The amount of noise added to the init image, the higher it is the less the video will look like the init image. Increase it for more motion.
+            decode_chunk_size (`int`, *optional*):
+                The number of frames to decode at a time. The higher the chunk size, the higher the temporal consistency
+                between frames, but also the higher the memory consumption. By default, the decoder will decode all frames at once
+                for maximal quality. Reduce `decode_chunk_size` to reduce memory usage.
+            num_videos_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+                generation deterministic.
+            latents (`torch.FloatTensor`, *optional*):
+                Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor is generated by sampling using the supplied random `generator`.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+            callback_on_step_end (`Callable`, *optional*):
+                A function that calls at the end of each denoising steps during the inference. The function is called
+                with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+                callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+                `callback_on_step_end_tensor_inputs`.
+            callback_on_step_end_tensor_inputs (`List`, *optional*):
+                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+                `._callback_tensor_inputs` attribute of your pipeline class.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+                plain tuple.
+
+        Returns:
+            [`~pipelines.stable_diffusion.StableVideoDiffusionPipelineOutput`] or `tuple`:
+                If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableVideoDiffusionPipelineOutput`] is returned,
+                otherwise a `tuple` is returned where the first element is a list of list with the generated frames.
+
+        Examples:
+
+        ```py
+        from diffusers import StableVideoDiffusionPipeline
+        from diffusers.utils import load_image, export_to_video
+
+        pipe = StableVideoDiffusionPipeline.from_pretrained("stabilityai/stable-video-diffusion-img2vid-xt", torch_dtype=torch.float16, variant="fp16")
+        pipe.to("cuda")
+
+        image = load_image("https://lh3.googleusercontent.com/y-iFOHfLTwkuQSUegpwDdgKmOjRSTvPxat63dQLB25xkTs4lhIbRUFeNBWZzYf370g=s1200")
+        image = image.resize((1024, 576))
+
+        frames = pipe(image, num_frames=25, decode_chunk_size=8).frames[0]
+        export_to_video(frames, "generated.mp4", fps=7)
+        ```
+        """
+        # 0. Default height and width to unet
+        height = height or self.unet.config.sample_size * self.vae_scale_factor
+        width = width or self.unet.config.sample_size * self.vae_scale_factor
+
+
+        num_frames = num_frames if num_frames is not None else self.unet.config.num_frames
+        decode_chunk_size = decode_chunk_size if decode_chunk_size is not None else num_frames
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(ref_image, height, width)
+
+        # 2. Define call parameters
+        if isinstance(ref_image, PIL.Image.Image):
+            batch_size = 1
+        elif isinstance(ref_image, list):
+            batch_size = len(ref_image)
+        else:
+            batch_size = ref_image.shape[0]
+
+        device = self._execution_device
+        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+        # corresponds to doing no classifier free guidance.
+        do_classifier_free_guidance = True
+
+        # 3. Prepare clip image embeds
+        image_embeddings, audio_prompts = self._clip_encode_image(
+            clip_image, 
+            audio_prompts,
+            uncond_audio_prompts,
+            num_frames,
+            device, 
+            num_videos_per_prompt, 
+            do_classifier_free_guidance,
+            frames_per_batch)        
+        motion_buckets = torch.stack(motion_buckets, dim=0).to(device=device)
+        motion_buckets = motion_buckets.unsqueeze(0)
+        # NOTE: Stable Diffusion Video was conditioned on fps - 1, which
+        # is why it is reduced here.
