Create Code

.gitignore

@@ -0,0 +1,167 @@
+# Byte-compiled / optimized / DLL files
+models*
+output*
+logs*
+taming*
+samples*
+datasets*
+asset*
+__pycache__/
+*.py[cod]
+*$py.class
+
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+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
+*.py,cover
+.hypothesis/
+.pytest_cache/
+cover/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+db.sqlite3-journal
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+.pybuilder/
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# IPython
+profile_default/
+ipython_config.py
+
+# pyenv
+#   For a library or package, you might want to ignore these files since the code is
+#   intended to run in multiple environments; otherwise, check them in:
+# .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
+
+# poetry
+#   Similar to Pipfile.lock, it is generally recommended to include poetry.lock in version control.
+#   This is especially recommended for binary packages to ensure reproducibility, and is more
+#   commonly ignored for libraries.
+#   https://python-poetry.org/docs/basic-usage/#commit-your-poetrylock-file-to-version-control
+#poetry.lock
+
+# pdm
+#   Similar to Pipfile.lock, it is generally recommended to include pdm.lock in version control.
+#pdm.lock
+#   pdm stores project-wide configurations in .pdm.toml, but it is recommended to not include it
+#   in version control.
+#   https://pdm.fming.dev/#use-with-ide
+.pdm.toml
+
+# PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm
+__pypackages__/
+
+# Celery stuff
+celerybeat-schedule
+celerybeat.pid
+
+# 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/
+
+# pytype static type analyzer
+.pytype/
+
+# Cython debug symbols
+cython_debug/
+
+# PyCharm
+#  JetBrains specific template is maintained in a separate JetBrains.gitignore that can
+#  be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore
+#  and can be added to the global gitignore or merged into this file.  For a more nuclear
+#  option (not recommended) you can uncomment the following to ignore the entire idea folder.
+#.idea/

LICENSE

@@ -0,0 +1,201 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

README.md

@@ -11,3 +11,10 @@ license: other
 ---
 
 Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+
+# License
+This project is licensed under the [Apache License (Version 2.0)](https://github.com/modelscope/modelscope/blob/master/LICENSE).
+
+The CogVideoX-2B model (including its corresponding Transformers module and VAE module) is released under the [Apache 2.0 License](LICENSE).
+
+The CogVideoX-5B model (Transformers module) is released under the [CogVideoX LICENSE](https://huggingface.co/THUDM/CogVideoX-5b/blob/main/LICENSE).

app.py

@@ -0,0 +1,46 @@
+import time 
+import torch
+
+from cogvideox.api.api import infer_forward_api, update_diffusion_transformer_api, update_edition_api
+from cogvideox.ui.ui import ui_modelscope, ui_eas, ui
+
+if __name__ == "__main__":
+    # Choose the ui mode  
+    ui_mode = "eas"
+    
+    # Low gpu memory mode, this is used when the GPU memory is under 16GB
+    low_gpu_memory_mode = False
+    # Use torch.float16 if GPU does not support torch.bfloat16
+    # ome graphics cards, such as v100, 2080ti, do not support torch.bfloat16
+    weight_dtype = torch.bfloat16
+
+    # Server ip
+    server_name = "0.0.0.0"
+    server_port = 7860
+
+    # Params below is used when ui_mode = "modelscope"
+    model_name = "models/Diffusion_Transformer/CogVideoX-Fun-5b-InP"
+    savedir_sample = "samples"
+
+    if ui_mode == "modelscope":
+        demo, controller = ui_modelscope(model_name, savedir_sample, low_gpu_memory_mode, weight_dtype)
+    elif ui_mode == "eas":
+        demo, controller = ui_eas(model_name, savedir_sample)
+    else:
+        demo, controller = ui(low_gpu_memory_mode, weight_dtype)
+
+    # launch gradio
+    app, _, _ = demo.queue(status_update_rate=1).launch(
+        server_name=server_name,
+        server_port=server_port,
+        prevent_thread_lock=True
+    )
+    
+    # launch api
+    infer_forward_api(None, app, controller)
+    update_diffusion_transformer_api(None, app, controller)
+    update_edition_api(None, app, controller)
+    
+    # not close the python
+    while True:
+        time.sleep(5)

cogvideox/api/api.py

@@ -0,0 +1,149 @@
+import io
+import gc
+import base64
+import torch
+import gradio as gr
+import tempfile
+import hashlib
+import os
+
+from fastapi import FastAPI
+from io import BytesIO
+from PIL import Image
+
+# Function to encode a file to Base64
+def encode_file_to_base64(file_path):
+    with open(file_path, "rb") as file:
+        # Encode the data to Base64
+        file_base64 = base64.b64encode(file.read())
+        return file_base64
+
+def update_edition_api(_: gr.Blocks, app: FastAPI, controller):
+    @app.post("/cogvideox_fun/update_edition")
+    def _update_edition_api(
+        datas: dict,
+    ):
+        edition = datas.get('edition', 'v2')
+
+        try:
+            controller.update_edition(
+                edition
+            )
+            comment = "Success"
+        except Exception as e:
+            torch.cuda.empty_cache()
+            comment = f"Error. error information is {str(e)}"
+
+        return {"message": comment}
+
+def update_diffusion_transformer_api(_: gr.Blocks, app: FastAPI, controller):
+    @app.post("/cogvideox_fun/update_diffusion_transformer")
+    def _update_diffusion_transformer_api(
+        datas: dict,
+    ):
+        diffusion_transformer_path = datas.get('diffusion_transformer_path', 'none')
+
+        try:
+            controller.update_diffusion_transformer(
+                diffusion_transformer_path
+            )
+            comment = "Success"
+        except Exception as e:
+            torch.cuda.empty_cache()
+            comment = f"Error. error information is {str(e)}"
+
+        return {"message": comment}
+
+def save_base64_video(base64_string):
+    video_data = base64.b64decode(base64_string)
+
+    md5_hash = hashlib.md5(video_data).hexdigest()
+    filename = f"{md5_hash}.mp4"  
+    
+    temp_dir = tempfile.gettempdir()
+    file_path = os.path.join(temp_dir, filename)
+
+    with open(file_path, 'wb') as video_file:
+        video_file.write(video_data)
+
+    return file_path
+
+def infer_forward_api(_: gr.Blocks, app: FastAPI, controller):
+    @app.post("/cogvideox_fun/infer_forward")
+    def _infer_forward_api(
+        datas: dict,
+    ):
+        base_model_path = datas.get('base_model_path', 'none')
+        lora_model_path = datas.get('lora_model_path', 'none')
+        lora_alpha_slider = datas.get('lora_alpha_slider', 0.55)
+        prompt_textbox = datas.get('prompt_textbox', None)
+        negative_prompt_textbox = datas.get('negative_prompt_textbox', 'The video is not of a high quality, it has a low resolution. Watermark present in each frame. Strange motion trajectory. ')
+        sampler_dropdown = datas.get('sampler_dropdown', 'Euler')
+        sample_step_slider = datas.get('sample_step_slider', 30)
+        resize_method = datas.get('resize_method', "Generate by")
+        width_slider = datas.get('width_slider', 672)
+        height_slider = datas.get('height_slider', 384)
+        base_resolution = datas.get('base_resolution', 512)
+        is_image = datas.get('is_image', False)
+        generation_method = datas.get('generation_method', False)
+        length_slider = datas.get('length_slider', 49)
+        overlap_video_length = datas.get('overlap_video_length', 4)
+        partial_video_length = datas.get('partial_video_length', 72)
+        cfg_scale_slider = datas.get('cfg_scale_slider', 6)
+        start_image = datas.get('start_image', None)
+        end_image = datas.get('end_image', None)
+        validation_video = datas.get('validation_video', None)
+        denoise_strength = datas.get('denoise_strength', 0.70)
+        seed_textbox = datas.get("seed_textbox", 43)
+
+        generation_method = "Image Generation" if is_image else generation_method
+
+        if start_image is not None:
+            start_image = base64.b64decode(start_image)
+            start_image = [Image.open(BytesIO(start_image))]
+        
+        if end_image is not None:
+            end_image = base64.b64decode(end_image)
+            end_image = [Image.open(BytesIO(end_image))]
+
+        if validation_video is not None:
+            validation_video = save_base64_video(validation_video)
+
+        try:
+            save_sample_path, comment = controller.generate(
+                "",
+                base_model_path,
+                lora_model_path, 
+                lora_alpha_slider,
+                prompt_textbox, 
+                negative_prompt_textbox, 
+                sampler_dropdown, 
+                sample_step_slider, 
+                resize_method,
+                width_slider, 
+                height_slider, 
+                base_resolution,
+                generation_method,
+                length_slider, 
+                overlap_video_length, 
+                partial_video_length, 
+                cfg_scale_slider, 
+                start_image,
+                end_image,
+                validation_video,
+                denoise_strength,
+                seed_textbox,
+                is_api = True,
+            )
+        except Exception as e:
+            gc.collect()
+            torch.cuda.empty_cache()
+            torch.cuda.ipc_collect()
+            save_sample_path = ""
+            comment = f"Error. error information is {str(e)}"
+            return {"message": comment}
+        
+        if save_sample_path != "":
+            return {"message": comment, "save_sample_path": save_sample_path, "base64_encoding": encode_file_to_base64(save_sample_path)}
+        else:
+            return {"message": comment, "save_sample_path": save_sample_path}

cogvideox/api/post_infer.py

@@ -0,0 +1,89 @@
+import base64
+import json
+import sys
+import time
+from datetime import datetime
+from io import BytesIO
+
+import cv2
+import requests
+import base64
+
+
+def post_diffusion_transformer(diffusion_transformer_path, url='http://127.0.0.1:7860'):
+    datas = json.dumps({
+        "diffusion_transformer_path": diffusion_transformer_path
+    })
+    r = requests.post(f'{url}/cogvideox_fun/update_diffusion_transformer', data=datas, timeout=1500)
+    data = r.content.decode('utf-8')
+    return data
+
+def post_update_edition(edition, url='http://0.0.0.0:7860'):
+    datas = json.dumps({
+        "edition": edition
+    })
+    r = requests.post(f'{url}/cogvideox_fun/update_edition', data=datas, timeout=1500)
+    data = r.content.decode('utf-8')
+    return data
+
+def post_infer(generation_method, length_slider, url='http://127.0.0.1:7860'):
+    datas = json.dumps({
+        "base_model_path": "none",
+        "motion_module_path": "none",
+        "lora_model_path": "none", 
+        "lora_alpha_slider": 0.55, 
+        "prompt_textbox": "A young woman with beautiful and clear eyes and blonde hair standing and white dress in a forest wearing a crown. She seems to be lost in thought, and the camera focuses on her face. The video is of high quality, and the view is very clear. High quality, masterpiece, best quality, highres, ultra-detailed, fantastic.", 
+        "negative_prompt_textbox": "The video is not of a high quality, it has a low resolution. Watermark present in each frame. Strange motion trajectory. ", 
+        "sampler_dropdown": "Euler", 
+        "sample_step_slider": 50, 
+        "width_slider": 672, 
+        "height_slider": 384, 
+        "generation_method": "Video Generation",
+        "length_slider": length_slider,
+        "cfg_scale_slider": 6,
+        "seed_textbox": 43,
+    })
+    r = requests.post(f'{url}/cogvideox_fun/infer_forward', data=datas, timeout=1500)
+    data = r.content.decode('utf-8')
+    return data
+
+if __name__ == '__main__':
+    # initiate time
+    now_date    = datetime.now()
+    time_start  = time.time()  
+    
+    # -------------------------- #
+    #  Step 1: update edition
+    # -------------------------- #
+    diffusion_transformer_path = "models/Diffusion_Transformer/CogVideoX-Fun-2b-InP"
+    outputs = post_diffusion_transformer(diffusion_transformer_path)
+    print('Output update edition: ', outputs)
+
+    # -------------------------- #
+    #  Step 2: infer
+    # -------------------------- #
+    # "Video Generation" and "Image Generation"
+    generation_method = "Video Generation"
+    length_slider = 49
+    outputs = post_infer(generation_method, length_slider)
+    
+    # Get decoded data
+    outputs = json.loads(outputs)
+    base64_encoding = outputs["base64_encoding"]
+    decoded_data = base64.b64decode(base64_encoding)
+
+    is_image = True if generation_method == "Image Generation" else False
+    if is_image or length_slider == 1:
+        file_path = "1.png"
+    else:
+        file_path = "1.mp4"
+    with open(file_path, "wb") as file:
+        file.write(decoded_data)
+        
+    # End of record time
+    # The calculated time difference is the execution time of the program, expressed in seconds / s
+    time_end = time.time()  
+    time_sum = (time_end - time_start) % 60 
+    print('# --------------------------------------------------------- #')
+    print(f'#   Total expenditure: {time_sum}s')
+    print('# --------------------------------------------------------- #')

cogvideox/data/bucket_sampler.py

@@ -0,0 +1,379 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+import os
+from typing import (Generic, Iterable, Iterator, List, Optional, Sequence,
+                    Sized, TypeVar, Union)
+
+import cv2
+import numpy as np
+import torch
+from PIL import Image
+from torch.utils.data import BatchSampler, Dataset, Sampler
+
+ASPECT_RATIO_512 = {
+    '0.25': [256.0, 1024.0], '0.26': [256.0, 992.0], '0.27': [256.0, 960.0], '0.28': [256.0, 928.0],
+    '0.32': [288.0, 896.0], '0.33': [288.0, 864.0], '0.35': [288.0, 832.0], '0.4': [320.0, 800.0],
+    '0.42': [320.0, 768.0], '0.48': [352.0, 736.0], '0.5': [352.0, 704.0], '0.52': [352.0, 672.0],
+    '0.57': [384.0, 672.0], '0.6': [384.0, 640.0], '0.68': [416.0, 608.0], '0.72': [416.0, 576.0],
+    '0.78': [448.0, 576.0], '0.82': [448.0, 544.0], '0.88': [480.0, 544.0], '0.94': [480.0, 512.0],
+    '1.0': [512.0, 512.0], '1.07': [512.0, 480.0], '1.13': [544.0, 480.0], '1.21': [544.0, 448.0],
+    '1.29': [576.0, 448.0], '1.38': [576.0, 416.0], '1.46': [608.0, 416.0], '1.67': [640.0, 384.0],
+    '1.75': [672.0, 384.0], '2.0': [704.0, 352.0], '2.09': [736.0, 352.0], '2.4': [768.0, 320.0],
+    '2.5': [800.0, 320.0], '2.89': [832.0, 288.0], '3.0': [864.0, 288.0], '3.11': [896.0, 288.0],
+    '3.62': [928.0, 256.0], '3.75': [960.0, 256.0], '3.88': [992.0, 256.0], '4.0': [1024.0, 256.0]
+}
+ASPECT_RATIO_RANDOM_CROP_512 = {
+    '0.42': [320.0, 768.0], '0.5': [352.0, 704.0], 
+    '0.57': [384.0, 672.0], '0.68': [416.0, 608.0], '0.78': [448.0, 576.0], '0.88': [480.0, 544.0], 
+    '0.94': [480.0, 512.0], '1.0': [512.0, 512.0], '1.07': [512.0, 480.0], 
+    '1.13': [544.0, 480.0], '1.29': [576.0, 448.0], '1.46': [608.0, 416.0], '1.75': [672.0, 384.0], 
+    '2.0': [704.0, 352.0],  '2.4': [768.0, 320.0]
+}
+ASPECT_RATIO_RANDOM_CROP_PROB = [
+    1, 2,
+    4, 4, 4, 4,
+    8, 8, 8,
+    4, 4, 4, 4,
+    2, 1
+]
+ASPECT_RATIO_RANDOM_CROP_PROB = np.array(ASPECT_RATIO_RANDOM_CROP_PROB) / sum(ASPECT_RATIO_RANDOM_CROP_PROB)
+
+def get_closest_ratio(height: float, width: float, ratios: dict = ASPECT_RATIO_512):
+    aspect_ratio = height / width
+    closest_ratio = min(ratios.keys(), key=lambda ratio: abs(float(ratio) - aspect_ratio))
+    return ratios[closest_ratio], float(closest_ratio)
+
+def get_image_size_without_loading(path):
+    with Image.open(path) as img:
+        return img.size  # (width, height)
+
+class RandomSampler(Sampler[int]):
+    r"""Samples elements randomly. If without replacement, then sample from a shuffled dataset.
+
+    If with replacement, then user can specify :attr:`num_samples` to draw.
+
+    Args:
+        data_source (Dataset): dataset to sample from
+        replacement (bool): samples are drawn on-demand with replacement if ``True``, default=``False``
+        num_samples (int): number of samples to draw, default=`len(dataset)`.
+        generator (Generator): Generator used in sampling.
+    """
+
+    data_source: Sized
+    replacement: bool
+
+    def __init__(self, data_source: Sized, replacement: bool = False,
+                 num_samples: Optional[int] = None, generator=None) -> None:
+        self.data_source = data_source
+        self.replacement = replacement
+        self._num_samples = num_samples
+        self.generator = generator
+        self._pos_start = 0
+
+        if not isinstance(self.replacement, bool):
+            raise TypeError(f"replacement should be a boolean value, but got replacement={self.replacement}")
+
+        if not isinstance(self.num_samples, int) or self.num_samples <= 0:
+            raise ValueError(f"num_samples should be a positive integer value, but got num_samples={self.num_samples}")
+
+    @property
+    def num_samples(self) -> int:
+        # dataset size might change at runtime
+        if self._num_samples is None:
+            return len(self.data_source)
+        return self._num_samples
+
+    def __iter__(self) -> Iterator[int]:
+        n = len(self.data_source)
+        if self.generator is None:
+            seed = int(torch.empty((), dtype=torch.int64).random_().item())
+            generator = torch.Generator()
+            generator.manual_seed(seed)
+        else:
+            generator = self.generator
+
+        if self.replacement:
+            for _ in range(self.num_samples // 32):
+                yield from torch.randint(high=n, size=(32,), dtype=torch.int64, generator=generator).tolist()
+            yield from torch.randint(high=n, size=(self.num_samples % 32,), dtype=torch.int64, generator=generator).tolist()
+        else:
+            for _ in range(self.num_samples // n):
+                xx = torch.randperm(n, generator=generator).tolist()
+                if self._pos_start >= n:
+                    self._pos_start = 0
+                print("xx top 10", xx[:10], self._pos_start)
+                for idx in range(self._pos_start, n):
+                    yield xx[idx]
+                    self._pos_start = (self._pos_start + 1) % n
+                self._pos_start = 0
+            yield from torch.randperm(n, generator=generator).tolist()[:self.num_samples % n]
+
+    def __len__(self) -> int:
+        return self.num_samples
+
+class AspectRatioBatchImageSampler(BatchSampler):
+    """A sampler wrapper for grouping images with similar aspect ratio into a same batch.
+
+    Args:
+        sampler (Sampler): Base sampler.
+        dataset (Dataset): Dataset providing data information.
+        batch_size (int): Size of mini-batch.
+        drop_last (bool): If ``True``, the sampler will drop the last batch if
+            its size would be less than ``batch_size``.
+        aspect_ratios (dict): The predefined aspect ratios.
+    """
+    def __init__(
+        self,
+        sampler: Sampler,
+        dataset: Dataset,
+        batch_size: int,
+        train_folder: str = None,
+        aspect_ratios: dict = ASPECT_RATIO_512,
+        drop_last: bool = False,
+        config=None,
+        **kwargs
+    ) -> None:
+        if not isinstance(sampler, Sampler):
+            raise TypeError('sampler should be an instance of ``Sampler``, '
+                            f'but got {sampler}')
+        if not isinstance(batch_size, int) or batch_size <= 0:
+            raise ValueError('batch_size should be a positive integer value, '
+                             f'but got batch_size={batch_size}')
+        self.sampler = sampler
+        self.dataset = dataset
+        self.train_folder = train_folder
+        self.batch_size = batch_size
+        self.aspect_ratios = aspect_ratios
+        self.drop_last = drop_last
+        self.config = config
+        # buckets for each aspect ratio 
+        self._aspect_ratio_buckets = {ratio: [] for ratio in aspect_ratios}
+        # [str(k) for k, v in aspect_ratios] 
+        self.current_available_bucket_keys = list(aspect_ratios.keys())
+
+    def __iter__(self):
+        for idx in self.sampler:
+            try:
+                image_dict = self.dataset[idx]
+
+                width, height = image_dict.get("width", None), image_dict.get("height", None)
+                if width is None or height is None:
+                    image_id, name = image_dict['file_path'], image_dict['text']
+                    if self.train_folder is None:
+                        image_dir = image_id
+                    else:
+                        image_dir = os.path.join(self.train_folder, image_id)
+
+                    width, height = get_image_size_without_loading(image_dir)
+
+                    ratio = height / width # self.dataset[idx]
+                else:
+                    height = int(height)
+                    width = int(width)
+                    ratio = height / width # self.dataset[idx]
+            except Exception as e:
+                print(e)
+                continue
+            # find the closest aspect ratio
+            closest_ratio = min(self.aspect_ratios.keys(), key=lambda r: abs(float(r) - ratio))
+            if closest_ratio not in self.current_available_bucket_keys:
+                continue
+            bucket = self._aspect_ratio_buckets[closest_ratio]
+            bucket.append(idx)
+            # yield a batch of indices in the same aspect ratio group
+            if len(bucket) == self.batch_size:
+                yield bucket[:]
+                del bucket[:]
+
+class AspectRatioBatchSampler(BatchSampler):
+    """A sampler wrapper for grouping images with similar aspect ratio into a same batch.
+
+    Args:
+        sampler (Sampler): Base sampler.
+        dataset (Dataset): Dataset providing data information.
+        batch_size (int): Size of mini-batch.
+        drop_last (bool): If ``True``, the sampler will drop the last batch if
+            its size would be less than ``batch_size``.
+        aspect_ratios (dict): The predefined aspect ratios.
+    """
+    def __init__(
+        self,
+        sampler: Sampler,
+        dataset: Dataset,
+        batch_size: int,
+        video_folder: str = None,
+        train_data_format: str = "webvid",
+        aspect_ratios: dict = ASPECT_RATIO_512,
+        drop_last: bool = False,
+        config=None,
+        **kwargs
+    ) -> None:
+        if not isinstance(sampler, Sampler):
+            raise TypeError('sampler should be an instance of ``Sampler``, '
+                            f'but got {sampler}')
+        if not isinstance(batch_size, int) or batch_size <= 0:
+            raise ValueError('batch_size should be a positive integer value, '
+                             f'but got batch_size={batch_size}')
+        self.sampler = sampler
+        self.dataset = dataset
+        self.video_folder = video_folder
+        self.train_data_format = train_data_format
+        self.batch_size = batch_size
+        self.aspect_ratios = aspect_ratios
+        self.drop_last = drop_last
+        self.config = config
+        # buckets for each aspect ratio 
+        self._aspect_ratio_buckets = {ratio: [] for ratio in aspect_ratios}
+        # [str(k) for k, v in aspect_ratios] 
+        self.current_available_bucket_keys = list(aspect_ratios.keys())
+
+    def __iter__(self):
+        for idx in self.sampler:
+            try:
+                video_dict = self.dataset[idx]
+                width, more = video_dict.get("width", None), video_dict.get("height", None)
+
+                if width is None or height is None:
+                    if self.train_data_format == "normal":
+                        video_id, name = video_dict['file_path'], video_dict['text']
+                        if self.video_folder is None:
+                            video_dir = video_id
+                        else:
+                            video_dir = os.path.join(self.video_folder, video_id)
+                    else:
+                        videoid, name, page_dir = video_dict['videoid'], video_dict['name'], video_dict['page_dir']
+                        video_dir = os.path.join(self.video_folder, f"{videoid}.mp4")
+                    cap = cv2.VideoCapture(video_dir)
+
+                    # 获取视频尺寸
+                    width  = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))   # 浮点数转换为整数
+                    height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))  # 浮点数转换为整数
+                    
+                    ratio = height / width # self.dataset[idx]
+                else:
+                    height = int(height)
+                    width = int(width)
+                    ratio = height / width # self.dataset[idx]
+            except Exception as e:
+                print(e)
+                continue
+            # find the closest aspect ratio
+            closest_ratio = min(self.aspect_ratios.keys(), key=lambda r: abs(float(r) - ratio))
+            if closest_ratio not in self.current_available_bucket_keys:
+                continue
+            bucket = self._aspect_ratio_buckets[closest_ratio]
+            bucket.append(idx)
+            # yield a batch of indices in the same aspect ratio group
+            if len(bucket) == self.batch_size:
+                yield bucket[:]
+                del bucket[:]
+
+class AspectRatioBatchImageVideoSampler(BatchSampler):
+    """A sampler wrapper for grouping images with similar aspect ratio into a same batch.
+
+    Args:
+        sampler (Sampler): Base sampler.
+        dataset (Dataset): Dataset providing data information.
+        batch_size (int): Size of mini-batch.
+        drop_last (bool): If ``True``, the sampler will drop the last batch if
+            its size would be less than ``batch_size``.
+        aspect_ratios (dict): The predefined aspect ratios.
+    """
+
+    def __init__(self,
+                 sampler: Sampler,
+                 dataset: Dataset,
+                 batch_size: int,
+                 train_folder: str = None,
+                 aspect_ratios: dict = ASPECT_RATIO_512,
+                 drop_last: bool = False
+                ) -> None:
+        if not isinstance(sampler, Sampler):
+            raise TypeError('sampler should be an instance of ``Sampler``, '
+                            f'but got {sampler}')
+        if not isinstance(batch_size, int) or batch_size <= 0:
+            raise ValueError('batch_size should be a positive integer value, '
+                             f'but got batch_size={batch_size}')
+        self.sampler = sampler
+        self.dataset = dataset
+        self.train_folder = train_folder
+        self.batch_size = batch_size
+        self.aspect_ratios = aspect_ratios
+        self.drop_last = drop_last
+
+        # buckets for each aspect ratio
+        self.current_available_bucket_keys = list(aspect_ratios.keys())
+        self.bucket = {
+            'image':{ratio: [] for ratio in aspect_ratios}, 
+            'video':{ratio: [] for ratio in aspect_ratios}
+        }
+
+    def __iter__(self):
+        for idx in self.sampler:
+            content_type = self.dataset[idx].get('type', 'image')
+            if content_type == 'image':
+                try:
+                    image_dict = self.dataset[idx]
+
+                    width, height = image_dict.get("width", None), image_dict.get("height", None)
+                    if width is None or height is None:
+                        image_id, name = image_dict['file_path'], image_dict['text']
+                        if self.train_folder is None:
+                            image_dir = image_id
+                        else:
+                            image_dir = os.path.join(self.train_folder, image_id)
+
+                        width, height = get_image_size_without_loading(image_dir)
+
+                        ratio = height / width # self.dataset[idx]
+                    else:
+                        height = int(height)
+                        width = int(width)
+                        ratio = height / width # self.dataset[idx]
+                except Exception as e:
+                    print(e)
+                    continue
+                # find the closest aspect ratio
+                closest_ratio = min(self.aspect_ratios.keys(), key=lambda r: abs(float(r) - ratio))
+                if closest_ratio not in self.current_available_bucket_keys:
+                    continue
+                bucket = self.bucket['image'][closest_ratio]
+                bucket.append(idx)
+                # yield a batch of indices in the same aspect ratio group
+                if len(bucket) == self.batch_size:
+                    yield bucket[:]
+                    del bucket[:]
+            else:
+                try:
+                    video_dict = self.dataset[idx]
+                    width, height = video_dict.get("width", None), video_dict.get("height", None)
+
+                    if width is None or height is None:
+                        video_id, name = video_dict['file_path'], video_dict['text']
+                        if self.train_folder is None:
+                            video_dir = video_id
+                        else:
+                            video_dir = os.path.join(self.train_folder, video_id)
+                        cap = cv2.VideoCapture(video_dir)
+
+                        # 获取视频尺寸
+                        width  = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))   # 浮点数转换为整数
+                        height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))  # 浮点数转换为整数
+                        
+                        ratio = height / width # self.dataset[idx]
+                    else:
+                        height = int(height)
+                        width = int(width)
+                        ratio = height / width # self.dataset[idx]
+                except Exception as e:
+                    print(e)
+                    continue
+                # find the closest aspect ratio
+                closest_ratio = min(self.aspect_ratios.keys(), key=lambda r: abs(float(r) - ratio))
+                if closest_ratio not in self.current_available_bucket_keys:
+                    continue
+                bucket = self.bucket['video'][closest_ratio]
+                bucket.append(idx)
+                # yield a batch of indices in the same aspect ratio group
+                if len(bucket) == self.batch_size:
+                    yield bucket[:]
+                    del bucket[:]

cogvideox/data/dataset_image.py

@@ -0,0 +1,76 @@
+import json
+import os
+import random
+
+import numpy as np
+import torch
+import torchvision.transforms as transforms
+from PIL import Image
+from torch.utils.data.dataset import Dataset
+
+
+class CC15M(Dataset):
+    def __init__(
+            self,
+            json_path, 
+            video_folder=None,
+            resolution=512,
+            enable_bucket=False,
+        ):
+        print(f"loading annotations from {json_path} ...")
+        self.dataset = json.load(open(json_path, 'r'))
+        self.length = len(self.dataset)
+        print(f"data scale: {self.length}")
+        
+        self.enable_bucket = enable_bucket
+        self.video_folder = video_folder
+
+        resolution = tuple(resolution) if not isinstance(resolution, int) else (resolution, resolution)
+        self.pixel_transforms = transforms.Compose([
+            transforms.Resize(resolution[0]),
+            transforms.CenterCrop(resolution),
+            transforms.ToTensor(),
+            transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5], inplace=True),
+        ])
+    
+    def get_batch(self, idx):
+        video_dict = self.dataset[idx]
+        video_id, name = video_dict['file_path'], video_dict['text']
+
+        if self.video_folder is None:
+            video_dir = video_id
+        else:
+            video_dir = os.path.join(self.video_folder, video_id)
+
+        pixel_values = Image.open(video_dir).convert("RGB")
+        return pixel_values, name
+
+    def __len__(self):
+        return self.length
+
+    def __getitem__(self, idx):
+        while True:
+            try:
+                pixel_values, name = self.get_batch(idx)
+                break
+            except Exception as e:
+                print(e)
+                idx = random.randint(0, self.length-1)
+
+        if not self.enable_bucket:
+            pixel_values = self.pixel_transforms(pixel_values)
+        else:
+            pixel_values = np.array(pixel_values)
+
+        sample = dict(pixel_values=pixel_values, text=name)
+        return sample
+
+if __name__ == "__main__":
+    dataset = CC15M(
+        csv_path="/mnt_wg/zhoumo.xjq/CCUtils/cc15m_add_index.json",
+        resolution=512,
+    )
+    
+    dataloader = torch.utils.data.DataLoader(dataset, batch_size=4, num_workers=0,)
+    for idx, batch in enumerate(dataloader):
+        print(batch["pixel_values"].shape, len(batch["text"]))

