Reproduce “STRIKE: Learning Visual State Transitions for Physical World Modeling” 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
Wenbin Teng, Tianshuo Xu, Depu Meng, Yuelei Li, Quentin Herau, Yihan Hu, Yajie Zhao, Wei Zhan
STRIKE predicts interaction outcomes as timed visual states, then conditions a video model to generate motion. Its image-transition model uses the current image, a local interaction description, and elapsed time to predict each next state. On PhyGenBench, STRIKE’s CogVideoX-5B variant scores 76.91, versus CausalMotion’s 70.21. CausalMotion scores higher in Mechanics, and these benchmark results do not establish general physical laws.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.