PhyWorldBench: benchmarking physical realism in text-to-video
A benchmark evaluating how well text-to-video models obey physical laws, covering a broad range of scenarios rather than scoring on general visual quality.
Benchmarks decide what gets optimised. As long as models are ranked on how good a still frame looks, that is what improves. Measuring physical realism directly is how the field starts fixing the thing that actually costs production time.
Most video model leaderboards reward what looks good in a single frame or a short clip. PhyWorldBench asks a narrower and more useful question: does the motion obey physics? For anyone who has watched a generated shot fall apart the moment a glass is set down or a body shifts its weight, that is the question that decides whether a shot survives review, and it has been oddly underserved by the evaluations labs cite when they launch.
The paper, from a team including Ming-Yu Liu and Xin Eric Wang, organises physical phenomena into levels. At the base are fundamentals such as object motion and energy conservation; above them sit more complex cases like rigid body interactions and human or animal motion. The authors evaluate 12 text-to-video models across 1,050 curated prompts spanning fundamental, composite, and what they call anti-physics scenarios, using large-scale human evaluation alongside a zero-shot method in which multimodal language models judge physical realism. The work was first posted in July 2025 and revised in May 2026, which suggests it is being maintained rather than published and left.
The anti-physics category is the most interesting design choice for production. These prompts deliberately ask for things that violate real-world physics, testing whether a model can follow the instruction while keeping the result internally consistent. That is a commercial brief, not an edge case. A product that floats, liquid that pours upward, a dancer who hangs in the air a beat too long: advertising and music video work routinely asks for controlled impossibility, and a model that can only reproduce plausible physics is as limited as one that cannot reproduce it at all. Measuring both is the right instinct.
In pipeline terms, physics failures are expensive because they are discovered late. A frame can pass lookdev and still fail in motion, and the fix is usually comp, cleanup, or a regenerate-and-hope loop that eats schedule. A benchmark that separates failure types gives supervisors a vocabulary for that conversation: whether a model struggles with collisions, with weight transfer in human motion, or with conservation over time tells you which shots to keep away from it. The authors also derive prompt recommendations for physical fidelity, which is practical, though we would treat prompt tuning as a short-term workaround rather than a fix.
The limits are worth stating. The benchmark is text-to-video, while much professional work is conditioned on reference images, plates, or previous shots, so the scores do not translate directly to image-to-video or reference-driven workflows. Using multimodal models as judges scales evaluation but inherits their own blind spots on subtle physics, which is presumably why the human evaluation remains central. And any benchmark, once it matters, becomes a target.
That last point is also the optimistic one. If physical realism becomes a number labs compete on, models will improve on it, which is the argument in the takeaway above. Read alongside the Phantom paper on physics-infused generation in April and the addressable-memory work on world models, the direction is consistent: evaluation is moving from how a frame looks toward whether a scene behaves, which is closer to how a director or VFX supervisor actually judges footage.