Feature caching accelerates video diffusion by reusing intermediate network outputs across denoising steps, avoiding redundant forward passes.
Feedback-driven scheduling beats heuristics for caching in video diffusion acceleration
AlignCache calibrates caching thresholds against measured approximation error, adapts budgets per denoising phase, and outperforms prior accelerators on Wan2.1, Open-Sora, and HunyuanVideo.
Chinese Tech
Jiyue Jiang · Zhiyu Huo · Yuhan Chen · Yanruisheng Shao · Gong Zhang · Lei Deng
The Chinese University of Hong Kong · Huawei Technology Co. Ltd
Research Digest··3 min read
Jiang et al.
Why this paper
From Huawei Technology Co. Ltd and The Chinese University of Hong Kong
In one line
AlignCache uses counterfactual error feedback to adaptively cache features in video diffusion, improving speed without quality loss.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors (2 noted)
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§