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

TL;DR - OmniPhys is a 1,551-sample, knowledge-graph-grounded benchmark for diagnosing physical commonsense failures in text-to-image models. Its OmniPrompt framework improves physical consistency by aggregating feedback across stochastic generations and multiple queries.

• Builds scenarios from PhET simulations and standard curricula using a Physical Knowledge Graph.
• Uses dual-path verification to test specific physical principles rather than coarse descriptions.
• Evaluations across 12 text-to-image models reveal shared physical-reasoning bottlenecks.
• OmniPrompt filters generation noise through per-query feedback buffers and batch-level meta-policy updates.