+        # See: https://github.com/Stability-AI/generative-models/blob/ed0997173f98eaf8f4edf7ba5fe8f15c6b877fd3/scripts/sampling/simple_video_sample.py#L188
+        # fps = fps - 1
+
+        # 4. Encode input image using VAE
+        # needs_upcasting = (self.vae.dtype == torch.float16 or self.vae.dtype == torch.bfloat16) and self.vae.config.force_upcast
+        needs_upcasting = False
+        vae_dtype = self.vae.dtype
+        if needs_upcasting:
+            self.vae.to(dtype=torch.float32)
+        
+        # Prepare ref image latents
+        ref_image_tensor = ref_image.to(
+            dtype=self.vae.dtype, device=self.vae.device
+        )
+ 
+        ref_image_latents = self.vae.encode(ref_image_tensor).latent_dist.mean
+        ref_image_latents = ref_image_latents * 0.18215  # (b, 4, h, w)
+
+        noise = randn_tensor(
+            ref_image_tensor.shape, 
+            generator=generator, 
+            device=self.vae.device, 
+            dtype=self.vae.dtype)
+        
+        ref_image_tensor = ref_image_tensor + noise_aug_strength * noise
+
+        image_latents = self._encode_vae_image(
+            ref_image_tensor,
+            device=device,
+            num_videos_per_prompt=num_videos_per_prompt,
+            do_classifier_free_guidance=do_classifier_free_guidance,
+        )
+        image_latents = image_latents.to(image_embeddings.dtype)
+        ref_image_latents = ref_image_latents.to(image_embeddings.dtype)
+        
+        # cast back to fp16 if needed
+        if needs_upcasting:
+            self.vae.to(dtype=vae_dtype)
+        
+        # Repeat the image latents for each frame so we can concatenate them with the noise
+        # image_latents [batch, channels, height, width] ->[batch, num_frames, channels, height, width]
+        image_latents = image_latents.unsqueeze(1).repeat(1, num_frames, 1, 1, 1)        
+
+        motion_buckets = motion_buckets * motion_bucket_scale
+        
+        # 4. Prepare timesteps
+        self.scheduler.set_timesteps(num_inference_steps, device=device)
+        timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, i2i_noise_strength, device)
+        latent_timestep = timesteps[:1].repeat(batch_size * num_videos_per_prompt)
+
+
+        # 5. Prepare latent variables
+        num_channels_latents = self.unet.config.in_channels
+        latents = self.prepare_latents(
+            batch_size * num_videos_per_prompt,
+            num_frames,
+            num_channels_latents,
+            height,
+            width,
+            image_embeddings.dtype,
+            device,
+            generator,
+            latents,
+            ref_image_latents,
+            timestep=latent_timestep
+        )
+
+        # Prepare a list of pose condition images
+
+
+        face_mask = face_mask.to(
+        device=device, dtype=self.unet.dtype
+        )[:,:1]
+
+        # 7. Prepare guidance scale
+        guidance_scale = torch.linspace(
+            min_guidance_scale1, 
+            max_guidance_scale1, 
+            num_inference_steps)
+        guidance_scale1 = guidance_scale.to(device, latents.dtype)
+
+        guidance_scale = torch.linspace(
+            min_guidance_scale2, 
+            max_guidance_scale2, 
+            num_inference_steps)
+        guidance_scale2 = guidance_scale.to(device, latents.dtype)
+
+        self._guidance_scale1 = guidance_scale1
+        self._guidance_scale2 = guidance_scale2
+
+        # 8. Denoising loop
+        latents_all = latents # for any-frame generation
+
+        num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+        self._num_timesteps = len(timesteps)
+        shift = 0
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+
+                # init 
+                pred_latents = torch.zeros_like(
+                    latents_all,
+                    dtype=self.unet.dtype,
+                )
+                counter = torch.zeros(
+                    (latents_all.shape[0], num_frames, 1, 1, 1),
+                    dtype=self.unet.dtype,
+                ).to(device=latents_all.device)
+
+                for batch, index_start in enumerate(range(0, num_frames, frames_per_batch - overlap)):
+                    self.scheduler._step_index = None
+                    index_start -= shift
+                    def indice_slice(tensor, idx_list):
+                        tensor_list = []
+                        for idx in idx_list:
+                            idx = idx % tensor.shape[1]
+                            tensor_list.append(tensor[:,idx])