cogvideox/data/dataset_image_video.py

@@ -0,0 +1,324 @@
+import csv
+import io
+import json
+import math
+import os
+import random
+from threading import Thread
+
+import albumentations
+import cv2
+import gc
+import numpy as np
+import torch
+import torchvision.transforms as transforms
+
+from func_timeout import func_timeout, FunctionTimedOut
+from decord import VideoReader
+from PIL import Image
+from torch.utils.data import BatchSampler, Sampler
+from torch.utils.data.dataset import Dataset
+from contextlib import contextmanager
+
+VIDEO_READER_TIMEOUT = 20
+
+def get_random_mask(shape):
+    f, c, h, w = shape
+    
+    if f != 1:
+        mask_index = np.random.choice([0, 1, 2, 3, 4], p = [0.05, 0.3, 0.3, 0.3, 0.05]) # np.random.randint(0, 5)
+    else:
+        mask_index = np.random.choice([0, 1], p = [0.2, 0.8]) # np.random.randint(0, 2)
+    mask = torch.zeros((f, 1, h, w), dtype=torch.uint8)
+
+    if mask_index == 0:
+        center_x = torch.randint(0, w, (1,)).item()
+        center_y = torch.randint(0, h, (1,)).item()
+        block_size_x = torch.randint(w // 4, w // 4 * 3, (1,)).item()  # 方块的宽度范围
+        block_size_y = torch.randint(h // 4, h // 4 * 3, (1,)).item()  # 方块的高度范围
+
+        start_x = max(center_x - block_size_x // 2, 0)
+        end_x = min(center_x + block_size_x // 2, w)
+        start_y = max(center_y - block_size_y // 2, 0)
+        end_y = min(center_y + block_size_y // 2, h)
+        mask[:, :, start_y:end_y, start_x:end_x] = 1
+    elif mask_index == 1:
+        mask[:, :, :, :] = 1
+    elif mask_index == 2:
+        mask_frame_index = np.random.randint(1, 5)
+        mask[mask_frame_index:, :, :, :] = 1
+    elif mask_index == 3:
+        mask_frame_index = np.random.randint(1, 5)
+        mask[mask_frame_index:-mask_frame_index, :, :, :] = 1
+    elif mask_index == 4:
+        center_x = torch.randint(0, w, (1,)).item()
+        center_y = torch.randint(0, h, (1,)).item()
+        block_size_x = torch.randint(w // 4, w // 4 * 3, (1,)).item()  # 方块的宽度范围
+        block_size_y = torch.randint(h // 4, h // 4 * 3, (1,)).item()  # 方块的高度范围
+
+        start_x = max(center_x - block_size_x // 2, 0)
+        end_x = min(center_x + block_size_x // 2, w)
+        start_y = max(center_y - block_size_y // 2, 0)
+        end_y = min(center_y + block_size_y // 2, h)
+
+        mask_frame_before = np.random.randint(0, f // 2)
+        mask_frame_after = np.random.randint(f // 2, f)
+        mask[mask_frame_before:mask_frame_after, :, start_y:end_y, start_x:end_x] = 1
+    else:
+        raise ValueError(f"The mask_index {mask_index} is not define")
+    return mask
+
+class ImageVideoSampler(BatchSampler):
+    """A sampler wrapper for grouping images with similar aspect ratio into a same batch.
+
+    Args:
+        sampler (Sampler): Base sampler.
+        dataset (Dataset): Dataset providing data information.
+        batch_size (int): Size of mini-batch.
+        drop_last (bool): If ``True``, the sampler will drop the last batch if
+            its size would be less than ``batch_size``.
+        aspect_ratios (dict): The predefined aspect ratios.
+    """
+
+    def __init__(self,
+                 sampler: Sampler,
+                 dataset: Dataset,
+                 batch_size: int,
+                 drop_last: bool = False
+                ) -> None:
+        if not isinstance(sampler, Sampler):
+            raise TypeError('sampler should be an instance of ``Sampler``, '
+                            f'but got {sampler}')
+        if not isinstance(batch_size, int) or batch_size <= 0:
+            raise ValueError('batch_size should be a positive integer value, '
+                             f'but got batch_size={batch_size}')
+        self.sampler = sampler
+        self.dataset = dataset
+        self.batch_size = batch_size
+        self.drop_last = drop_last
+
+        # buckets for each aspect ratio
+        self.bucket = {'image':[], 'video':[]}
+
+    def __iter__(self):
+        for idx in self.sampler:
+            content_type = self.dataset.dataset[idx].get('type', 'image')
+            self.bucket[content_type].append(idx)
+
+            # yield a batch of indices in the same aspect ratio group
+            if len(self.bucket['video']) == self.batch_size:
+                bucket = self.bucket['video']
+                yield bucket[:]
+                del bucket[:]
+            elif len(self.bucket['image']) == self.batch_size:
+                bucket = self.bucket['image']
+                yield bucket[:]
+                del bucket[:]
+
+@contextmanager
+def VideoReader_contextmanager(*args, **kwargs):
+    vr = VideoReader(*args, **kwargs)
+    try:
+        yield vr
+    finally:
+        del vr
+        gc.collect()
+
+def get_video_reader_batch(video_reader, batch_index):
+    frames = video_reader.get_batch(batch_index).asnumpy()
+    return frames
+
+def resize_frame(frame, target_short_side):
+    h, w, _ = frame.shape
+    if h < w:
+        if target_short_side > h:
+            return frame
+        new_h = target_short_side
+        new_w = int(target_short_side * w / h)
+    else:
+        if target_short_side > w:
+            return frame
+        new_w = target_short_side
+        new_h = int(target_short_side * h / w)
+    
+    resized_frame = cv2.resize(frame, (new_w, new_h))
+    return resized_frame
+
+class ImageVideoDataset(Dataset):
+    def __init__(
+            self,
+            ann_path, data_root=None,
+            video_sample_size=512, video_sample_stride=4, video_sample_n_frames=16,
+            image_sample_size=512,
+            video_repeat=0,
+            text_drop_ratio=-1,
+            enable_bucket=False,
+            video_length_drop_start=0.1, 
+            video_length_drop_end=0.9,
+            enable_inpaint=False,
+        ):
+        # Loading annotations from files
+        print(f"loading annotations from {ann_path} ...")
+        if ann_path.endswith('.csv'):
+            with open(ann_path, 'r') as csvfile:
+                dataset = list(csv.DictReader(csvfile))
+        elif ann_path.endswith('.json'):
+            dataset = json.load(open(ann_path))
+    
+        self.data_root = data_root
+
+        # It's used to balance num of images and videos.
+        self.dataset = []
+        for data in dataset:
+            if data.get('type', 'image') != 'video':
+                self.dataset.append(data)
+        if video_repeat > 0:
+            for _ in range(video_repeat):
+                for data in dataset:
+                    if data.get('type', 'image') == 'video':
+                        self.dataset.append(data)
+        del dataset
+
+        self.length = len(self.dataset)
+        print(f"data scale: {self.length}")
+        # TODO: enable bucket training
+        self.enable_bucket = enable_bucket
+        self.text_drop_ratio = text_drop_ratio
+        self.enable_inpaint  = enable_inpaint
+
+        self.video_length_drop_start = video_length_drop_start
+        self.video_length_drop_end = video_length_drop_end
+
+        # Video params
+        self.video_sample_stride    = video_sample_stride
+        self.video_sample_n_frames  = video_sample_n_frames
+        self.video_sample_size = tuple(video_sample_size) if not isinstance(video_sample_size, int) else (video_sample_size, video_sample_size)
+        self.video_transforms = transforms.Compose(
+            [
+                transforms.Resize(min(self.video_sample_size)),
+                transforms.CenterCrop(self.video_sample_size),
+                transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5], inplace=True),
+            ]
+        )
+
+        # Image params
+        self.image_sample_size  = tuple(image_sample_size) if not isinstance(image_sample_size, int) else (image_sample_size, image_sample_size)
+        self.image_transforms   = transforms.Compose([
+            transforms.Resize(min(self.image_sample_size)),
+            transforms.CenterCrop(self.image_sample_size),
+            transforms.ToTensor(),
+            transforms.Normalize([0.5, 0.5, 0.5],[0.5, 0.5, 0.5])
+        ])
+
+        self.larger_side_of_image_and_video = max(min(self.image_sample_size), min(self.video_sample_size))
+
+    def get_batch(self, idx):
+        data_info = self.dataset[idx % len(self.dataset)]
+        
+        if data_info.get('type', 'image')=='video':
+            video_id, text = data_info['file_path'], data_info['text']
+
+            if self.data_root is None:
+                video_dir = video_id
+            else:
+                video_dir = os.path.join(self.data_root, video_id)
+
+            with VideoReader_contextmanager(video_dir, num_threads=2) as video_reader:
+                min_sample_n_frames = min(
+                    self.video_sample_n_frames, 
+                    int(len(video_reader) * (self.video_length_drop_end - self.video_length_drop_start) // self.video_sample_stride)
+                )
+                if min_sample_n_frames == 0:
+                    raise ValueError(f"No Frames in video.")
+
+                video_length = int(self.video_length_drop_end * len(video_reader))
+                clip_length = min(video_length, (min_sample_n_frames - 1) * self.video_sample_stride + 1)
+                start_idx   = random.randint(int(self.video_length_drop_start * video_length), video_length - clip_length) if video_length != clip_length else 0
+                batch_index = np.linspace(start_idx, start_idx + clip_length - 1, min_sample_n_frames, dtype=int)
+
+                try:
+                    sample_args = (video_reader, batch_index)
+                    pixel_values = func_timeout(
+                        VIDEO_READER_TIMEOUT, get_video_reader_batch, args=sample_args
+                    )
+                    resized_frames = []
+                    for i in range(len(pixel_values)):
+                        frame = pixel_values[i]
+                        resized_frame = resize_frame(frame, self.larger_side_of_image_and_video)
+                        resized_frames.append(resized_frame)
+                    pixel_values = np.array(resized_frames)
+                except FunctionTimedOut:
+                    raise ValueError(f"Read {idx} timeout.")
+                except Exception as e:
+                    raise ValueError(f"Failed to extract frames from video. Error is {e}.")
+
+                if not self.enable_bucket:
+                    pixel_values = torch.from_numpy(pixel_values).permute(0, 3, 1, 2).contiguous()
+                    pixel_values = pixel_values / 255.
+                    del video_reader
+                else:
+                    pixel_values = pixel_values
+
+                if not self.enable_bucket:
+                    pixel_values = self.video_transforms(pixel_values)
+                
+                # Random use no text generation
+                if random.random() < self.text_drop_ratio:
+                    text = ''
+            return pixel_values, text, 'video'
+        else:
+            image_path, text = data_info['file_path'], data_info['text']
+            if self.data_root is not None:
+                image_path = os.path.join(self.data_root, image_path)
+            image = Image.open(image_path).convert('RGB')
+            if not self.enable_bucket:
+                image = self.image_transforms(image).unsqueeze(0)
+            else:
+                image = np.expand_dims(np.array(image), 0)
+            if random.random() < self.text_drop_ratio:
+                text = ''
+            return image, text, 'image'
+
+    def __len__(self):
+        return self.length
+
+    def __getitem__(self, idx):
+        data_info = self.dataset[idx % len(self.dataset)]
+        data_type = data_info.get('type', 'image')
+        while True:
+            sample = {}
+            try:
+                data_info_local = self.dataset[idx % len(self.dataset)]
+                data_type_local = data_info_local.get('type', 'image')
+                if data_type_local != data_type:
+                    raise ValueError("data_type_local != data_type")
+
+                pixel_values, name, data_type = self.get_batch(idx)
+                sample["pixel_values"] = pixel_values
+                sample["text"] = name
+                sample["data_type"] = data_type
+                sample["idx"] = idx
+                
+                if len(sample) > 0:
+                    break
+            except Exception as e:
+                print(e, self.dataset[idx % len(self.dataset)])
+                idx = random.randint(0, self.length-1)
+
+        if self.enable_inpaint and not self.enable_bucket:
+            mask = get_random_mask(pixel_values.size())
+            mask_pixel_values = pixel_values * (1 - mask) + torch.ones_like(pixel_values) * -1 * mask
+            sample["mask_pixel_values"] = mask_pixel_values
+            sample["mask"] = mask
+
+            clip_pixel_values = sample["pixel_values"][0].permute(1, 2, 0).contiguous()
+            clip_pixel_values = (clip_pixel_values * 0.5 + 0.5) * 255
+            sample["clip_pixel_values"] = clip_pixel_values
+
+            ref_pixel_values = sample["pixel_values"][0].unsqueeze(0)
+            if (mask == 1).all():
+                ref_pixel_values = torch.ones_like(ref_pixel_values) * -1
+            sample["ref_pixel_values"] = ref_pixel_values
+
+        return sample
+

cogvideox/data/dataset_video.py

@@ -0,0 +1,262 @@
+import csv
+import gc
+import io
+import json
+import math
+import os
+import random
+from contextlib import contextmanager
+from threading import Thread
+
+import albumentations
+import cv2
+import numpy as np
+import torch
+import torchvision.transforms as transforms
+from decord import VideoReader
+from einops import rearrange
+from func_timeout import FunctionTimedOut, func_timeout
+from PIL import Image
+from torch.utils.data import BatchSampler, Sampler
+from torch.utils.data.dataset import Dataset
+
+VIDEO_READER_TIMEOUT = 20
+
+def get_random_mask(shape):
+    f, c, h, w = shape
+    
+    mask_index = np.random.randint(0, 4)
+    mask = torch.zeros((f, 1, h, w), dtype=torch.uint8)
+    if mask_index == 0:
+        mask[1:, :, :, :] = 1
+    elif mask_index == 1:
+        mask_frame_index = 1
+        mask[mask_frame_index:-mask_frame_index, :, :, :] = 1
+    elif mask_index == 2:
+        center_x = torch.randint(0, w, (1,)).item()
+        center_y = torch.randint(0, h, (1,)).item()
+        block_size_x = torch.randint(w // 4, w // 4 * 3, (1,)).item()  # 方块的宽度范围
+        block_size_y = torch.randint(h // 4, h // 4 * 3, (1,)).item()  # 方块的高度范围
+
+        start_x = max(center_x - block_size_x // 2, 0)
+        end_x = min(center_x + block_size_x // 2, w)
+        start_y = max(center_y - block_size_y // 2, 0)
+        end_y = min(center_y + block_size_y // 2, h)
+        mask[:, :, start_y:end_y, start_x:end_x] = 1
+    elif mask_index == 3:
+        center_x = torch.randint(0, w, (1,)).item()
+        center_y = torch.randint(0, h, (1,)).item()
+        block_size_x = torch.randint(w // 4, w // 4 * 3, (1,)).item()  # 方块的宽度范围
+        block_size_y = torch.randint(h // 4, h // 4 * 3, (1,)).item()  # 方块的高度范围
+
+        start_x = max(center_x - block_size_x // 2, 0)
+        end_x = min(center_x + block_size_x // 2, w)
+        start_y = max(center_y - block_size_y // 2, 0)
+        end_y = min(center_y + block_size_y // 2, h)
+
+        mask_frame_before = np.random.randint(0, f // 2)
+        mask_frame_after = np.random.randint(f // 2, f)
+        mask[mask_frame_before:mask_frame_after, :, start_y:end_y, start_x:end_x] = 1
+    else:
+        raise ValueError(f"The mask_index {mask_index} is not define")
+    return mask
+
+
+@contextmanager
+def VideoReader_contextmanager(*args, **kwargs):
+    vr = VideoReader(*args, **kwargs)
+    try:
+        yield vr
+    finally:
+        del vr
+        gc.collect()
+
+
+def get_video_reader_batch(video_reader, batch_index):
+    frames = video_reader.get_batch(batch_index).asnumpy()
+    return frames
+
+
+class WebVid10M(Dataset):
+    def __init__(
+            self,
+            csv_path, video_folder,
+            sample_size=256, sample_stride=4, sample_n_frames=16,
+            enable_bucket=False, enable_inpaint=False, is_image=False,
+        ):
+        print(f"loading annotations from {csv_path} ...")
+        with open(csv_path, 'r') as csvfile:
+            self.dataset = list(csv.DictReader(csvfile))
+        self.length = len(self.dataset)
+        print(f"data scale: {self.length}")
+
+        self.video_folder    = video_folder
+        self.sample_stride   = sample_stride
+        self.sample_n_frames = sample_n_frames
+        self.enable_bucket   = enable_bucket
+        self.enable_inpaint  = enable_inpaint
+        self.is_image        = is_image
+        
+        sample_size = tuple(sample_size) if not isinstance(sample_size, int) else (sample_size, sample_size)
+        self.pixel_transforms = transforms.Compose([
+            transforms.Resize(sample_size[0]),
+            transforms.CenterCrop(sample_size),
+            transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5], inplace=True),
+        ])
+    
+    def get_batch(self, idx):
+        video_dict = self.dataset[idx]
+        videoid, name, page_dir = video_dict['videoid'], video_dict['name'], video_dict['page_dir']
+        
+        video_dir    = os.path.join(self.video_folder, f"{videoid}.mp4")
+        video_reader = VideoReader(video_dir)
+        video_length = len(video_reader)
+        
+        if not self.is_image:
+            clip_length = min(video_length, (self.sample_n_frames - 1) * self.sample_stride + 1)
+            start_idx   = random.randint(0, video_length - clip_length)
+            batch_index = np.linspace(start_idx, start_idx + clip_length - 1, self.sample_n_frames, dtype=int)
+        else:
+            batch_index = [random.randint(0, video_length - 1)]
+
+        if not self.enable_bucket:
+            pixel_values = torch.from_numpy(video_reader.get_batch(batch_index).asnumpy()).permute(0, 3, 1, 2).contiguous()
+            pixel_values = pixel_values / 255.
+            del video_reader
+        else:
+            pixel_values = video_reader.get_batch(batch_index).asnumpy()
+
+        if self.is_image:
+            pixel_values = pixel_values[0]
+        return pixel_values, name
+
+    def __len__(self):
+        return self.length
+
+    def __getitem__(self, idx):
+        while True:
+            try:
+                pixel_values, name = self.get_batch(idx)
+                break
+
+            except Exception as e:
+                print("Error info:", e)
+                idx = random.randint(0, self.length-1)
+
+        if not self.enable_bucket:
+            pixel_values = self.pixel_transforms(pixel_values)
+        if self.enable_inpaint:
+            mask = get_random_mask(pixel_values.size())
+            mask_pixel_values = pixel_values * (1 - mask) + torch.ones_like(pixel_values) * -1 * mask
+            sample = dict(pixel_values=pixel_values, mask_pixel_values=mask_pixel_values, mask=mask, text=name)
+        else:
+            sample = dict(pixel_values=pixel_values, text=name)
+        return sample
+
+
+class VideoDataset(Dataset):
+    def __init__(
+        self,
+        json_path, video_folder=None,
+        sample_size=256, sample_stride=4, sample_n_frames=16,
+        enable_bucket=False, enable_inpaint=False
+    ):
+        print(f"loading annotations from {json_path} ...")
+        self.dataset = json.load(open(json_path, 'r'))
+        self.length = len(self.dataset)
+        print(f"data scale: {self.length}")
+
+        self.video_folder    = video_folder
+        self.sample_stride   = sample_stride
+        self.sample_n_frames = sample_n_frames
+        self.enable_bucket   = enable_bucket
+        self.enable_inpaint  = enable_inpaint
+        
+        sample_size = tuple(sample_size) if not isinstance(sample_size, int) else (sample_size, sample_size)
+        self.pixel_transforms = transforms.Compose(
+            [
+                transforms.Resize(sample_size[0]),
+                transforms.CenterCrop(sample_size),
+                transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5], inplace=True),
+            ]
+        )
+    
+    def get_batch(self, idx):
+        video_dict = self.dataset[idx]
+        video_id, name = video_dict['file_path'], video_dict['text']
+
+        if self.video_folder is None:
+            video_dir = video_id
+        else:
+            video_dir = os.path.join(self.video_folder, video_id)
+
+        with VideoReader_contextmanager(video_dir, num_threads=2) as video_reader:
+            video_length = len(video_reader)
+        
+            clip_length = min(video_length, (self.sample_n_frames - 1) * self.sample_stride + 1)
+            start_idx   = random.randint(0, video_length - clip_length)
+            batch_index = np.linspace(start_idx, start_idx + clip_length - 1, self.sample_n_frames, dtype=int)
+
+            try:
+                sample_args = (video_reader, batch_index)
+                pixel_values = func_timeout(
+                    VIDEO_READER_TIMEOUT, get_video_reader_batch, args=sample_args
+                )
+            except FunctionTimedOut:
+                raise ValueError(f"Read {idx} timeout.")
+            except Exception as e:
+                raise ValueError(f"Failed to extract frames from video. Error is {e}.")
+
+            if not self.enable_bucket:
+                pixel_values = torch.from_numpy(pixel_values).permute(0, 3, 1, 2).contiguous()
+                pixel_values = pixel_values / 255.
+                del video_reader
+            else:
+                pixel_values = pixel_values
+
+            return pixel_values, name
+
+    def __len__(self):
+        return self.length
+
+    def __getitem__(self, idx):
+        while True:
+            try:
+                pixel_values, name = self.get_batch(idx)
+                break
+
+            except Exception as e:
+                print("Error info:", e)
+                idx = random.randint(0, self.length-1)
+
+        if not self.enable_bucket:
+            pixel_values = self.pixel_transforms(pixel_values)
+        if self.enable_inpaint:
+            mask = get_random_mask(pixel_values.size())
+            mask_pixel_values = pixel_values * (1 - mask) + torch.ones_like(pixel_values) * -1 * mask
+            sample = dict(pixel_values=pixel_values, mask_pixel_values=mask_pixel_values, mask=mask, text=name)
+        else:
+            sample = dict(pixel_values=pixel_values, text=name)
+        return sample
+
+
+if __name__ == "__main__":
+    if 1:
+        dataset = VideoDataset(
+            json_path="/home/zhoumo.xjq/disk3/datasets/webvidval/results_2M_val.json",
+            sample_size=256,
+            sample_stride=4, sample_n_frames=16,
+        )
+
+    if 0:
+        dataset = WebVid10M(
+            csv_path="/mnt/petrelfs/guoyuwei/projects/datasets/webvid/results_2M_val.csv",
+            video_folder="/mnt/petrelfs/guoyuwei/projects/datasets/webvid/2M_val",
+            sample_size=256,
+            sample_stride=4, sample_n_frames=16,
+            is_image=False,
+        )
+
+    dataloader = torch.utils.data.DataLoader(dataset, batch_size=4, num_workers=0,)
+    for idx, batch in enumerate(dataloader):
+        print(batch["pixel_values"].shape, len(batch["text"]))

