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

<div class="is-size-5 publication-authors", align="center",> <span class="author-block"> <a href="https://liewfeng.github.io" target="_blank">Feng Liu</a>¹^*,&nbsp; </span> <span class="author-block"> <a href="https://scholar.google.com.hk/citations?user=ZO3OQ-8AAAAJ" target="_blank">Shiwei Zhang</a>²^†,&nbsp; </span> <span class="author-block"> <a href="https://jeffwang987.github.io" target="_blank">Xiaofeng Wang</a>^1,3,&nbsp; </span> <span class="author-block"> <a href="https://weilllllls.github.io" target="_blank">Yujie Wei</a>⁴,&nbsp; </span> <span class="author-block"> <a href="http://haonanqiu.com" target="_blank">Haonan Qiu</a>⁵ </span> <br> <span class="author-block"> <a href="https://callsys.github.io/zhaoyuzhong.github.io-main" target="_blank">Yuzhong Zhao</a>¹,&nbsp; </span> <span class="author-block"> <a href="https://scholar.google.com.sg/citations?user=16RDSEUAAAAJ" target="_blank">Yingya Zhang</a>²,&nbsp; </span> <span class="author-block"> <a href="https://scholar.google.com/citations?user=tjEfgsEAAAAJ&hl=en&oi=ao" target="_blank">Qixiang Ye</a>¹,&nbsp; </span> <span class="author-block"> <a href="https://scholar.google.com/citations?user=0IKavloAAAAJ&hl=en&oi=ao" target="_blank">Fang Wan</a>¹^‡ </span> </div> <div class="is-size-5 publication-authors", align="center"> <span class="author-block">¹University of Chinese Academy of Sciences,&nbsp;</span> <span class="author-block">²Alibaba Group</span> <br> <span class="author-block">³Institute of Automation, Chinese Academy of Sciences</span> <br> <span class="author-block">⁴Fudan University,&nbsp;</span> <span class="author-block">⁵Nanyang Technological University</span> </div> <div class="is-size-5 publication-authors", align="center"> (* Work was done during internship at Alibaba Group. † Project Leader. ‡ CorresCorresponding author.) </div> <h5 align="center">

</h5>

🫖 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

• If you like our project, please give us a star ⭐ on GitHub for the latest update.
• [2025/06/08] 🔥 Update coefficients of Lumina-Image-2.0. Thanks @spawner1145.
• [2025/05/26] 🔥 Support Lumina-Image-2.0. Thanks @spawner1145.
• [2025/05/25] 🔥 Support HiDream-I1. Thanks @YunjieYu.
• [2025/04/14] 🔥 Update coefficients of CogVideoX1.5. Thanks @zishen-ucap.
• [2025/04/05] 🎉 Recommended as a highlight in CVPR 2025, top 16.8% in accepted papers and top 3.7% in all papers.
• [2025/03/13] 🔥 Optimized TeaCache for Wan2.1. Thanks @zishen-ucap.
• [2025/03/05] 🔥 Support Wan2.1 for both T2V and I2V.
• [2025/02/27] 🎉 Accepted in CVPR 2025.
• [2025/01/24] 🔥 Support Cosmos for both T2V and I2V. Thanks @zishen-ucap.
• [2025/01/20] 🔥 Support CogVideoX1.5-5B for both T2V and I2V. Thanks @zishen-ucap.
• [2025/01/07] 🔥 Support TangoFlux. TeaCache works well for Audio Diffusion Models!
• [2024/12/30] 🔥 Support Mochi and LTX-Video for Video Diffusion Models. Support Lumina-T2X for Image Diffusion Models.
• [2024/12/27] 🔥 Support FLUX. TeaCache works well for Image Diffusion Models!
• [2024/12/26] 🔥 Support ConsisID. Thanks @SHYuanBest.
• [2024/12/24] 🔥 Support HunyuanVideo.
• [2024/12/19] 🔥 Support CogVideoX.
• [2024/12/06] 🎉 Release the code of TeaCache. Support Open-Sora, Open-Sora-Plan and Latte.
• [2024/11/28] 🎉 Release the paper of TeaCache.

🧩 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

• FramePack supports TeaCache. Thanks @lllyasviel.
• FastVideo supports TeaCache. Thanks @BrianChen1129 and @jzhang38.
• EasyAnimate supports TeaCache. Thanks @hkunzhe and @bubbliiiing.
• Ruyi-Models supports TeaCache. Thanks @cellzero.
• ConsisID supports TeaCache. Thanks @SHYuanBest.

ComfyUI

• ComfyUI-TeaCache for TeaCache. Thanks @YunjieYu.
• ComfyUI-WanVideoWrapper supports TeaCache4Wan2.1. Thanks @kijai.
• ComfyUI-TangoFlux supports TeaCache. Thanks @LucipherDev.
• ComfyUI_Patches_ll supports TeaCache. Thanks @lldacing.
• Comfyui_TTP_Toolset supports TeaCache. Thanks @TTPlanetPig.
• ComfyUI-TeaCacheHunyuanVideo for TeaCache4HunyuanVideo. Thanks @facok.
• ComfyUI-HunyuanVideoWrapper supports TeaCache4HunyuanVideo. Thanks @kijai, ctf05 and DarioFT.

Parallelism

• Teacache-xDiT for multi-gpu inference. Thanks @MingXiangL.

Engine

• SD.Next supports TeaCache. Thanks @vladmandic.
• DiffSynth Studio supports TeaCache. Thanks @Artiprocher.

🎉 Supported Models

Text to Video

• TeaCache4Wan2.1
• TeaCache4Cosmos
• TeaCache4CogVideoX1.5
• TeaCache4LTX-Video
• TeaCache4Mochi
• TeaCache4HunyuanVideo
• TeaCache4CogVideoX
• TeaCache4Open-Sora
• TeaCache4Open-Sora-Plan
• TeaCache4Latte

Image to Video

• TeaCache4Wan2.1
• TeaCache4Cosmos
• TeaCache4CogVideoX1.5
• TeaCache4ConsisID

Text to Image

• TeaCache4Lumina2
• TeaCache4HiDream-I1
• TeaCache4FLUX
• TeaCache4Lumina-T2X

Text to Audio

• TeaCache4TangoFlux

🤖 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 d