In this paper, we present PhyWorldBench, a comprehensive benchmark for evaluating video generation models based on their adherence to the laws of physics. PhyWorldBench covers a wide range of physical phenomena, from basic principles such as object motion and energy conservation to more complex scenarios involving rigid-body interactions and human or animal movements. We also introduce an “Anti-Physics” category that uses prompts that intentionally violate the laws of physics in the real world, to assess whether models can follow these instructions while remaining logically consistent. In addition to large-scale human evaluation, we present a simple yet effective way to evaluate physical realism in a zero-shot manner by leveraging current MLLMs. We evaluate 12 state-of-the-art text-to-video generation models (five open source, five proprietary) and conduct a detailed comparative analysis to identify significant challenges models face in complying with the laws of physics in the real world. We conduct extensive testing on 1,050 curated prompts (basic, complex, and semi-physical scenarios) to rigorously examine performance on a wide range of physical phenomena with a variety of prompt types, and derive targeted guidelines for writing prompts that enhance fidelity to physical principles.
