PhyEditBench: leaderboard
PhyEditBench (2026) tests physics-aware instruction-based image editing on real-world physical processes extracted from videos, with outputs scored by a GPT-4o judge from 1 to 10 on consistency, instruction following, physical plausibility and image quality. The board is the weighted overall score (0.2, 0.3, 0.4 and 0.1 on the four dimensions) on the 238 real-world instances, averaged over single-step, multi-step and global edits; higher is better.
Metric: Overall judge score (1-10 scale) on the 238 real-world physical-process instances: 0.2 consistency + 0.3 instruction following + 0.4 physical plausibility + 0.1 image quality, averaged over five run types (three single-step edits between consecutive states, three steps applied jointly, one global edit); 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 | Overall rank |
|---|---|---|---|
| 1 | ChronoEdit-14B | 8.51 | |
| 2 | Seedream4.0 | 8.47 | |
| 3 | GPT-Image-1.5 | 8.23 | |
| 4 | UniWorld-V2 | 7.08 | |
| 5 | Step1X-Edit-v1p1 | 6.79 | |
| 6 | Qwen-Image-Edit | 6.71 | |
| 7 | gemini-2.5-flash-image | 6.51 | |
| 8 | BAGEL-7B-MoT (Thinking) | 6.43 | #5470 (BAGEL-7B-MoT) |
| 9 | BAGEL-7B-MoT | 6.09 | #5470 |
| 10 | InstructPix2Pix | 5.61 |
Interactive version: theaggregate.ai/benchmark?slug=phyeditbench · How It Works · Data refreshed daily, snapshot 2026-11-13.