TeaCache

Timestep Embedding Tells: It's Time to Cache for Video Diffusion Model

ali-vilab
869
36
Python

[CVPR 2025 Highlight] Timestep Embedding Tells: It’s Time to Cache for Video Diffusion Model

Feng Liu1*,  Shiwei Zhang2†,  Xiaofeng Wang1,3,  Yujie Wei4,  Haonan Qiu5
Yuzhong Zhao1,  Yingya Zhang2,  Qixiang Ye1,  Fang Wan1‡
1University of Chinese Academy of Sciences,  2Alibaba Group
3Institute of Automation, Chinese Academy of Sciences
4Fudan University,  5Nanyang Technological University
(* Work was done during internship at Alibaba Group. † Project Leader. ‡ CorresCorresponding author.)

hf_paper
arXiv
Home Page
License
github

visualization

🫖 Introduction

We introduce Timestep Embedding Aware Cache (TeaCache), a training-free caching approach that estimates and leverages the fluctuating differences among model outputs across timesteps, thereby accelerating the inference. TeaCache works well for Video Diffusion Models, Image Diffusion models and Audio Diffusion Models. For more details and results, please visit our project page.

🔥 Latest News

🧩 Community Contributions

If you develop/use TeaCache in your projects and you would like more people to see it, please inform us.(liufeng20@mails.ucas.ac.cn)

Model

ComfyUI

Parallelism

Engine

🎉 Supported Models

Text to Video

Image to Video

Text to Image

Text to Audio

🤖 Instructions for Supporting Other Models

  • Welcome for PRs to support other models.
  • If the custom model is based on or has similar model structure to the models we’ve supported, you can try to directly transfer TeaCache to the custom model. For example, rescaling coefficients for CogVideoX-5B can be directly applied to CogVideoX1.5, ConsisID and rescaling coefficients for FLUX can be directly applied to TangoFlux.
  • Otherwise, you can refer to these successful attempts, e.g., 1, 2.

💐 Acknowledgement

This repository is built based on VideoSys, Diffusers, Open-Sora, Open-Sora-Plan, Latte, CogVideoX, HunyuanVideo, ConsisID, FLUX, Mochi, LTX-Video, Lumina-T2X, TangoFlux, Cosmos, Wan2.1, HiDream-I1 and Lumina-Image-2.0. Thanks for their contributions!

🔒 License

📖 Citation

If you find TeaCache is useful in your research or applications, please consider giving us a star ⭐ and citing it by the following BibTeX entry.

@article{liu2024timestep,
  title={Timestep Embedding Tells: It's Time to Cache for Video Diffusion Model},
  author={Liu, Feng and Zhang, Shiwei and Wang, Xiaofeng and Wei, Yujie and Qiu, Haonan and Zhao, Yuzhong and Zhang, Yingya and Ye, Qixiang and Wan, Fang},
  journal={arXiv preprint arXiv:2411.19108},
  year={2024}
}