SkyReels-A2: Compose Anything in Video Diffusion Transformers

🔥 For more results, visit our homepage 🔥

👋 Join our Discord

This repo, named SkyReels-A2, contains the official PyTorch implementation of our paper SkyReels-A2: Compose Anything in Video Diffusion Transformers.

🎉 News

• Jun 1, 2025: 🎉 We published the technical report, SkyReels-Audio: Omni Audio-Conditioned Talking Portraits in Video Diffusion Transformers.
• Apr 21, 2025: 👋 We release the Infinite-Length Film Generative Model, SkyReels-V2 Series Models and the video captioning model SkyCaptioner-V1 .
• Apr 8, 2025: 🎉 We open the A2-Bench online evaluation and leaderboard. Test it on A2-Bench Leaderboard !
• Apr 7, 2025: 🔥 ComfyUI is now available.
• Apr 3, 2025: 🔥 We release pre-view version of checkpoints, code of model inference and gradio demo.
• Previously, we released SkyReels-A1. This is an open-sourced and effective framework for portrait image animation.

📑 TODO List

• Support A2-Bench evaluation and leaderboard.
• ComfyUI
• Parallel Inference on Multi-GPUs
• TeaCache for faster inference
• User-Level GPU Inference on RTX4090
• Release all model sequence, including infinity-long version.
• Diffusers

🪄 Models

Models	Download Link	Video Size
A2-Wan2.1-14B-Preview	Huggingface 🤗	~ 81 x 480 x 832
A2-Wan2.1-14B	To be released	~ 81 x 480 x 832
A2-Wan2.1-14B-Pro	To be released	~ 97 x 544 x 960
A2-Wan2.1-14B-Infinity	To be released	~ Inf x 720 x 1080

1. Getting Started 🏁

1.1 Clone the code and prepare the environment 🛠️

First git clone the repository with code:

git clone https://github.com/SkyworkAI/SkyReels-A2.git
cd SkyReels-A2

# create env using conda
conda create -n skyreels-a2 python=3.10
conda activate skyreels-a2

Then, install the remaining dependencies:

pip install -r requirements.txt

1.2 Download pretrained weights 📥

You can download the pretrained weights from HuggingFace as:

# !pip install -U "huggingface_hub[cli]"
huggingface-cli download Skywork/SkyReels-A2 --local-dir local_path --exclude "*.git*" "README.md" "docs"

or download from webpage mannually.

1.3 Inference 🚀

You can first set the model path and reference images path and then simply run the inference scripts as:

python infer.py

If the script runs successfully, you will get an output mp4 file. This file includes the following results: driving video, input image or video, and generated result.

We also support multi-GPU inference scripts for faster inference, as:

torchrun --nproc_per_node=$GPU_NUM infer_MGPU.py

Set the offload_switch of infer_MGPU.py to True, and you can run it on RTX4090

Gradio Interface 🤗

We also provide a Gradio interface for a better user experience, just run by:

python app.py

The graphical interactive interface is shown as below.

2. A2-Bench Evaluation 👓

We public the evaluation data in Huggingface, you can infer with results and then submit to leaderboard to obtain the results automatically. More detail about metric computation code will coming soon.

Acknowledgements 💐

We would like to thank the contributors of Wan and finetrainers repositories, for their open research and contributions.

Citation 💖

If you find SkyReels-A2 useful for your research, welcome to 🌟 this repo and cite our work using the following BibTeX:

@article{fei2025skyreels,
  title={SkyReels-A2: Compose Anything in Video Diffusion Transformers},
  author={Fei, Zhengcong and Li, Debang and Qiu, Di and Wang, Jiahua and Dou, Yikun and Wang, Rui and Xu, Jingtao and Fan, Mingyuan and Chen, Guibin and Li, Yang and others},
  journal={arXiv preprint arXiv:2504.02436},
  year={2025}
}

Star History