cogvideox/models/autoencoder_magvit.py

@@ -0,0 +1,1296 @@
+# Copyright 2024 The CogVideoX team, Tsinghua University & ZhipuAI and The HuggingFace Team.
+# All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Optional, Tuple, Union
+
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.loaders.single_file_model import FromOriginalModelMixin
+from diffusers.utils import logging
+from diffusers.utils.accelerate_utils import apply_forward_hook
+from diffusers.models.activations import get_activation
+from diffusers.models.downsampling import CogVideoXDownsample3D
+from diffusers.models.modeling_outputs import AutoencoderKLOutput
+from diffusers.models.modeling_utils import ModelMixin
+from diffusers.models.upsampling import CogVideoXUpsample3D
+from diffusers.models.autoencoders.vae import DecoderOutput, DiagonalGaussianDistribution
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+class CogVideoXSafeConv3d(nn.Conv3d):
+    r"""
+    A 3D convolution layer that splits the input tensor into smaller parts to avoid OOM in CogVideoX Model.
+    """
+
+    def forward(self, input: torch.Tensor) -> torch.Tensor:
+        memory_count = torch.prod(torch.tensor(input.shape)).item() * 2 / 1024**3
+
+        # Set to 2GB, suitable for CuDNN
+        if memory_count > 2:
+            kernel_size = self.kernel_size[0]
+            part_num = int(memory_count / 2) + 1
+            input_chunks = torch.chunk(input, part_num, dim=2)
+
+            if kernel_size > 1:
+                input_chunks = [input_chunks[0]] + [
+                    torch.cat((input_chunks[i - 1][:, :, -kernel_size + 1 :], input_chunks[i]), dim=2)
+                    for i in range(1, len(input_chunks))
+                ]
+
+            output_chunks = []
+            for input_chunk in input_chunks:
+                output_chunks.append(super().forward(input_chunk))
+            output = torch.cat(output_chunks, dim=2)
+            return output
+        else:
+            return super().forward(input)
+
+
+class CogVideoXCausalConv3d(nn.Module):
+    r"""A 3D causal convolution layer that pads the input tensor to ensure causality in CogVideoX Model.
+
+    Args:
+        in_channels (`int`): Number of channels in the input tensor.
+        out_channels (`int`): Number of output channels produced by the convolution.
+        kernel_size (`int` or `Tuple[int, int, int]`): Kernel size of the convolutional kernel.
+        stride (`int`, defaults to `1`): Stride of the convolution.
+        dilation (`int`, defaults to `1`): Dilation rate of the convolution.
+        pad_mode (`str`, defaults to `"constant"`): Padding mode.
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        kernel_size: Union[int, Tuple[int, int, int]],
+        stride: int = 1,
+        dilation: int = 1,
+        pad_mode: str = "constant",
+    ):
+        super().__init__()
+
+        if isinstance(kernel_size, int):
+            kernel_size = (kernel_size,) * 3
+
+        time_kernel_size, height_kernel_size, width_kernel_size = kernel_size
+
+        self.pad_mode = pad_mode
+        time_pad = dilation * (time_kernel_size - 1) + (1 - stride)
+        height_pad = height_kernel_size // 2
+        width_pad = width_kernel_size // 2
+
+        self.height_pad = height_pad
+        self.width_pad = width_pad
+        self.time_pad = time_pad
+        self.time_causal_padding = (width_pad, width_pad, height_pad, height_pad, time_pad, 0)
+
+        self.temporal_dim = 2
+        self.time_kernel_size = time_kernel_size
+
+        stride = (stride, 1, 1)
+        dilation = (dilation, 1, 1)
+        self.conv = CogVideoXSafeConv3d(
+            in_channels=in_channels,
+            out_channels=out_channels,
+            kernel_size=kernel_size,
+            stride=stride,
+            dilation=dilation,
+        )
+
+        self.conv_cache = None
+
+    def fake_context_parallel_forward(self, inputs: torch.Tensor) -> torch.Tensor:
+        kernel_size = self.time_kernel_size
+        if kernel_size > 1:
+            cached_inputs = (
+                [self.conv_cache] if self.conv_cache is not None else [inputs[:, :, :1]] * (kernel_size - 1)
+            )
+            inputs = torch.cat(cached_inputs + [inputs], dim=2)
+        return inputs
+
+    def _clear_fake_context_parallel_cache(self):
+        del self.conv_cache
+        self.conv_cache = None
+
+    def forward(self, inputs: torch.Tensor) -> torch.Tensor:
+        inputs = self.fake_context_parallel_forward(inputs)
+
+        self._clear_fake_context_parallel_cache()
+        # Note: we could move these to the cpu for a lower maximum memory usage but its only a few
+        # hundred megabytes and so let's not do it for now
+        self.conv_cache = inputs[:, :, -self.time_kernel_size + 1 :].clone()
+
+        padding_2d = (self.width_pad, self.width_pad, self.height_pad, self.height_pad)
+        inputs = F.pad(inputs, padding_2d, mode="constant", value=0)
+
+        output = self.conv(inputs)
+        return output
+
+
+class CogVideoXSpatialNorm3D(nn.Module):
+    r"""
+    Spatially conditioned normalization as defined in https://arxiv.org/abs/2209.09002. This implementation is specific
+    to 3D-video like data.
+
+    CogVideoXSafeConv3d is used instead of nn.Conv3d to avoid OOM in CogVideoX Model.
+
+    Args:
+        f_channels (`int`):
+            The number of channels for input to group normalization layer, and output of the spatial norm layer.
+        zq_channels (`int`):
+            The number of channels for the quantized vector as described in the paper.
+        groups (`int`):
+            Number of groups to separate the channels into for group normalization.
+    """
+
+    def __init__(
+        self,
+        f_channels: int,
+        zq_channels: int,
+        groups: int = 32,
+    ):
+        super().__init__()
+        self.norm_layer = nn.GroupNorm(num_channels=f_channels, num_groups=groups, eps=1e-6, affine=True)
+        self.conv_y = CogVideoXCausalConv3d(zq_channels, f_channels, kernel_size=1, stride=1)
+        self.conv_b = CogVideoXCausalConv3d(zq_channels, f_channels, kernel_size=1, stride=1)
+
+    def forward(self, f: torch.Tensor, zq: torch.Tensor) -> torch.Tensor:
+        if f.shape[2] > 1 and f.shape[2] % 2 == 1:
+            f_first, f_rest = f[:, :, :1], f[:, :, 1:]
+            f_first_size, f_rest_size = f_first.shape[-3:], f_rest.shape[-3:]
+            z_first, z_rest = zq[:, :, :1], zq[:, :, 1:]
+            z_first = F.interpolate(z_first, size=f_first_size)
+            z_rest = F.interpolate(z_rest, size=f_rest_size)
+            zq = torch.cat([z_first, z_rest], dim=2)
+        else:
+            zq = F.interpolate(zq, size=f.shape[-3:])
+
+        norm_f = self.norm_layer(f)
+        new_f = norm_f * self.conv_y(zq) + self.conv_b(zq)
+        return new_f
+
+
+class CogVideoXResnetBlock3D(nn.Module):
+    r"""
+    A 3D ResNet block used in the CogVideoX model.
+
+    Args:
+        in_channels (`int`):
+            Number of input channels.
+        out_channels (`int`, *optional*):
+            Number of output channels. If None, defaults to `in_channels`.
+        dropout (`float`, defaults to `0.0`):
+            Dropout rate.
+        temb_channels (`int`, defaults to `512`):
+            Number of time embedding channels.
+        groups (`int`, defaults to `32`):
+            Number of groups to separate the channels into for group normalization.
+        eps (`float`, defaults to `1e-6`):
+            Epsilon value for normalization layers.
+        non_linearity (`str`, defaults to `"swish"`):
+            Activation function to use.
+        conv_shortcut (bool, defaults to `False`):
+            Whether or not to use a convolution shortcut.
+        spatial_norm_dim (`int`, *optional*):
+            The dimension to use for spatial norm if it is to be used instead of group norm.
+        pad_mode (str, defaults to `"first"`):
+            Padding mode.
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: Optional[int] = None,
+        dropout: float = 0.0,
+        temb_channels: int = 512,
+        groups: int = 32,
+        eps: float = 1e-6,
+        non_linearity: str = "swish",
+        conv_shortcut: bool = False,
+        spatial_norm_dim: Optional[int] = None,
+        pad_mode: str = "first",
+    ):
+        super().__init__()
+
+        out_channels = out_channels or in_channels
+
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.nonlinearity = get_activation(non_linearity)
+        self.use_conv_shortcut = conv_shortcut
+
+        if spatial_norm_dim is None:
+            self.norm1 = nn.GroupNorm(num_channels=in_channels, num_groups=groups, eps=eps)
+            self.norm2 = nn.GroupNorm(num_channels=out_channels, num_groups=groups, eps=eps)
+        else:
+            self.norm1 = CogVideoXSpatialNorm3D(
+                f_channels=in_channels,
+                zq_channels=spatial_norm_dim,
+                groups=groups,
+            )
+            self.norm2 = CogVideoXSpatialNorm3D(
+                f_channels=out_channels,
+                zq_channels=spatial_norm_dim,
+                groups=groups,
+            )
+
+        self.conv1 = CogVideoXCausalConv3d(
+            in_channels=in_channels, out_channels=out_channels, kernel_size=3, pad_mode=pad_mode
+        )
+
+        if temb_channels > 0:
+            self.temb_proj = nn.Linear(in_features=temb_channels, out_features=out_channels)
+
+        self.dropout = nn.Dropout(dropout)
+        self.conv2 = CogVideoXCausalConv3d(
+            in_channels=out_channels, out_channels=out_channels, kernel_size=3, pad_mode=pad_mode
+        )
+
+        if self.in_channels != self.out_channels:
+            if self.use_conv_shortcut:
+                self.conv_shortcut = CogVideoXCausalConv3d(
+                    in_channels=in_channels, out_channels=out_channels, kernel_size=3, pad_mode=pad_mode
+                )
+            else:
+                self.conv_shortcut = CogVideoXSafeConv3d(
+                    in_channels=in_channels, out_channels=out_channels, kernel_size=1, stride=1, padding=0
+                )
+
+    def forward(
+        self,
+        inputs: torch.Tensor,
+        temb: Optional[torch.Tensor] = None,
+        zq: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        hidden_states = inputs
+
+        if zq is not None:
+            hidden_states = self.norm1(hidden_states, zq)
+        else:
+            hidden_states = self.norm1(hidden_states)
+
+        hidden_states = self.nonlinearity(hidden_states)
+        hidden_states = self.conv1(hidden_states)
+
+        if temb is not None:
+            hidden_states = hidden_states + self.temb_proj(self.nonlinearity(temb))[:, :, None, None, None]
+
+        if zq is not None:
+            hidden_states = self.norm2(hidden_states, zq)
+        else:
+            hidden_states = self.norm2(hidden_states)
+
+        hidden_states = self.nonlinearity(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.conv2(hidden_states)
+
+        if self.in_channels != self.out_channels:
+            inputs = self.conv_shortcut(inputs)
+
+        hidden_states = hidden_states + inputs
+        return hidden_states
+
+
+class CogVideoXDownBlock3D(nn.Module):
+    r"""
+    A downsampling block used in the CogVideoX model.
+
+    Args:
+        in_channels (`int`):
+            Number of input channels.
+        out_channels (`int`, *optional*):
+            Number of output channels. If None, defaults to `in_channels`.
+        temb_channels (`int`, defaults to `512`):
+            Number of time embedding channels.
+        num_layers (`int`, defaults to `1`):
+            Number of resnet layers.
+        dropout (`float`, defaults to `0.0`):
+            Dropout rate.
+        resnet_eps (`float`, defaults to `1e-6`):
+            Epsilon value for normalization layers.
+        resnet_act_fn (`str`, defaults to `"swish"`):
+            Activation function to use.
+        resnet_groups (`int`, defaults to `32`):
+            Number of groups to separate the channels into for group normalization.
+        add_downsample (`bool`, defaults to `True`):
+            Whether or not to use a downsampling layer. If not used, output dimension would be same as input dimension.
+        compress_time (`bool`, defaults to `False`):
+            Whether or not to downsample across temporal dimension.
+        pad_mode (str, defaults to `"first"`):
+            Padding mode.
+    """
+
+    _supports_gradient_checkpointing = True
+
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        add_downsample: bool = True,
+        downsample_padding: int = 0,
+        compress_time: bool = False,
+        pad_mode: str = "first",
+    ):
+        super().__init__()
+
+        resnets = []
+        for i in range(num_layers):
+            in_channel = in_channels if i == 0 else out_channels
+            resnets.append(
+                CogVideoXResnetBlock3D(
+                    in_channels=in_channel,
+                    out_channels=out_channels,
+                    dropout=dropout,
+                    temb_channels=temb_channels,
+                    groups=resnet_groups,
+                    eps=resnet_eps,
+                    non_linearity=resnet_act_fn,
+                    pad_mode=pad_mode,
+                )
+            )
+
+        self.resnets = nn.ModuleList(resnets)
+        self.downsamplers = None
+
+        if add_downsample:
+            self.downsamplers = nn.ModuleList(
+                [
+                    CogVideoXDownsample3D(
+                        out_channels, out_channels, padding=downsample_padding, compress_time=compress_time
+                    )
+                ]
+            )
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        temb: Optional[torch.Tensor] = None,
+        zq: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        for resnet in self.resnets:
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module):
+                    def create_forward(*inputs):
+                        return module(*inputs)
+
+                    return create_forward
+
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(resnet), hidden_states, temb, zq
+                )
+            else:
+                hidden_states = resnet(hidden_states, temb, zq)
+
+        if self.downsamplers is not None:
+            for downsampler in self.downsamplers:
+                hidden_states = downsampler(hidden_states)
+
+        return hidden_states
+
+
+class CogVideoXMidBlock3D(nn.Module):
+    r"""
+    A middle block used in the CogVideoX model.
+
+    Args:
+        in_channels (`int`):
+            Number of input channels.
+        temb_channels (`int`, defaults to `512`):
+            Number of time embedding channels.
+        dropout (`float`, defaults to `0.0`):
+            Dropout rate.
+        num_layers (`int`, defaults to `1`):
+            Number of resnet layers.
+        resnet_eps (`float`, defaults to `1e-6`):
+            Epsilon value for normalization layers.
+        resnet_act_fn (`str`, defaults to `"swish"`):
+            Activation function to use.
+        resnet_groups (`int`, defaults to `32`):
+            Number of groups to separate the channels into for group normalization.
+        spatial_norm_dim (`int`, *optional*):
+            The dimension to use for spatial norm if it is to be used instead of group norm.
+        pad_mode (str, defaults to `"first"`):
+            Padding mode.
+    """
+
+    _supports_gradient_checkpointing = True
+
+    def __init__(
+        self,
+        in_channels: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        spatial_norm_dim: Optional[int] = None,
+        pad_mode: str = "first",
+    ):
+        super().__init__()
+
+        resnets = []
+        for _ in range(num_layers):
+            resnets.append(
+                CogVideoXResnetBlock3D(
+                    in_channels=in_channels,
+                    out_channels=in_channels,
+                    dropout=dropout,
+                    temb_channels=temb_channels,
+                    groups=resnet_groups,
+                    eps=resnet_eps,
+                    spatial_norm_dim=spatial_norm_dim,
+                    non_linearity=resnet_act_fn,
+                    pad_mode=pad_mode,
+                )
+            )
+        self.resnets = nn.ModuleList(resnets)
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        temb: Optional[torch.Tensor] = None,
+        zq: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        for resnet in self.resnets:
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module):
+                    def create_forward(*inputs):
+                        return module(*inputs)
+
+                    return create_forward
+
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(resnet), hidden_states, temb, zq
+                )
+            else:
+                hidden_states = resnet(hidden_states, temb, zq)
+
+        return hidden_states
+
+
+class CogVideoXUpBlock3D(nn.Module):
+    r"""
+    An upsampling block used in the CogVideoX model.
+
+    Args:
+        in_channels (`int`):
+            Number of input channels.
+        out_channels (`int`, *optional*):
+            Number of output channels. If None, defaults to `in_channels`.
+        temb_channels (`int`, defaults to `512`):
+            Number of time embedding channels.
+        dropout (`float`, defaults to `0.0`):
+            Dropout rate.
+        num_layers (`int`, defaults to `1`):
+            Number of resnet layers.
+        resnet_eps (`float`, defaults to `1e-6`):
+            Epsilon value for normalization layers.
+        resnet_act_fn (`str`, defaults to `"swish"`):
+            Activation function to use.
+        resnet_groups (`int`, defaults to `32`):
+            Number of groups to separate the channels into for group normalization.
+        spatial_norm_dim (`int`, defaults to `16`):
+            The dimension to use for spatial norm if it is to be used instead of group norm.
+        add_upsample (`bool`, defaults to `True`):
+            Whether or not to use a upsampling layer. If not used, output dimension would be same as input dimension.
+        compress_time (`bool`, defaults to `False`):
+            Whether or not to downsample across temporal dimension.
+        pad_mode (str, defaults to `"first"`):
+            Padding mode.
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        spatial_norm_dim: int = 16,
+        add_upsample: bool = True,
+        upsample_padding: int = 1,
+        compress_time: bool = False,
+        pad_mode: str = "first",
+    ):
+        super().__init__()
+
+        resnets = []
+        for i in range(num_layers):
+            in_channel = in_channels if i == 0 else out_channels
+            resnets.append(
+                CogVideoXResnetBlock3D(
+                    in_channels=in_channel,
+                    out_channels=out_channels,
+                    dropout=dropout,
+                    temb_channels=temb_channels,
+                    groups=resnet_groups,
+                    eps=resnet_eps,
+                    non_linearity=resnet_act_fn,
+                    spatial_norm_dim=spatial_norm_dim,
+                    pad_mode=pad_mode,
+                )
+            )
+
+        self.resnets = nn.ModuleList(resnets)
+        self.upsamplers = None
+
+        if add_upsample:
+            self.upsamplers = nn.ModuleList(
+                [
+                    CogVideoXUpsample3D(
+                        out_channels, out_channels, padding=upsample_padding, compress_time=compress_time
+                    )
+                ]
+            )
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        temb: Optional[torch.Tensor] = None,
+        zq: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        r"""Forward method of the `CogVideoXUpBlock3D` class."""
+        for resnet in self.resnets:
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module):
+                    def create_forward(*inputs):
+                        return module(*inputs)
+
+                    return create_forward
+
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(resnet), hidden_states, temb, zq
+                )
+            else:
+                hidden_states = resnet(hidden_states, temb, zq)
+
+        if self.upsamplers is not None:
+            for upsampler in self.upsamplers:
+                hidden_states = upsampler(hidden_states)
+
+        return hidden_states
+
+
+class CogVideoXEncoder3D(nn.Module):
+    r"""
+    The `CogVideoXEncoder3D` layer of a variational autoencoder that encodes its input into a latent representation.
+
+    Args:
+        in_channels (`int`, *optional*, defaults to 3):
+            The number of input channels.
+        out_channels (`int`, *optional*, defaults to 3):
+            The number of output channels.
+        down_block_types (`Tuple[str, ...]`, *optional*, defaults to `("DownEncoderBlock2D",)`):
+            The types of down blocks to use. See `~diffusers.models.unet_2d_blocks.get_down_block` for available
+            options.
+        block_out_channels (`Tuple[int, ...]`, *optional*, defaults to `(64,)`):
+            The number of output channels for each block.
+        act_fn (`str`, *optional*, defaults to `"silu"`):
+            The activation function to use. See `~diffusers.models.activations.get_activation` for available options.
+        layers_per_block (`int`, *optional*, defaults to 2):
+            The number of layers per block.
+        norm_num_groups (`int`, *optional*, defaults to 32):
+            The number of groups for normalization.
+    """
+
+    _supports_gradient_checkpointing = True
+
+    def __init__(
+        self,
+        in_channels: int = 3,
+        out_channels: int = 16,
+        down_block_types: Tuple[str, ...] = (
+            "CogVideoXDownBlock3D",
+            "CogVideoXDownBlock3D",
+            "CogVideoXDownBlock3D",
+            "CogVideoXDownBlock3D",
+        ),
+        block_out_channels: Tuple[int, ...] = (128, 256, 256, 512),
+        layers_per_block: int = 3,
+        act_fn: str = "silu",
+        norm_eps: float = 1e-6,
+        norm_num_groups: int = 32,
+        dropout: float = 0.0,
+        pad_mode: str = "first",
+        temporal_compression_ratio: float = 4,
+    ):
+        super().__init__()
+
+        # log2 of temporal_compress_times
+        temporal_compress_level = int(np.log2(temporal_compression_ratio))
+
+        self.conv_in = CogVideoXCausalConv3d(in_channels, block_out_channels[0], kernel_size=3, pad_mode=pad_mode)
+        self.down_blocks = nn.ModuleList([])
+
+        # down blocks
+        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
+            compress_time = i < temporal_compress_level
+
+            if down_block_type == "CogVideoXDownBlock3D":
+                down_block = CogVideoXDownBlock3D(
+                    in_channels=input_channel,
+                    out_channels=output_channel,
+                    temb_channels=0,
+                    dropout=dropout,
+                    num_layers=layers_per_block,
+                    resnet_eps=norm_eps,
+                    resnet_act_fn=act_fn,
+                    resnet_groups=norm_num_groups,
+                    add_downsample=not is_final_block,
+                    compress_time=compress_time,
+                )
+            else:
+                raise ValueError("Invalid `down_block_type` encountered. Must be `CogVideoXDownBlock3D`")
+
+            self.down_blocks.append(down_block)
+
+        # mid block
+        self.mid_block = CogVideoXMidBlock3D(
+            in_channels=block_out_channels[-1],
+            temb_channels=0,
+            dropout=dropout,
+            num_layers=2,
+            resnet_eps=norm_eps,
+            resnet_act_fn=act_fn,
+            resnet_groups=norm_num_groups,
+            pad_mode=pad_mode,
+        )
+
+        self.norm_out = nn.GroupNorm(norm_num_groups, block_out_channels[-1], eps=1e-6)
+        self.conv_act = nn.SiLU()
+        self.conv_out = CogVideoXCausalConv3d(
+            block_out_channels[-1], 2 * out_channels, kernel_size=3, pad_mode=pad_mode
+        )
+
+        self.gradient_checkpointing = False
+
+    def forward(self, sample: torch.Tensor, temb: Optional[torch.Tensor] = None) -> torch.Tensor:
+        r"""The forward method of the `CogVideoXEncoder3D` class."""
+        hidden_states = self.conv_in(sample)
+
+        if self.training and self.gradient_checkpointing:
+
+            def create_custom_forward(module):
+                def custom_forward(*inputs):
+                    return module(*inputs)
+
+                return custom_forward
+
+            # 1. Down
+            for down_block in self.down_blocks:
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(down_block), hidden_states, temb, None
+                )
+
+            # 2. Mid
+            hidden_states = torch.utils.checkpoint.checkpoint(
+                create_custom_forward(self.mid_block), hidden_states, temb, None
+            )
+        else:
+            # 1. Down
+            for down_block in self.down_blocks:
+                hidden_states = down_block(hidden_states, temb, None)
+
+            # 2. Mid
+            hidden_states = self.mid_block(hidden_states, temb, None)
+
+        # 3. Post-process
+        hidden_states = self.norm_out(hidden_states)
+        hidden_states = self.conv_act(hidden_states)
+        hidden_states = self.conv_out(hidden_states)
+        return hidden_states
+
+
+class CogVideoXDecoder3D(nn.Module):
+    r"""
+    The `CogVideoXDecoder3D` layer of a variational autoencoder that decodes its latent representation into an output
+    sample.
+
+    Args:
+        in_channels (`int`, *optional*, defaults to 3):
+            The number of input channels.
+        out_channels (`int`, *optional*, defaults to 3):
+            The number of output channels.
+        up_block_types (`Tuple[str, ...]`, *optional*, defaults to `("UpDecoderBlock2D",)`):
+            The types of up blocks to use. See `~diffusers.models.unet_2d_blocks.get_up_block` for available options.
+        block_out_channels (`Tuple[int, ...]`, *optional*, defaults to `(64,)`):
+            The number of output channels for each block.
+        act_fn (`str`, *optional*, defaults to `"silu"`):
+            The activation function to use. See `~diffusers.models.activations.get_activation` for available options.
+        layers_per_block (`int`, *optional*, defaults to 2):
+            The number of layers per block.
+        norm_num_groups (`int`, *optional*, defaults to 32):
+            The number of groups for normalization.
+    """
+
+    _supports_gradient_checkpointing = True
+
+    def __init__(
+        self,
+        in_channels: int = 16,
+        out_channels: int = 3,
+        up_block_types: Tuple[str, ...] = (
+            "CogVideoXUpBlock3D",
+            "CogVideoXUpBlock3D",
+            "CogVideoXUpBlock3D",
+            "CogVideoXUpBlock3D",
+        ),
+        block_out_channels: Tuple[int, ...] = (128, 256, 256, 512),
+        layers_per_block: int = 3,
+        act_fn: str = "silu",
+        norm_eps: float = 1e-6,
+        norm_num_groups: int = 32,
+        dropout: float = 0.0,
+        pad_mode: str = "first",
+        temporal_compression_ratio: float = 4,
+    ):
+        super().__init__()
+
+        reversed_block_out_channels = list(reversed(block_out_channels))
+
+        self.conv_in = CogVideoXCausalConv3d(
+            in_channels, reversed_block_out_channels[0], kernel_size=3, pad_mode=pad_mode
+        )
+
+        # mid block
+        self.mid_block = CogVideoXMidBlock3D(
+            in_channels=reversed_block_out_channels[0],
+            temb_channels=0,
+            num_layers=2,
+            resnet_eps=norm_eps,
+            resnet_act_fn=act_fn,
+            resnet_groups=norm_num_groups,
+            spatial_norm_dim=in_channels,
+            pad_mode=pad_mode,
+        )
+
+        # up blocks
+        self.up_blocks = nn.ModuleList([])
+
+        output_channel = reversed_block_out_channels[0]
+        temporal_compress_level = int(np.log2(temporal_compression_ratio))
+
+        for i, up_block_type in enumerate(up_block_types):
+            prev_output_channel = output_channel
+            output_channel = reversed_block_out_channels[i]
+            is_final_block = i == len(block_out_channels) - 1
+            compress_time = i < temporal_compress_level
+
+            if up_block_type == "CogVideoXUpBlock3D":
+                up_block = CogVideoXUpBlock3D(
+                    in_channels=prev_output_channel,
+                    out_channels=output_channel,
+                    temb_channels=0,
+                    dropout=dropout,
+                    num_layers=layers_per_block + 1,
+                    resnet_eps=norm_eps,
+                    resnet_act_fn=act_fn,
+                    resnet_groups=norm_num_groups,
+                    spatial_norm_dim=in_channels,
+                    add_upsample=not is_final_block,
+                    compress_time=compress_time,
+                    pad_mode=pad_mode,
+                )
+                prev_output_channel = output_channel
+            else:
+                raise ValueError("Invalid `up_block_type` encountered. Must be `CogVideoXUpBlock3D`")
+
+            self.up_blocks.append(up_block)
+
+        self.norm_out = CogVideoXSpatialNorm3D(reversed_block_out_channels[-1], in_channels, groups=norm_num_groups)
+        self.conv_act = nn.SiLU()
+        self.conv_out = CogVideoXCausalConv3d(
+            reversed_block_out_channels[-1], out_channels, kernel_size=3, pad_mode=pad_mode
+        )
+
+        self.gradient_checkpointing = False
+
+    def forward(self, sample: torch.Tensor, temb: Optional[torch.Tensor] = None) -> torch.Tensor:
+        r"""The forward method of the `CogVideoXDecoder3D` class."""
+        hidden_states = self.conv_in(sample)
+
+        if self.training and self.gradient_checkpointing:
+
+            def create_custom_forward(module):
+                def custom_forward(*inputs):
+                    return module(*inputs)
+
+                return custom_forward
+
+            # 1. Mid
+            hidden_states = torch.utils.checkpoint.checkpoint(
+                create_custom_forward(self.mid_block), hidden_states, temb, sample
+            )
+
+            # 2. Up
+            for up_block in self.up_blocks:
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(up_block), hidden_states, temb, sample
+                )
+        else:
+            # 1. Mid
+            hidden_states = self.mid_block(hidden_states, temb, sample)
+
+            # 2. Up
+            for up_block in self.up_blocks:
+                hidden_states = up_block(hidden_states, temb, sample)
+
+        # 3. Post-process
+        hidden_states = self.norm_out(hidden_states, sample)
+        hidden_states = self.conv_act(hidden_states)
+        hidden_states = self.conv_out(hidden_states)
+        return hidden_states
+
+
+class AutoencoderKLCogVideoX(ModelMixin, ConfigMixin, FromOriginalModelMixin):
+    r"""
+    A VAE model with KL loss for encoding images into latents and decoding latent representations into images. Used in
+    [CogVideoX](https://github.com/THUDM/CogVideo).
+
+    This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented
+    for all models (such as downloading or saving).
+
+    Parameters:
+        in_channels (int, *optional*, defaults to 3): Number of channels in the input image.
+        out_channels (int,  *optional*, defaults to 3): Number of channels in the output.
+        down_block_types (`Tuple[str]`, *optional*, defaults to `("DownEncoderBlock2D",)`):
+            Tuple of downsample block types.
+        up_block_types (`Tuple[str]`, *optional*, defaults to `("UpDecoderBlock2D",)`):
+            Tuple of upsample block types.
+        block_out_channels (`Tuple[int]`, *optional*, defaults to `(64,)`):
+            Tuple of block output channels.
+        act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use.
+        sample_size (`int`, *optional*, defaults to `32`): Sample input size.
+        scaling_factor (`float`, *optional*, defaults to `1.15258426`):
+            The component-wise standard deviation of the trained latent space computed using the first batch of the
+            training set. This is used to scale the latent space to have unit variance when training the diffusion
+            model. The latents are scaled with the formula `z = z * scaling_factor` before being passed to the
+            diffusion model. When decoding, the latents are scaled back to the original scale with the formula: `z = 1
+            / scaling_factor * z`. For more details, refer to sections 4.3.2 and D.1 of the [High-Resolution Image
+            Synthesis with Latent Diffusion Models](https://arxiv.org/abs/2112.10752) paper.
+        force_upcast (`bool`, *optional*, default to `True`):
+            If enabled it will force the VAE to run in float32 for high image resolution pipelines, such as SD-XL. VAE
+            can be fine-tuned / trained to a lower range without loosing too much precision in which case
+            `force_upcast` can be set to `False` - see: https://huggingface.co/madebyollin/sdxl-vae-fp16-fix
+    """
+
+    _supports_gradient_checkpointing = True
+    _no_split_modules = ["CogVideoXResnetBlock3D"]
+
+    @register_to_config
+    def __init__(
+        self,
+        in_channels: int = 3,
+        out_channels: int = 3,
+        down_block_types: Tuple[str] = (
+            "CogVideoXDownBlock3D",
+            "CogVideoXDownBlock3D",
+            "CogVideoXDownBlock3D",
+            "CogVideoXDownBlock3D",
+        ),
+        up_block_types: Tuple[str] = (
+            "CogVideoXUpBlock3D",
+            "CogVideoXUpBlock3D",
+            "CogVideoXUpBlock3D",
+            "CogVideoXUpBlock3D",
+        ),
+        block_out_channels: Tuple[int] = (128, 256, 256, 512),
+        latent_channels: int = 16,
+        layers_per_block: int = 3,
+        act_fn: str = "silu",
+        norm_eps: float = 1e-6,
+        norm_num_groups: int = 32,
+        temporal_compression_ratio: float = 4,
+        sample_height: int = 480,
+        sample_width: int = 720,
+        scaling_factor: float = 1.15258426,
+        shift_factor: Optional[float] = None,
+        latents_mean: Optional[Tuple[float]] = None,
+        latents_std: Optional[Tuple[float]] = None,
+        force_upcast: float = True,
+        use_quant_conv: bool = False,
+        use_post_quant_conv: bool = False,
+    ):
+        super().__init__()
+
+        self.encoder = CogVideoXEncoder3D(
+            in_channels=in_channels,
+            out_channels=latent_channels,
+            down_block_types=down_block_types,
+            block_out_channels=block_out_channels,
+            layers_per_block=layers_per_block,
+            act_fn=act_fn,
+            norm_eps=norm_eps,
+            norm_num_groups=norm_num_groups,
+            temporal_compression_ratio=temporal_compression_ratio,
+        )
+        self.decoder = CogVideoXDecoder3D(
+            in_channels=latent_channels,
+            out_channels=out_channels,
+            up_block_types=up_block_types,
+            block_out_channels=block_out_channels,
+            layers_per_block=layers_per_block,
+            act_fn=act_fn,
+            norm_eps=norm_eps,
+            norm_num_groups=norm_num_groups,
+            temporal_compression_ratio=temporal_compression_ratio,
+        )
+        self.quant_conv = CogVideoXSafeConv3d(2 * out_channels, 2 * out_channels, 1) if use_quant_conv else None
+        self.post_quant_conv = CogVideoXSafeConv3d(out_channels, out_channels, 1) if use_post_quant_conv else None
+
+        self.use_slicing = False
+        self.use_tiling = False
+
+        # Can be increased to decode more latent frames at once, but comes at a reasonable memory cost and it is not
+        # recommended because the temporal parts of the VAE, here, are tricky to understand.
+        # If you decode X latent frames together, the number of output frames is:
+        #     (X + (2 conv cache) + (2 time upscale_1) + (4 time upscale_2) - (2 causal conv downscale)) => X + 6 frames
+        #
+        # Example with num_latent_frames_batch_size = 2:
+        #     - 12 latent frames: (0, 1), (2, 3), (4, 5), (6, 7), (8, 9), (10, 11) are processed together
+        #         => (12 // 2 frame slices) * ((2 num_latent_frames_batch_size) + (2 conv cache) + (2 time upscale_1) + (4 time upscale_2) - (2 causal conv downscale))
+        #         => 6 * 8 = 48 frames
+        #     - 13 latent frames: (0, 1, 2) (special case), (3, 4), (5, 6), (7, 8), (9, 10), (11, 12) are processed together
+        #         => (1 frame slice) * ((3 num_latent_frames_batch_size) + (2 conv cache) + (2 time upscale_1) + (4 time upscale_2) - (2 causal conv downscale)) +
+        #            ((13 - 3) // 2) * ((2 num_latent_frames_batch_size) + (2 conv cache) + (2 time upscale_1) + (4 time upscale_2) - (2 causal conv downscale))
+        #         => 1 * 9 + 5 * 8 = 49 frames
+        # It has been implemented this way so as to not have "magic values" in the code base that would be hard to explain. Note that
+        # setting it to anything other than 2 would give poor results because the VAE hasn't been trained to be adaptive with different
+        # number of temporal frames.
+        self.num_latent_frames_batch_size = 2
+
+        # We make the minimum height and width of sample for tiling half that of the generally supported
+        self.tile_sample_min_height = sample_height // 2
+        self.tile_sample_min_width = sample_width // 2
+        self.tile_latent_min_height = int(
+            self.tile_sample_min_height / (2 ** (len(self.config.block_out_channels) - 1))
+        )
+        self.tile_latent_min_width = int(self.tile_sample_min_width / (2 ** (len(self.config.block_out_channels) - 1)))
+
+        # These are experimental overlap factors that were chosen based on experimentation and seem to work best for
+        # 720x480 (WxH) resolution. The above resolution is the strongly recommended generation resolution in CogVideoX
+        # and so the tiling implementation has only been tested on those specific resolutions.
+        self.tile_overlap_factor_height = 1 / 6
+        self.tile_overlap_factor_width = 1 / 5
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if isinstance(module, (CogVideoXEncoder3D, CogVideoXDecoder3D)):
+            module.gradient_checkpointing = value
+
+    def _clear_fake_context_parallel_cache(self):
+        for name, module in self.named_modules():
+            if isinstance(module, CogVideoXCausalConv3d):
+                logger.debug(f"Clearing fake Context Parallel cache for layer: {name}")
+                module._clear_fake_context_parallel_cache()
+
+    def enable_tiling(
+        self,
+        tile_sample_min_height: Optional[int] = None,
+        tile_sample_min_width: Optional[int] = None,
+        tile_overlap_factor_height: Optional[float] = None,
+        tile_overlap_factor_width: Optional[float] = None,
+    ) -> None:
+        r"""
+        Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
+        compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
+        processing larger images.
+
+        Args:
+            tile_sample_min_height (`int`, *optional*):
+                The minimum height required for a sample to be separated into tiles across the height dimension.
+            tile_sample_min_width (`int`, *optional*):
+                The minimum width required for a sample to be separated into tiles across the width dimension.
+            tile_overlap_factor_height (`int`, *optional*):
+                The minimum amount of overlap between two consecutive vertical tiles. This is to ensure that there are
+                no tiling artifacts produced across the height dimension. Must be between 0 and 1. Setting a higher
+                value might cause more tiles to be processed leading to slow down of the decoding process.
+            tile_overlap_factor_width (`int`, *optional*):
+                The minimum amount of overlap between two consecutive horizontal tiles. This is to ensure that there
+                are no tiling artifacts produced across the width dimension. Must be between 0 and 1. Setting a higher
+                value might cause more tiles to be processed leading to slow down of the decoding process.
+        """
+        self.use_tiling = True
+        self.tile_sample_min_height = tile_sample_min_height or self.tile_sample_min_height
+        self.tile_sample_min_width = tile_sample_min_width or self.tile_sample_min_width
+        self.tile_latent_min_height = int(
+            self.tile_sample_min_height / (2 ** (len(self.config.block_out_channels) - 1))
+        )
+        self.tile_latent_min_width = int(self.tile_sample_min_width / (2 ** (len(self.config.block_out_channels) - 1)))
+        self.tile_overlap_factor_height = tile_overlap_factor_height or self.tile_overlap_factor_height
+        self.tile_overlap_factor_width = tile_overlap_factor_width or self.tile_overlap_factor_width
+
+    def disable_tiling(self) -> None:
+        r"""
+        Disable tiled VAE decoding. If `enable_tiling` was previously enabled, this method will go back to computing
+        decoding in one step.
+        """
+        self.use_tiling = False
+
+    def enable_slicing(self) -> None:
+        r"""
+        Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
+        compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
+        """
+        self.use_slicing = True
+
+    def disable_slicing(self) -> None:
+        r"""
+        Disable sliced VAE decoding. If `enable_slicing` was previously enabled, this method will go back to computing
+        decoding in one step.
+        """
+        self.use_slicing = False
+
+    @apply_forward_hook
+    def encode(
+        self, x: torch.Tensor, return_dict: bool = True
+    ) -> Union[AutoencoderKLOutput, Tuple[DiagonalGaussianDistribution]]:
+        """
+        Encode a batch of images into latents.
+
+        Args:
+            x (`torch.Tensor`): Input batch of images.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether to return a [`~models.autoencoder_kl.AutoencoderKLOutput`] instead of a plain tuple.
+
+        Returns:
+                The latent representations of the encoded images. If `return_dict` is True, a
+                [`~models.autoencoder_kl.AutoencoderKLOutput`] is returned, otherwise a plain `tuple` is returned.
+        """
+        batch_size, num_channels, num_frames, height, width = x.shape
+        if num_frames == 1:
+            h = self.encoder(x)
+            if self.quant_conv is not None:
+                h = self.quant_conv(h)
+            posterior = DiagonalGaussianDistribution(h)
+        else:
+            frame_batch_size = 4
+            h = []
+            for i in range(num_frames // frame_batch_size):
+                remaining_frames = num_frames % frame_batch_size
+                start_frame = frame_batch_size * i + (0 if i == 0 else remaining_frames)
+                end_frame = frame_batch_size * (i + 1) + remaining_frames
+                z_intermediate = x[:, :, start_frame:end_frame]
+                z_intermediate = self.encoder(z_intermediate)
+                if self.quant_conv is not None:
+                    z_intermediate = self.quant_conv(z_intermediate)
+                h.append(z_intermediate)
+            self._clear_fake_context_parallel_cache()
+            h = torch.cat(h, dim=2)
+            posterior = DiagonalGaussianDistribution(h)
+        if not return_dict:
+            return (posterior,)
+        return AutoencoderKLOutput(latent_dist=posterior)
+
+    def _decode(self, z: torch.Tensor, return_dict: bool = True) -> Union[DecoderOutput, torch.Tensor]:
+        batch_size, num_channels, num_frames, height, width = z.shape
+
+        if self.use_tiling and (width > self.tile_latent_min_width or height > self.tile_latent_min_height):
+            return self.tiled_decode(z, return_dict=return_dict)
+
+        if num_frames == 1:
+            dec = []
+            z_intermediate = z
+            if self.post_quant_conv is not None:
+                z_intermediate = self.post_quant_conv(z_intermediate)
+            z_intermediate = self.decoder(z_intermediate)
+            dec.append(z_intermediate)
+        else:
+            frame_batch_size = self.num_latent_frames_batch_size
+            dec = []
+            for i in range(num_frames // frame_batch_size):
+                remaining_frames = num_frames % frame_batch_size
+                start_frame = frame_batch_size * i + (0 if i == 0 else remaining_frames)
+                end_frame = frame_batch_size * (i + 1) + remaining_frames
+                z_intermediate = z[:, :, start_frame:end_frame]
+                if self.post_quant_conv is not None:
+                    z_intermediate = self.post_quant_conv(z_intermediate)
+                z_intermediate = self.decoder(z_intermediate)
+                dec.append(z_intermediate)
+
+        self._clear_fake_context_parallel_cache()
+        dec = torch.cat(dec, dim=2)
+
+        if not return_dict:
+            return (dec,)
+
+        return DecoderOutput(sample=dec)
+
+    @apply_forward_hook
+    def decode(self, z: torch.Tensor, return_dict: bool = True) -> Union[DecoderOutput, torch.Tensor]:
+        """
+        Decode a batch of images.
+
+        Args:
+            z (`torch.Tensor`): Input batch of latent vectors.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether to return a [`~models.vae.DecoderOutput`] instead of a plain tuple.
+
+        Returns:
+            [`~models.vae.DecoderOutput`] or `tuple`:
+                If return_dict is True, a [`~models.vae.DecoderOutput`] is returned, otherwise a plain `tuple` is
+                returned.
+        """
+        if self.use_slicing and z.shape[0] > 1:
+            decoded_slices = [self._decode(z_slice).sample for z_slice in z.split(1)]
+            decoded = torch.cat(decoded_slices)
+        else:
+            decoded = self._decode(z).sample
+
+        if not return_dict:
+            return (decoded,)
+        return DecoderOutput(sample=decoded)
+
+    def blend_v(self, a: torch.Tensor, b: torch.Tensor, blend_extent: int) -> torch.Tensor:
+        blend_extent = min(a.shape[3], b.shape[3], blend_extent)
+        for y in range(blend_extent):
+            b[:, :, :, y, :] = a[:, :, :, -blend_extent + y, :] * (1 - y / blend_extent) + b[:, :, :, y, :] * (
+                y / blend_extent
+            )
+        return b
+
+    def blend_h(self, a: torch.Tensor, b: torch.Tensor, blend_extent: int) -> torch.Tensor:
+        blend_extent = min(a.shape[4], b.shape[4], blend_extent)
+        for x in range(blend_extent):
+            b[:, :, :, :, x] = a[:, :, :, :, -blend_extent + x] * (1 - x / blend_extent) + b[:, :, :, :, x] * (
+                x / blend_extent
+            )
+        return b
+
+    def tiled_decode(self, z: torch.Tensor, return_dict: bool = True) -> Union[DecoderOutput, torch.Tensor]:
+        r"""
+        Decode a batch of images using a tiled decoder.
+
+        Args:
+            z (`torch.Tensor`): Input batch of latent vectors.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~models.vae.DecoderOutput`] instead of a plain tuple.
+
+        Returns:
+            [`~models.vae.DecoderOutput`] or `tuple`:
+                If return_dict is True, a [`~models.vae.DecoderOutput`] is returned, otherwise a plain `tuple` is
+                returned.
+        """
+        # Rough memory assessment:
+        #   - In CogVideoX-2B, there are a total of 24 CausalConv3d layers.
+        #   - The biggest intermediate dimensions are: [1, 128, 9, 480, 720].
+        #   - Assume fp16 (2 bytes per value).
+        # Memory required: 1 * 128 * 9 * 480 * 720 * 24 * 2 / 1024**3 = 17.8 GB
+        #
+        # Memory assessment when using tiling:
+        #   - Assume everything as above but now HxW is 240x360 by tiling in half
+        # Memory required: 1 * 128 * 9 * 240 * 360 * 24 * 2 / 1024**3 = 4.5 GB
+
+        batch_size, num_channels, num_frames, height, width = z.shape
+
+        overlap_height = int(self.tile_latent_min_height * (1 - self.tile_overlap_factor_height))
+        overlap_width = int(self.tile_latent_min_width * (1 - self.tile_overlap_factor_width))
+        blend_extent_height = int(self.tile_sample_min_height * self.tile_overlap_factor_height)
+        blend_extent_width = int(self.tile_sample_min_width * self.tile_overlap_factor_width)
+        row_limit_height = self.tile_sample_min_height - blend_extent_height
+        row_limit_width = self.tile_sample_min_width - blend_extent_width
+        frame_batch_size = self.num_latent_frames_batch_size
+
+        # Split z into overlapping tiles and decode them separately.
+        # The tiles have an overlap to avoid seams between tiles.
+        rows = []
+        for i in range(0, height, overlap_height):
+            row = []
+            for j in range(0, width, overlap_width):
+                time = []
+                for k in range(num_frames // frame_batch_size):
+                    remaining_frames = num_frames % frame_batch_size
+                    start_frame = frame_batch_size * k + (0 if k == 0 else remaining_frames)
+                    end_frame = frame_batch_size * (k + 1) + remaining_frames
+                    tile = z[
+                        :,
+                        :,
+                        start_frame:end_frame,
+                        i : i + self.tile_latent_min_height,
+                        j : j + self.tile_latent_min_width,
+                    ]
+                    if self.post_quant_conv is not None:
+                        tile = self.post_quant_conv(tile)
+                    tile = self.decoder(tile)
+                    time.append(tile)
+                self._clear_fake_context_parallel_cache()
+                row.append(torch.cat(time, dim=2))
+            rows.append(row)
+
+        result_rows = []
+        for i, row in enumerate(rows):
+            result_row = []
+            for j, tile in enumerate(row):
+                # blend the above tile and the left tile
+                # to the current tile and add the current tile to the result row
+                if i > 0:
+                    tile = self.blend_v(rows[i - 1][j], tile, blend_extent_height)
+                if j > 0:
+                    tile = self.blend_h(row[j - 1], tile, blend_extent_width)
+                result_row.append(tile[:, :, :, :row_limit_height, :row_limit_width])
+            result_rows.append(torch.cat(result_row, dim=4))
+
+        dec = torch.cat(result_rows, dim=3)
+
+        if not return_dict:
+            return (dec,)
+
+        return DecoderOutput(sample=dec)
+
+    def forward(
+        self,
+        sample: torch.Tensor,
+        sample_posterior: bool = False,
+        return_dict: bool = True,
+        generator: Optional[torch.Generator] = None,
+    ) -> Union[torch.Tensor, torch.Tensor]:
+        x = sample
+        posterior = self.encode(x).latent_dist
+        if sample_posterior:
+            z = posterior.sample(generator=generator)
+        else:
+            z = posterior.mode()
+        dec = self.decode(z)
+        if not return_dict:
+            return (dec,)
+        return dec

