Reproduce “What One Sample Cannot See: Why Fewer Denoising Steps Can Score Higher on Physics Benchmarks” 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

Yutao Li, Chen Wang

The authors evaluate video generators’ forecasts of physical events with Kernel Physics-IQ, a two-sample score that compares generated futures with each other and a recorded reference. It separates predictive fit from sample variation. Across 66 scenes, CogVideoX’s official score rises from 0.301 to 0.356 with two-step sampling, while Kernel Physics-IQ falls from 0.737 to 0.607; Kernel scores fall in all five models, but official-score rises are point estimates for only four.

Summary by alphaXiv.

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