RE-Edit - Referential: leaderboard
Metric: Pass rate (%) on the referential-reasoning subset of RE-Edit's 1,000 curated edit requests (which object the request refers to): a Qwen3-VL-30B judge marks each edited image pass or fail against the case rationale; higher is better. Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 12 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Seedream 4.0 | 53.4 |
| 2 | FLUX.2-dev | 50.5 |
| 3 | Qwen-Image-Edit | 50 |
| 4 | Nano Banana | 48 |
| 5 | Step1X-Edit-v1p2-preview | 47.5 |
| 6 | Ovis-U1-3B | 39.2 |
| 7 | Step1X-Edit-v1p1 | 38.2 |
| 8 | OmniGen2 | 38.2 |
| 9 | DreamOmni2 | 36.8 |
| 10 | FLUX.1 Kontext [dev] | 35.3 |
Interactive version: theaggregate.ai/benchmark?slug=re-edit-referential · How It Works · Data refreshed daily, snapshot 2026-09-29.