cogvideox/models/transformer3d.py

@@ -0,0 +1,567 @@
+# Copyright 2024 The CogVideoX team, Tsinghua University & ZhipuAI and The HuggingFace Team.
+# All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Any, Dict, Optional, Tuple, Union
+
+import os
+import json
+import torch
+import glob
+import torch.nn.functional as F
+from torch import nn
+
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.utils import is_torch_version, logging
+from diffusers.utils.torch_utils import maybe_allow_in_graph
+from diffusers.models.attention import Attention, FeedForward
+from diffusers.models.attention_processor import AttentionProcessor, CogVideoXAttnProcessor2_0, FusedCogVideoXAttnProcessor2_0
+from diffusers.models.embeddings import CogVideoXPatchEmbed, TimestepEmbedding, Timesteps, get_3d_sincos_pos_embed
+from diffusers.models.modeling_outputs import Transformer2DModelOutput
+from diffusers.models.modeling_utils import ModelMixin
+from diffusers.models.normalization import AdaLayerNorm, CogVideoXLayerNormZero
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+@maybe_allow_in_graph
+class CogVideoXBlock(nn.Module):
+    r"""
+    Transformer block used in [CogVideoX](https://github.com/THUDM/CogVideo) model.
+
+    Parameters:
+        dim (`int`):
+            The number of channels in the input and output.
+        num_attention_heads (`int`):
+            The number of heads to use for multi-head attention.
+        attention_head_dim (`int`):
+            The number of channels in each head.
+        time_embed_dim (`int`):
+            The number of channels in timestep embedding.
+        dropout (`float`, defaults to `0.0`):
+            The dropout probability to use.
+        activation_fn (`str`, defaults to `"gelu-approximate"`):
+            Activation function to be used in feed-forward.
+        attention_bias (`bool`, defaults to `False`):
+            Whether or not to use bias in attention projection layers.
+        qk_norm (`bool`, defaults to `True`):
+            Whether or not to use normalization after query and key projections in Attention.
+        norm_elementwise_affine (`bool`, defaults to `True`):
+            Whether to use learnable elementwise affine parameters for normalization.
+        norm_eps (`float`, defaults to `1e-5`):
+            Epsilon value for normalization layers.
+        final_dropout (`bool` defaults to `False`):
+            Whether to apply a final dropout after the last feed-forward layer.
+        ff_inner_dim (`int`, *optional*, defaults to `None`):
+            Custom hidden dimension of Feed-forward layer. If not provided, `4 * dim` is used.
+        ff_bias (`bool`, defaults to `True`):
+            Whether or not to use bias in Feed-forward layer.
+        attention_out_bias (`bool`, defaults to `True`):
+            Whether or not to use bias in Attention output projection layer.
+    """
+
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        time_embed_dim: int,
+        dropout: float = 0.0,
+        activation_fn: str = "gelu-approximate",
+        attention_bias: bool = False,
+        qk_norm: bool = True,
+        norm_elementwise_affine: bool = True,
+        norm_eps: float = 1e-5,
+        final_dropout: bool = True,
+        ff_inner_dim: Optional[int] = None,
+        ff_bias: bool = True,
+        attention_out_bias: bool = True,
+    ):
+        super().__init__()
+
+        # 1. Self Attention
+        self.norm1 = CogVideoXLayerNormZero(time_embed_dim, dim, norm_elementwise_affine, norm_eps, bias=True)
+
+        self.attn1 = Attention(
+            query_dim=dim,
+            dim_head=attention_head_dim,
+            heads=num_attention_heads,
+            qk_norm="layer_norm" if qk_norm else None,
+            eps=1e-6,
+            bias=attention_bias,
+            out_bias=attention_out_bias,
+            processor=CogVideoXAttnProcessor2_0(),
+        )
+
+        # 2. Feed Forward
+        self.norm2 = CogVideoXLayerNormZero(time_embed_dim, dim, norm_elementwise_affine, norm_eps, bias=True)
+
+        self.ff = FeedForward(
+            dim,
+            dropout=dropout,
+            activation_fn=activation_fn,
+            final_dropout=final_dropout,
+            inner_dim=ff_inner_dim,
+            bias=ff_bias,
+        )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        temb: torch.Tensor,
+        image_rotary_emb: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+    ) -> torch.Tensor:
+        text_seq_length = encoder_hidden_states.size(1)
+
+        # norm & modulate
+        norm_hidden_states, norm_encoder_hidden_states, gate_msa, enc_gate_msa = self.norm1(
+            hidden_states, encoder_hidden_states, temb
+        )
+
+        # attention
+        attn_hidden_states, attn_encoder_hidden_states = self.attn1(
+            hidden_states=norm_hidden_states,
+            encoder_hidden_states=norm_encoder_hidden_states,
+            image_rotary_emb=image_rotary_emb,
+        )
+
+        hidden_states = hidden_states + gate_msa * attn_hidden_states
+        encoder_hidden_states = encoder_hidden_states + enc_gate_msa * attn_encoder_hidden_states
+
+        # norm & modulate
+        norm_hidden_states, norm_encoder_hidden_states, gate_ff, enc_gate_ff = self.norm2(
+            hidden_states, encoder_hidden_states, temb
+        )
+
+        # feed-forward
+        norm_hidden_states = torch.cat([norm_encoder_hidden_states, norm_hidden_states], dim=1)
+        ff_output = self.ff(norm_hidden_states)
+
+        hidden_states = hidden_states + gate_ff * ff_output[:, text_seq_length:]
+        encoder_hidden_states = encoder_hidden_states + enc_gate_ff * ff_output[:, :text_seq_length]
+
+        return hidden_states, encoder_hidden_states
+
+
+class CogVideoXTransformer3DModel(ModelMixin, ConfigMixin):
+    """
+    A Transformer model for video-like data in [CogVideoX](https://github.com/THUDM/CogVideo).
+
+    Parameters:
+        num_attention_heads (`int`, defaults to `30`):
+            The number of heads to use for multi-head attention.
+        attention_head_dim (`int`, defaults to `64`):
+            The number of channels in each head.
+        in_channels (`int`, defaults to `16`):
+            The number of channels in the input.
+        out_channels (`int`, *optional*, defaults to `16`):
+            The number of channels in the output.
+        flip_sin_to_cos (`bool`, defaults to `True`):
+            Whether to flip the sin to cos in the time embedding.
+        time_embed_dim (`int`, defaults to `512`):
+            Output dimension of timestep embeddings.
+        text_embed_dim (`int`, defaults to `4096`):
+            Input dimension of text embeddings from the text encoder.
+        num_layers (`int`, defaults to `30`):
+            The number of layers of Transformer blocks to use.
+        dropout (`float`, defaults to `0.0`):
+            The dropout probability to use.
+        attention_bias (`bool`, defaults to `True`):
+            Whether or not to use bias in the attention projection layers.
+        sample_width (`int`, defaults to `90`):
+            The width of the input latents.
+        sample_height (`int`, defaults to `60`):
+            The height of the input latents.
+        sample_frames (`int`, defaults to `49`):
+            The number of frames in the input latents. Note that this parameter was incorrectly initialized to 49
+            instead of 13 because CogVideoX processed 13 latent frames at once in its default and recommended settings,
+            but cannot be changed to the correct value to ensure backwards compatibility. To create a transformer with
+            K latent frames, the correct value to pass here would be: ((K - 1) * temporal_compression_ratio + 1).
+        patch_size (`int`, defaults to `2`):
+            The size of the patches to use in the patch embedding layer.
+        temporal_compression_ratio (`int`, defaults to `4`):
+            The compression ratio across the temporal dimension. See documentation for `sample_frames`.
+        max_text_seq_length (`int`, defaults to `226`):
+            The maximum sequence length of the input text embeddings.
+        activation_fn (`str`, defaults to `"gelu-approximate"`):
+            Activation function to use in feed-forward.
+        timestep_activation_fn (`str`, defaults to `"silu"`):
+            Activation function to use when generating the timestep embeddings.
+        norm_elementwise_affine (`bool`, defaults to `True`):
+            Whether or not to use elementwise affine in normalization layers.
+        norm_eps (`float`, defaults to `1e-5`):
+            The epsilon value to use in normalization layers.
+        spatial_interpolation_scale (`float`, defaults to `1.875`):
+            Scaling factor to apply in 3D positional embeddings across spatial dimensions.
+        temporal_interpolation_scale (`float`, defaults to `1.0`):
+            Scaling factor to apply in 3D positional embeddings across temporal dimensions.
+    """
+
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        num_attention_heads: int = 30,
+        attention_head_dim: int = 64,
+        in_channels: int = 16,
+        out_channels: Optional[int] = 16,
+        flip_sin_to_cos: bool = True,
+        freq_shift: int = 0,
+        time_embed_dim: int = 512,
+        text_embed_dim: int = 4096,
+        num_layers: int = 30,
+        dropout: float = 0.0,
+        attention_bias: bool = True,
+        sample_width: int = 90,
+        sample_height: int = 60,
+        sample_frames: int = 49,
+        patch_size: int = 2,
+        temporal_compression_ratio: int = 4,
+        max_text_seq_length: int = 226,
+        activation_fn: str = "gelu-approximate",
+        timestep_activation_fn: str = "silu",
+        norm_elementwise_affine: bool = True,
+        norm_eps: float = 1e-5,
+        spatial_interpolation_scale: float = 1.875,
+        temporal_interpolation_scale: float = 1.0,
+        use_rotary_positional_embeddings: bool = False,
+    ):
+        super().__init__()
+        inner_dim = num_attention_heads * attention_head_dim
+
+        post_patch_height = sample_height // patch_size
+        post_patch_width = sample_width // patch_size
+        post_time_compression_frames = (sample_frames - 1) // temporal_compression_ratio + 1
+        self.num_patches = post_patch_height * post_patch_width * post_time_compression_frames
+        self.post_patch_height = post_patch_height
+        self.post_patch_width = post_patch_width
+        self.post_time_compression_frames = post_time_compression_frames
+        self.patch_size = patch_size
+
+        # 1. Patch embedding
+        self.patch_embed = CogVideoXPatchEmbed(patch_size, in_channels, inner_dim, text_embed_dim, bias=True)
+        self.embedding_dropout = nn.Dropout(dropout)
+
+        # 2. 3D positional embeddings
+        spatial_pos_embedding = get_3d_sincos_pos_embed(
+            inner_dim,
+            (post_patch_width, post_patch_height),
+            post_time_compression_frames,
+            spatial_interpolation_scale,
+            temporal_interpolation_scale,
+        )
+        spatial_pos_embedding = torch.from_numpy(spatial_pos_embedding).flatten(0, 1)
+        pos_embedding = torch.zeros(1, max_text_seq_length + self.num_patches, inner_dim, requires_grad=False)
+        pos_embedding.data[:, max_text_seq_length:].copy_(spatial_pos_embedding)
+        self.register_buffer("pos_embedding", pos_embedding, persistent=False)
+
+        # 3. Time embeddings
+        self.time_proj = Timesteps(inner_dim, flip_sin_to_cos, freq_shift)
+        self.time_embedding = TimestepEmbedding(inner_dim, time_embed_dim, timestep_activation_fn)
+
+        # 4. Define spatio-temporal transformers blocks
+        self.transformer_blocks = nn.ModuleList(
+            [
+                CogVideoXBlock(
+                    dim=inner_dim,
+                    num_attention_heads=num_attention_heads,
+                    attention_head_dim=attention_head_dim,
+                    time_embed_dim=time_embed_dim,
+                    dropout=dropout,
+                    activation_fn=activation_fn,
+                    attention_bias=attention_bias,
+                    norm_elementwise_affine=norm_elementwise_affine,
+                    norm_eps=norm_eps,
+                )
+                for _ in range(num_layers)
+            ]
+        )
+        self.norm_final = nn.LayerNorm(inner_dim, norm_eps, norm_elementwise_affine)
+
+        # 5. Output blocks
+        self.norm_out = AdaLayerNorm(
+            embedding_dim=time_embed_dim,
+            output_dim=2 * inner_dim,
+            norm_elementwise_affine=norm_elementwise_affine,
+            norm_eps=norm_eps,
+            chunk_dim=1,
+        )
+        self.proj_out = nn.Linear(inner_dim, patch_size * patch_size * out_channels)
+
+        self.gradient_checkpointing = False
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        self.gradient_checkpointing = value
+
+    @property
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
+    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
+
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
+    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)
+
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.fuse_qkv_projections with FusedAttnProcessor2_0->FusedCogVideoXAttnProcessor2_0
+    def fuse_qkv_projections(self):
+        """
+        Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query, key, value)
+        are fused. For cross-attention modules, key and value projection matrices are fused.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+        """
+        self.original_attn_processors = None
+
+        for _, attn_processor in self.attn_processors.items():
+            if "Added" in str(attn_processor.__class__.__name__):
+                raise ValueError("`fuse_qkv_projections()` is not supported for models having added KV projections.")
+
+        self.original_attn_processors = self.attn_processors
+
+        for module in self.modules():
+            if isinstance(module, Attention):
+                module.fuse_projections(fuse=True)
+
+        self.set_attn_processor(FusedCogVideoXAttnProcessor2_0())
+
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.unfuse_qkv_projections
+    def unfuse_qkv_projections(self):
+        """Disables the fused QKV projection if enabled.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+
+        """
+        if self.original_attn_processors is not None:
+            self.set_attn_processor(self.original_attn_processors)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        timestep: Union[int, float, torch.LongTensor],
+        timestep_cond: Optional[torch.Tensor] = None,
+        inpaint_latents: Optional[torch.Tensor] = None,
+        image_rotary_emb: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        return_dict: bool = True,
+    ):
+        batch_size, num_frames, channels, height, width = hidden_states.shape
+
+        # 1. Time embedding
+        timesteps = timestep
+        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=hidden_states.dtype)
+        emb = self.time_embedding(t_emb, timestep_cond)
+
+        # 2. Patch embedding
+        if inpaint_latents is not None:
+            hidden_states = torch.concat([hidden_states, inpaint_latents], 2)
+        hidden_states = self.patch_embed(encoder_hidden_states, hidden_states)
+
+        # 3. Position embedding
+        text_seq_length = encoder_hidden_states.shape[1]
+        if not self.config.use_rotary_positional_embeddings:
+            seq_length = height * width * num_frames // (self.config.patch_size**2)
+            # pos_embeds = self.pos_embedding[:, : text_seq_length + seq_length]
+            pos_embeds = self.pos_embedding
+            emb_size = hidden_states.size()[-1]
+            pos_embeds_without_text = pos_embeds[:, text_seq_length: ].view(1, self.post_time_compression_frames, self.post_patch_height, self.post_patch_width, emb_size)
+            pos_embeds_without_text = pos_embeds_without_text.permute([0, 4, 1, 2, 3])
+            pos_embeds_without_text = F.interpolate(pos_embeds_without_text,size=[self.post_time_compression_frames, height // self.config.patch_size, width // self.config.patch_size],mode='trilinear',align_corners=False)
+            pos_embeds_without_text = pos_embeds_without_text.permute([0, 2, 3, 4, 1]).view(1, -1, emb_size)
+            pos_embeds = torch.cat([pos_embeds[:, :text_seq_length], pos_embeds_without_text], dim = 1)
+            pos_embeds = pos_embeds[:, : text_seq_length + seq_length]
+            hidden_states = hidden_states + pos_embeds
+            hidden_states = self.embedding_dropout(hidden_states)
+
+        encoder_hidden_states = hidden_states[:, :text_seq_length]
+        hidden_states = hidden_states[:, text_seq_length:]
+
+        # 4. Transformer blocks
+        for i, block in enumerate(self.transformer_blocks):
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        return module(*inputs)
+
+                    return custom_forward
+
+                ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+                hidden_states, encoder_hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(block),
+                    hidden_states,
+                    encoder_hidden_states,
+                    emb,
+                    image_rotary_emb,
+                    **ckpt_kwargs,
+                )
+            else:
+                hidden_states, encoder_hidden_states = block(
+                    hidden_states=hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    temb=emb,
+                    image_rotary_emb=image_rotary_emb,
+                )
+
+        if not self.config.use_rotary_positional_embeddings:
+            # CogVideoX-2B
+            hidden_states = self.norm_final(hidden_states)
+        else:
+            # CogVideoX-5B
+            hidden_states = torch.cat([encoder_hidden_states, hidden_states], dim=1)
+            hidden_states = self.norm_final(hidden_states)
+            hidden_states = hidden_states[:, text_seq_length:]
+
+        # 5. Final block
+        hidden_states = self.norm_out(hidden_states, temb=emb)
+        hidden_states = self.proj_out(hidden_states)
+
+        # 6. Unpatchify
+        p = self.config.patch_size
+        output = hidden_states.reshape(batch_size, num_frames, height // p, width // p, channels, p, p)
+        output = output.permute(0, 1, 4, 2, 5, 3, 6).flatten(5, 6).flatten(3, 4)
+
+        if not return_dict:
+            return (output,)
+        return Transformer2DModelOutput(sample=output)
+
+    @classmethod
+    def from_pretrained_2d(cls, pretrained_model_path, subfolder=None, transformer_additional_kwargs={}):
+        if subfolder is not None:
+            pretrained_model_path = os.path.join(pretrained_model_path, subfolder)
+        print(f"loaded 3D transformer's pretrained weights from {pretrained_model_path} ...")
+
+        config_file = os.path.join(pretrained_model_path, 'config.json')
+        if not os.path.isfile(config_file):
+            raise RuntimeError(f"{config_file} does not exist")
+        with open(config_file, "r") as f:
+            config = json.load(f)
+
+        from diffusers.utils import WEIGHTS_NAME
+        model = cls.from_config(config, **transformer_additional_kwargs)
+        model_file = os.path.join(pretrained_model_path, WEIGHTS_NAME)
+        model_file_safetensors = model_file.replace(".bin", ".safetensors")
+        if os.path.exists(model_file):
+            state_dict = torch.load(model_file, map_location="cpu")
+        elif os.path.exists(model_file_safetensors):
+            from safetensors.torch import load_file, safe_open
+            state_dict = load_file(model_file_safetensors)
+        else:
+            from safetensors.torch import load_file, safe_open
+            model_files_safetensors = glob.glob(os.path.join(pretrained_model_path, "*.safetensors"))
+            state_dict = {}
+            for model_file_safetensors in model_files_safetensors:
+                _state_dict = load_file(model_file_safetensors)
+                for key in _state_dict:
+                    state_dict[key] = _state_dict[key]
+        
+        if model.state_dict()['patch_embed.proj.weight'].size() != state_dict['patch_embed.proj.weight'].size():
+            new_shape   = model.state_dict()['patch_embed.proj.weight'].size()
+            if len(new_shape) == 5:
+                state_dict['patch_embed.proj.weight'] = state_dict['patch_embed.proj.weight'].unsqueeze(2).expand(new_shape).clone()
+                state_dict['patch_embed.proj.weight'][:, :, :-1] = 0
+            else:
+                if model.state_dict()['patch_embed.proj.weight'].size()[1] > state_dict['patch_embed.proj.weight'].size()[1]:
+                    model.state_dict()['patch_embed.proj.weight'][:, :state_dict['patch_embed.proj.weight'].size()[1], :, :] = state_dict['patch_embed.proj.weight']
+                    model.state_dict()['patch_embed.proj.weight'][:, state_dict['patch_embed.proj.weight'].size()[1]:, :, :] = 0
+                    state_dict['patch_embed.proj.weight'] = model.state_dict()['patch_embed.proj.weight']
+                else:
+                    model.state_dict()['patch_embed.proj.weight'][:, :, :, :] = state_dict['patch_embed.proj.weight'][:, :model.state_dict()['patch_embed.proj.weight'].size()[1], :, :]
+                    state_dict['patch_embed.proj.weight'] = model.state_dict()['patch_embed.proj.weight']
+
+        tmp_state_dict = {} 
+        for key in state_dict:
+            if key in model.state_dict().keys() and model.state_dict()[key].size() == state_dict[key].size():
+                tmp_state_dict[key] = state_dict[key]
+            else:
+                print(key, "Size don't match, skip")
+        state_dict = tmp_state_dict
+
+        m, u = model.load_state_dict(state_dict, strict=False)
+        print(f"### missing keys: {len(m)}; \n### unexpected keys: {len(u)};")
+        print(m)
+        
+        params = [p.numel() if "mamba" in n else 0 for n, p in model.named_parameters()]
+        print(f"### Mamba Parameters: {sum(params) / 1e6} M")
+
+        params = [p.numel() if "attn1." in n else 0 for n, p in model.named_parameters()]
+        print(f"### attn1 Parameters: {sum(params) / 1e6} M")
+        
+        return model