+                        return torch.stack(tensor_list, 1)
+                    idx_list = list(range(index_start, index_start+frames_per_batch))
+                    latents = indice_slice(latents_all, idx_list)
+                    image_latents_input = indice_slice(image_latents, idx_list)
+                    batch_image_embeddings = indice_slice(image_embeddings, idx_list)
+                    batch_audio_prompts = indice_slice(audio_prompts, idx_list)
+
+                    cross_attention_kwargs = {'ip_adapter_masks': [face_mask]}
+                    latent_model_input = torch.cat([latents] * 3) if do_classifier_free_guidance else latents
+                    latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+                    # Concatenate image_latents over channels dimention
+                    latent_model_input = torch.cat([
+                        latent_model_input, 
+                        image_latents_input], dim=2)
+                    
+                    motion_bucket = indice_slice(motion_buckets, idx_list)
+                    motion_bucket = torch.mean(motion_bucket, dim=1).squeeze()
+                    motion_bucket_id = motion_bucket[0]
+                    motion_bucket_id_exp = motion_bucket[1]
+                    added_time_ids = self._get_add_time_ids(
+                        fps,
+                        motion_bucket_id,
+                        motion_bucket_id_exp,
+                        image_embeddings.dtype,
+                        batch_size,
+                        num_videos_per_prompt,
+                        do_classifier_free_guidance,
+                    )
+                    added_time_ids = added_time_ids.to(device, dtype=self.unet.dtype)
+
+                    # predict the noise residual
+                    noise_pred = self.unet(
+                        latent_model_input,
+                        t,
+                        encoder_hidden_states=(batch_image_embeddings.flatten(0,1), [batch_audio_prompts.flatten(0,1)]),
+                        cross_attention_kwargs=cross_attention_kwargs,
+                        added_time_ids=added_time_ids,
+                        return_dict=False,
+                    )[0]        
+                    # perform guidance
+                    if do_classifier_free_guidance:
+                        noise_pred_uncond, noise_pred_drop_audio, noise_pred_cond = noise_pred.chunk(3)
+                        noise_pred = noise_pred_uncond + self.guidance_scale1[i] * (noise_pred_drop_audio - noise_pred_uncond) + self.guidance_scale2[i] * (noise_pred_cond - noise_pred_drop_audio)
+
+                    # compute the previous noisy sample x_t -> x_t-1
+                    latents = self.scheduler.step(noise_pred, t.to(self.unet.dtype), latents).prev_sample
+
+                    if callback_on_step_end is not None:
+                        callback_kwargs = {}
+                        for k in callback_on_step_end_tensor_inputs:
+                            callback_kwargs[k] = locals()[k]
+                        callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+                        latents = callback_outputs.pop("latents", latents)
+
+                    # if batch == 0:
+                    for iii in range(frames_per_batch):
+                        p = (index_start + iii) % pred_latents.shape[1]
+                        pred_latents[:, p] += latents[:, iii]
+                        counter[:, p] += 1
+                shift += shift_offset
+
+                pred_latents  = pred_latents / counter
+                latents_all = pred_latents
+
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+
+        latents = latents_all
+        if not output_type == "latent":
+            # cast back to fp16 if needed
+            if needs_upcasting:
+                self.vae.to(dtype=vae_dtype)
+            frames = self.decode_latents(latents, num_frames, decode_chunk_size)
+        else:
+            frames = latents
+
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return frames
+        return Pose2VideoSVDPipelineOutput(frames=frames)

src/utils/RIFE/IFNet_HDv3.py

@@ -0,0 +1,115 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from .warplayer import warp
+
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+def conv(in_planes, out_planes, kernel_size=3, stride=1, padding=1, dilation=1):
+    return nn.Sequential(
+        nn.Conv2d(in_planes, out_planes, kernel_size=kernel_size, stride=stride,
+                  padding=padding, dilation=dilation, bias=True),        
+        nn.PReLU(out_planes)
