Reproduce “CausalAlign: Causal Alignment of Video Generation Models for Consistent Ego-Motion” 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
Kai Zhu, Yintianrun Zhang, Shuwei Jiang, Ruolin Zuo, Yixin Chen, Duomin Wang, Miao Liu, Yin Li and 2 more
CausalAlign identifies layers and denoising stages where the camera-motion signal becomes less readable and influential, then uses a 3D teacher to align frame-to-frame transitions at those locations. On the tested video model, reprojection error falls from 8.4 to 6.5 pixels and trajectory smoothness rises from 0.71 to 0.88. These measure intrinsic consistency, not agreement with a required camera path.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.