cogvideox/pipeline/pipeline_cogvideox.py

@@ -0,0 +1,751 @@
+# Copyright 2024 The CogVideoX team, Tsinghua University & ZhipuAI and The HuggingFace Team.
+# All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+import math
+from dataclasses import dataclass
+from typing import Callable, Dict, List, Optional, Tuple, Union
+
+import torch
+from transformers import T5EncoderModel, T5Tokenizer
+
+from diffusers.callbacks import MultiPipelineCallbacks, PipelineCallback
+from diffusers.models import AutoencoderKLCogVideoX, CogVideoXTransformer3DModel
+from diffusers.models.embeddings import get_3d_rotary_pos_embed
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline
+from diffusers.schedulers import CogVideoXDDIMScheduler, CogVideoXDPMScheduler
+from diffusers.utils import BaseOutput, logging, replace_example_docstring
+from diffusers.utils.torch_utils import randn_tensor
+from diffusers.video_processor import VideoProcessor
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```python
+        >>> import torch
+        >>> from diffusers import CogVideoX_Fun_Pipeline
+        >>> from diffusers.utils import export_to_video
+
+        >>> # Models: "THUDM/CogVideoX-2b" or "THUDM/CogVideoX-5b"
+        >>> pipe = CogVideoX_Fun_Pipeline.from_pretrained("THUDM/CogVideoX-2b", torch_dtype=torch.float16).to("cuda")
+        >>> prompt = (
+        ...     "A panda, dressed in a small, red jacket and a tiny hat, sits on a wooden stool in a serene bamboo forest. "
+        ...     "The panda's fluffy paws strum a miniature acoustic guitar, producing soft, melodic tunes. Nearby, a few other "
+        ...     "pandas gather, watching curiously and some clapping in rhythm. Sunlight filters through the tall bamboo, "
+        ...     "casting a gentle glow on the scene. The panda's face is expressive, showing concentration and joy as it plays. "
+        ...     "The background includes a small, flowing stream and vibrant green foliage, enhancing the peaceful and magical "
+        ...     "atmosphere of this unique musical performance."
+        ... )
+        >>> video = pipe(prompt=prompt, guidance_scale=6, num_inference_steps=50).frames[0]
+        >>> export_to_video(video, "output.mp4", fps=8)
+        ```
+"""
+
+
+# Similar to diffusers.pipelines.hunyuandit.pipeline_hunyuandit.get_resize_crop_region_for_grid
+def get_resize_crop_region_for_grid(src, tgt_width, tgt_height):
+    tw = tgt_width
+    th = tgt_height
+    h, w = src
+    r = h / w
+    if r > (th / tw):
+        resize_height = th
+        resize_width = int(round(th / h * w))
+    else:
+        resize_width = tw
+        resize_height = int(round(tw / w * h))
+
+    crop_top = int(round((th - resize_height) / 2.0))
+    crop_left = int(round((tw - resize_width) / 2.0))
+
+    return (crop_top, crop_left), (crop_top + resize_height, crop_left + resize_width)
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
+def retrieve_timesteps(
+    scheduler,
+    num_inference_steps: Optional[int] = None,
+    device: Optional[Union[str, torch.device]] = None,
+    timesteps: Optional[List[int]] = None,
+    sigmas: Optional[List[float]] = None,
+    **kwargs,
+):
+    """
+    Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
+    custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
+
+    Args:
+        scheduler (`SchedulerMixin`):
+            The scheduler to get timesteps from.
+        num_inference_steps (`int`):
+            The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps`
+            must be `None`.
+        device (`str` or `torch.device`, *optional*):
+            The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+        timesteps (`List[int]`, *optional*):
+            Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed,
+            `num_inference_steps` and `sigmas` must be `None`.
+        sigmas (`List[float]`, *optional*):
+            Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed,
+            `num_inference_steps` and `timesteps` must be `None`.
+
+    Returns:
+        `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
+        second element is the number of inference steps.
+    """
+    if timesteps is not None and sigmas is not None:
+        raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values")
+    if timesteps is not None:
+        accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accepts_timesteps:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" timestep schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    elif sigmas is not None:
+        accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accept_sigmas:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" sigmas schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    else:
+        scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+    return timesteps, num_inference_steps
+
+
+@dataclass
+class CogVideoX_Fun_PipelineOutput(BaseOutput):
+    r"""
+    Output class for CogVideo pipelines.
+
+    Args:
+        video (`torch.Tensor`, `np.ndarray`, or List[List[PIL.Image.Image]]):
+            List of video outputs - It can be a nested list of length `batch_size,` with each sub-list containing
+            denoised PIL image sequences of length `num_frames.` It can also be a NumPy array or Torch tensor of shape
+            `(batch_size, num_frames, channels, height, width)`.
+    """
+
+    videos: torch.Tensor
+
+
+class CogVideoX_Fun_Pipeline(DiffusionPipeline):
+    r"""
+    Pipeline for text-to-video generation using CogVideoX_Fun.
+
+    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+    library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+    Args:
+        vae ([`AutoencoderKL`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode videos to and from latent representations.
+        text_encoder ([`T5EncoderModel`]):
+            Frozen text-encoder. CogVideoX uses
+            [T5](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5EncoderModel); specifically the
+            [t5-v1_1-xxl](https://huggingface.co/PixArt-alpha/PixArt-alpha/tree/main/t5-v1_1-xxl) variant.
+        tokenizer (`T5Tokenizer`):
+            Tokenizer of class
+            [T5Tokenizer](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5Tokenizer).
+        transformer ([`CogVideoXTransformer3DModel`]):
+            A text conditioned `CogVideoXTransformer3DModel` to denoise the encoded video latents.
+        scheduler ([`SchedulerMixin`]):
+            A scheduler to be used in combination with `transformer` to denoise the encoded video latents.
+    """
+
+    _optional_components = []
+    model_cpu_offload_seq = "text_encoder->transformer->vae"
+
+    _callback_tensor_inputs = [
+        "latents",
+        "prompt_embeds",
+        "negative_prompt_embeds",
+    ]
+
+    def __init__(
+        self,
+        tokenizer: T5Tokenizer,
+        text_encoder: T5EncoderModel,
+        vae: AutoencoderKLCogVideoX,
+        transformer: CogVideoXTransformer3DModel,
+        scheduler: Union[CogVideoXDDIMScheduler, CogVideoXDPMScheduler],
+    ):
+        super().__init__()
+
+        self.register_modules(
+            tokenizer=tokenizer, text_encoder=text_encoder, vae=vae, transformer=transformer, scheduler=scheduler
+        )
+        self.vae_scale_factor_spatial = (
+            2 ** (len(self.vae.config.block_out_channels) - 1) if hasattr(self, "vae") and self.vae is not None else 8
+        )
+        self.vae_scale_factor_temporal = (
+            self.vae.config.temporal_compression_ratio if hasattr(self, "vae") and self.vae is not None else 4
+        )
+
+        self.video_processor = VideoProcessor(vae_scale_factor=self.vae_scale_factor_spatial)
+
+    def _get_t5_prompt_embeds(
+        self,
+        prompt: Union[str, List[str]] = None,
+        num_videos_per_prompt: int = 1,
+        max_sequence_length: int = 226,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        device = device or self._execution_device
+        dtype = dtype or self.text_encoder.dtype
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        batch_size = len(prompt)
+
+        text_inputs = self.tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=max_sequence_length,
+            truncation=True,
+            add_special_tokens=True,
+            return_tensors="pt",
+        )
+        text_input_ids = text_inputs.input_ids
+        untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+        if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids):
+            removed_text = self.tokenizer.batch_decode(untruncated_ids[:, max_sequence_length - 1 : -1])
+            logger.warning(
+                "The following part of your input was truncated because `max_sequence_length` is set to "
+                f" {max_sequence_length} tokens: {removed_text}"
+            )
+
+        prompt_embeds = self.text_encoder(text_input_ids.to(device))[0]
+        prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
+
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        _, seq_len, _ = prompt_embeds.shape
+        prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1)
+
+        return prompt_embeds
+
+    def encode_prompt(
+        self,
+        prompt: Union[str, List[str]],
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        do_classifier_free_guidance: bool = True,
+        num_videos_per_prompt: int = 1,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        max_sequence_length: int = 226,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        r"""
+        Encodes the prompt into text encoder hidden states.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                prompt to be encoded
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            do_classifier_free_guidance (`bool`, *optional*, defaults to `True`):
+                Whether to use classifier free guidance or not.
+            num_videos_per_prompt (`int`, *optional*, defaults to 1):
+                Number of videos that should be generated per prompt. torch device to place the resulting embeddings on
+            prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            device: (`torch.device`, *optional*):
+                torch device
+            dtype: (`torch.dtype`, *optional*):
+                torch dtype
+        """
+        device = device or self._execution_device
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        if prompt is not None:
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        if prompt_embeds is None:
+            prompt_embeds = self._get_t5_prompt_embeds(
+                prompt=prompt,
+                num_videos_per_prompt=num_videos_per_prompt,
+                max_sequence_length=max_sequence_length,
+                device=device,
+                dtype=dtype,
+            )
+
+        if do_classifier_free_guidance and negative_prompt_embeds is None:
+            negative_prompt = negative_prompt or ""
+            negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt
+
+            if prompt is not None and type(prompt) is not type(negative_prompt):
+                raise TypeError(
+                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+                    f" {type(prompt)}."
+                )
+            elif batch_size != len(negative_prompt):
+                raise ValueError(
+                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                    " the batch size of `prompt`."
+                )
+
+            negative_prompt_embeds = self._get_t5_prompt_embeds(
+                prompt=negative_prompt,
+                num_videos_per_prompt=num_videos_per_prompt,
+                max_sequence_length=max_sequence_length,
+                device=device,
+                dtype=dtype,
+            )
+
+        return prompt_embeds, negative_prompt_embeds
+
+    def prepare_latents(
+        self, batch_size, num_channels_latents, num_frames, height, width, dtype, device, generator, latents=None
+    ):
+        shape = (
+            batch_size,
+            (num_frames - 1) // self.vae_scale_factor_temporal + 1,
+            num_channels_latents,
+            height // self.vae_scale_factor_spatial,
+            width // self.vae_scale_factor_spatial,
+        )
+        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:
+            latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+        else:
+            latents = latents.to(device)
+
+        # scale the initial noise by the standard deviation required by the scheduler
+        latents = latents * self.scheduler.init_noise_sigma
+        return latents
+
+    def decode_latents(self, latents: torch.Tensor) -> torch.Tensor:
+        latents = latents.permute(0, 2, 1, 3, 4)  # [batch_size, num_channels, num_frames, height, width]
+        latents = 1 / self.vae.config.scaling_factor * latents
+
+        frames = self.vae.decode(latents).sample
+        frames = (frames / 2 + 0.5).clamp(0, 1)
+        # we always cast to float32 as this does not cause significant overhead and is compatible with bfloa16
+        frames = frames.cpu().float().numpy()
+        return frames
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+    def prepare_extra_step_kwargs(self, generator, eta):
+        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+        # and should be between [0, 1]
+
+        accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        extra_step_kwargs = {}
+        if accepts_eta:
+            extra_step_kwargs["eta"] = eta
+
+        # check if the scheduler accepts generator
+        accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        if accepts_generator:
+            extra_step_kwargs["generator"] = generator
+        return extra_step_kwargs
+
+    # Copied from diffusers.pipelines.latte.pipeline_latte.LattePipeline.check_inputs
+    def check_inputs(
+        self,
+        prompt,
+        height,
+        width,
+        negative_prompt,
+        callback_on_step_end_tensor_inputs,
+        prompt_embeds=None,
+        negative_prompt_embeds=None,
+    ):
+        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}.")
+
+        if callback_on_step_end_tensor_inputs is not None and not all(
+            k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+        ):
+            raise ValueError(
+                f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+            )
+        if prompt is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt is None and prompt_embeds is None:
+            raise ValueError(
+                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+            )
+        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+        if prompt is not None and negative_prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt`: {prompt} and `negative_prompt_embeds`:"
+                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+            )
+
+        if negative_prompt is not None and negative_prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+            )
+
+        if prompt_embeds is not None and negative_prompt_embeds is not None:
+            if prompt_embeds.shape != negative_prompt_embeds.shape:
+                raise ValueError(
+                    "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+                    f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+                    f" {negative_prompt_embeds.shape}."
+                )
+
+    def fuse_qkv_projections(self) -> None:
+        r"""Enables fused QKV projections."""
+        self.fusing_transformer = True
+        self.transformer.fuse_qkv_projections()
+
+    def unfuse_qkv_projections(self) -> None:
+        r"""Disable QKV projection fusion if enabled."""
+        if not self.fusing_transformer:
+            logger.warning("The Transformer was not initially fused for QKV projections. Doing nothing.")
+        else:
+            self.transformer.unfuse_qkv_projections()
+            self.fusing_transformer = False
+
+    def _prepare_rotary_positional_embeddings(
+        self,
+        height: int,
+        width: int,
+        num_frames: int,
+        device: torch.device,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        grid_height = height // (self.vae_scale_factor_spatial * self.transformer.config.patch_size)
+        grid_width = width // (self.vae_scale_factor_spatial * self.transformer.config.patch_size)
+        base_size_width = 720 // (self.vae_scale_factor_spatial * self.transformer.config.patch_size)
+        base_size_height = 480 // (self.vae_scale_factor_spatial * self.transformer.config.patch_size)
+
+        grid_crops_coords = get_resize_crop_region_for_grid(
+            (grid_height, grid_width), base_size_width, base_size_height
+        )
+        freqs_cos, freqs_sin = get_3d_rotary_pos_embed(
+            embed_dim=self.transformer.config.attention_head_dim,
+            crops_coords=grid_crops_coords,
+            grid_size=(grid_height, grid_width),
+            temporal_size=num_frames,
+            use_real=True,
+        )
+
+        freqs_cos = freqs_cos.to(device=device)
+        freqs_sin = freqs_sin.to(device=device)
+        return freqs_cos, freqs_sin
+
+    @property
+    def guidance_scale(self):
+        return self._guidance_scale
+
+    @property
+    def num_timesteps(self):
+        return self._num_timesteps
+
+    @property
+    def interrupt(self):
+        return self._interrupt
+
+    @torch.no_grad()
+    @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+        self,
+        prompt: Optional[Union[str, List[str]]] = None,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        height: int = 480,
+        width: int = 720,
+        num_frames: int = 49,
+        num_inference_steps: int = 50,
+        timesteps: Optional[List[int]] = None,
+        guidance_scale: float = 6,
+        use_dynamic_cfg: bool = False,
+        num_videos_per_prompt: int = 1,
+        eta: float = 0.0,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        output_type: str = "numpy",
+        return_dict: bool = False,
+        callback_on_step_end: Optional[
+            Union[Callable[[int, int, Dict], None], PipelineCallback, MultiPipelineCallbacks]
+        ] = None,
+        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+        max_sequence_length: int = 226,
+    ) -> Union[CogVideoX_Fun_PipelineOutput, Tuple]:
+        """
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+                instead.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+                The height in pixels of the generated image. This is set to 1024 by default for the best results.
+            width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+                The width in pixels of the generated image. This is set to 1024 by default for the best results.
+            num_frames (`int`, defaults to `48`):
+                Number of frames to generate. Must be divisible by self.vae_scale_factor_temporal. Generated video will
+                contain 1 extra frame because CogVideoX is conditioned with (num_seconds * fps + 1) frames where
+                num_seconds is 6 and fps is 4. However, since videos can be saved at any fps, the only condition that
+                needs to be satisfied is that of divisibility mentioned above.
+            num_inference_steps (`int`, *optional*, defaults to 50):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            timesteps (`List[int]`, *optional*):
+                Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
+                in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
+                passed will be used. Must be in descending order.
+            guidance_scale (`float`, *optional*, defaults to 7.0):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+                usually at the expense of lower image quality.
+            num_videos_per_prompt (`int`, *optional*, defaults to 1):
+                The number of videos to generate per prompt.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                One or a list of [torch generator(s)](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 will ge generated by sampling using the supplied random `generator`.
+            prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generate image. Choose between
+                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] instead
+                of a plain tuple.
+            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.
+            max_sequence_length (`int`, defaults to `226`):
+                Maximum sequence length in encoded prompt. Must be consistent with
+                `self.transformer.config.max_text_seq_length` otherwise may lead to poor results.
+
+        Examples:
+
+        Returns:
+            [`~pipelines.cogvideo.pipeline_cogvideox.CogVideoX_Fun_PipelineOutput`] or `tuple`:
+            [`~pipelines.cogvideo.pipeline_cogvideox.CogVideoX_Fun_PipelineOutput`] if `return_dict` is True, otherwise a
+            `tuple`. When returning a tuple, the first element is a list with the generated images.
+        """
+
+        if num_frames > 49:
+            raise ValueError(
+                "The number of frames must be less than 49 for now due to static positional embeddings. This will be updated in the future to remove this limitation."
+            )
+
+        if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)):
+            callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs
+
+        height = height or self.transformer.config.sample_size * self.vae_scale_factor_spatial
+        width = width or self.transformer.config.sample_size * self.vae_scale_factor_spatial
+        num_videos_per_prompt = 1
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(
+            prompt,
+            height,
+            width,
+            negative_prompt,
+            callback_on_step_end_tensor_inputs,
+            prompt_embeds,
+            negative_prompt_embeds,
+        )
+        self._guidance_scale = guidance_scale
+        self._interrupt = False
+
+        # 2. Default call parameters
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.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 = guidance_scale > 1.0
+
+        # 3. Encode input prompt
+        prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+            prompt,
+            negative_prompt,
+            do_classifier_free_guidance,
+            num_videos_per_prompt=num_videos_per_prompt,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            max_sequence_length=max_sequence_length,
+            device=device,
+        )
+        if do_classifier_free_guidance:
+            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
+
+        # 4. Prepare timesteps
+        timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps)
+        self._num_timesteps = len(timesteps)
+
+        # 5. Prepare latents.
+        latent_channels = self.transformer.config.in_channels
+        latents = self.prepare_latents(
+            batch_size * num_videos_per_prompt,
+            latent_channels,
+            num_frames,
+            height,
+            width,
+            prompt_embeds.dtype,
+            device,
+            generator,
+            latents,
+        )
+
+        # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+        # 7. Create rotary embeds if required
+        image_rotary_emb = (
+            self._prepare_rotary_positional_embeddings(height, width, latents.size(1), device)
+            if self.transformer.config.use_rotary_positional_embeddings
+            else None
+        )
+
+        # 8. Denoising loop
+        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
+
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            # for DPM-solver++
+            old_pred_original_sample = None
+            for i, t in enumerate(timesteps):
+                if self.interrupt:
+                    continue
+
+                latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+                latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+                # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+                timestep = t.expand(latent_model_input.shape[0])
+
+                # predict noise model_output
+                noise_pred = self.transformer(
+                    hidden_states=latent_model_input,
+                    encoder_hidden_states=prompt_embeds,
+                    timestep=timestep,
+                    image_rotary_emb=image_rotary_emb,
+                    return_dict=False,
+                )[0]
+                noise_pred = noise_pred.float()
+
+                # perform guidance
+                if use_dynamic_cfg:
+                    self._guidance_scale = 1 + guidance_scale * (
+                        (1 - math.cos(math.pi * ((num_inference_steps - t.item()) / num_inference_steps) ** 5.0)) / 2
+                    )
+                if do_classifier_free_guidance:
+                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                    noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+                # compute the previous noisy sample x_t -> x_t-1
+                if not isinstance(self.scheduler, CogVideoXDPMScheduler):
+                    latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
+                else:
+                    latents, old_pred_original_sample = self.scheduler.step(
+                        noise_pred,
+                        old_pred_original_sample,
+                        t,
+                        timesteps[i - 1] if i > 0 else None,
+                        latents,
+                        **extra_step_kwargs,
+                        return_dict=False,
+                    )
+                latents = latents.to(prompt_embeds.dtype)
+
+                # call the callback, if provided
+                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)
+                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+                    negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+
+        if output_type == "numpy":
+            video = self.decode_latents(latents)
+        elif not output_type == "latent":
+            video = self.decode_latents(latents)
+            video = self.video_processor.postprocess_video(video=video, output_type=output_type)
+        else:
+            video = latents
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            video = torch.from_numpy(video)
+
+        return CogVideoX_Fun_PipelineOutput(videos=video)

cogvideox/pipeline/pipeline_cogvideox_inpaint.py

@@ -0,0 +1,1003 @@
+# Copyright 2024 The CogVideoX team, Tsinghua University & ZhipuAI and The HuggingFace Team.
+# All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+import math
+from dataclasses import dataclass
+from typing import Callable, Dict, List, Optional, Tuple, Union
+
+import torch
+import torch.nn.functional as F
+from einops import rearrange
+from transformers import T5EncoderModel, T5Tokenizer
+
+from diffusers.callbacks import MultiPipelineCallbacks, PipelineCallback
+from diffusers.models import AutoencoderKLCogVideoX, CogVideoXTransformer3DModel
+from diffusers.models.embeddings import get_3d_rotary_pos_embed
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline
+from diffusers.schedulers import CogVideoXDDIMScheduler, CogVideoXDPMScheduler
+from diffusers.utils import BaseOutput, logging, replace_example_docstring
+from diffusers.utils.torch_utils import randn_tensor
+from diffusers.video_processor import VideoProcessor
+from diffusers.image_processor import VaeImageProcessor
+from einops import rearrange
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```python
+        >>> import torch
+        >>> from diffusers import CogVideoX_Fun_Pipeline
+        >>> from diffusers.utils import export_to_video
+
+        >>> # Models: "THUDM/CogVideoX-2b" or "THUDM/CogVideoX-5b"
+        >>> pipe = CogVideoX_Fun_Pipeline.from_pretrained("THUDM/CogVideoX-2b", torch_dtype=torch.float16).to("cuda")
+        >>> prompt = (
+        ...     "A panda, dressed in a small, red jacket and a tiny hat, sits on a wooden stool in a serene bamboo forest. "
+        ...     "The panda's fluffy paws strum a miniature acoustic guitar, producing soft, melodic tunes. Nearby, a few other "
+        ...     "pandas gather, watching curiously and some clapping in rhythm. Sunlight filters through the tall bamboo, "
+        ...     "casting a gentle glow on the scene. The panda's face is expressive, showing concentration and joy as it plays. "
+        ...     "The background includes a small, flowing stream and vibrant green foliage, enhancing the peaceful and magical "
+        ...     "atmosphere of this unique musical performance."
+        ... )
+        >>> video = pipe(prompt=prompt, guidance_scale=6, num_inference_steps=50).frames[0]
+        >>> export_to_video(video, "output.mp4", fps=8)
+        ```
+"""
+
+
+# Similar to diffusers.pipelines.hunyuandit.pipeline_hunyuandit.get_resize_crop_region_for_grid
+def get_resize_crop_region_for_grid(src, tgt_width, tgt_height):
+    tw = tgt_width
+    th = tgt_height
+    h, w = src
+    r = h / w
+    if r > (th / tw):
+        resize_height = th
+        resize_width = int(round(th / h * w))
+    else:
+        resize_width = tw
+        resize_height = int(round(tw / w * h))
+
+    crop_top = int(round((th - resize_height) / 2.0))
+    crop_left = int(round((tw - resize_width) / 2.0))
+
+    return (crop_top, crop_left), (crop_top + resize_height, crop_left + resize_width)
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
+def retrieve_timesteps(
+    scheduler,
+    num_inference_steps: Optional[int] = None,
+    device: Optional[Union[str, torch.device]] = None,
+    timesteps: Optional[List[int]] = None,
+    sigmas: Optional[List[float]] = None,
+    **kwargs,
+):
+    """
+    Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
+    custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
+
+    Args:
+        scheduler (`SchedulerMixin`):
+            The scheduler to get timesteps from.
+        num_inference_steps (`int`):
+            The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps`
+            must be `None`.
+        device (`str` or `torch.device`, *optional*):
+            The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+        timesteps (`List[int]`, *optional*):
+            Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed,
+            `num_inference_steps` and `sigmas` must be `None`.
+        sigmas (`List[float]`, *optional*):
+            Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed,
+            `num_inference_steps` and `timesteps` must be `None`.
+
+    Returns:
+        `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
+        second element is the number of inference steps.
+    """
+    if timesteps is not None and sigmas is not None:
+        raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values")
+    if timesteps is not None:
+        accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accepts_timesteps:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" timestep schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    elif sigmas is not None:
+        accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accept_sigmas:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" sigmas schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    else:
+        scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+    return timesteps, num_inference_steps
+
+
+def resize_mask(mask, latent, process_first_frame_only=True):
+    latent_size = latent.size()
+    batch_size, channels, num_frames, height, width = mask.shape
+
+    if process_first_frame_only:
+        target_size = list(latent_size[2:])
+        target_size[0] = 1
+        first_frame_resized = F.interpolate(
+            mask[:, :, 0:1, :, :],
+            size=target_size,
+            mode='trilinear',
+            align_corners=False
+        )
+        
+        target_size = list(latent_size[2:])
+        target_size[0] = target_size[0] - 1
+        if target_size[0] != 0:
+            remaining_frames_resized = F.interpolate(
+                mask[:, :, 1:, :, :],
+                size=target_size,
+                mode='trilinear',
+                align_corners=False
+            )
+            resized_mask = torch.cat([first_frame_resized, remaining_frames_resized], dim=2)
+        else:
+            resized_mask = first_frame_resized
+    else:
+        target_size = list(latent_size[2:])
+        resized_mask = F.interpolate(
+            mask,
+            size=target_size,
+            mode='trilinear',
+            align_corners=False
+        )
+    return resized_mask
+
+
+@dataclass
+class CogVideoX_Fun_PipelineOutput(BaseOutput):
+    r"""
+    Output class for CogVideo pipelines.
+
+    Args:
+        video (`torch.Tensor`, `np.ndarray`, or List[List[PIL.Image.Image]]):
+            List of video outputs - It can be a nested list of length `batch_size,` with each sub-list containing
+            denoised PIL image sequences of length `num_frames.` It can also be a NumPy array or Torch tensor of shape
+            `(batch_size, num_frames, channels, height, width)`.
+    """
+
+    videos: torch.Tensor
+
+
+class CogVideoX_Fun_Pipeline_Inpaint(DiffusionPipeline):
+    r"""
+    Pipeline for text-to-video generation using CogVideoX.
+
+    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+    library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+    Args:
+        vae ([`AutoencoderKL`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode videos to and from latent representations.
+        text_encoder ([`T5EncoderModel`]):
+            Frozen text-encoder. CogVideoX_Fun uses
+            [T5](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5EncoderModel); specifically the
+            [t5-v1_1-xxl](https://huggingface.co/PixArt-alpha/PixArt-alpha/tree/main/t5-v1_1-xxl) variant.
+        tokenizer (`T5Tokenizer`):
+            Tokenizer of class
+            [T5Tokenizer](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5Tokenizer).
+        transformer ([`CogVideoXTransformer3DModel`]):
+            A text conditioned `CogVideoXTransformer3DModel` to denoise the encoded video latents.
+        scheduler ([`SchedulerMixin`]):
+            A scheduler to be used in combination with `transformer` to denoise the encoded video latents.
+    """
+
+    _optional_components = []
+    model_cpu_offload_seq = "text_encoder->vae->transformer->vae"
+
+    _callback_tensor_inputs = [
+        "latents",
+        "prompt_embeds",
+        "negative_prompt_embeds",
+    ]
+
+    def __init__(
+        self,
+        tokenizer: T5Tokenizer,
+        text_encoder: T5EncoderModel,
+        vae: AutoencoderKLCogVideoX,
+        transformer: CogVideoXTransformer3DModel,
+        scheduler: Union[CogVideoXDDIMScheduler, CogVideoXDPMScheduler],
+    ):
+        super().__init__()
+
+        self.register_modules(
+            tokenizer=tokenizer, text_encoder=text_encoder, vae=vae, transformer=transformer, scheduler=scheduler
+        )
+        self.vae_scale_factor_spatial = (
+            2 ** (len(self.vae.config.block_out_channels) - 1) if hasattr(self, "vae") and self.vae is not None else 8
+        )
+        self.vae_scale_factor_temporal = (
+            self.vae.config.temporal_compression_ratio if hasattr(self, "vae") and self.vae is not None else 4
+        )
+
+        self.video_processor = VideoProcessor(vae_scale_factor=self.vae_scale_factor_spatial)
+
+        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+        self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+        self.mask_processor = VaeImageProcessor(
+            vae_scale_factor=self.vae_scale_factor, do_normalize=False, do_binarize=True, do_convert_grayscale=True
+        )
+
+    def _get_t5_prompt_embeds(
+        self,
+        prompt: Union[str, List[str]] = None,
+        num_videos_per_prompt: int = 1,
+        max_sequence_length: int = 226,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        device = device or self._execution_device
+        dtype = dtype or self.text_encoder.dtype
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        batch_size = len(prompt)
+
+        text_inputs = self.tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=max_sequence_length,
+            truncation=True,
+            add_special_tokens=True,
+            return_tensors="pt",
+        )
+        text_input_ids = text_inputs.input_ids
+        untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+        if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids):
+            removed_text = self.tokenizer.batch_decode(untruncated_ids[:, max_sequence_length - 1 : -1])
+            logger.warning(
+                "The following part of your input was truncated because `max_sequence_length` is set to "
+                f" {max_sequence_length} tokens: {removed_text}"
+            )
+
+        prompt_embeds = self.text_encoder(text_input_ids.to(device))[0]
+        prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
+
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        _, seq_len, _ = prompt_embeds.shape
+        prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1)
+
+        return prompt_embeds
+
+    def encode_prompt(
+        self,
+        prompt: Union[str, List[str]],
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        do_classifier_free_guidance: bool = True,
+        num_videos_per_prompt: int = 1,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        max_sequence_length: int = 226,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        r"""
+        Encodes the prompt into text encoder hidden states.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                prompt to be encoded
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            do_classifier_free_guidance (`bool`, *optional*, defaults to `True`):
+                Whether to use classifier free guidance or not.
+            num_videos_per_prompt (`int`, *optional*, defaults to 1):
+                Number of videos that should be generated per prompt. torch device to place the resulting embeddings on
+            prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            device: (`torch.device`, *optional*):
+                torch device
+            dtype: (`torch.dtype`, *optional*):
+                torch dtype
+        """
+        device = device or self._execution_device
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        if prompt is not None:
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        if prompt_embeds is None:
+            prompt_embeds = self._get_t5_prompt_embeds(
+                prompt=prompt,
+                num_videos_per_prompt=num_videos_per_prompt,
+                max_sequence_length=max_sequence_length,
+                device=device,
+                dtype=dtype,
+            )
+
+        if do_classifier_free_guidance and negative_prompt_embeds is None:
+            negative_prompt = negative_prompt or ""
+            negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt
+
+            if prompt is not None and type(prompt) is not type(negative_prompt):
+                raise TypeError(
+                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+                    f" {type(prompt)}."
+                )
+            elif batch_size != len(negative_prompt):
+                raise ValueError(
+                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                    " the batch size of `prompt`."
+                )
+
+            negative_prompt_embeds = self._get_t5_prompt_embeds(
+                prompt=negative_prompt,
+                num_videos_per_prompt=num_videos_per_prompt,
+                max_sequence_length=max_sequence_length,
+                device=device,
+                dtype=dtype,
+            )
+
+        return prompt_embeds, negative_prompt_embeds
+
+    def prepare_latents(
+        self, 
+        batch_size,
+        num_channels_latents,
+        height,
+        width,
+        video_length,
+        dtype,
+        device,
+        generator,
+        latents=None,
+        video=None,
+        timestep=None,
+        is_strength_max=True,
+        return_noise=False,
+        return_video_latents=False,
+    ):
+        shape = (
+            batch_size,
+            (video_length - 1) // self.vae_scale_factor_temporal + 1,
+            num_channels_latents,
+            height // self.vae_scale_factor_spatial,
+            width // self.vae_scale_factor_spatial,
+        )
+        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 return_video_latents or (latents is None and not is_strength_max):
+            video = video.to(device=device, dtype=self.vae.dtype)
+            
+            bs = 1
+            new_video = []
+            for i in range(0, video.shape[0], bs):
+                video_bs = video[i : i + bs]
+                video_bs = self.vae.encode(video_bs)[0]
+                video_bs = video_bs.sample()
+                new_video.append(video_bs)
+            video = torch.cat(new_video, dim = 0)
+            video = video * self.vae.config.scaling_factor
+
+            video_latents = video.repeat(batch_size // video.shape[0], 1, 1, 1, 1)
+            video_latents = video_latents.to(device=device, dtype=dtype)
+            video_latents = rearrange(video_latents, "b c f h w -> b f c h w")
+
+        if latents is None:
+            noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+            # if strength is 1. then initialise the latents to noise, else initial to image + noise
+            latents = noise if is_strength_max else self.scheduler.add_noise(video_latents, noise, timestep)
+            # if pure noise then scale the initial latents by the  Scheduler's init sigma
+            latents = latents * self.scheduler.init_noise_sigma if is_strength_max else latents
+        else:
+            noise = latents.to(device)
+            latents = noise * self.scheduler.init_noise_sigma
+
+        # scale the initial noise by the standard deviation required by the scheduler
+        outputs = (latents,)
+
+        if return_noise:
+            outputs += (noise,)
+
+        if return_video_latents:
+            outputs += (video_latents,)
+
+        return outputs
+
+    def prepare_mask_latents(
+        self, mask, masked_image, batch_size, height, width, dtype, device, generator, do_classifier_free_guidance
+    ):
+        # resize the mask to latents shape as we concatenate the mask to the latents
+        # we do that before converting to dtype to avoid breaking in case we're using cpu_offload
+        # and half precision
+
+        if mask is not None:
+            mask = mask.to(device=device, dtype=self.vae.dtype)
+            bs = 1
+            new_mask = []
+            for i in range(0, mask.shape[0], bs):
+                mask_bs = mask[i : i + bs]
+                mask_bs = self.vae.encode(mask_bs)[0]
+                mask_bs = mask_bs.mode()
+                new_mask.append(mask_bs)
+            mask = torch.cat(new_mask, dim = 0)
+            mask = mask * self.vae.config.scaling_factor
+
+        if masked_image is not None:
+            masked_image = masked_image.to(device=device, dtype=self.vae.dtype)
+            bs = 1
+            new_mask_pixel_values = []
+            for i in range(0, masked_image.shape[0], bs):
+                mask_pixel_values_bs = masked_image[i : i + bs]
+                mask_pixel_values_bs = self.vae.encode(mask_pixel_values_bs)[0]
+                mask_pixel_values_bs = mask_pixel_values_bs.mode()
+                new_mask_pixel_values.append(mask_pixel_values_bs)
+            masked_image_latents = torch.cat(new_mask_pixel_values, dim = 0)
+            masked_image_latents = masked_image_latents * self.vae.config.scaling_factor
+        else:
+            masked_image_latents = None
+
+        return mask, masked_image_latents
+
+    def decode_latents(self, latents: torch.Tensor) -> torch.Tensor:
+        latents = latents.permute(0, 2, 1, 3, 4)  # [batch_size, num_channels, num_frames, height, width]
+        latents = 1 / self.vae.config.scaling_factor * latents
+
+        frames = self.vae.decode(latents).sample
+        frames = (frames / 2 + 0.5).clamp(0, 1)
+        # we always cast to float32 as this does not cause significant overhead and is compatible with bfloa16
+        frames = frames.cpu().float().numpy()
+        return frames
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+    def prepare_extra_step_kwargs(self, generator, eta):
+        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+        # and should be between [0, 1]
+
+        accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        extra_step_kwargs = {}
+        if accepts_eta:
+            extra_step_kwargs["eta"] = eta
+
+        # check if the scheduler accepts generator
+        accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        if accepts_generator:
+            extra_step_kwargs["generator"] = generator
+        return extra_step_kwargs
+
+    # Copied from diffusers.pipelines.latte.pipeline_latte.LattePipeline.check_inputs
+    def check_inputs(
+        self,
+        prompt,
+        height,
+        width,
+        negative_prompt,
+        callback_on_step_end_tensor_inputs,
+        prompt_embeds=None,
+        negative_prompt_embeds=None,
+    ):
+        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}.")
+
+        if callback_on_step_end_tensor_inputs is not None and not all(
+            k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+        ):
+            raise ValueError(
+                f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+            )
+        if prompt is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt is None and prompt_embeds is None:
+            raise ValueError(
+                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+            )
+        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+        if prompt is not None and negative_prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt`: {prompt} and `negative_prompt_embeds`:"
+                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+            )
+
+        if negative_prompt is not None and negative_prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+            )
+
+        if prompt_embeds is not None and negative_prompt_embeds is not None:
+            if prompt_embeds.shape != negative_prompt_embeds.shape:
+                raise ValueError(
+                    "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+                    f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+                    f" {negative_prompt_embeds.shape}."
+                )
+
+    def fuse_qkv_projections(self) -> None:
+        r"""Enables fused QKV projections."""
+        self.fusing_transformer = True
+        self.transformer.fuse_qkv_projections()
+
+    def unfuse_qkv_projections(self) -> None:
+        r"""Disable QKV projection fusion if enabled."""
+        if not self.fusing_transformer:
+            logger.warning("The Transformer was not initially fused for QKV projections. Doing nothing.")
+        else:
+            self.transformer.unfuse_qkv_projections()
+            self.fusing_transformer = False
+
+    def _prepare_rotary_positional_embeddings(
+        self,
+        height: int,
+        width: int,
+        num_frames: int,
+        device: torch.device,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        grid_height = height // (self.vae_scale_factor_spatial * self.transformer.config.patch_size)
+        grid_width = width // (self.vae_scale_factor_spatial * self.transformer.config.patch_size)
+        base_size_width = 720 // (self.vae_scale_factor_spatial * self.transformer.config.patch_size)
+        base_size_height = 480 // (self.vae_scale_factor_spatial * self.transformer.config.patch_size)
+
+        grid_crops_coords = get_resize_crop_region_for_grid(
+            (grid_height, grid_width), base_size_width, base_size_height
+        )
+        freqs_cos, freqs_sin = get_3d_rotary_pos_embed(
+            embed_dim=self.transformer.config.attention_head_dim,
+            crops_coords=grid_crops_coords,
+            grid_size=(grid_height, grid_width),
+            temporal_size=num_frames,
+            use_real=True,
+        )
+
+        freqs_cos = freqs_cos.to(device=device)
+        freqs_sin = freqs_sin.to(device=device)
+        return freqs_cos, freqs_sin
+
+    @property
+    def guidance_scale(self):
+        return self._guidance_scale
+
+    @property
+    def num_timesteps(self):
+        return self._num_timesteps
+
+    @property
+    def interrupt(self):
+        return self._interrupt
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.StableDiffusionImg2ImgPipeline.get_timesteps
+    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
+
+    @torch.no_grad()
+    @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+        self,
+        prompt: Optional[Union[str, List[str]]] = None,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        height: int = 480,
+        width: int = 720,
+        video: Union[torch.FloatTensor] = None,
+        mask_video: Union[torch.FloatTensor] = None,
+        masked_video_latents: Union[torch.FloatTensor] = None,
+        num_frames: int = 49,
+        num_inference_steps: int = 50,
+        timesteps: Optional[List[int]] = None,
+        guidance_scale: float = 6,
+        use_dynamic_cfg: bool = False,
+        num_videos_per_prompt: int = 1,
+        eta: float = 0.0,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        output_type: str = "numpy",
+        return_dict: bool = False,
+        callback_on_step_end: Optional[
+            Union[Callable[[int, int, Dict], None], PipelineCallback, MultiPipelineCallbacks]
+        ] = None,
+        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+        max_sequence_length: int = 226,
+        strength: float = 1,
+        comfyui_progressbar: bool = False,
+    ) -> Union[CogVideoX_Fun_PipelineOutput, Tuple]:
+        """
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+                instead.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+                The height in pixels of the generated image. This is set to 1024 by default for the best results.
+            width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+                The width in pixels of the generated image. This is set to 1024 by default for the best results.
+            num_frames (`int`, defaults to `48`):
+                Number of frames to generate. Must be divisible by self.vae_scale_factor_temporal. Generated video will
+                contain 1 extra frame because CogVideoX_Fun is conditioned with (num_seconds * fps + 1) frames where
+                num_seconds is 6 and fps is 4. However, since videos can be saved at any fps, the only condition that
+                needs to be satisfied is that of divisibility mentioned above.
+            num_inference_steps (`int`, *optional*, defaults to 50):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            timesteps (`List[int]`, *optional*):
+                Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
+                in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
+                passed will be used. Must be in descending order.
+            guidance_scale (`float`, *optional*, defaults to 7.0):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+                usually at the expense of lower image quality.
+            num_videos_per_prompt (`int`, *optional*, defaults to 1):
+                The number of videos to generate per prompt.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                One or a list of [torch generator(s)](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 will ge generated by sampling using the supplied random `generator`.
+            prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generate image. Choose between
+                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] instead
+                of a plain tuple.
+            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.
+            max_sequence_length (`int`, defaults to `226`):
+                Maximum sequence length in encoded prompt. Must be consistent with
+                `self.transformer.config.max_text_seq_length` otherwise may lead to poor results.
+
+        Examples:
+
+        Returns:
+            [`~pipelines.cogvideo.pipeline_cogvideox.CogVideoX_Fun_PipelineOutput`] or `tuple`:
+            [`~pipelines.cogvideo.pipeline_cogvideox.CogVideoX_Fun_PipelineOutput`] if `return_dict` is True, otherwise a
+            `tuple`. When returning a tuple, the first element is a list with the generated images.
+        """
+
+        if num_frames > 49:
+            raise ValueError(
+                "The number of frames must be less than 49 for now due to static positional embeddings. This will be updated in the future to remove this limitation."
+            )
+
+        if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)):
+            callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs
+
+        height = height or self.transformer.config.sample_size * self.vae_scale_factor_spatial
+        width = width or self.transformer.config.sample_size * self.vae_scale_factor_spatial
+        num_videos_per_prompt = 1
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(
+            prompt,
+            height,
+            width,
+            negative_prompt,
+            callback_on_step_end_tensor_inputs,
+            prompt_embeds,
+            negative_prompt_embeds,
+        )
+        self._guidance_scale = guidance_scale
+        self._interrupt = False
+
+        # 2. Default call parameters
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.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 = guidance_scale > 1.0
+
+        # 3. Encode input prompt
+        prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+            prompt,
+            negative_prompt,
+            do_classifier_free_guidance,
+            num_videos_per_prompt=num_videos_per_prompt,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            max_sequence_length=max_sequence_length,
+            device=device,
+        )
+        if do_classifier_free_guidance:
+            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
+
+        # 4. set timesteps
+        self.scheduler.set_timesteps(num_inference_steps, device=device)
+        timesteps, num_inference_steps = self.get_timesteps(
+            num_inference_steps=num_inference_steps, strength=strength, device=device
+        )
+        self._num_timesteps = len(timesteps)
+        if comfyui_progressbar:
+            from comfy.utils import ProgressBar
+            pbar = ProgressBar(num_inference_steps + 2)
+        # at which timestep to set the initial noise (n.b. 50% if strength is 0.5)
+        latent_timestep = timesteps[:1].repeat(batch_size * num_videos_per_prompt)
+        # create a boolean to check if the strength is set to 1. if so then initialise the latents with pure noise
+        is_strength_max = strength == 1.0
+
+        # 5. Prepare latents.
+        if video is not None:
+            video_length = video.shape[2]
+            init_video = self.image_processor.preprocess(rearrange(video, "b c f h w -> (b f) c h w"), height=height, width=width) 
+            init_video = init_video.to(dtype=torch.float32)
+            init_video = rearrange(init_video, "(b f) c h w -> b c f h w", f=video_length)
+        else:
+            init_video = None
+
+        num_channels_latents = self.vae.config.latent_channels
+        num_channels_transformer = self.transformer.config.in_channels
+        return_image_latents = num_channels_transformer == num_channels_latents
+
+        latents_outputs = self.prepare_latents(
+            batch_size * num_videos_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            video_length,
+            prompt_embeds.dtype,
+            device,
+            generator,
+            latents,
+            video=init_video,
+            timestep=latent_timestep,
+            is_strength_max=is_strength_max,
+            return_noise=True,
+            return_video_latents=return_image_latents,
+        )
+        if return_image_latents:
+            latents, noise, image_latents = latents_outputs
+        else:
+            latents, noise = latents_outputs
+        if comfyui_progressbar:
+            pbar.update(1)
+        
+        if mask_video is not None:
+            if (mask_video == 255).all():
+                mask_latents = torch.zeros_like(latents)[:, :, :1].to(latents.device, latents.dtype)
+                masked_video_latents = torch.zeros_like(latents).to(latents.device, latents.dtype)
+
+                mask_input = torch.cat([mask_latents] * 2) if do_classifier_free_guidance else mask_latents
+                masked_video_latents_input = (
+                    torch.cat([masked_video_latents] * 2) if do_classifier_free_guidance else masked_video_latents
+                )
+                inpaint_latents = torch.cat([mask_input, masked_video_latents_input], dim=2).to(latents.dtype)
+            else:
+                # Prepare mask latent variables
+                video_length = video.shape[2]
+                mask_condition = self.mask_processor.preprocess(rearrange(mask_video, "b c f h w -> (b f) c h w"), height=height, width=width) 
+                mask_condition = mask_condition.to(dtype=torch.float32)
+                mask_condition = rearrange(mask_condition, "(b f) c h w -> b c f h w", f=video_length)
+
+                if num_channels_transformer != num_channels_latents:
+                    mask_condition_tile = torch.tile(mask_condition, [1, 3, 1, 1, 1])
+                    if masked_video_latents is None:
+                        masked_video = init_video * (mask_condition_tile < 0.5) + torch.ones_like(init_video) * (mask_condition_tile > 0.5) * -1
+                    else:
+                        masked_video = masked_video_latents
+
+                    _, masked_video_latents = self.prepare_mask_latents(
+                        None,
+                        masked_video,
+                        batch_size,
+                        height,
+                        width,
+                        prompt_embeds.dtype,
+                        device,
+                        generator,
+                        do_classifier_free_guidance,
+                    )
+                    mask_latents = resize_mask(1 - mask_condition, masked_video_latents)
+                    mask_latents = mask_latents.to(masked_video_latents.device) * self.vae.config.scaling_factor
+
+                    mask = torch.tile(mask_condition, [1, num_channels_latents, 1, 1, 1])
+                    mask = F.interpolate(mask, size=latents.size()[-3:], mode='trilinear', align_corners=True).to(latents.device, latents.dtype)
+                    
+                    mask_input = torch.cat([mask_latents] * 2) if do_classifier_free_guidance else mask_latents
+                    masked_video_latents_input = (
+                        torch.cat([masked_video_latents] * 2) if do_classifier_free_guidance else masked_video_latents
+                    )
+
+                    mask = rearrange(mask, "b c f h w -> b f c h w")
+                    mask_input = rearrange(mask_input, "b c f h w -> b f c h w")
+                    masked_video_latents_input = rearrange(masked_video_latents_input, "b c f h w -> b f c h w")
+
+                    inpaint_latents = torch.cat([mask_input, masked_video_latents_input], dim=2).to(latents.dtype)
+                else:
+                    mask = torch.tile(mask_condition, [1, num_channels_latents, 1, 1, 1])
+                    mask = F.interpolate(mask, size=latents.size()[-3:], mode='trilinear', align_corners=True).to(latents.device, latents.dtype)
+                    mask = rearrange(mask, "b c f h w -> b f c h w")
+                    
+                    inpaint_latents = None
+        else:
+            if num_channels_transformer != num_channels_latents:
+                mask = torch.zeros_like(latents).to(latents.device, latents.dtype)
+                masked_video_latents = torch.zeros_like(latents).to(latents.device, latents.dtype)
+
+                mask_input = torch.cat([mask] * 2) if do_classifier_free_guidance else mask
+                masked_video_latents_input = (
+                    torch.cat([masked_video_latents] * 2) if do_classifier_free_guidance else masked_video_latents
+                )
+                inpaint_latents = torch.cat([mask_input, masked_video_latents_input], dim=1).to(latents.dtype)
+            else:
+                mask = torch.zeros_like(init_video[:, :1])
+                mask = torch.tile(mask, [1, num_channels_latents, 1, 1, 1])
+                mask = F.interpolate(mask, size=latents.size()[-3:], mode='trilinear', align_corners=True).to(latents.device, latents.dtype)
+                mask = rearrange(mask, "b c f h w -> b f c h w")
+
+                inpaint_latents = None
+        if comfyui_progressbar:
+            pbar.update(1)
+
+        # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+        # 7. Create rotary embeds if required
+        image_rotary_emb = (
+            self._prepare_rotary_positional_embeddings(height, width, latents.size(1), device)
+            if self.transformer.config.use_rotary_positional_embeddings
+            else None
+        )
+
+        # 8. Denoising loop
+        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
+
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            # for DPM-solver++
+            old_pred_original_sample = None
+            for i, t in enumerate(timesteps):
+                if self.interrupt:
+                    continue
+
+                latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+                latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+                # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+                timestep = t.expand(latent_model_input.shape[0])
+
+                # predict noise model_output
+                noise_pred = self.transformer(
+                    hidden_states=latent_model_input,
+                    encoder_hidden_states=prompt_embeds,
+                    timestep=timestep,
+                    image_rotary_emb=image_rotary_emb,
+                    return_dict=False,
+                    inpaint_latents=inpaint_latents,
+                )[0]
+                noise_pred = noise_pred.float()
+
+                # perform guidance
+                if use_dynamic_cfg:
+                    self._guidance_scale = 1 + guidance_scale * (
+                        (1 - math.cos(math.pi * ((num_inference_steps - t.item()) / num_inference_steps) ** 5.0)) / 2
+                    )
+                if do_classifier_free_guidance:
+                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                    noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+                # compute the previous noisy sample x_t -> x_t-1
+                if not isinstance(self.scheduler, CogVideoXDPMScheduler):
+                    latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
+                else:
+                    latents, old_pred_original_sample = self.scheduler.step(
+                        noise_pred,
+                        old_pred_original_sample,
+                        t,
+                        timesteps[i - 1] if i > 0 else None,
+                        latents,
+                        **extra_step_kwargs,
+                        return_dict=False,
+                    )
+                latents = latents.to(prompt_embeds.dtype)
+
+                # call the callback, if provided
+                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)
+                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+                    negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+                if comfyui_progressbar:
+                    pbar.update(1)
+
+        if output_type == "numpy":
+            video = self.decode_latents(latents)
+        elif not output_type == "latent":
+            video = self.decode_latents(latents)
+            video = self.video_processor.postprocess_video(video=video, output_type=output_type)
+        else:
+            video = latents
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            video = torch.from_numpy(video)
+
+        return CogVideoX_Fun_PipelineOutput(videos=video)

