A video can look convincing while quietly violating gravity, momentum, or shadow consistency. That makes a single overall “realism” score hard to interpret: it tells you a model failed, but not what kind of failure to fix. This benchmark’s central value is turning physical plausibility into a set of inspectable, law-specific measurements.
What the Measurements Reveal
The taxonomy spans solid-body mechanics, fluid dynamics, and optics, with separate criteria for laws such as collision, fluid displacement, reflection, and shadows. Each clip is rated on applicable laws as well as semantic alignment, physical temporal validity, and object persistence. This makes it possible to distinguish, for example, a prompt-following failure from an implausible interaction. The released annotations preserve human judgments on a 1–5 scale, while the accompanying PhyJudge-9B model supports automated scoring against the same rubric.
Who Should Use It—and Where It Falls Short
A good fit for researchers comparing video generators, studying human–judge agreement, or developing physics-aware evaluation and training methods. It is less suitable as evidence of genuine causal understanding or real-world safety: its 13 observable laws are not an exhaustive physics model, and results depend on the sampled prompts, model versions, and inference settings. The ratings are diagnostic signals, not a substitute for simulation or physical validation.
