Benchmarks
Generative Physicshigher is better
Physics-IQ
Tests whether generative video models predict physically plausible continuations (solid/fluid mechanics, optics, thermodynamics, magnetism). The Physics-IQ Score is normalized to 0–100, where 100 is the variance between two real videos; even the best model scores far below.
Benchmark source- Domain
- Generative Physics
- Metric
- pts
- Orientation
- Higher is better
- Results
- 2
Ranking
| # | Model | Score | Source | Status |
|---|---|---|---|---|
| 1 | Sora 2 OpenAI | 42.3 | Physics-IQ leaderboard (google-deepmind/physics-IQ-benchmark)3rd-party | unverified |
| 2 | Sora OpenAI | 8.7 | Motamed et al. — Physics-IQ (arXiv:2501.09038)paper | unverified |