TeaCache4LTX-Video

TeaCache can speedup LTX-Video 2x without much visual quality degradation, in a training-free manner. The following video presents the videos generated by TeaCache-LTX-Video with various rel_l1_thresh values: 0 (original), 0.03 (1.6x speedup), 0.05 (2.1x speedup), 0.6 (2.0x speedup), and 0.8 (2.25x speedup).

https://github.com/user-attachments/assets/1f4cf26c-b8c6-45e3-b402-840bcd6ba00e

📈 Inference Latency Comparisons on a Single A800

LTX-Video-0.9.1	TeaCache (0.03)	TeaCache (0.05)
~32 s	~20 s	~16 s

Installation

pip install --upgrade diffusers[torch] transformers protobuf tokenizers sentencepiece imageio

Usage

You can modify the thresh in line 187 to obtain your desired trade-off between latency and visul quality. For single-gpu inference, you can use the following command:

python teacache_ltx.py

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}
}

Acknowledgements

We would like to thank the contributors to the LTX-Video and Diffusers.

1.8 KiB Raw Blame History