Benchmark RadarAI BENCHMARK PROFILE
OmniPhys
OmniPhys is a benchmark of 1,551 text-to-image generation samples grounded in a Physical Knowledge Graph, aligned with PhET simulations and curricula. It evaluates physical commonsense in generated images using a dual-path verification protocol.
- Released
- 2026-07-28
- Readiness
- Runnable
- Primary field
- General AI
Why it matters
Existing benchmarks use coarse descriptions and fail to diagnose specific physical principles. OmniPhys provides a fine-grained, curriculum-aligned evaluation to identify systemic physical reasoning gaps in image generation models, supporting targeted improvement.
Motivation
While text-to-image models exhibit remarkable visual fidelity, they frequently violate fundamental physical commonsense.
Primary resources
Benchmark Radar records only publicly supported details and links back to primary sources for verification.