Reproduce “DecoupAttn: Decoupled Sparse Attention for Training-Free Acceleration of Video Diffusion” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yanjin Li, Bingbing Qiu, Qiang Huang, Rui Qian, Yanfeng Li, Jian Wang, Xiaoying Tang, Mingda Zhang
DecoupAttn accelerates long-video attention by fixing offline the blocks receiving exact computation and using semantic key clusters to approximate skipped keys. At inference, it refreshes assignments and combines exact and estimated contributions under one softmax normalization. At a 0.20 sparse-unit density target, its mean PSNR is 30.02 dB versus 27.84 dB for SVG-EAR, measuring agreement with dense outputs rather than ground-truth quality.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.