量子位 (微信公众号)T2·Models
78 pts
A team at HKUST (Guangzhou) introduces Motion Forcing, a framework that decomposes video generation into three layers: points (trajectories), shape (depth), and appearance (color), deciding motion before rendering visuals to eliminate vehicle interpenetration and physical implausibility. On real Waymo driving data it ranks first on both motion coherence and physical plausibility, surpassing Seedance 2.0 and Wan 2.6, with code and an interactive demo open-sourced.