cogvideox/ui/ui.py

@@ -0,0 +1,1403 @@
+"""Modified from https://github.com/guoyww/AnimateDiff/blob/main/app.py
+"""
+import base64
+import gc
+import json
+import os
+import random
+from datetime import datetime
+from glob import glob
+
+import cv2
+import gradio as gr
+import numpy as np
+import pkg_resources
+import requests
+import torch
+from diffusers import (AutoencoderKL, AutoencoderKLCogVideoX,
+                       CogVideoXDDIMScheduler, DDIMScheduler,
+                       DPMSolverMultistepScheduler,
+                       EulerAncestralDiscreteScheduler, EulerDiscreteScheduler,
+                       PNDMScheduler)
+from diffusers.utils.import_utils import is_xformers_available
+from omegaconf import OmegaConf
+from PIL import Image
+from safetensors import safe_open
+from transformers import (CLIPImageProcessor, CLIPVisionModelWithProjection,
+                          T5EncoderModel, T5Tokenizer)
+
+from cogvideox.data.bucket_sampler import ASPECT_RATIO_512, get_closest_ratio
+from ..models.autoencoder_magvit import AutoencoderKLCogVideoX
+from cogvideox.models.transformer3d import CogVideoXTransformer3DModel
+from cogvideox.pipeline.pipeline_cogvideox import CogVideoX_Fun_Pipeline
+from cogvideox.pipeline.pipeline_cogvideox_inpaint import \
+    CogVideoX_Fun_Pipeline_Inpaint
+from cogvideox.utils.lora_utils import merge_lora, unmerge_lora
+from cogvideox.utils.utils import (
+    get_image_to_video_latent, get_video_to_video_latent,
+    get_width_and_height_from_image_and_base_resolution, save_videos_grid)
+
+scheduler_dict = {
+    "Euler": EulerDiscreteScheduler,
+    "Euler A": EulerAncestralDiscreteScheduler,
+    "DPM++": DPMSolverMultistepScheduler, 
+    "PNDM": PNDMScheduler,
+    "DDIM_Cog": CogVideoXDDIMScheduler,
+    "DDIM_Origin": DDIMScheduler,
+}
+
+gradio_version = pkg_resources.get_distribution("gradio").version
+gradio_version_is_above_4 = True if int(gradio_version.split('.')[0]) >= 4 else False
+
+css = """
+.toolbutton {
+    margin-buttom: 0em 0em 0em 0em;
+    max-width: 2.5em;
+    min-width: 2.5em !important;
+    height: 2.5em;
+}
+"""
+
+class CogVideoX_I2VController:
+    def __init__(self, low_gpu_memory_mode, weight_dtype):
+        # config dirs
+        self.basedir                    = os.getcwd()
+        self.config_dir                 = os.path.join(self.basedir, "config")
+        self.diffusion_transformer_dir  = os.path.join(self.basedir, "models", "Diffusion_Transformer")
+        self.motion_module_dir          = os.path.join(self.basedir, "models", "Motion_Module")
+        self.personalized_model_dir     = os.path.join(self.basedir, "models", "Personalized_Model")
+        self.savedir                    = os.path.join(self.basedir, "samples", datetime.now().strftime("Gradio-%Y-%m-%dT%H-%M-%S"))
+        self.savedir_sample             = os.path.join(self.savedir, "sample")
+        os.makedirs(self.savedir, exist_ok=True)
+
+        self.diffusion_transformer_list = []
+        self.motion_module_list      = []
+        self.personalized_model_list = []
+        
+        self.refresh_diffusion_transformer()
+        self.refresh_motion_module()
+        self.refresh_personalized_model()
+
+        # config models
+        self.tokenizer             = None
+        self.text_encoder          = None
+        self.vae                   = None
+        self.transformer           = None
+        self.pipeline              = None
+        self.motion_module_path    = "none"
+        self.base_model_path       = "none"
+        self.lora_model_path       = "none"
+        self.low_gpu_memory_mode   = low_gpu_memory_mode
+        
+        self.weight_dtype = weight_dtype
+
+    def refresh_diffusion_transformer(self):
+        self.diffusion_transformer_list = sorted(glob(os.path.join(self.diffusion_transformer_dir, "*/")))
+
+    def refresh_motion_module(self):
+        motion_module_list = sorted(glob(os.path.join(self.motion_module_dir, "*.safetensors")))
+        self.motion_module_list = [os.path.basename(p) for p in motion_module_list]
+
+    def refresh_personalized_model(self):
+        personalized_model_list = sorted(glob(os.path.join(self.personalized_model_dir, "*.safetensors")))
+        self.personalized_model_list = [os.path.basename(p) for p in personalized_model_list]
+
+    def update_diffusion_transformer(self, diffusion_transformer_dropdown):
+        print("Update diffusion transformer")
+        if diffusion_transformer_dropdown == "none":
+            return gr.update()
+        self.vae = AutoencoderKLCogVideoX.from_pretrained(
+            diffusion_transformer_dropdown, 
+            subfolder="vae", 
+        ).to(self.weight_dtype)
+
+        # Get Transformer
+        self.transformer = CogVideoXTransformer3DModel.from_pretrained_2d(
+            diffusion_transformer_dropdown, 
+            subfolder="transformer", 
+        ).to(self.weight_dtype)
+        
+        # Get pipeline
+        if self.transformer.config.in_channels != self.vae.config.latent_channels:
+            self.pipeline = CogVideoX_Fun_Pipeline_Inpaint.from_pretrained(
+                diffusion_transformer_dropdown,
+                vae=self.vae, 
+                transformer=self.transformer,
+                scheduler=scheduler_dict["Euler"].from_pretrained(diffusion_transformer_dropdown, subfolder="scheduler"),
+                torch_dtype=self.weight_dtype
+            )
+        else:
+            self.pipeline = CogVideoX_Fun_Pipeline.from_pretrained(
+                diffusion_transformer_dropdown,
+                vae=self.vae, 
+                transformer=self.transformer,
+                scheduler=scheduler_dict["Euler"].from_pretrained(diffusion_transformer_dropdown, subfolder="scheduler"),
+                torch_dtype=self.weight_dtype
+            )
+
+        if self.low_gpu_memory_mode:
+            self.pipeline.enable_sequential_cpu_offload()
+        else:
+            self.pipeline.enable_model_cpu_offload()
+        print("Update diffusion transformer done")
+        return gr.update()
+
+    def update_base_model(self, base_model_dropdown):
+        self.base_model_path = base_model_dropdown
+        print("Update base model")
+        if base_model_dropdown == "none":
+            return gr.update()
+        if self.transformer is None:
+            gr.Info(f"Please select a pretrained model path.")
+            return gr.update(value=None)
+        else:
+            base_model_dropdown = os.path.join(self.personalized_model_dir, base_model_dropdown)
+            base_model_state_dict = {}
+            with safe_open(base_model_dropdown, framework="pt", device="cpu") as f:
+                for key in f.keys():
+                    base_model_state_dict[key] = f.get_tensor(key)
+            self.transformer.load_state_dict(base_model_state_dict, strict=False)
+            print("Update base done")
+            return gr.update()
+
+    def update_lora_model(self, lora_model_dropdown):
+        print("Update lora model")
+        if lora_model_dropdown == "none":
+            self.lora_model_path = "none"
+            return gr.update()
+        lora_model_dropdown = os.path.join(self.personalized_model_dir, lora_model_dropdown)
+        self.lora_model_path = lora_model_dropdown
+        return gr.update()
+
+    def generate(
+        self,
+        diffusion_transformer_dropdown,
+        base_model_dropdown,
+        lora_model_dropdown, 
+        lora_alpha_slider,
+        prompt_textbox, 
+        negative_prompt_textbox, 
+        sampler_dropdown, 
+        sample_step_slider, 
+        resize_method,
+        width_slider, 
+        height_slider, 
+        base_resolution, 
+        generation_method, 
+        length_slider, 
+        overlap_video_length, 
+        partial_video_length, 
+        cfg_scale_slider, 
+        start_image, 
+        end_image, 
+        validation_video,
+        denoise_strength,
+        seed_textbox,
+        is_api = False,
+    ):
+        gc.collect()
+        torch.cuda.empty_cache()
+        torch.cuda.ipc_collect()
+
+        if self.transformer is None:
+            raise gr.Error(f"Please select a pretrained model path.")
+
+        if self.base_model_path != base_model_dropdown:
+            self.update_base_model(base_model_dropdown)
+
+        if self.lora_model_path != lora_model_dropdown:
+            print("Update lora model")
+            self.update_lora_model(lora_model_dropdown)
+        
+        if resize_method == "Resize according to Reference":
+            if start_image is None and validation_video is None:
+                if is_api:
+                    return "", f"Please upload an image when using \"Resize according to Reference\"."
+                else:
+                    raise gr.Error(f"Please upload an image when using \"Resize according to Reference\".")
+
+            aspect_ratio_sample_size    = {key : [x / 512 * base_resolution for x in ASPECT_RATIO_512[key]] for key in ASPECT_RATIO_512.keys()}
+            
+            if validation_video is not None:
+                original_width, original_height = Image.fromarray(cv2.VideoCapture(validation_video).read()[1]).size
+            else:
+                original_width, original_height = start_image[0].size if type(start_image) is list else Image.open(start_image).size
+            closest_size, closest_ratio = get_closest_ratio(original_height, original_width, ratios=aspect_ratio_sample_size)
+            height_slider, width_slider = [int(x / 16) * 16 for x in closest_size]
+
+        if self.transformer.config.in_channels == self.vae.config.latent_channels and start_image is not None:
+            if is_api:
+                return "", f"Please select an image to video pretrained model while using image to video."
+            else:
+                raise gr.Error(f"Please select an image to video pretrained model while using image to video.")
+
+        if self.transformer.config.in_channels == self.vae.config.latent_channels and generation_method == "Long Video Generation":
+            if is_api:
+                return "", f"Please select an image to video pretrained model while using long video generation."
+            else:
+                raise gr.Error(f"Please select an image to video pretrained model while using long video generation.")
+        
+        if start_image is None and end_image is not None:
+            if is_api:
+                return "", f"If specifying the ending image of the video, please specify a starting image of the video."
+            else:
+                raise gr.Error(f"If specifying the ending image of the video, please specify a starting image of the video.")
+
+        is_image = True if generation_method == "Image Generation" else False
+
+        self.pipeline.scheduler = scheduler_dict[sampler_dropdown].from_config(self.pipeline.scheduler.config)
+        if self.lora_model_path != "none":
+            # lora part
+            self.pipeline = merge_lora(self.pipeline, self.lora_model_path, multiplier=lora_alpha_slider)
+
+        if int(seed_textbox) != -1 and seed_textbox != "": torch.manual_seed(int(seed_textbox))
+        else: seed_textbox = np.random.randint(0, 1e10)
+        generator = torch.Generator(device="cuda").manual_seed(int(seed_textbox))
+        
+        try:
+            if self.transformer.config.in_channels != self.vae.config.latent_channels:
+                if generation_method == "Long Video Generation":
+                    if validation_video is not None:
+                        raise gr.Error(f"Video to Video is not Support Long Video Generation now.")
+                    init_frames = 0
+                    last_frames = init_frames + partial_video_length
+                    while init_frames < length_slider:
+                        if last_frames >= length_slider:
+                            _partial_video_length = length_slider - init_frames
+                            _partial_video_length = int((_partial_video_length - 1) // self.vae.config.temporal_compression_ratio * self.vae.config.temporal_compression_ratio) + 1
+                            
+                            if _partial_video_length <= 0:
+                                break
+                        else:
+                            _partial_video_length = partial_video_length
+
+                        if last_frames >= length_slider:
+                            input_video, input_video_mask, clip_image = get_image_to_video_latent(start_image, end_image, video_length=_partial_video_length, sample_size=(height_slider, width_slider))
+                        else:
+                            input_video, input_video_mask, clip_image = get_image_to_video_latent(start_image, None, video_length=_partial_video_length, sample_size=(height_slider, width_slider))
+
+                        with torch.no_grad():
+                            sample = self.pipeline(
+                                prompt_textbox, 
+                                negative_prompt     = negative_prompt_textbox,
+                                num_inference_steps = sample_step_slider,
+                                guidance_scale      = cfg_scale_slider,
+                                width               = width_slider,
+                                height              = height_slider,
+                                num_frames          = _partial_video_length,
+                                generator           = generator,
+
+                                video        = input_video,
+                                mask_video   = input_video_mask,
+                                strength     = 1,
+                            ).videos
+                        
+                        if init_frames != 0:
+                            mix_ratio = torch.from_numpy(
+                                np.array([float(_index) / float(overlap_video_length) for _index in range(overlap_video_length)], np.float32)
+                            ).unsqueeze(0).unsqueeze(0).unsqueeze(-1).unsqueeze(-1)
+                            
+                            new_sample[:, :, -overlap_video_length:] = new_sample[:, :, -overlap_video_length:] * (1 - mix_ratio) + \
+                                sample[:, :, :overlap_video_length] * mix_ratio
+                            new_sample = torch.cat([new_sample, sample[:, :, overlap_video_length:]], dim = 2)
+
+                            sample = new_sample
+                        else:
+                            new_sample = sample
+
+                        if last_frames >= length_slider:
+                            break
+
+                        start_image = [
+                            Image.fromarray(
+                                (sample[0, :, _index].transpose(0, 1).transpose(1, 2) * 255).numpy().astype(np.uint8)
+                            ) for _index in range(-overlap_video_length, 0)
+                        ]
+
+                        init_frames = init_frames + _partial_video_length - overlap_video_length
+                        last_frames = init_frames + _partial_video_length
+                else:
+                    if validation_video is not None:
+                        input_video, input_video_mask, clip_image = get_video_to_video_latent(validation_video, length_slider if not is_image else 1, sample_size=(height_slider, width_slider))
+                        strength = denoise_strength
+                    else:
+                        input_video, input_video_mask, clip_image = get_image_to_video_latent(start_image, end_image, length_slider if not is_image else 1, sample_size=(height_slider, width_slider))
+                        strength = 1
+
+                    sample = self.pipeline(
+                        prompt_textbox,
+                        negative_prompt     = negative_prompt_textbox,
+                        num_inference_steps = sample_step_slider,
+                        guidance_scale      = cfg_scale_slider,
+                        width               = width_slider,
+                        height              = height_slider,
+                        num_frames          = length_slider if not is_image else 1,
+                        generator           = generator,
+
+                        video        = input_video,
+                        mask_video   = input_video_mask,
+                        strength     = strength,
+                    ).videos
+            else:
+                sample = self.pipeline(
+                    prompt_textbox,
+                    negative_prompt     = negative_prompt_textbox,
+                    num_inference_steps = sample_step_slider,
+                    guidance_scale      = cfg_scale_slider,
+                    width               = width_slider,
+                    height              = height_slider,
+                    num_frames          = length_slider if not is_image else 1,
+                    generator           = generator
+                ).videos
+        except Exception as e:
+            gc.collect()
+            torch.cuda.empty_cache()
+            torch.cuda.ipc_collect()
+            if self.lora_model_path != "none":
+                self.pipeline = unmerge_lora(self.pipeline, self.lora_model_path, multiplier=lora_alpha_slider)
+            if is_api:
+                return "", f"Error. error information is {str(e)}"
+            else:
+                return gr.update(), gr.update(), f"Error. error information is {str(e)}"
+
+        gc.collect()
+        torch.cuda.empty_cache()
+        torch.cuda.ipc_collect()
+
+        # lora part
+        if self.lora_model_path != "none":
+            self.pipeline = unmerge_lora(self.pipeline, self.lora_model_path, multiplier=lora_alpha_slider)
+
+        sample_config = {
+            "prompt": prompt_textbox,
+            "n_prompt": negative_prompt_textbox,
+            "sampler": sampler_dropdown,
+            "num_inference_steps": sample_step_slider,
+            "guidance_scale": cfg_scale_slider,
+            "width": width_slider,
+            "height": height_slider,
+            "video_length": length_slider,
+            "seed_textbox": seed_textbox
+        }
+        json_str = json.dumps(sample_config, indent=4)
+        with open(os.path.join(self.savedir, "logs.json"), "a") as f:
+            f.write(json_str)
+            f.write("\n\n")
+            
+        if not os.path.exists(self.savedir_sample):
+            os.makedirs(self.savedir_sample, exist_ok=True)
+        index = len([path for path in os.listdir(self.savedir_sample)]) + 1
+        prefix = str(index).zfill(3)
+
+        gc.collect()
+        torch.cuda.empty_cache()
+        torch.cuda.ipc_collect()
+        if is_image or length_slider == 1:
+            save_sample_path = os.path.join(self.savedir_sample, prefix + f".png")
+
+            image = sample[0, :, 0]
+            image = image.transpose(0, 1).transpose(1, 2)
+            image = (image * 255).numpy().astype(np.uint8)
+            image = Image.fromarray(image)
+            image.save(save_sample_path)
+
+            if is_api:
+                return save_sample_path, "Success"
+            else:
+                if gradio_version_is_above_4:
+                    return gr.Image(value=save_sample_path, visible=True), gr.Video(value=None, visible=False), "Success"
+                else:
+                    return gr.Image.update(value=save_sample_path, visible=True), gr.Video.update(value=None, visible=False), "Success"
+        else:
+            save_sample_path = os.path.join(self.savedir_sample, prefix + f".mp4")
+            save_videos_grid(sample, save_sample_path, fps=8)
+
+            if is_api:
+                return save_sample_path, "Success"
+            else:
+                if gradio_version_is_above_4:
+                    return gr.Image(visible=False, value=None), gr.Video(value=save_sample_path, visible=True), "Success"
+                else:
+                    return gr.Image.update(visible=False, value=None), gr.Video.update(value=save_sample_path, visible=True), "Success"
+
+
+def ui(low_gpu_memory_mode, weight_dtype):
+    controller = CogVideoX_I2VController(low_gpu_memory_mode, weight_dtype)
+
+    with gr.Blocks(css=css) as demo:
+        gr.Markdown(
+            """
+            # CogVideoX-Fun:
+
+            A CogVideoX with more flexible generation conditions, capable of producing videos of different resolutions, around 6 seconds, and fps 8 (frames 1 to 49), as well as image generated videos. 
+
+            [Github](https://github.com/aigc-apps/CogVideoX-Fun/)
+            """
+        )
+        with gr.Column(variant="panel"):
+            gr.Markdown(
+                """
+                ### 1. Model checkpoints (模型路径).
+                """
+            )
+            with gr.Row():
+                diffusion_transformer_dropdown = gr.Dropdown(
+                    label="Pretrained Model Path (预训练模型路径)",
+                    choices=controller.diffusion_transformer_list,
+                    value="none",
+                    interactive=True,
+                )
+                diffusion_transformer_dropdown.change(
+                    fn=controller.update_diffusion_transformer, 
+                    inputs=[diffusion_transformer_dropdown], 
+                    outputs=[diffusion_transformer_dropdown]
+                )
+                
+                diffusion_transformer_refresh_button = gr.Button(value="\U0001F503", elem_classes="toolbutton")
+                def refresh_diffusion_transformer():
+                    controller.refresh_diffusion_transformer()
+                    return gr.update(choices=controller.diffusion_transformer_list)
+                diffusion_transformer_refresh_button.click(fn=refresh_diffusion_transformer, inputs=[], outputs=[diffusion_transformer_dropdown])
+
+            with gr.Row():
+                base_model_dropdown = gr.Dropdown(
+                    label="Select base Dreambooth model (选择基模型[非必需])",
+                    choices=controller.personalized_model_list,
+                    value="none",
+                    interactive=True,
+                )
+                
+                lora_model_dropdown = gr.Dropdown(
+                    label="Select LoRA model (选择LoRA模型[非必需])",
+                    choices=["none"] + controller.personalized_model_list,
+                    value="none",
+                    interactive=True,
+                )
+
+                lora_alpha_slider = gr.Slider(label="LoRA alpha (LoRA权重)", value=0.55, minimum=0, maximum=2, interactive=True)
+                
+                personalized_refresh_button = gr.Button(value="\U0001F503", elem_classes="toolbutton")
+                def update_personalized_model():
+                    controller.refresh_personalized_model()
+                    return [
+                        gr.update(choices=controller.personalized_model_list),
+                        gr.update(choices=["none"] + controller.personalized_model_list)
+                    ]
+                personalized_refresh_button.click(fn=update_personalized_model, inputs=[], outputs=[base_model_dropdown, lora_model_dropdown])
+
+        with gr.Column(variant="panel"):
+            gr.Markdown(
+                """
+                ### 2. Configs for Generation (生成参数配置).
+                """
+            )
+            
+            prompt_textbox = gr.Textbox(label="Prompt (正向提示词)", lines=2, value="A young woman with beautiful and clear eyes and blonde hair standing and white dress in a forest wearing a crown. She seems to be lost in thought, and the camera focuses on her face. The video is of high quality, and the view is very clear. High quality, masterpiece, best quality, highres, ultra-detailed, fantastic.")
+            negative_prompt_textbox = gr.Textbox(label="Negative prompt (负向提示词)", lines=2, value="The video is not of a high quality, it has a low resolution. Watermark present in each frame. Strange motion trajectory. " )
+                
+            with gr.Row():
+                with gr.Column():
+                    with gr.Row():
+                        sampler_dropdown   = gr.Dropdown(label="Sampling method (采样器种类)", choices=list(scheduler_dict.keys()), value=list(scheduler_dict.keys())[0])
+                        sample_step_slider = gr.Slider(label="Sampling steps (生成步数)", value=50, minimum=10, maximum=100, step=1)
+                        
+                    resize_method = gr.Radio(
+                        ["Generate by", "Resize according to Reference"],
+                        value="Generate by",
+                        show_label=False,
+                    )
+                    width_slider     = gr.Slider(label="Width (视频宽度)",            value=672, minimum=128, maximum=1344, step=16)
+                    height_slider    = gr.Slider(label="Height (视频高度)",           value=384, minimum=128, maximum=1344, step=16)
+                    base_resolution  = gr.Radio(label="Base Resolution of Pretrained Models", value=512, choices=[512, 768, 960], visible=False)
+
+                    with gr.Group():
+                        generation_method = gr.Radio(
+                            ["Video Generation", "Image Generation", "Long Video Generation"],
+                            value="Video Generation",
+                            show_label=False,
+                        )
+                        with gr.Row():
+                            length_slider = gr.Slider(label="Animation length (视频帧数)", value=49, minimum=1,   maximum=49,  step=4)
+                            overlap_video_length = gr.Slider(label="Overlap length (视频续写的重叠帧数)", value=4, minimum=1,   maximum=4,  step=1, visible=False)
+                            partial_video_length = gr.Slider(label="Partial video generation length (每个部分的视频生成帧数)", value=25, minimum=5,   maximum=49,  step=4, visible=False)
+                    
+                    source_method = gr.Radio(
+                        ["Text to Video (文本到视频)", "Image to Video (图片到视频)", "Video to Video (视频到视频)"],
+                        value="Text to Video (文本到视频)",
+                        show_label=False,
+                    )
+                    with gr.Column(visible = False) as image_to_video_col:
+                        start_image = gr.Image(
+                            label="The image at the beginning of the video (图片到视频的开始图片)",  show_label=True, 
+                            elem_id="i2v_start", sources="upload", type="filepath", 
+                        )
+                        
+                        template_gallery_path = ["asset/1.png", "asset/2.png", "asset/3.png", "asset/4.png", "asset/5.png"]
+                        def select_template(evt: gr.SelectData):
+                            text = {
+                                "asset/1.png": "The dog is looking at camera and smiling. The video is of high quality, and the view is very clear. High quality, masterpiece, best quality, highres, ultra-detailed, fantastic.", 
+                                "asset/2.png": "a sailboat sailing in rough seas with a dramatic sunset. The video is of high quality, and the view is very clear. High quality, masterpiece, best quality, highres, ultra-detailed, fantastic.", 
+                                "asset/3.png": "a beautiful woman with long hair and a dress blowing in the wind. The video is of high quality, and the view is very clear. High quality, masterpiece, best quality, highres, ultra-detailed, fantastic.", 
+                                "asset/4.png": "a man in an astronaut suit playing a guitar. The video is of high quality, and the view is very clear. High quality, masterpiece, best quality, highres, ultra-detailed, fantastic.", 
+                                "asset/5.png": "fireworks display over night city. The video is of high quality, and the view is very clear. High quality, masterpiece, best quality, highres, ultra-detailed, fantastic.", 
+                            }[template_gallery_path[evt.index]]
+                            return template_gallery_path[evt.index], text
+
+                        template_gallery = gr.Gallery(
+                            template_gallery_path,
+                            columns=5, rows=1,
+                            height=140,
+                            allow_preview=False,
+                            container=False,
+                            label="Template Examples",
+                        )
+                        template_gallery.select(select_template, None, [start_image, prompt_textbox])
+                        
+                        with gr.Accordion("The image at the ending of the video (图片到视频的结束图片[非必需, Optional])", open=False):
+                            end_image   = gr.Image(label="The image at the ending of the video (图片到视频的结束图片[非必需, Optional])", show_label=False, elem_id="i2v_end", sources="upload", type="filepath")
+
+                    with gr.Column(visible = False) as video_to_video_col:
+                        validation_video = gr.Video(
+                            label="The video to convert (视频转视频的参考视频)",  show_label=True, 
+                            elem_id="v2v", sources="upload", 
+                        )
+                        denoise_strength = gr.Slider(label="Denoise strength (重绘系数)", value=0.70, minimum=0.10, maximum=0.95, step=0.01)
+
+                    cfg_scale_slider  = gr.Slider(label="CFG Scale (引导系数)",        value=7.0, minimum=0,   maximum=20)
+                    
+                    with gr.Row():
+                        seed_textbox = gr.Textbox(label="Seed (随机种子)", value=43)
+                        seed_button  = gr.Button(value="\U0001F3B2", elem_classes="toolbutton")
+                        seed_button.click(
+                            fn=lambda: gr.Textbox(value=random.randint(1, 1e8)) if gradio_version_is_above_4 else gr.Textbox.update(value=random.randint(1, 1e8)), 
+                            inputs=[], 
+                            outputs=[seed_textbox]
+                        )
+
+                    generate_button = gr.Button(value="Generate (生成)", variant='primary')
+                    
+                with gr.Column():
+                    result_image = gr.Image(label="Generated Image (生成图片)", interactive=False, visible=False)
+                    result_video = gr.Video(label="Generated Animation (生成视频)", interactive=False)
+                    infer_progress = gr.Textbox(
+                        label="Generation Info (生成信息)",
+                        value="No task currently",
+                        interactive=False
+                    )
+
+            def upload_generation_method(generation_method):
+                if generation_method == "Video Generation":
+                    return [gr.update(visible=True, maximum=49, value=49), gr.update(visible=False), gr.update(visible=False)]
+                elif generation_method == "Image Generation":
+                    return [gr.update(visible=False), gr.update(visible=False), gr.update(visible=False)]
+                else:
+                    return [gr.update(visible=True, maximum=1344), gr.update(visible=True), gr.update(visible=True)]
+            generation_method.change(
+                upload_generation_method, generation_method, [length_slider, overlap_video_length, partial_video_length]
+            )
+
+            def upload_source_method(source_method):
+                if source_method == "Text to Video (文本到视频)":
+                    return [gr.update(visible=False), gr.update(visible=False), gr.update(value=None), gr.update(value=None), gr.update(value=None)]
+                elif source_method == "Image to Video (图片到视频)":
+                    return [gr.update(visible=True), gr.update(visible=False), gr.update(), gr.update(), gr.update(value=None)]
+                else:
+                    return [gr.update(visible=False), gr.update(visible=True), gr.update(value=None), gr.update(value=None), gr.update()]
+            source_method.change(
+                upload_source_method, source_method, [image_to_video_col, video_to_video_col, start_image, end_image, validation_video]
+            )
+
+            def upload_resize_method(resize_method):
+                if resize_method == "Generate by":
+                    return [gr.update(visible=True), gr.update(visible=True), gr.update(visible=False)]
+                else:
+                    return [gr.update(visible=False), gr.update(visible=False), gr.update(visible=True)]
+            resize_method.change(
+                upload_resize_method, resize_method, [width_slider, height_slider, base_resolution]
+            )
+
+            generate_button.click(
+                fn=controller.generate,
+                inputs=[
+                    diffusion_transformer_dropdown,
+                    base_model_dropdown,
+                    lora_model_dropdown,
+                    lora_alpha_slider,
+                    prompt_textbox, 
+                    negative_prompt_textbox, 
+                    sampler_dropdown, 
+                    sample_step_slider, 
+                    resize_method,
+                    width_slider, 
+                    height_slider, 
+                    base_resolution, 
+                    generation_method, 
+                    length_slider, 
+                    overlap_video_length, 
+                    partial_video_length, 
+                    cfg_scale_slider, 
+                    start_image, 
+                    end_image, 
+                    validation_video,
+                    denoise_strength, 
+                    seed_textbox,
+                ],
+                outputs=[result_image, result_video, infer_progress]
+            )
+    return demo, controller
+
+
+class CogVideoX_I2VController_Modelscope:
+    def __init__(self, model_name, savedir_sample, low_gpu_memory_mode, weight_dtype):
+        # Basic dir
+        self.basedir                    = os.getcwd()
+        self.personalized_model_dir     = os.path.join(self.basedir, "models", "Personalized_Model")
+        self.lora_model_path            = "none"
+        self.savedir_sample             = savedir_sample
+        self.refresh_personalized_model()
+        os.makedirs(self.savedir_sample, exist_ok=True)
+
+        # model path
+        self.weight_dtype = weight_dtype
+        
+        self.vae = AutoencoderKLCogVideoX.from_pretrained(
+            model_name, 
+            subfolder="vae", 
+        ).to(self.weight_dtype)
+
+        # Get Transformer
+        self.transformer = CogVideoXTransformer3DModel.from_pretrained_2d(
+            model_name, 
+            subfolder="transformer", 
+        ).to(self.weight_dtype)
+        
+        # Get pipeline
+        if self.transformer.config.in_channels != self.vae.config.latent_channels:
+            self.pipeline = CogVideoX_Fun_Pipeline_Inpaint.from_pretrained(
+                model_name,
+                vae=self.vae, 
+                transformer=self.transformer,
+                scheduler=scheduler_dict["Euler"].from_pretrained(model_name, subfolder="scheduler"),
+                torch_dtype=self.weight_dtype
+            )
+        else:
+            self.pipeline = CogVideoX_Fun_Pipeline.from_pretrained(
+                model_name,
+                vae=self.vae, 
+                transformer=self.transformer,
+                scheduler=scheduler_dict["Euler"].from_pretrained(model_name, subfolder="scheduler"),
+                torch_dtype=self.weight_dtype
+            )
+
+        if low_gpu_memory_mode:
+            self.pipeline.enable_sequential_cpu_offload()
+        else:
+            self.pipeline.enable_model_cpu_offload()
+        print("Update diffusion transformer done")
+
+
+    def refresh_personalized_model(self):
+        personalized_model_list = sorted(glob(os.path.join(self.personalized_model_dir, "*.safetensors")))
+        self.personalized_model_list = [os.path.basename(p) for p in personalized_model_list]
+
+
+    def update_lora_model(self, lora_model_dropdown):
+        print("Update lora model")
+        if lora_model_dropdown == "none":
+            self.lora_model_path = "none"
+            return gr.update()
+        lora_model_dropdown = os.path.join(self.personalized_model_dir, lora_model_dropdown)
+        self.lora_model_path = lora_model_dropdown
+        return gr.update()
+
+    
+    def generate(
+        self,
+        diffusion_transformer_dropdown,
+        base_model_dropdown,
+        lora_model_dropdown, 
+        lora_alpha_slider,
+        prompt_textbox, 
+        negative_prompt_textbox, 
+        sampler_dropdown, 
+        sample_step_slider, 
+        resize_method,
+        width_slider, 
+        height_slider, 
+        base_resolution, 