+    )
+
+def conv_bn(in_planes, out_planes, kernel_size=3, stride=1, padding=1, dilation=1):
+    return nn.Sequential(
+        nn.Conv2d(in_planes, out_planes, kernel_size=kernel_size, stride=stride,
+                  padding=padding, dilation=dilation, bias=False),
+        nn.BatchNorm2d(out_planes),
+        nn.PReLU(out_planes)
+    )
+
+class IFBlock(nn.Module):
+    def __init__(self, in_planes, c=64):
+        super(IFBlock, self).__init__()
+        self.conv0 = nn.Sequential(
+            conv(in_planes, c//2, 3, 2, 1),
+            conv(c//2, c, 3, 2, 1),
+            )
+        self.convblock0 = nn.Sequential(
+            conv(c, c),
+            conv(c, c)
+        )
+        self.convblock1 = nn.Sequential(
+            conv(c, c),
+            conv(c, c)
+        )
+        self.convblock2 = nn.Sequential(
+            conv(c, c),
+            conv(c, c)
+        )
+        self.convblock3 = nn.Sequential(
+            conv(c, c),
+            conv(c, c)
+        )
+        self.conv1 = nn.Sequential(
+            nn.ConvTranspose2d(c, c//2, 4, 2, 1),
+            nn.PReLU(c//2),
+            nn.ConvTranspose2d(c//2, 4, 4, 2, 1),
+        )
+        self.conv2 = nn.Sequential(
+            nn.ConvTranspose2d(c, c//2, 4, 2, 1),
+            nn.PReLU(c//2),
+            nn.ConvTranspose2d(c//2, 1, 4, 2, 1),
+        )
+
+    def forward(self, x, flow, scale=1):
+        x = F.interpolate(x, scale_factor= 1. / scale, mode="bilinear", align_corners=False, recompute_scale_factor=False)
+        flow = F.interpolate(flow, scale_factor= 1. / scale, mode="bilinear", align_corners=False, recompute_scale_factor=False) * 1. / scale
+        feat = self.conv0(torch.cat((x, flow), 1))
+        feat = self.convblock0(feat) + feat
+        feat = self.convblock1(feat) + feat
+        feat = self.convblock2(feat) + feat
+        feat = self.convblock3(feat) + feat        
+        flow = self.conv1(feat)
+        mask = self.conv2(feat)
+        flow = F.interpolate(flow, scale_factor=scale, mode="bilinear", align_corners=False, recompute_scale_factor=False) * scale
+        mask = F.interpolate(mask, scale_factor=scale, mode="bilinear", align_corners=False, recompute_scale_factor=False)
+        return flow, mask
+        
+class IFNet(nn.Module):
+    def __init__(self):
+        super(IFNet, self).__init__()
+        self.block0 = IFBlock(7+4, c=90)
+        self.block1 = IFBlock(7+4, c=90)
+        self.block2 = IFBlock(7+4, c=90)
+        self.block_tea = IFBlock(10+4, c=90)
+        # self.contextnet = Contextnet()
+        # self.unet = Unet()
+
+    def forward(self, x, scale_list=[4, 2, 1], training=False):
+        if training == False:
+            channel = x.shape[1] // 2
+            img0 = x[:, :channel]
+            img1 = x[:, channel:]
+        flow_list = []
+        merged = []
+        mask_list = []
+        warped_img0 = img0
+        warped_img1 = img1
+        flow = (x[:, :4]).detach() * 0
+        mask = (x[:, :1]).detach() * 0
+        loss_cons = 0
+        block = [self.block0, self.block1, self.block2]
+        for i in range(3):
+            f0, m0 = block[i](torch.cat((warped_img0[:, :3], warped_img1[:, :3], mask), 1), flow, scale=scale_list[i])
+            f1, m1 = block[i](torch.cat((warped_img1[:, :3], warped_img0[:, :3], -mask), 1), torch.cat((flow[:, 2:4], flow[:, :2]), 1), scale=scale_list[i])
+            flow = flow + (f0 + torch.cat((f1[:, 2:4], f1[:, :2]), 1)) / 2
+            mask = mask + (m0 + (-m1)) / 2
+            mask_list.append(mask)
+            flow_list.append(flow)
+            warped_img0 = warp(img0, flow[:, :2])
+            warped_img1 = warp(img1, flow[:, 2:4])
+            merged.append((warped_img0, warped_img1))
+        '''
+        c0 = self.contextnet(img0, flow[:, :2])
+        c1 = self.contextnet(img1, flow[:, 2:4])
+        tmp = self.unet(img0, img1, warped_img0, warped_img1, mask, flow, c0, c1)
+        res = tmp[:, 1:4] * 2 - 1
+        '''
+        for i in range(3):
+            mask_list[i] = torch.sigmoid(mask_list[i])
+            merged[i] = merged[i][0] * mask_list[i] + merged[i][1] * (1 - mask_list[i])
+            # merged[i] = torch.clamp(merged[i] + res, 0, 1)        
+        return flow_list, mask_list[2], merged

src/utils/RIFE/RIFE_HDv3.py

@@ -0,0 +1,37 @@
+import torch
+from .IFNet_HDv3 import *
+import torch.nn.functional as F
+    
+class RIFEModel:
+    def __init__(self, device=None):
+        if device is None:
+            self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+        else:
