Reproduce “ScoutCache: Efficient Few-Step Autoregressive Video Diffusion via Conditional Readout and Selective Caching” 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

Hao Tan, Jinghui Qin, Xuming Hu

ScoutCache accelerates autoregressive video diffusion by estimating an intermediate denoising update from early signals, allocating attention reads by output error, and selecting spatial tokens before cache construction. It calibrates on early chunks without changing pretrained weights. In 120-second Self-Forcing MovieGen generation, latency falls from 264.32 to 115.78 seconds, a 2.28× speedup, while VBench-Long Total Score rises from 75.03 to 79.11.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.