+        generation_method, 
+        length_slider, 
+        overlap_video_length, 
+        partial_video_length, 
+        cfg_scale_slider, 
+        start_image, 
+        end_image, 
+        validation_video,
+        denoise_strength,
+        seed_textbox,
+        is_api = False,
+    ):    
+        gc.collect()
+        torch.cuda.empty_cache()
+        torch.cuda.ipc_collect()
+
+        if self.transformer is None:
+            raise gr.Error(f"Please select a pretrained model path.")
+
+        if self.lora_model_path != lora_model_dropdown:
+            print("Update lora model")
+            self.update_lora_model(lora_model_dropdown)
+
+        if resize_method == "Resize according to Reference":
+            if start_image is None and validation_video is None:
+                raise gr.Error(f"Please upload an image when using \"Resize according to Reference\".")
+        
+            aspect_ratio_sample_size = {key : [x / 512 * base_resolution for x in ASPECT_RATIO_512[key]] for key in ASPECT_RATIO_512.keys()}
+            
+            if validation_video is not None:
+                original_width, original_height = Image.fromarray(cv2.VideoCapture(validation_video).read()[1]).size
+            else:
+                original_width, original_height = start_image[0].size if type(start_image) is list else Image.open(start_image).size
+            closest_size, closest_ratio = get_closest_ratio(original_height, original_width, ratios=aspect_ratio_sample_size)
+            height_slider, width_slider = [int(x / 16) * 16 for x in closest_size]
+
+        if self.transformer.config.in_channels == self.vae.config.latent_channels and start_image is not None:
+            raise gr.Error(f"Please select an image to video pretrained model while using image to video.")
+        
+        if start_image is None and end_image is not None:
+            raise gr.Error(f"If specifying the ending image of the video, please specify a starting image of the video.")
+
+        is_image = True if generation_method == "Image Generation" else False
+
+        self.pipeline.scheduler = scheduler_dict[sampler_dropdown].from_config(self.pipeline.scheduler.config)
+        if self.lora_model_path != "none":
+            # lora part
+            self.pipeline = merge_lora(self.pipeline, self.lora_model_path, multiplier=lora_alpha_slider)
+
+        if int(seed_textbox) != -1 and seed_textbox != "": torch.manual_seed(int(seed_textbox))
+        else: seed_textbox = np.random.randint(0, 1e10)
+        generator = torch.Generator(device="cuda").manual_seed(int(seed_textbox))
+        
+        try:
+            if self.transformer.config.in_channels != self.vae.config.latent_channels:
+                if validation_video is not None:
+                    input_video, input_video_mask, clip_image = get_video_to_video_latent(validation_video, length_slider if not is_image else 1, sample_size=(height_slider, width_slider))
+                    strength = denoise_strength
+                else:
+                    input_video, input_video_mask, clip_image = get_image_to_video_latent(start_image, end_image, length_slider if not is_image else 1, sample_size=(height_slider, width_slider))
+                    strength = 1
+
+                sample = self.pipeline(
+                    prompt_textbox,
+                    negative_prompt     = negative_prompt_textbox,
+                    num_inference_steps = sample_step_slider,
+                    guidance_scale      = cfg_scale_slider,
+                    width               = width_slider,
+                    height              = height_slider,
+                    num_frames          = length_slider if not is_image else 1,
+                    generator           = generator,
+
+                    video        = input_video,
+                    mask_video   = input_video_mask,
+                    strength     = strength,
+                ).videos
+            else:
+                sample = self.pipeline(
+                    prompt_textbox,
+                    negative_prompt     = negative_prompt_textbox,
+                    num_inference_steps = sample_step_slider,
+                    guidance_scale      = cfg_scale_slider,
+                    width               = width_slider,
+                    height              = height_slider,
+                    num_frames          = length_slider if not is_image else 1,
+                    generator           = generator
+                ).videos
+        except Exception as e:
+            gc.collect()
+            torch.cuda.empty_cache()
+            torch.cuda.ipc_collect()
+            if self.lora_model_path != "none":
+                self.pipeline = unmerge_lora(self.pipeline, self.lora_model_path, multiplier=lora_alpha_slider)
+            if is_api:
+                return "", f"Error. error information is {str(e)}"
+            else:
+                return gr.update(), gr.update(), f"Error. error information is {str(e)}"
+
+        gc.collect()
+        torch.cuda.empty_cache()
+        torch.cuda.ipc_collect()
+        
+        # lora part
+        if self.lora_model_path != "none":
+            self.pipeline = unmerge_lora(self.pipeline, self.lora_model_path, multiplier=lora_alpha_slider)
+
+        if not os.path.exists(self.savedir_sample):
+            os.makedirs(self.savedir_sample, exist_ok=True)
+        index = len([path for path in os.listdir(self.savedir_sample)]) + 1
+        prefix = str(index).zfill(3)
+        
+        gc.collect()
+        torch.cuda.empty_cache()
+        torch.cuda.ipc_collect()
+        if is_image or length_slider == 1:
+            save_sample_path = os.path.join(self.savedir_sample, prefix + f".png")
+
+            image = sample[0, :, 0]
+            image = image.transpose(0, 1).transpose(1, 2)
+            image = (image * 255).numpy().astype(np.uint8)
+            image = Image.fromarray(image)
+            image.save(save_sample_path)
+            if is_api:
+                return save_sample_path, "Success"
+            else:
+                if gradio_version_is_above_4:
+                    return gr.Image(value=save_sample_path, visible=True), gr.Video(value=None, visible=False), "Success"
+                else:
+                    return gr.Image.update(value=save_sample_path, visible=True), gr.Video.update(value=None, visible=False), "Success"
+        else:
+            save_sample_path = os.path.join(self.savedir_sample, prefix + f".mp4")
+            save_videos_grid(sample, save_sample_path, fps=8)
+            if is_api:
+                return save_sample_path, "Success"
+            else:
+                if gradio_version_is_above_4:
+                    return gr.Image(visible=False, value=None), gr.Video(value=save_sample_path, visible=True), "Success"
+                else:
+                    return gr.Image.update(visible=False, value=None), gr.Video.update(value=save_sample_path, visible=True), "Success"
+
+
+def ui_modelscope(model_name, savedir_sample, low_gpu_memory_mode, weight_dtype):
+    controller = CogVideoX_I2VController_Modelscope(model_name, savedir_sample, low_gpu_memory_mode, weight_dtype)
+
+    with gr.Blocks(css=css) as demo:
+        gr.Markdown(
+            """
+            # CogVideoX-Fun
+
+            A CogVideoX with more flexible generation conditions, capable of producing videos of different resolutions, around 6 seconds, and fps 8 (frames 1 to 49), as well as image generated videos. 
+
+            [Github](https://github.com/aigc-apps/CogVideoX-Fun/)
+            """
+        )
+        with gr.Column(variant="panel"):
+            gr.Markdown(
+                """
+                ### 1. Model checkpoints (模型路径).
+                """
+            )
+            with gr.Row():
+                diffusion_transformer_dropdown = gr.Dropdown(
+                    label="Pretrained Model Path (预训练模型路径)",
+                    choices=[model_name],
+                    value=model_name,
+                    interactive=False,
+                )
+            with gr.Row():
+                base_model_dropdown = gr.Dropdown(
+                    label="Select base Dreambooth model (选择基模型[非必需])",
+                    choices=["none"],
+                    value="none",
+                    interactive=False,
+                    visible=False
+                )
+                with gr.Column(visible=False):
+                    gr.Markdown(
+                        """
+                        ### Minimalism is an example portrait of Lora, triggered by specific prompt words. More details can be found on [Wiki](https://github.com/aigc-apps/CogVideoX-Fun/wiki/Training-Lora).
+                        """
+                    )
+                    with gr.Row():
+                        lora_model_dropdown = gr.Dropdown(
+                            label="Select LoRA model",
+                            choices=["none"],
+                            value="none",
+                            interactive=True,
+                        )
+
+                        lora_alpha_slider = gr.Slider(label="LoRA alpha (LoRA权重)", value=0.55, minimum=0, maximum=2, interactive=True)
+                
+        with gr.Column(variant="panel"):
+            gr.Markdown(
+                """
+                ### 2. Configs for Generation (生成参数配置).
+                """
+            )
+
+            prompt_textbox = gr.Textbox(label="Prompt (正向提示词)", lines=2, value="A young woman with beautiful and clear eyes and blonde hair standing and white dress in a forest wearing a crown. She seems to be lost in thought, and the camera focuses on her face. The video is of high quality, and the view is very clear. High quality, masterpiece, best quality, highres, ultra-detailed, fantastic.")
+            negative_prompt_textbox = gr.Textbox(label="Negative prompt (负向提示词)", lines=2, value="The video is not of a high quality, it has a low resolution. Watermark present in each frame. Strange motion trajectory. " )
+                
+            with gr.Row():
+                with gr.Column():
+                    with gr.Row():
+                        sampler_dropdown   = gr.Dropdown(label="Sampling method (采样器种类)", choices=list(scheduler_dict.keys()), value=list(scheduler_dict.keys())[0])
+                        sample_step_slider = gr.Slider(label="Sampling steps (生成步数)", value=50, minimum=10, maximum=50, step=1, interactive=False)
+                    
+                    resize_method = gr.Radio(
+                        ["Generate by", "Resize according to Reference"],
+                        value="Generate by",
+                        show_label=False,
+                    )                        
+                    width_slider     = gr.Slider(label="Width (视频宽度)",            value=672, minimum=128, maximum=1280, step=16, interactive=False)
+                    height_slider    = gr.Slider(label="Height (视频高度)",           value=384, minimum=128, maximum=1280, step=16, interactive=False)
+                    base_resolution  = gr.Radio(label="Base Resolution of Pretrained Models", value=512, choices=[512, 768, 960], interactive=False, visible=False)
+
+                    with gr.Group():
+                        generation_method = gr.Radio(
+                            ["Video Generation", "Image Generation"],
+                            value="Video Generation",
+                            show_label=False,
+                            visible=True,
+                        )
+                        length_slider = gr.Slider(label="Animation length (视频帧数)", value=49, minimum=5,   maximum=49,  step=4)
+                        overlap_video_length = gr.Slider(label="Overlap length (视频续写的重叠帧数)", value=4, minimum=1,   maximum=4,  step=1, visible=False)
+                        partial_video_length = gr.Slider(label="Partial video generation length (每个部分的视频生成帧数)", value=25, minimum=5,   maximum=49,  step=4, visible=False)
+                        
+                    source_method = gr.Radio(
+                        ["Text to Video (文本到视频)", "Image to Video (图片到视频)", "Video to Video (视频到视频)"],
+                        value="Text to Video (文本到视频)",
+                        show_label=False,
+                    )
+                    with gr.Column(visible = False) as image_to_video_col:
+                        with gr.Row():
+                            start_image = gr.Image(label="The image at the beginning of the video (图片到视频的开始图片)", show_label=True, elem_id="i2v_start", sources="upload", type="filepath")
+                        
+                        template_gallery_path = ["asset/1.png", "asset/2.png", "asset/3.png", "asset/4.png", "asset/5.png"]
+                        def select_template(evt: gr.SelectData):
+                            text = {
+                                "asset/1.png": "The dog is looking at camera and smiling. The video is of high quality, and the view is very clear. High quality, masterpiece, best quality, highres, ultra-detailed, fantastic.", 
+                                "asset/2.png": "a sailboat sailing in rough seas with a dramatic sunset. The video is of high quality, and the view is very clear. High quality, masterpiece, best quality, highres, ultra-detailed, fantastic.", 
+                                "asset/3.png": "a beautiful woman with long hair and a dress blowing in the wind. The video is of high quality, and the view is very clear. High quality, masterpiece, best quality, highres, ultra-detailed, fantastic.", 
+                                "asset/4.png": "a man in an astronaut suit playing a guitar. The video is of high quality, and the view is very clear. High quality, masterpiece, best quality, highres, ultra-detailed, fantastic.", 
+                                "asset/5.png": "fireworks display over night city. The video is of high quality, and the view is very clear. High quality, masterpiece, best quality, highres, ultra-detailed, fantastic.", 
+                            }[template_gallery_path[evt.index]]
+                            return template_gallery_path[evt.index], text
+
+                        template_gallery = gr.Gallery(
+                            template_gallery_path,
+                            columns=5, rows=1,
+                            height=140,
+                            allow_preview=False,
+                            container=False,
+                            label="Template Examples",
+                        )
+                        template_gallery.select(select_template, None, [start_image, prompt_textbox])
+
+                        with gr.Accordion("The image at the ending of the video (图片到视频的结束图片[非必需, Optional])", open=False):
+                            end_image   = gr.Image(label="The image at the ending of the video (图片到视频的结束图片[非必需, Optional])", show_label=False, elem_id="i2v_end", sources="upload", type="filepath")
+
+                    with gr.Column(visible = False) as video_to_video_col:
+                        validation_video = gr.Video(
+                            label="The video to convert (视频转视频的参考视频)",  show_label=True, 
+                            elem_id="v2v", sources="upload", 
+                        )
+                        denoise_strength = gr.Slider(label="Denoise strength (重绘系数)", value=0.70, minimum=0.10, maximum=0.95, step=0.01)
+
+                    cfg_scale_slider  = gr.Slider(label="CFG Scale (引导系数)",        value=7.0, minimum=0,   maximum=20)
+                    
+                    with gr.Row():
+                        seed_textbox = gr.Textbox(label="Seed (随机种子)", value=43)
+                        seed_button  = gr.Button(value="\U0001F3B2", elem_classes="toolbutton")
+                        seed_button.click(
+                            fn=lambda: gr.Textbox(value=random.randint(1, 1e8)) if gradio_version_is_above_4 else gr.Textbox.update(value=random.randint(1, 1e8)), 
+                            inputs=[], 
+                            outputs=[seed_textbox]
+                        )
+
+                    generate_button = gr.Button(value="Generate (生成)", variant='primary')
+                    
+                with gr.Column():
+                    result_image = gr.Image(label="Generated Image (生成图片)", interactive=False, visible=False)
+                    result_video = gr.Video(label="Generated Animation (生成视频)", interactive=False)
+                    infer_progress = gr.Textbox(
+                        label="Generation Info (生成信息)",
+                        value="No task currently",
+                        interactive=False
+                    )
+
+            def upload_generation_method(generation_method):
+                if generation_method == "Video Generation":
+                    return gr.update(visible=True, minimum=8, maximum=49, value=49, interactive=True)
+                elif generation_method == "Image Generation":
+                    return gr.update(minimum=1, maximum=1, value=1, interactive=False)
+            generation_method.change(
+                upload_generation_method, generation_method, [length_slider]
+            )
+
+            def upload_source_method(source_method):
+                if source_method == "Text to Video (文本到视频)":
+                    return [gr.update(visible=False), gr.update(visible=False), gr.update(value=None), gr.update(value=None), gr.update(value=None)]
+                elif source_method == "Image to Video (图片到视频)":
+                    return [gr.update(visible=True), gr.update(visible=False), gr.update(), gr.update(), gr.update(value=None)]
+                else:
+                    return [gr.update(visible=False), gr.update(visible=True), gr.update(value=None), gr.update(value=None), gr.update()]
+            source_method.change(
+                upload_source_method, source_method, [image_to_video_col, video_to_video_col, start_image, end_image, validation_video]
+            )
+
+            def upload_resize_method(resize_method):
+                if resize_method == "Generate by":
+                    return [gr.update(visible=True), gr.update(visible=True), gr.update(visible=False)]
+                else:
+                    return [gr.update(visible=False), gr.update(visible=False), gr.update(visible=True)]
+            resize_method.change(
+                upload_resize_method, resize_method, [width_slider, height_slider, base_resolution]
+            )
+
+            generate_button.click(
+                fn=controller.generate,
+                inputs=[
+                    diffusion_transformer_dropdown,
+                    base_model_dropdown,
+                    lora_model_dropdown, 
+                    lora_alpha_slider,
+                    prompt_textbox, 
+                    negative_prompt_textbox, 
+                    sampler_dropdown, 
+                    sample_step_slider, 
+                    resize_method,
+                    width_slider, 
+                    height_slider, 
+                    base_resolution, 
+                    generation_method, 
+                    length_slider, 
+                    overlap_video_length, 
+                    partial_video_length, 
+                    cfg_scale_slider, 
+                    start_image, 
+                    end_image, 
+                    validation_video,
+                    denoise_strength, 
+                    seed_textbox,
+                ],
+                outputs=[result_image, result_video, infer_progress]
+            )
+    return demo, controller
+
+
+def post_eas(
+    diffusion_transformer_dropdown,
+    base_model_dropdown, lora_model_dropdown, lora_alpha_slider,
+    prompt_textbox, negative_prompt_textbox, 
+    sampler_dropdown, sample_step_slider, resize_method, width_slider, height_slider,
+    base_resolution, generation_method, length_slider, cfg_scale_slider, 
+    start_image, end_image, validation_video, denoise_strength, seed_textbox,
+):
+    if start_image is not None:
+        with open(start_image, 'rb') as file:
+            file_content = file.read()
+            start_image_encoded_content = base64.b64encode(file_content)
+            start_image = start_image_encoded_content.decode('utf-8')
+
+    if end_image is not None:
+        with open(end_image, 'rb') as file:
+            file_content = file.read()
+            end_image_encoded_content = base64.b64encode(file_content)
+            end_image = end_image_encoded_content.decode('utf-8')
+
+    if validation_video is not None:
+        with open(validation_video, 'rb') as file:
+            file_content = file.read()
+            validation_video_encoded_content = base64.b64encode(file_content)
+            validation_video = validation_video_encoded_content.decode('utf-8')
+
+    datas = {
+        "base_model_path": base_model_dropdown,
+        "lora_model_path": lora_model_dropdown, 
+        "lora_alpha_slider": lora_alpha_slider, 
+        "prompt_textbox": prompt_textbox, 
+        "negative_prompt_textbox": negative_prompt_textbox, 
+        "sampler_dropdown": sampler_dropdown, 
+        "sample_step_slider": sample_step_slider, 
+        "resize_method": resize_method,
+        "width_slider": width_slider, 
+        "height_slider": height_slider, 
+        "base_resolution": base_resolution,
+        "generation_method": generation_method,
+        "length_slider": length_slider,
+        "cfg_scale_slider": cfg_scale_slider,
+        "start_image": start_image,
+        "end_image": end_image,
+        "validation_video": validation_video,
+        "denoise_strength": denoise_strength,
+        "seed_textbox": seed_textbox,
+    }
+
+    session = requests.session()
+    session.headers.update({"Authorization": os.environ.get("EAS_TOKEN")})
+
+    response = session.post(url=f'{os.environ.get("EAS_URL")}/cogvideox_fun/infer_forward', json=datas, timeout=300)
+
+    outputs = response.json()
+    return outputs
+
+
+class CogVideoX_I2VController_EAS:
+    def __init__(self, edition, config_path, model_name, savedir_sample):
+        self.savedir_sample = savedir_sample
+        os.makedirs(self.savedir_sample, exist_ok=True)
+
+    def generate(
+        self,
+        diffusion_transformer_dropdown,
+        base_model_dropdown,
+        lora_model_dropdown, 
+        lora_alpha_slider,
+        prompt_textbox, 
+        negative_prompt_textbox, 
+        sampler_dropdown, 
+        sample_step_slider, 
+        resize_method,
+        width_slider, 
+        height_slider, 
+        base_resolution, 
+        generation_method, 
+        length_slider, 
+        cfg_scale_slider, 
+        start_image, 
+        end_image, 
+        validation_video, 
+        denoise_strength,
+        seed_textbox
+    ):
+        is_image = True if generation_method == "Image Generation" else False
+
+        outputs = post_eas(
+            diffusion_transformer_dropdown,
+            base_model_dropdown, lora_model_dropdown, lora_alpha_slider,
+            prompt_textbox, negative_prompt_textbox, 
+            sampler_dropdown, sample_step_slider, resize_method, width_slider, height_slider,
+            base_resolution, generation_method, length_slider, cfg_scale_slider, 
+            start_image, end_image, validation_video, denoise_strength, 
+            seed_textbox
+        )
+        try:
+            base64_encoding = outputs["base64_encoding"]
+        except:
+            return gr.Image(visible=False, value=None), gr.Video(None, visible=True), outputs["message"]
+            
+        decoded_data = base64.b64decode(base64_encoding)
+
+        if not os.path.exists(self.savedir_sample):
+            os.makedirs(self.savedir_sample, exist_ok=True)
+        index = len([path for path in os.listdir(self.savedir_sample)]) + 1
+        prefix = str(index).zfill(3)
+        
+        if is_image or length_slider == 1:
+            save_sample_path = os.path.join(self.savedir_sample, prefix + f".png")
+            with open(save_sample_path, "wb") as file:
+                file.write(decoded_data)
+            if gradio_version_is_above_4:
+                return gr.Image(value=save_sample_path, visible=True), gr.Video(value=None, visible=False), "Success"
+            else:
+                return gr.Image.update(value=save_sample_path, visible=True), gr.Video.update(value=None, visible=False), "Success"
+        else:
+            save_sample_path = os.path.join(self.savedir_sample, prefix + f".mp4")
+            with open(save_sample_path, "wb") as file:
+                file.write(decoded_data)
+            if gradio_version_is_above_4:
+                return gr.Image(visible=False, value=None), gr.Video(value=save_sample_path, visible=True), "Success"
+            else:
+                return gr.Image.update(visible=False, value=None), gr.Video.update(value=save_sample_path, visible=True), "Success"
+
+
+def ui_eas(model_name, savedir_sample):
+    controller = CogVideoX_I2VController_EAS(model_name, savedir_sample)
+
+    with gr.Blocks(css=css) as demo:
+        gr.Markdown(
+            """
+            # CogVideoX-Fun
+
+            A CogVideoX with more flexible generation conditions, capable of producing videos of different resolutions, around 6 seconds, and fps 8 (frames 1 to 49), as well as image generated videos. 
+
+            [Github](https://github.com/aigc-apps/CogVideoX-Fun/)
+            """
+        )
+        with gr.Column(variant="panel"):
+            gr.Markdown(
+                """
+                ### 1. Model checkpoints.
+                """
+            )
+            with gr.Row():
+                diffusion_transformer_dropdown = gr.Dropdown(
+                    label="Pretrained Model Path",
+                    choices=[model_name],
+                    value=model_name,
+                    interactive=False,
+                )
+            with gr.Row():
+                base_model_dropdown = gr.Dropdown(
+                    label="Select base Dreambooth model",
+                    choices=["none"],
+                    value="none",
+                    interactive=False,
+                    visible=False
+                )
+                with gr.Column(visible=False):
+                    gr.Markdown(
+                        """
+                        ### Minimalism is an example portrait of Lora, triggered by specific prompt words. More details can be found on [Wiki](https://github.com/aigc-apps/CogVideoX-Fun/wiki/Training-Lora).
+                        """
+                    )
+                    with gr.Row():
+                        lora_model_dropdown = gr.Dropdown(
+                            label="Select LoRA model",
+                            choices=["none"],
+                            value="none",
+                            interactive=True,
+                        )
+
+                        lora_alpha_slider = gr.Slider(label="LoRA alpha (LoRA权重)", value=0.55, minimum=0, maximum=2, interactive=True)
+                
+        with gr.Column(variant="panel"):
+            gr.Markdown(
+                """
+                ### 2. Configs for Generation.
+                """
+            )
+            
+            prompt_textbox = gr.Textbox(label="Prompt", lines=2, value="A young woman with beautiful and clear eyes and blonde hair standing and white dress in a forest wearing a crown. She seems to be lost in thought, and the camera focuses on her face. The video is of high quality, and the view is very clear. High quality, masterpiece, best quality, highres, ultra-detailed, fantastic.")
+            negative_prompt_textbox = gr.Textbox(label="Negative prompt", lines=2, value="The video is not of a high quality, it has a low resolution. Watermark present in each frame. Strange motion trajectory.  " )
+                
+            with gr.Row():
+                with gr.Column():
+                    with gr.Row():
+                        sampler_dropdown   = gr.Dropdown(label="Sampling method", choices=list(scheduler_dict.keys()), value=list(scheduler_dict.keys())[0])
+                        sample_step_slider = gr.Slider(label="Sampling steps", value=50, minimum=10, maximum=50, step=1, interactive=False)
+                    
+                    resize_method = gr.Radio(
+                        ["Generate by", "Resize according to Reference"],
+                        value="Generate by",
+                        show_label=False,
+                    )
+                    width_slider     = gr.Slider(label="Width (视频宽度)",            value=672, minimum=128, maximum=1280, step=16, interactive=False)
+                    height_slider    = gr.Slider(label="Height (视频高度)",           value=384, minimum=128, maximum=1280, step=16, interactive=False)
+                    base_resolution  = gr.Radio(label="Base Resolution of Pretrained Models", value=512, choices=[512, 768, 960], interactive=False, visible=False)
+
+                    with gr.Group():
+                        generation_method = gr.Radio(
+                            ["Video Generation", "Image Generation"],
+                            value="Video Generation",
+                            show_label=False,
+                            visible=True,
+                        )
+                        length_slider = gr.Slider(label="Animation length (视频帧数)", value=49, minimum=5,   maximum=49,  step=4)
+                    
+                    source_method = gr.Radio(
+                        ["Text to Video (文本到视频)", "Image to Video (图片到视频)", "Video to Video (视频到视频)"],
+                        value="Text to Video (文本到视频)",
+                        show_label=False,
+                    )
+                    with gr.Column(visible = False) as image_to_video_col:
+                        start_image = gr.Image(label="The image at the beginning of the video", show_label=True, elem_id="i2v_start", sources="upload", type="filepath")
+                        
+                        template_gallery_path = ["asset/1.png", "asset/2.png", "asset/3.png", "asset/4.png", "asset/5.png"]
+                        def select_template(evt: gr.SelectData):
+                            text = {
+                                "asset/1.png": "The dog is looking at camera and smiling. The video is of high quality, and the view is very clear. High quality, masterpiece, best quality, highres, ultra-detailed, fantastic.", 
+                                "asset/2.png": "a sailboat sailing in rough seas with a dramatic sunset. The video is of high quality, and the view is very clear. High quality, masterpiece, best quality, highres, ultra-detailed, fantastic.", 
+                                "asset/3.png": "a beautiful woman with long hair and a dress blowing in the wind. The video is of high quality, and the view is very clear. High quality, masterpiece, best quality, highres, ultra-detailed, fantastic.", 
+                                "asset/4.png": "a man in an astronaut suit playing a guitar. The video is of high quality, and the view is very clear. High quality, masterpiece, best quality, highres, ultra-detailed, fantastic.", 
+                                "asset/5.png": "fireworks display over night city. The video is of high quality, and the view is very clear. High quality, masterpiece, best quality, highres, ultra-detailed, fantastic.", 
+                            }[template_gallery_path[evt.index]]
+                            return template_gallery_path[evt.index], text
+
+                        template_gallery = gr.Gallery(
+                            template_gallery_path,
+                            columns=5, rows=1,
+                            height=140,
+                            allow_preview=False,
+                            container=False,
+                            label="Template Examples",
+                        )
+                        template_gallery.select(select_template, None, [start_image, prompt_textbox])
+
+                        with gr.Accordion("The image at the ending of the video (Optional)", open=False):
+                            end_image   = gr.Image(label="The image at the ending of the video (Optional)", show_label=True, elem_id="i2v_end", sources="upload", type="filepath")
+                    
+                    with gr.Column(visible = False) as video_to_video_col:
+                        validation_video = gr.Video(
+                            label="The video to convert (视频转视频的参考视频)",  show_label=True, 
+                            elem_id="v2v", sources="upload", 
+                        )
+                        denoise_strength = gr.Slider(label="Denoise strength (重绘系数)", value=0.70, minimum=0.10, maximum=0.95, step=0.01)
+
+                    cfg_scale_slider  = gr.Slider(label="CFG Scale (引导系数)",        value=7.0, minimum=0,   maximum=20)
+                    
+                    with gr.Row():
+                        seed_textbox = gr.Textbox(label="Seed", value=43)
+                        seed_button  = gr.Button(value="\U0001F3B2", elem_classes="toolbutton")
+                        seed_button.click(
+                            fn=lambda: gr.Textbox(value=random.randint(1, 1e8)) if gradio_version_is_above_4 else gr.Textbox.update(value=random.randint(1, 1e8)), 
+                            inputs=[], 
+                            outputs=[seed_textbox]
+                        )
+
+                    generate_button = gr.Button(value="Generate", variant='primary')
+                    
+                with gr.Column():
+                    result_image = gr.Image(label="Generated Image", interactive=False, visible=False)
+                    result_video = gr.Video(label="Generated Animation", interactive=False)
+                    infer_progress = gr.Textbox(
+                        label="Generation Info",
+                        value="No task currently",
+                        interactive=False
+                    )
+
+            def upload_generation_method(generation_method):
+                if generation_method == "Video Generation":
+                    return gr.update(visible=True, minimum=5, maximum=49, value=49, interactive=True)
+                elif generation_method == "Image Generation":
+                    return gr.update(minimum=1, maximum=1, value=1, interactive=False)
+            generation_method.change(
+                upload_generation_method, generation_method, [length_slider]
+            )
+
+            def upload_source_method(source_method):
+                if source_method == "Text to Video (文本到视频)":
+                    return [gr.update(visible=False), gr.update(visible=False), gr.update(value=None), gr.update(value=None), gr.update(value=None)]
+                elif source_method == "Image to Video (图片到视频)":
+                    return [gr.update(visible=True), gr.update(visible=False), gr.update(), gr.update(), gr.update(value=None)]
+                else:
+                    return [gr.update(visible=False), gr.update(visible=True), gr.update(value=None), gr.update(value=None), gr.update()]
+            source_method.change(
+                upload_source_method, source_method, [image_to_video_col, video_to_video_col, start_image, end_image, validation_video]
+            )
+
+            def upload_resize_method(resize_method):
+                if resize_method == "Generate by":
+                    return [gr.update(visible=True), gr.update(visible=True), gr.update(visible=False)]
+                else:
+                    return [gr.update(visible=False), gr.update(visible=False), gr.update(visible=True)]
+            resize_method.change(
+                upload_resize_method, resize_method, [width_slider, height_slider, base_resolution]
+            )
+
+            generate_button.click(
+                fn=controller.generate,
+                inputs=[
+                    diffusion_transformer_dropdown,
+                    base_model_dropdown,
+                    lora_model_dropdown, 
+                    lora_alpha_slider,
+                    prompt_textbox, 
+                    negative_prompt_textbox, 
+                    sampler_dropdown, 
+                    sample_step_slider, 
+                    resize_method,
+                    width_slider, 
+                    height_slider, 
+                    base_resolution, 
+                    generation_method, 
+                    length_slider, 
+                    cfg_scale_slider, 
+                    start_image, 
+                    end_image, 
+                    validation_video,
+                    denoise_strength, 
+                    seed_textbox,
+                ],
+                outputs=[result_image, result_video, infer_progress]
+            )
+    return demo, controller

