[Project Page] [中文主页]

This project aims to create a simple and scalable repo, to reproduce Sora (OpenAI, but we prefer to call it "CloseAI" ) and build knowledge about Video-VQVAE (VideoGPT) + DiT at scale. However, we have limited resources, we deeply wish all open-source community can contribute to this project. Pull request are welcome!!!

本项目希望通过开源社区的力量复现Sora，由北大-兔展AIGC联合实验室共同发起，当前我们资源有限仅搭建了基础架构，无法进行完整训练，希望通过开源社区逐步增加模块并筹集资源进行训练，当前版本离目标差距巨大，仍需持续完善和快速迭代，欢迎Pull request！！！

Project stages:

• Primary

1. Setup the codebase and train a un-conditional model on landscape dataset.
2. Train models that boost resolution and duration.

• Extensions

3. Conduct text2video experiments on landscape dataset.
4. Train the 1080p model on video2text dataset.
5. Control model with more condition.

[2024.03.05] See our latest todo, welcome to pull request.

[2024.03.04] We re-organize and modulize our codes and make it easy to contribute to the project, please see the Repo structure.

[2024.03.03] We open some discussions and clarify several issues.

[2024.03.01] Training codes are available now! Learn more in our project page. Please feel free to watch 👀 this repository for the latest updates.

• Setup repo-structure.
• Add Video-VQGAN model, which is borrowed from VideoGPT.
• Support variable aspect ratios, resolutions, durations training on DiT.
• Support Dynamic mask input inspired FiT.
• Add class-conditioning on embeddings.
• Incorporating Latte as main codebase.
• Add VAE model, which is borrowed from Stable Diffusion.
• Joint dynamic mask input with VAE.
• Make the codebase ready for the cluster training. Add SLURM scripts.
• Add sampling script.
• Incorporating SiT.

• Add PI to support out-of-domain size.
• Add frame interpolation model.

• Finish data loading, pre-processing utils.
• Add CLIP and T5 support.
• Add text2image training script.
• Add prompt captioner.

• Looking for a suitable dataset, welcome to discuss and recommend.
• Finish data loading, pre-processing utils.
• Support memory friendly training.
  • Add flash-attention2 from pytorch.
  • Add xformers.
  • Add accelerate to automatically manage training, e.g. mixed precision training.
  • Add gradient checkpoint.
  • Train using the deepspeed engine.

• Load pretrained weight from PixArt-α.
• Incorporating ControlNet.

├── README.md
├── docs
│   ├── Data.md                    -> Datasets description.
│   ├── Contribution_Guidelines.md -> Contribution guidelines description.
├── scripts                        -> All training scripts.
│   └── train.sh
├── sora
│   ├── dataset                    -> Dataset code to read videos
│   ├── models 
│   │   ├── captioner               
│   │   ├── super_resolution        
│   ├── modules
│   │   ├── ae                     -> compress videos to latents
│   │   │   ├── vqvae
│   │   │   ├── vae
│   │   ├── diffusion              -> denoise latents
│   │   │   ├── dit
│   │   │   ├── unet
|   ├── utils.py                   
│   ├── train.py                   -> Training code

The recommended requirements are as follows.

• Python >= 3.8
• Pytorch >= 1.13.1
• CUDA Version >= 11.7
• Install required packages:

git clone https://github.com/PKU-YuanGroup/Open-Sora-Plan
cd Open-Sora-Plan
conda create -n opensora python=3.8 -y
conda activate opensora
pip install torch==1.13.1+cu117 torchvision==0.14.1+cu117 torchaudio==0.13.1 --extra-index-url https://download.pytorch.org/whl/cu117
pip install -r requirements.txt
cd VideoGPT
pip install -e .
cd ..

Refer to Data.md

cd src/sora/modules/ae/vqvae/videogpt

Refer to origin repo. Use the scripts/train_vqvae.py script to train a Video-VQVAE. Execute python scripts/train_vqvae.py -h for information on all available training settings. A subset of more relevant settings are listed below, along with default values.

• --embedding_dim: number of dimensions for codebooks embeddings
• --n_codes 2048: number of codes in the codebook
• --n_hiddens 240: number of hidden features in the residual blocks
• --n_res_layers 4: number of residual blocks
• --downsample 4 4 4: T H W downsampling stride of the encoder

• --gpus 2: number of gpus for distributed training
• --sync_batchnorm: uses SyncBatchNorm instead of BatchNorm3d when using > 1 gpu
• --gradient_clip_val 1: gradient clipping threshold for training
• --batch_size 16: batch size per gpu
• --num_workers 8: number of workers for each DataLoader

• --data_path <path>: path to an hdf5 file or a folder containing train and test folders with subdirectories of videos
• --resolution 128: spatial resolution to train on
• --sequence_length 16: temporal resolution, or video clip length

python rec_video.py --video-path "assets/origin_video_0.mp4" --rec-path "rec_video_0.mp4" --num-frames 500 --sample-rate 1

python rec_video.py --video-path "assets/origin_video_1.mp4" --rec-path "rec_video_1.mp4" --resolution 196 --num-frames 600 --sample-rate 1

We present four reconstructed videos in this demonstration, arranged from left to right as follows:

3s 596x336	10s 256x256	18s 196x196	24s 168x96

sh scripts/train.sh

Coming soon.

We greatly appreciate your contributions to the Open-Sora Plan open-source community and helping us make it even better than it is now!

For more details, please refer to the Contribution Guidelines

• DiT: Scalable Diffusion Models with Transformers.
• VideoGPT: Video Generation using VQ-VAE and Transformers.
• FiT: Flexible Vision Transformer for Diffusion Model.
• Positional Interpolation: Extending Context Window of Large Language Models via Positional Interpolation.

• The service is a research preview intended for non-commercial use only. See LICENSE.txt for details.