Physics-IQ (Do generative video models understand physical principles?)

Physics-IQ 物理理解基准Advanced

A benchmark of real filmed footage that tests whether video-generation models actually understand physics, not just look realistic.

Physics-IQ was released in January 2025 by researchers at INSAIT (Sofia University, Bulgaria) and Google DeepMind, with the paper appearing at WACV 2026. Every clip is real footage: 66 physical scenarios, each filmed twice from 3 camera angles, for 396 eight-second videos spanning solid mechanics, fluids, optics, thermodynamics, and magnetism. At test time a model only sees the beginning (one frame for image-to-video models, three seconds for video-continuation models) and has to generate the next 5 seconds, which is then compared against the real footage using spatial IoU (intersection-over-union, checking whether things moved to the right place), spatiotemporal IoU, weighted spatial IoU, and mean squared error, combined into a Physics-IQ score; the difference between two real takes of the same scenario is calibrated to 100%. In the original paper Sora scored only 10%, leading to the conclusion that looking realistic is not the same as understanding physics. It is often used to test whether video models can serve as world models; an improved-prompt, improved-metric Physics-IQ Verified version followed in 2026, with a leaderboard updated continuously.

ExampleShow a model the moment the first domino has just been tipped over but hasn't yet touched the second one, and ask it to generate the next 5 seconds; check whether the dominoes fall in order, matching the real filmed video pixel by pixel.

Also called: Physics-IQ Benchmark, Physics-IQ Verified
Related: Intuitive Physics · Video Generation Model · World Model · VBench: Comprehensive Benchmark Suite for Video Generative Models · Sora · WorldScore: A Unified Evaluation Benchmark for World Generation
Sources: Do generative video models understand physical principles? (arXiv 2501.09038)
Physics-IQ benchmark GitHub（含 Original / Verified 排行榜） (Chinese)
Physics-IQ 项目主页 (Chinese)
As of: 2026-09