TeaCache4HiDream-I1

TeaCache can speedup HiDream-I1 2x without much visual quality degradation, in a training-free manner. The following image shows the results generated by TeaCache-HiDream-I1-Full with various rel_l1_thresh values: 0 (original), 0.17 (1.5x speedup), 0.25 (1.7x speedup), 0.3 (2.0x speedup), and 0.45 (2.6x speedup).

📈 Inference Latency Comparisons on a Single A100

HiDream-I1-Full	TeaCache (0.17)	TeaCache (0.25)	TeaCache (0.3)	TeaCache (0.45)
~50 s	~34 s	~29 s	~25 s	~19 s

Installation

pip install git+https://github.com/huggingface/diffusers
pip install --upgrade transformers protobuf tiktoken tokenizers sentencepiece

Usage

You can modify the rel_l1_thresh in line 297 to obtain your desired trade-off between latency and visul quality. For single-gpu inference, you can use the following command:

python teacache_hidream_i1.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 HiDream-I1 and Diffusers.

2.0 KiB Raw Blame History