PhyEditBench - Anti-Physics: leaderboard
Metric: Overall judge score (1-10 scale) on the 35 synthetic anti-physics instances whose edit prompt imposes a counterfactual physical rule (same weighting); GPT-4o judge scores each output 1-10 on consistency, instruction following, physical plausibility and image quality against the instruction, physical explanation, invariants and reference frames; higher is better. Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 12 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | gemini-2.5-flash-image | 7.07 |
| 2 | GPT-Image-1.5 | 7.04 |
| 3 | Seedream4.0 | 6.95 |
| 4 | Qwen-Image-Edit | 6.16 |
| 5 | BAGEL-7B-MoT (Thinking) | 6.16 |
| 6 | FLUX.1-Kontext-dev | 6.07 |
| 7 | UniWorld-V2 | 5.79 |
| 8 | Step1X-Edit-v1p1 | 5.34 |
| 9 | BAGEL-7B-MoT | 5.11 |
| 10 | ChronoEdit-14B | 4.99 |
Interactive version: theaggregate.ai/benchmark?slug=phyeditbench-anti-physics · How It Works · Data refreshed daily, snapshot 2026-09-29.