cogvideox/utils/lora_utils.py

@@ -0,0 +1,477 @@
+# LoRA network module
+# reference:
+# https://github.com/microsoft/LoRA/blob/main/loralib/layers.py
+# https://github.com/cloneofsimo/lora/blob/master/lora_diffusion/lora.py
+# https://github.com/bmaltais/kohya_ss
+
+import hashlib
+import math
+import os
+from collections import defaultdict
+from io import BytesIO
+from typing import List, Optional, Type, Union
+
+import safetensors.torch
+import torch
+import torch.utils.checkpoint
+from diffusers.models.lora import LoRACompatibleConv, LoRACompatibleLinear
+from safetensors.torch import load_file
+from transformers import T5EncoderModel
+
+
+class LoRAModule(torch.nn.Module):
+    """
+    replaces forward method of the original Linear, instead of replacing the original Linear module.
+    """
+
+    def __init__(
+        self,
+        lora_name,
+        org_module: torch.nn.Module,
+        multiplier=1.0,
+        lora_dim=4,
+        alpha=1,
+        dropout=None,
+        rank_dropout=None,
+        module_dropout=None,
+    ):
+        """if alpha == 0 or None, alpha is rank (no scaling)."""
+        super().__init__()
+        self.lora_name = lora_name
+
+        if org_module.__class__.__name__ == "Conv2d":
+            in_dim = org_module.in_channels
+            out_dim = org_module.out_channels
+        else:
+            in_dim = org_module.in_features
+            out_dim = org_module.out_features
+
+        self.lora_dim = lora_dim
+        if org_module.__class__.__name__ == "Conv2d":
+            kernel_size = org_module.kernel_size
+            stride = org_module.stride
+            padding = org_module.padding
+            self.lora_down = torch.nn.Conv2d(in_dim, self.lora_dim, kernel_size, stride, padding, bias=False)
+            self.lora_up = torch.nn.Conv2d(self.lora_dim, out_dim, (1, 1), (1, 1), bias=False)
+        else:
+            self.lora_down = torch.nn.Linear(in_dim, self.lora_dim, bias=False)
+            self.lora_up = torch.nn.Linear(self.lora_dim, out_dim, bias=False)
+
+        if type(alpha) == torch.Tensor:
+            alpha = alpha.detach().float().numpy()  # without casting, bf16 causes error
+        alpha = self.lora_dim if alpha is None or alpha == 0 else alpha
+        self.scale = alpha / self.lora_dim
+        self.register_buffer("alpha", torch.tensor(alpha))
+
+        # same as microsoft's
+        torch.nn.init.kaiming_uniform_(self.lora_down.weight, a=math.sqrt(5))
+        torch.nn.init.zeros_(self.lora_up.weight)
+
+        self.multiplier = multiplier
+        self.org_module = org_module  # remove in applying
+        self.dropout = dropout
+        self.rank_dropout = rank_dropout
+        self.module_dropout = module_dropout
+
+    def apply_to(self):
+        self.org_forward = self.org_module.forward
+        self.org_module.forward = self.forward
+        del self.org_module
+
+    def forward(self, x, *args, **kwargs):
+        weight_dtype = x.dtype
+        org_forwarded = self.org_forward(x)
+
+        # module dropout
+        if self.module_dropout is not None and self.training:
+            if torch.rand(1) < self.module_dropout:
+                return org_forwarded
+
+        lx = self.lora_down(x.to(self.lora_down.weight.dtype))
+
+        # normal dropout
+        if self.dropout is not None and self.training:
+            lx = torch.nn.functional.dropout(lx, p=self.dropout)
+
+        # rank dropout
+        if self.rank_dropout is not None and self.training:
+            mask = torch.rand((lx.size(0), self.lora_dim), device=lx.device) > self.rank_dropout
+            if len(lx.size()) == 3:
+                mask = mask.unsqueeze(1)  # for Text Encoder
+            elif len(lx.size()) == 4:
+                mask = mask.unsqueeze(-1).unsqueeze(-1)  # for Conv2d
+            lx = lx * mask
+
+            # scaling for rank dropout: treat as if the rank is changed
+            scale = self.scale * (1.0 / (1.0 - self.rank_dropout))  # redundant for readability
+        else:
+            scale = self.scale
+
+        lx = self.lora_up(lx)
+
+        return org_forwarded.to(weight_dtype) + lx.to(weight_dtype) * self.multiplier * scale
+
+
+def addnet_hash_legacy(b):
+    """Old model hash used by sd-webui-additional-networks for .safetensors format files"""
+    m = hashlib.sha256()
+
+    b.seek(0x100000)
+    m.update(b.read(0x10000))
+    return m.hexdigest()[0:8]
+
+
+def addnet_hash_safetensors(b):
+    """New model hash used by sd-webui-additional-networks for .safetensors format files"""
+    hash_sha256 = hashlib.sha256()
+    blksize = 1024 * 1024
+
+    b.seek(0)
+    header = b.read(8)
+    n = int.from_bytes(header, "little")
+
+    offset = n + 8
+    b.seek(offset)
+    for chunk in iter(lambda: b.read(blksize), b""):
+        hash_sha256.update(chunk)
+
+    return hash_sha256.hexdigest()
+
+
+def precalculate_safetensors_hashes(tensors, metadata):
+    """Precalculate the model hashes needed by sd-webui-additional-networks to
+    save time on indexing the model later."""
+
+    # Because writing user metadata to the file can change the result of
+    # sd_models.model_hash(), only retain the training metadata for purposes of
+    # calculating the hash, as they are meant to be immutable
+    metadata = {k: v for k, v in metadata.items() if k.startswith("ss_")}
+
+    bytes = safetensors.torch.save(tensors, metadata)
+    b = BytesIO(bytes)
+
+    model_hash = addnet_hash_safetensors(b)
+    legacy_hash = addnet_hash_legacy(b)
+    return model_hash, legacy_hash
+
+
+class LoRANetwork(torch.nn.Module):
+    TRANSFORMER_TARGET_REPLACE_MODULE = ["CogVideoXTransformer3DModel"]
+    TEXT_ENCODER_TARGET_REPLACE_MODULE = ["T5LayerSelfAttention", "T5LayerFF", "BertEncoder"]
+    LORA_PREFIX_TRANSFORMER = "lora_unet"
+    LORA_PREFIX_TEXT_ENCODER = "lora_te"
+    def __init__(
+        self,
+        text_encoder: Union[List[T5EncoderModel], T5EncoderModel],
+        unet,
+        multiplier: float = 1.0,
+        lora_dim: int = 4,
+        alpha: float = 1,
+        dropout: Optional[float] = None,
+        module_class: Type[object] = LoRAModule,
+        add_lora_in_attn_temporal: bool = False,
+        varbose: Optional[bool] = False,
+    ) -> None:
+        super().__init__()
+        self.multiplier = multiplier
+
+        self.lora_dim = lora_dim
+        self.alpha = alpha
+        self.dropout = dropout
+
+        print(f"create LoRA network. base dim (rank): {lora_dim}, alpha: {alpha}")
+        print(f"neuron dropout: p={self.dropout}")
+
+        # create module instances
+        def create_modules(
+            is_unet: bool,
+            root_module: torch.nn.Module,
+            target_replace_modules: List[torch.nn.Module],
+        ) -> List[LoRAModule]:
+            prefix = (
+                self.LORA_PREFIX_TRANSFORMER
+                if is_unet
+                else self.LORA_PREFIX_TEXT_ENCODER
+            )
+            loras = []
+            skipped = []
+            for name, module in root_module.named_modules():
+                if module.__class__.__name__ in target_replace_modules:
+                    for child_name, child_module in module.named_modules():
+                        is_linear = child_module.__class__.__name__ == "Linear" or child_module.__class__.__name__ == "LoRACompatibleLinear"
+                        is_conv2d = child_module.__class__.__name__ == "Conv2d" or child_module.__class__.__name__ == "LoRACompatibleConv"
+                        is_conv2d_1x1 = is_conv2d and child_module.kernel_size == (1, 1)
+                        
+                        if not add_lora_in_attn_temporal:
+                            if "attn_temporal" in child_name:
+                                continue
+
+                        if is_linear or is_conv2d:
+                            lora_name = prefix + "." + name + "." + child_name
+                            lora_name = lora_name.replace(".", "_")
+
+                            dim = None
+                            alpha = None
+
+                            if is_linear or is_conv2d_1x1:
+                                dim = self.lora_dim
+                                alpha = self.alpha
+
+                            if dim is None or dim == 0:
+                                if is_linear or is_conv2d_1x1:
+                                    skipped.append(lora_name)
+                                continue
+
+                            lora = module_class(
+                                lora_name,
+                                child_module,
+                                self.multiplier,
+                                dim,
+                                alpha,
+                                dropout=dropout,
+                            )
+                            loras.append(lora)
+            return loras, skipped
+
+        text_encoders = text_encoder if type(text_encoder) == list else [text_encoder]
+
+        self.text_encoder_loras = []
+        skipped_te = []
+        for i, text_encoder in enumerate(text_encoders):
+            if text_encoder is not None:
+                text_encoder_loras, skipped = create_modules(False, text_encoder, LoRANetwork.TEXT_ENCODER_TARGET_REPLACE_MODULE)
+                self.text_encoder_loras.extend(text_encoder_loras)
+                skipped_te += skipped
+        print(f"create LoRA for Text Encoder: {len(self.text_encoder_loras)} modules.")
+
+        self.unet_loras, skipped_un = create_modules(True, unet, LoRANetwork.TRANSFORMER_TARGET_REPLACE_MODULE)
+        print(f"create LoRA for U-Net: {len(self.unet_loras)} modules.")
+
+        # assertion
+        names = set()
+        for lora in self.text_encoder_loras + self.unet_loras:
+            assert lora.lora_name not in names, f"duplicated lora name: {lora.lora_name}"
+            names.add(lora.lora_name)
+
+    def apply_to(self, text_encoder, unet, apply_text_encoder=True, apply_unet=True):
+        if apply_text_encoder:
+            print("enable LoRA for text encoder")
+        else:
+            self.text_encoder_loras = []
+
+        if apply_unet:
+            print("enable LoRA for U-Net")
+        else:
+            self.unet_loras = []
+
+        for lora in self.text_encoder_loras + self.unet_loras:
+            lora.apply_to()
+            self.add_module(lora.lora_name, lora)
+
+    def set_multiplier(self, multiplier):
+        self.multiplier = multiplier
+        for lora in self.text_encoder_loras + self.unet_loras:
+            lora.multiplier = self.multiplier
+
+    def load_weights(self, file):
+        if os.path.splitext(file)[1] == ".safetensors":
+            from safetensors.torch import load_file
+
+            weights_sd = load_file(file)
+        else:
+            weights_sd = torch.load(file, map_location="cpu")
+        info = self.load_state_dict(weights_sd, False)
+        return info
+
+    def prepare_optimizer_params(self, text_encoder_lr, unet_lr, default_lr):
+        self.requires_grad_(True)
+        all_params = []
+
+        def enumerate_params(loras):
+            params = []
+            for lora in loras:
+                params.extend(lora.parameters())
+            return params
+
+        if self.text_encoder_loras:
+            param_data = {"params": enumerate_params(self.text_encoder_loras)}
+            if text_encoder_lr is not None:
+                param_data["lr"] = text_encoder_lr
+            all_params.append(param_data)
+
+        if self.unet_loras:
+            param_data = {"params": enumerate_params(self.unet_loras)}
+            if unet_lr is not None:
+                param_data["lr"] = unet_lr
+            all_params.append(param_data)
+
+        return all_params
+
+    def enable_gradient_checkpointing(self):
+        pass
+
+    def get_trainable_params(self):
+        return self.parameters()
+
+    def save_weights(self, file, dtype, metadata):
+        if metadata is not None and len(metadata) == 0:
+            metadata = None
+
+        state_dict = self.state_dict()
+
+        if dtype is not None:
+            for key in list(state_dict.keys()):
+                v = state_dict[key]
+                v = v.detach().clone().to("cpu").to(dtype)
+                state_dict[key] = v
+
+        if os.path.splitext(file)[1] == ".safetensors":
+            from safetensors.torch import save_file
+
+            # Precalculate model hashes to save time on indexing
+            if metadata is None:
+                metadata = {}
+            model_hash, legacy_hash = precalculate_safetensors_hashes(state_dict, metadata)
+            metadata["sshs_model_hash"] = model_hash
+            metadata["sshs_legacy_hash"] = legacy_hash
+
+            save_file(state_dict, file, metadata)
+        else:
+            torch.save(state_dict, file)
+
+def create_network(
+    multiplier: float,
+    network_dim: Optional[int],
+    network_alpha: Optional[float],
+    text_encoder: Union[T5EncoderModel, List[T5EncoderModel]],
+    transformer,
+    neuron_dropout: Optional[float] = None,
+    add_lora_in_attn_temporal: bool = False,
+    **kwargs,
+):
+    if network_dim is None:
+        network_dim = 4  # default
+    if network_alpha is None:
+        network_alpha = 1.0
+
+    network = LoRANetwork(
+        text_encoder,
+        transformer,
+        multiplier=multiplier,
+        lora_dim=network_dim,
+        alpha=network_alpha,
+        dropout=neuron_dropout,
+        add_lora_in_attn_temporal=add_lora_in_attn_temporal,
+        varbose=True,
+    )
+    return network
+
+def merge_lora(pipeline, lora_path, multiplier, device='cpu', dtype=torch.float32, state_dict=None, transformer_only=False):
+    LORA_PREFIX_TRANSFORMER = "lora_unet"
+    LORA_PREFIX_TEXT_ENCODER = "lora_te"
+    if state_dict is None:
+        state_dict = load_file(lora_path, device=device)
+    else:
+        state_dict = state_dict
+    updates = defaultdict(dict)
+    for key, value in state_dict.items():
+        layer, elem = key.split('.', 1)
+        updates[layer][elem] = value
+
+    for layer, elems in updates.items():
+
+        if "lora_te" in layer:
+            if transformer_only:
+                continue
+            else:
+                layer_infos = layer.split(LORA_PREFIX_TEXT_ENCODER + "_")[-1].split("_")
+                curr_layer = pipeline.text_encoder
+        else:
+            layer_infos = layer.split(LORA_PREFIX_TRANSFORMER + "_")[-1].split("_")
+            curr_layer = pipeline.transformer
+
+        temp_name = layer_infos.pop(0)
+        while len(layer_infos) > -1:
+            try:
+                curr_layer = curr_layer.__getattr__(temp_name)
+                if len(layer_infos) > 0:
+                    temp_name = layer_infos.pop(0)
+                elif len(layer_infos) == 0:
+                    break
+            except Exception:
+                if len(layer_infos) == 0:
+                    print('Error loading layer')
+                if len(temp_name) > 0:
+                    temp_name += "_" + layer_infos.pop(0)
+                else:
+                    temp_name = layer_infos.pop(0)
+
+        weight_up = elems['lora_up.weight'].to(dtype)
+        weight_down = elems['lora_down.weight'].to(dtype)
+        if 'alpha' in elems.keys():
+            alpha = elems['alpha'].item() / weight_up.shape[1]
+        else:
+            alpha = 1.0
+
+        curr_layer.weight.data = curr_layer.weight.data.to(device)
+        if len(weight_up.shape) == 4:
+            curr_layer.weight.data += multiplier * alpha * torch.mm(weight_up.squeeze(3).squeeze(2),
+                                                                    weight_down.squeeze(3).squeeze(2)).unsqueeze(
+                2).unsqueeze(3)
+        else:
+            curr_layer.weight.data += multiplier * alpha * torch.mm(weight_up, weight_down)
+
+    return pipeline
+
+# TODO: Refactor with merge_lora.
+def unmerge_lora(pipeline, lora_path, multiplier=1, device="cpu", dtype=torch.float32):
+    """Unmerge state_dict in LoRANetwork from the pipeline in diffusers."""
+    LORA_PREFIX_UNET = "lora_unet"
+    LORA_PREFIX_TEXT_ENCODER = "lora_te"
+    state_dict = load_file(lora_path, device=device)
+
+    updates = defaultdict(dict)
+    for key, value in state_dict.items():
+        layer, elem = key.split('.', 1)
+        updates[layer][elem] = value
+
+    for layer, elems in updates.items():
+
+        if "lora_te" in layer:
+            layer_infos = layer.split(LORA_PREFIX_TEXT_ENCODER + "_")[-1].split("_")
+            curr_layer = pipeline.text_encoder
+        else:
+            layer_infos = layer.split(LORA_PREFIX_UNET + "_")[-1].split("_")
+            curr_layer = pipeline.transformer
+
+        temp_name = layer_infos.pop(0)
+        while len(layer_infos) > -1:
+            try:
+                curr_layer = curr_layer.__getattr__(temp_name)
+                if len(layer_infos) > 0:
+                    temp_name = layer_infos.pop(0)
+                elif len(layer_infos) == 0:
+                    break
+            except Exception:
+                if len(layer_infos) == 0:
+                    print('Error loading layer')
+                if len(temp_name) > 0:
+                    temp_name += "_" + layer_infos.pop(0)
+                else:
+                    temp_name = layer_infos.pop(0)
+
+        weight_up = elems['lora_up.weight'].to(dtype)
+        weight_down = elems['lora_down.weight'].to(dtype)
+        if 'alpha' in elems.keys():
+            alpha = elems['alpha'].item() / weight_up.shape[1]
+        else:
+            alpha = 1.0
+
+        curr_layer.weight.data = curr_layer.weight.data.to(device)
+        if len(weight_up.shape) == 4:
+            curr_layer.weight.data -= multiplier * alpha * torch.mm(weight_up.squeeze(3).squeeze(2),
+                                                                    weight_down.squeeze(3).squeeze(2)).unsqueeze(2).unsqueeze(3)
+        else:
+            curr_layer.weight.data -= multiplier * alpha * torch.mm(weight_up, weight_down)
+
+    return pipeline

cogvideox/utils/utils.py

@@ -0,0 +1,189 @@
+import os
+import gc
+import imageio
+import numpy as np
+import torch
+import torchvision
+import cv2
+from einops import rearrange
+from PIL import Image
+
+def get_width_and_height_from_image_and_base_resolution(image, base_resolution):
+    target_pixels = int(base_resolution) * int(base_resolution)
+    original_width, original_height = Image.open(image).size
+    ratio = (target_pixels / (original_width * original_height)) ** 0.5
+    width_slider = round(original_width * ratio)
+    height_slider = round(original_height * ratio)
+    return height_slider, width_slider
+
+def color_transfer(sc, dc):
+    """
+    Transfer color distribution from of sc, referred to dc.
+
+    Args:
+        sc (numpy.ndarray): input image to be transfered.
+        dc (numpy.ndarray): reference image
+
+    Returns:
+        numpy.ndarray: Transferred color distribution on the sc.
+    """
+
+    def get_mean_and_std(img):
+        x_mean, x_std = cv2.meanStdDev(img)
+        x_mean = np.hstack(np.around(x_mean, 2))
+        x_std = np.hstack(np.around(x_std, 2))
+        return x_mean, x_std
+
+    sc = cv2.cvtColor(sc, cv2.COLOR_RGB2LAB)
+    s_mean, s_std = get_mean_and_std(sc)
+    dc = cv2.cvtColor(dc, cv2.COLOR_RGB2LAB)
+    t_mean, t_std = get_mean_and_std(dc)
+    img_n = ((sc - s_mean) * (t_std / s_std)) + t_mean
+    np.putmask(img_n, img_n > 255, 255)
+    np.putmask(img_n, img_n < 0, 0)
+    dst = cv2.cvtColor(cv2.convertScaleAbs(img_n), cv2.COLOR_LAB2RGB)
+    return dst
+
+def save_videos_grid(videos: torch.Tensor, path: str, rescale=False, n_rows=6, fps=12, imageio_backend=True, color_transfer_post_process=False):
+    videos = rearrange(videos, "b c t h w -> t b c h w")
+    outputs = []
+    for x in videos:
+        x = torchvision.utils.make_grid(x, nrow=n_rows)
+        x = x.transpose(0, 1).transpose(1, 2).squeeze(-1)
+        if rescale:
+            x = (x + 1.0) / 2.0  # -1,1 -> 0,1
+        x = (x * 255).numpy().astype(np.uint8)
+        outputs.append(Image.fromarray(x))
+
+    if color_transfer_post_process:
+        for i in range(1, len(outputs)):
+            outputs[i] = Image.fromarray(color_transfer(np.uint8(outputs[i]), np.uint8(outputs[0])))
+
+    os.makedirs(os.path.dirname(path), exist_ok=True)
+    if imageio_backend:
+        if path.endswith("mp4"):
+            imageio.mimsave(path, outputs, fps=fps)
+        else:
+            imageio.mimsave(path, outputs, duration=(1000 * 1/fps))
+    else:
+        if path.endswith("mp4"):
+            path = path.replace('.mp4', '.gif')
+        outputs[0].save(path, format='GIF', append_images=outputs, save_all=True, duration=100, loop=0)
+
+def get_image_to_video_latent(validation_image_start, validation_image_end, video_length, sample_size):
+    if validation_image_start is not None and validation_image_end is not None:
+        if type(validation_image_start) is str and os.path.isfile(validation_image_start):
+            image_start = clip_image = Image.open(validation_image_start).convert("RGB")
+            image_start = image_start.resize([sample_size[1], sample_size[0]])
+            clip_image = clip_image.resize([sample_size[1], sample_size[0]])
+        else:
+            image_start = clip_image = validation_image_start
+            image_start = [_image_start.resize([sample_size[1], sample_size[0]]) for _image_start in image_start]
+            clip_image = [_clip_image.resize([sample_size[1], sample_size[0]]) for _clip_image in clip_image]
+
+        if type(validation_image_end) is str and os.path.isfile(validation_image_end):
+            image_end = Image.open(validation_image_end).convert("RGB")
+            image_end = image_end.resize([sample_size[1], sample_size[0]])
+        else:
+            image_end = validation_image_end
+            image_end = [_image_end.resize([sample_size[1], sample_size[0]]) for _image_end in image_end]
+
+        if type(image_start) is list:
+            clip_image = clip_image[0]
+            start_video = torch.cat(
+                [torch.from_numpy(np.array(_image_start)).permute(2, 0, 1).unsqueeze(1).unsqueeze(0) for _image_start in image_start], 
+                dim=2
+            )
+            input_video = torch.tile(start_video[:, :, :1], [1, 1, video_length, 1, 1])
+            input_video[:, :, :len(image_start)] = start_video
+            
+            input_video_mask = torch.zeros_like(input_video[:, :1])
+            input_video_mask[:, :, len(image_start):] = 255
+        else:
+            input_video = torch.tile(
+                torch.from_numpy(np.array(image_start)).permute(2, 0, 1).unsqueeze(1).unsqueeze(0), 
+                [1, 1, video_length, 1, 1]
+            )
+            input_video_mask = torch.zeros_like(input_video[:, :1])
+            input_video_mask[:, :, 1:] = 255
+
+        if type(image_end) is list:
+            image_end = [_image_end.resize(image_start[0].size if type(image_start) is list else image_start.size) for _image_end in image_end]
+            end_video = torch.cat(
+                [torch.from_numpy(np.array(_image_end)).permute(2, 0, 1).unsqueeze(1).unsqueeze(0) for _image_end in image_end], 
+                dim=2
+            )
+            input_video[:, :, -len(end_video):] = end_video
+            
+            input_video_mask[:, :, -len(image_end):] = 0
+        else:
+            image_end = image_end.resize(image_start[0].size if type(image_start) is list else image_start.size)
+            input_video[:, :, -1:] = torch.from_numpy(np.array(image_end)).permute(2, 0, 1).unsqueeze(1).unsqueeze(0)
+            input_video_mask[:, :, -1:] = 0
+
+        input_video = input_video / 255
+
+    elif validation_image_start is not None:
+        if type(validation_image_start) is str and os.path.isfile(validation_image_start):
+            image_start = clip_image = Image.open(validation_image_start).convert("RGB")
+            image_start = image_start.resize([sample_size[1], sample_size[0]])
+            clip_image = clip_image.resize([sample_size[1], sample_size[0]])
+        else:
+            image_start = clip_image = validation_image_start
+            image_start = [_image_start.resize([sample_size[1], sample_size[0]]) for _image_start in image_start]
+            clip_image = [_clip_image.resize([sample_size[1], sample_size[0]]) for _clip_image in clip_image]
+        image_end = None
+        
+        if type(image_start) is list:
+            clip_image = clip_image[0]
+            start_video = torch.cat(
+                [torch.from_numpy(np.array(_image_start)).permute(2, 0, 1).unsqueeze(1).unsqueeze(0) for _image_start in image_start], 
+                dim=2
+            )
+            input_video = torch.tile(start_video[:, :, :1], [1, 1, video_length, 1, 1])
+            input_video[:, :, :len(image_start)] = start_video
+            input_video = input_video / 255
+            
+            input_video_mask = torch.zeros_like(input_video[:, :1])
+            input_video_mask[:, :, len(image_start):] = 255
+        else:
+            input_video = torch.tile(
+                torch.from_numpy(np.array(image_start)).permute(2, 0, 1).unsqueeze(1).unsqueeze(0), 
+                [1, 1, video_length, 1, 1]
+            ) / 255
+            input_video_mask = torch.zeros_like(input_video[:, :1])
+            input_video_mask[:, :, 1:, ] = 255
+    else:
+        image_start = None
+        image_end = None
+        input_video = torch.zeros([1, 3, video_length, sample_size[0], sample_size[1]])
+        input_video_mask = torch.ones([1, 1, video_length, sample_size[0], sample_size[1]]) * 255
+        clip_image = None
+
+    del image_start
+    del image_end
+    gc.collect()
+
+    return  input_video, input_video_mask, clip_image
+
+def get_video_to_video_latent(input_video_path, video_length, sample_size):
+    if type(input_video_path) is str:
+        cap = cv2.VideoCapture(input_video_path)
+        input_video = []
+        while True:
+            ret, frame = cap.read()
+            if not ret:
+                break
+            frame = cv2.resize(frame, (sample_size[1], sample_size[0]))
+            input_video.append(cv2.cvtColor(frame, cv2.COLOR_BGR2RGB))
+        cap.release()
+    else:
+        input_video = input_video_path
+
+    input_video = torch.from_numpy(np.array(input_video))[:video_length]
+    input_video = input_video.permute([3, 0, 1, 2]).unsqueeze(0) / 255
+
+    input_video_mask = torch.zeros_like(input_video[:, :1])
+    input_video_mask[:, :, :] = 255
+
+    return  input_video, input_video_mask, None

reports/report_v1.md

@@ -0,0 +1,36 @@
+# CogVideoX FUN v1 Report
+In CogVideoX-FUN, we trained on approximately 1.2 million data points based on CogVideoX, supporting image and video predictions. It accommodates pixel values for video generation across different resolutions of 512x512x49, 768x768x49, and 1024x1024x49, as well as videos with different aspect ratios. Moreover, we support the generation of videos from images and the reconstruction of videos from other videos.
+
+Compared to CogVideoX, CogVideoX FUN also highlights the following features:
+- Introduction of the InPaint model, enabling the generation of videos from images with specified starting and ending images.
+- Training the model based on token lengths. This allows for the implementation of various sizes and resolutions within the same model.
+
+## InPaint Model
+We used [CogVideoX](https://github.com/THUDM/CogVideo/) as the foundational structure, referencing [EasyAnimate](https://github.com/aigc-apps/EasyAnimate) for the model training to generate videos from images. 
+
+During video generation, the **reference video** is encoded using VAE, with the **black area in the above image representing the part to be reconstructed, and the white area representing the start image**. This is stacked with noise latents and input into the Transformer for video generation. We perform 3D resizing on the **masked area**, directly resizing it to fit the canvas size of the video that needs reconstruction. 
+
+Then, we concatenate the latent, the encoded reference video, and the masked area, inputting them into DiT for noise prediction to obtain the final video. 
+The pipeline structure of CogVideoX FUN is as follows:
+<img src="https://pai-aigc-photog.oss-cn-hangzhou.aliyuncs.com/cogvideox_fun/asset/v1/pipeline.jpg" alt="ui" style="zoom:50%;" />
+
+## Token Length-Based Model Training
+We collected approximately 1.2 million high-quality data for the training of CogVideoX-Fun. During the training, we resized the videos based on different token lengths. The entire training process is divided into three phases, with each phase corresponding to 13312 (for 512x512x49 videos), 29952 (for 768x768x49 videos), and 53248 (for 1024x1024x49 videos).
+
+Taking CogVideoX-Fun-2B as an example:
+- In the 13312 phase, the batch size is 128 with 7k training steps.
+- In the 29952 phase, the batch size is 256 with 6.5k training steps.
+- In the 53248 phase, the batch size is 128 with 5k training steps.
+
+During training, we combined high and low resolutions, enabling the model to support video generation from any resolution between 512 and 1280. For example, with a token length of 13312:
+- At a resolution of 512x512, the number of video frames is 49.
+- At a resolution of 768x768, the number of video frames is 21.
+- At a resolution of 1024x1024, the number of video frames is 9.
+
+These resolutions and corresponding lengths were mixed for training, allowing the model to generate videos at different resolutions.
+
+## Resize 3D Embedding
+In adapting CogVideoX-2B to the CogVideoX-Fun framework, it was found that the source code obtains 3D embeddings in a truncated manner. This approach only accommodates a single resolution; when the resolution changes, the embedding should also change.
+<img src="https://pai-aigc-photog.oss-cn-hangzhou.aliyuncs.com/cogvideox_fun/asset/v1/PE_Interpolation.jpg" alt="ui" style="zoom:50%;" />
+
+Referencing Pixart-Sigma, the above image is from the Pixart-Sigma paper. We used Positional Embeddings Interpolation (PE Interpolation) to resize 3D embeddings. PE Interpolation is more conducive to convergence than directly generating cosine and sine embeddings for different resolutions.

reports/report_v1_zh-CN.md

@@ -0,0 +1,43 @@
+# CogVideoX FUN v1 Report
+
+在CogVideoX-FUN中，我们基于CogVideoX在大约1.2m的数据上进行了训练，支持图片与视频预测，支持像素值从512x512x49、768x768x49、1024x1024x49与不同纵横比的视频生成。另外，我们支持图像到视频的生成与视频到视频的重建。
+
+对比与CogVideoX，CogVideoX FUN还突出了以下功能：
+
+- 引入InPaint模型，实现图生视频功能，可以通过首尾图指定视频生成。
+- 基于Token长度的模型训练。达成不同大小多分辨率在同一模型中的实现。
+
+## InPaint模型
+我们以[CogVideoX](https://github.com/THUDM/CogVideo/)作为基础结构，参考[EasyAnimate](https://github.com/aigc-apps/EasyAnimate)进行图生视频的模型训练。
+
+在进行视频生成的时候，将**参考视频**使用VAE进行encode，**上图黑色的部分代表需要重建的部分，白色的部分代表首图**，与噪声Latents一起堆叠后输入到Transformer中进行视频生成。
+
+我们对**被Mask的区域**进行3D Resize，直接Resize到需要重建的视频的画布大小。
+
+然后将Latent、Encode后的参考视频、被Mask的区域，concat后输入到DiT中进行噪声预测。获得最终的视频。
+
+CogVideoX FUN的Pipeline结构如下：
+<img src="https://pai-aigc-photog.oss-cn-hangzhou.aliyuncs.com/cogvideox_fun/asset/v1/pipeline.jpg" alt="ui" style="zoom:50%;" />
+
+## 基于Token长度的模型训练
+我们收集了大约高质量的1.2m数据进行CogVideoX-Fun的训练。
+
+在进行训练时，我们根据不同Token长度，对视频进行缩放后进行训练。整个训练过程分为三个阶段，每个阶段的13312（对应512x512x49的视频），29952（对应768x768x49的视频），53248（对应1024x1024x49的视频）。
+
+以CogVideoX-Fun-2B为例子，其中：
+- 13312阶段，Batch size为128，训练步数为7k
+- 29952阶段，Batch size为256，训练步数为6.5k。
+- 53248阶段，Batch size为128，训练步数为5k。
+
+训练时我们采用高低分辨率结合训练，因此模型支持从512到1280任意分辨率的视频生成，以13312 token长度为例：
+- 在512x512分辨率下，视频帧数为49；
+- 在768x768分辨率下，视频帧数为21；
+- 在1024x1024分辨率下，视频帧数为9；
+这些分辨率与对应长度混合训练，模型可以完成不同大小分辨率的视频生成。
+
+## Resize 3D Embedding
+在适配CogVideoX-2B到CogVideoX-Fun框架的途中，发现源码是以截断的方式去得到3D Embedding的，这样的方式只能适配单一分辨率，当分辨率发生变化时，Embedding也应当发生变化。
+
+<img src="https://pai-aigc-photog.oss-cn-hangzhou.aliyuncs.com/cogvideox_fun/asset/v1/PE_Interpolation.jpg" alt="ui" style="zoom:50%;" />
+
+参考Pixart-Sigma，上图来自于Pixart-Sigma论文，我们采用Positional Embeddings Interpolation（PE Interpolation）对3D embedding进行Resize，PE Interpolation相比于直接生成不同分辨率的Cos Sin Embedding更易收敛。

requirements.txt

@@ -0,0 +1,28 @@
+Pillow
+einops
+safetensors
+timm
+tomesd
+torch>=2.1.2
+torchdiffeq
+torchsde
+xformers
+decord
+datasets
+numpy
+scikit-image
+opencv-python
+omegaconf
+SentencePiece
+albumentations
+imageio[ffmpeg]
+imageio[pyav]
+tensorboard
+beautifulsoup4
+ftfy
+func_timeout
+deepspeed
+accelerate>=0.25.0
+gradio>=3.41.2
+diffusers>=0.28.2
+transformers>=4.37.2