OmniPhys: Knowledge-Graph-Driven Benchmarking and Collective Optimization for Physical Commonsense in Text-to-Image Generation

Researchers introduced OmniPhys, a 1,551-sample benchmark for probing physical commonsense in text-to-image generation.

Built from a Physical Knowledge Graph and aligned to PhET curricula, it diagnoses specific reasoning failures in generated images.

They also proposed OmniPrompt, an iterative prompt optimization method that aggregates multiple samples and batch feedback to reduce noise.

Tests across 12 text-to-image models found broad weaknesses in physical reasoning, while OmniPrompt improved consistency and physical alignment.