TeaCache4FLUX

TeaCache can speedup FLUX 2x without much visual quality degradation, in a training-free manner. The following image shows the results generated by TeaCache-FLUX with various rel_l1_thresh values: 0 (original), 0.25 (1.5x speedup), 0.4 (1.8x speedup), 0.6 (2.0x speedup), and 0.8 (2.25x speedup).

📈 Inference Latency Comparisons on a Single A800

FLUX.1 [dev]	TeaCache (0.25)	TeaCache (0.4)	TeaCache (0.6)	TeaCache (0.8)
~18 s	~12 s	~10 s	~9 s	~8 s

Installation

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

Usage

You can modify the rel_l1_thresh in line 320 to obtain your desired trade-off between latency and visul quality. For single-gpu inference, you can use the following command:

python teacache_flux.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 FLUX and Diffusers.