+            self.device = device
+        self.flownet = IFNet().to(self.device).eval()
+
+    def train(self):
+        self.flownet.train()
+
+    def eval(self):
+        self.flownet.eval()
+
+
+    def load_model(self, path, rank=-1):
+        def convert(param):
+            if rank == -1:
+                return {
+                    k.replace("module.", ""): v
+                    for k, v in param.items()
+                    if "module." in k
+                }
+            else:
+                return param
+        self.flownet.load_state_dict(convert(torch.load('{}/flownet.pkl'.format(path), map_location ='cpu')))
+                
+
+    def inference(self, img0, img1, scale=1.0):
+        imgs = torch.cat((img0, img1), 1)
+        scale_list = [4/scale, 2/scale, 1/scale]
+        flow, mask, merged = self.flownet(imgs, scale_list)
+        return merged[2]

src/utils/RIFE/warplayer.py

@@ -0,0 +1,22 @@
+import torch
+import torch.nn as nn
+
+backwarp_tenGrid = {}
+
+
+def warp(tenInput, tenFlow):
+    device = tenFlow.device
+    k = (str(tenFlow.device), str(tenFlow.size()))
+    if k not in backwarp_tenGrid:
+        tenHorizontal = torch.linspace(-1.0, 1.0, tenFlow.shape[3], device=device).view(
+            1, 1, 1, tenFlow.shape[3]).expand(tenFlow.shape[0], -1, tenFlow.shape[2], -1)
+        tenVertical = torch.linspace(-1.0, 1.0, tenFlow.shape[2], device=device).view(
+            1, 1, tenFlow.shape[2], 1).expand(tenFlow.shape[0], -1, -1, tenFlow.shape[3])
+        backwarp_tenGrid[k] = torch.cat(
+            [tenHorizontal, tenVertical], 1).to(device)
+
+    tenFlow = torch.cat([tenFlow[:, 0:1, :, :] / ((tenInput.shape[3] - 1.0) / 2.0),
+                         tenFlow[:, 1:2, :, :] / ((tenInput.shape[2] - 1.0) / 2.0)], 1)
+
+    g = (backwarp_tenGrid[k] + tenFlow).permute(0, 2, 3, 1)
+    return torch.nn.functional.grid_sample(input=tenInput, grid=g, mode='bilinear', padding_mode='border', align_corners=True)

src/utils/mask_processer.py

@@ -0,0 +1,117 @@
+
+import math
+import warnings
+from typing import List, Optional, Tuple, Union
+
+import numpy as np
+import PIL.Image
+import torch
+import torch.nn.functional as F
+from PIL import Image, ImageFilter, ImageOps
+
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.utils import CONFIG_NAME, PIL_INTERPOLATION, deprecate
+from diffusers.image_processor import VaeImageProcessor
+
+class IPAdapterMaskProcessor(VaeImageProcessor):
+    """
+    Image processor for IP Adapter image masks.
+
+    Args:
+        do_resize (`bool`, *optional*, defaults to `True`):
+            Whether to downscale the image's (height, width) dimensions to multiples of `vae_scale_factor`.
+        vae_scale_factor (`int`, *optional*, defaults to `8`):
+            VAE scale factor. If `do_resize` is `True`, the image is automatically resized to multiples of this factor.
+        resample (`str`, *optional*, defaults to `lanczos`):
+            Resampling filter to use when resizing the image.
+        do_normalize (`bool`, *optional*, defaults to `False`):
+            Whether to normalize the image to [-1,1].
+        do_binarize (`bool`, *optional*, defaults to `True`):
+            Whether to binarize the image to 0/1.
+        do_convert_grayscale (`bool`, *optional*, defaults to be `True`):
+            Whether to convert the images to grayscale format.
+
+    """
+
+    config_name = CONFIG_NAME
+
+    @register_to_config
+    def __init__(
+        self,
+        do_resize: bool = True,
+        vae_scale_factor: int = 8,
+        resample: str = "lanczos",
+        do_normalize: bool = False,
+        do_binarize: bool = True,
+        do_convert_grayscale: bool = True,
+    ):
+        super().__init__(
+            do_resize=do_resize,
+            vae_scale_factor=vae_scale_factor,
+            resample=resample,
+            do_normalize=do_normalize,
+            do_binarize=do_binarize,
+            do_convert_grayscale=do_convert_grayscale,
+        )
+
+    @staticmethod
+    def downsample(mask: torch.Tensor, batch_size: int, num_queries: int, value_embed_dim: int):
+        """
+        Downsamples the provided mask tensor to match the expected dimensions for scaled dot-product attention. If the
+        aspect ratio of the mask does not match the aspect ratio of the output image, a warning is issued.
+
+        Args:
+            mask (`torch.Tensor`):
+                The input mask tensor generated with `IPAdapterMaskProcessor.preprocess()`.
+            batch_size (`int`):
+                The batch size.
