Benchmark RadarRigidBench
RigidBench evaluates rigid-body physics in video generation models using a simulator-grounded benchmark with 100 examples and five tasks. It provides ten measurements covering motion, geometry, identity, background stability, and appearance, with per-frame data.
- Released
- 2026-08-16
- Readiness
- Runnable
- Primary field
- General AI
Why it matters
Video generation metrics often mix independent errors. RigidBench separates these aspects and shows rankings depend on what is measured, providing a more detailed evaluation for physics fidelity.
Motivation
Video models are increasingly used to predict what happens next in a scene, yet the metrics commonly used to compare their outputs say little about whether the predicted objects move correctly.
Primary resources
Benchmark Radar records only publicly supported details and links back to primary sources for verification.