PromptTea: Let Prompts Tell TeaCache the Optimal Threshold

🫖 Introduction

This repository contains the official implementation of our paper: PromptTea: Let Prompts Tell TeaCache the Optimal Threshold

🔥 Latest News

• If you like our project, please give us a star ⭐ on GitHub for the latest update.
• [2025/10/27] 🔥 Support multi-gpu parallel inference for the Wan2.2-I2V and Wan2.2-S2V.
• [2025/10/09] 🎉 The dataset of Complex-Simple prompt in the paper, as well as the dataset of the fitting model, have been made public. The dataset of the fitting model is from T2V-CompBench.
• [2025/08/06] 🔥 Support Wan2.2.
• [2025/07/30] 🔥 Support multi-gpu parallel inference for the Wan2.1.
• [2025/07/10] 🎉 Submitted to arXiv for review.
• [2025/07/10] 🔥 Released open-source code of PromptTea. Support CogVideoX1.5, HunyuanVideo and Wan2.1.

🎉 Supported Models

Text to Video

• PromptTea4CogVideoX1.5
• PromptTea4HunyuanVideo
• PromptTea4Wan2.1
• PromptTea4Wan2.2

Image to Video

• PromptTea4Wan2.2-I2V

Audio-Driven

• PromptTea4Wan2.2-S2V

💐 Acknowledgement

This repository is built based on TeaCache, TeaCache-xDit, FasterCache, Diffusers, CogVideoX, HunyuanVideo, Wan2.1 and Wan2.2. Thanks for their contributions!

📚 Citation

@article{huang2025prompttea,
  title={PromptTea: Let Prompts Tell TeaCache the Optimal Threshold},
  author={Huang, Zishen and Yang, Chunyu and Ren, Mengyuan},
  journal={arXiv preprint arXiv:2507.06739},
  year={2025}
}

📜 License

This project is licensed under the Apache 2.0 License.