MagCache4FramePack

MagCache can speedup FramePack 2.25x and FramePack-F1 2.45x without much visual quality degradation, in a training-free manner. The following video shows the results generated by MagCache-FramePack.

• FramePack demo

Prompt: The girl dances gracefully, with clear movements, full of charm.

The girl dances gracefully, with clear movements, full of charm.

• FramePack-F1 demo

Prompt: The girl dances gracefully, with clear movements, full of charm.

The girl dances gracefully, with clear movements, full of charm.

📈 Inference Latency Comparisons on a Single L20 GPU

Model	Original	TeaCache	MagCache (E010K3R02)	MagCache (E015K4R02)
FramePack 5s	9min	4min40s	4min20s	4min
FramePack-F1 5s	9min	4min	4min20s	3min40s

Usage

Follow FramePack to clone the repo and finish the installation, then copy 'magcache_demo_gradio.py' and 'magcache_demo_gradio_f1.py' in this repo to the FramePack repo. You can modify the 'magcache_thresh', 'magcache_K', and 'retention_ratio' in the web UI to obtain your desired trade-off between latency and visul quality.

# FramePack
python magcache_demo_gradio.py

# FramePack-F1
python magcache_demo_gradio_f1.py

If you want to recalibrate the mag_ratios in the function initialize_magcache, please replace magcache_framepack_forward in line 527 of magcache_demo_gradio.py with magcache_framepack_calibration and run the gradio demo with magcache. The mag_ratios will be printed in the shell.

Citation

If you find MagCache is useful in your research or applications, please consider giving us a star 🌟 and citing it by the following BibTeX entry.

@misc{ma2025magcachefastvideogeneration,
      title={MagCache: Fast Video Generation with Magnitude-Aware Cache}, 
      author={Zehong Ma and Longhui Wei and Feng Wang and Shiliang Zhang and Qi Tian},
      year={2025},
      eprint={2506.09045},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2506.09045}, 
}

Acknowledgements

We would like to thank the contributors to the FramePack and TeaCache.