+            num_queries (`int`):
+                The number of queries.
+            value_embed_dim (`int`):
+                The dimensionality of the value embeddings.
+
+        Returns:
+            `torch.Tensor`:
+                The downsampled mask tensor.
+
+        """
+        o_h = mask.shape[1]
+        o_w = mask.shape[2]
+        ratio = o_w / o_h
+        mask_h = int(torch.sqrt(torch.FloatTensor([num_queries / ratio]))[0])
+        mask_h = int(mask_h) + int((num_queries % int(mask_h)) != 0)
+        mask_w = num_queries // mask_h
+
+        mask_downsample = F.interpolate(mask.unsqueeze(0), size=(mask_h, mask_w), mode="bicubic").squeeze(0)
+
+        # Repeat batch_size times
+        if mask_downsample.shape[0] < batch_size:
+            mask_downsample = mask_downsample.repeat(batch_size, 1, 1)
+
+        mask_downsample = mask_downsample.view(mask_downsample.shape[0], -1)
+
+        downsampled_area = mask_h * mask_w
+        # If the output image and the mask do not have the same aspect ratio, tensor shapes will not match
+        # Pad tensor if downsampled_mask.shape[1] is smaller than num_queries
+        if downsampled_area < num_queries:
+            warnings.warn(
+                "The aspect ratio of the mask does not match the aspect ratio of the output image. "
+                "Please update your masks or adjust the output size for optimal performance.",
+                UserWarning,
+            )
+            mask_downsample = F.pad(mask_downsample, (0, num_queries - mask_downsample.shape[1]), value=0.0)
+        # Discard last embeddings if downsampled_mask.shape[1] is bigger than num_queries
+        if downsampled_area > num_queries:
+            warnings.warn(
+                "The aspect ratio of the mask does not match the aspect ratio of the output image. "
+                "Please update your masks or adjust the output size for optimal performance.",
+                UserWarning,
+            )
+            mask_downsample = mask_downsample[:, :num_queries]
+
+        # Repeat last dimension to match SDPA output shape
+        mask_downsample = mask_downsample.view(mask_downsample.shape[0], mask_downsample.shape[1], 1).repeat(
+            1, 1, value_embed_dim
+        )
+
+        return mask_downsample

src/utils/util.py

@@ -0,0 +1,67 @@
+import importlib
+import os
+import os.path as osp
+import shutil
+import sys
+from pathlib import Path
+
+import numpy as np
+import torch
+import torchvision
+from einops import rearrange
+from PIL import Image
+import imageio
+
+def seed_everything(seed):
+    import random
+    import numpy as np
+
+    torch.manual_seed(seed)
+    torch.cuda.manual_seed_all(seed)
+    np.random.seed(seed % (2**32))
+    random.seed(seed)
+
+
+def save_videos_from_pil(pil_images, path, fps=8):
+    save_fmt = Path(path).suffix
+    os.makedirs(os.path.dirname(path), exist_ok=True)
+
+    if save_fmt == ".mp4":
+        with imageio.get_writer(path, fps=fps) as writer:
+            for img in pil_images:
+                img_array = np.array(img)  # Convert PIL Image to numpy array
+                writer.append_data(img_array)
+
+    elif save_fmt == ".gif":
+        pil_images[0].save(
+            fp=path,
+            format="GIF",
+            append_images=pil_images[1:],
+            save_all=True,
+            duration=(1 / fps * 1000),
+            loop=0,
+            optimize=False,
+            lossless=True
+        )
+    else:
+        raise ValueError("Unsupported file type. Use .mp4 or .gif.")
+
+
+def save_videos_grid(videos: torch.Tensor, path: str, rescale=False, n_rows=6, fps=8):
+    videos = rearrange(videos, "b c t h w -> t b c h w")
+    height, width = videos.shape[-2:]
+    outputs = []
+
+    for i, x in enumerate(videos):
+        x = torchvision.utils.make_grid(x, nrow=n_rows)  # (c h w)
+        x = x.transpose(0, 1).transpose(1, 2).squeeze(-1)  # (h w c)
+        if rescale:
+            x = (x + 1.0) / 2.0  # -1,1 -> 0,1
+        x = (x * 255).numpy().astype(np.uint8)
+        x = Image.fromarray(x)
+        outputs.append(x)
+
+    os.makedirs(os.path.dirname(path), exist_ok=True)
+
+    save_videos_from_pil(outputs, path, fps)
+