Principia: Relational Physics Tests for Video Models
The benchmark says top video generators still fail basic relational physics tests.
Principia compares paired objects in the same scene, avoiding calibration problems from frame rate, scale, or camera setup. It covers eight Newtonian phenomena, including gravity, friction, momentum, pendulums, and mass-spring motion. Across thousands of generations from six state-of-the-art video models, none scored above 0.42 on Principia, even as they scored around 0.8 on VBench. Vision-language models also struggled to spot the violations, with the best reaching 67% accuracy and most near chance.
HF Daily Papers' note
Principia compares paired objects in the same scene, avoiding calibration problems from frame rate, scale, or camera setup. It covers eight Newtonian phenomena, including gravity, friction, momentum, pendulums, and mass-spring motion. Across thousands of generations from six state-of-the-art video models, none scored above 0.42 on Principia, even as they scored around 0.8 on VBench. Vision-language models also struggled to spot the violations, with the best reaching 67% accuracy and most near chance.
HF Daily Papers' note
score 5