GEOPHYS: The Geometry of Physical Plausibility

The paper introduces GEOPHYS, a novel approach to assessing physical plausibility in videos by leveraging five geometric properties of per-frame embeddings from frozen image encoders. GEOPHYS achieves state-of-the-art performance in physics-violation detection, with 98.3% accuracy on LikePhys and 93.3% on IntPhys2, significantly outperforming models like V-JEPA 2 and GPT-4o. Additionally, it enhances the performance of the MAGI-1 24B model on the PhysicsIQ benchmark, reducing computational costs while improving accuracy, making it a valuable tool for practitioners focused on video generation and analysis in AI.