SeaCache is a training-free diffusion cache schedule that bases reuse decisions on a spectrally aligned representations. By modeling spectral evolution along the denoising trajectory, we derive a Spectral-Evolution-Aware (SEA) filter that preserves content-relevant components while suppressing noise components.
- No additional parameters to tune. SeaCache introduces no additional hyperparameters (e.g., retention ratios or the coefficients). You just apply the scheduler-based SEA filtering and run inference.
- Minimal overhead. SEA filtering (including FFT/iFFT) is inexpensive yet highly effective, accounting only ~0.2% to total inference latency.
- Compatibility. SeaCache is compatible with orthogonal acceleration techniques, such as efficient attention and block-wise caching.
Figure 1: Conceptual illustration of SeaCache. SeaCache applies timestep-aligned, frequency-aware filtering to guide cache scheduling, achieving strong acceleration with negligible overhead.
- [2026-02-26] Support Wan2.1, HunyuanVideo, and FLUX.
- [2026-02-21] Released the paper, code, and project page of SeaCache.
ComfyUI support is planned for a future update.
- More open-source model support will be added over time.
- If you use SeaCache in your project and want it listed here, contact us at
jiwoo.jg@gmail.com.
This repository is built on top of TeaCache, VBench, Wan2.1, Diffusers, HunyuanVideo, and FLUX. We thank the authors and contributors for their efforts.
If SeaCache is useful for your research or applications, please consider starring this repository and citing:
@article{chung2026seacache,
title={SeaCache: Spectral-Evolution-Aware Cache for Accelerating Diffusion Models},
author={Chung, Jiwoo and Hyun, Sangeek and Lee, MinKyu and Han, Byeongju and Cha, Geonho and Wee, Dongyoon and Hong, Youngjun and Heo, Jae-Pil},
journal={arXiv preprint arXiv:2602.18993},
year={2026}
}