PaintBench - Symbolic Reasoning: leaderboard
PaintBench mIoU (%) averaged over its five symbolic reasoning tasks (comparison, ordering, patterns, counting and legends); edits are scored pixel by pixel against the single correct image, higher is better (NYU, 2026).
Metric: mIoU (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Nano-Banana-2 | 22.6 | |
| 2 | Nano Banana | 18 | |
| 3 | GPT-Image-2 | 15.9 | |
| 4 | Qwen-Image-Edit-2511 | 7.7 | |
| 5 | BAGEL-7B-MoT | 5.1 | #5515 |
| 6 | FLUX.2-dev | 4.1 | |
| 7 | FLUX.1-Kontext-dev | 3.6 | |
| 8 | LongCat-Image-Edit | 3.5 | |
| 9 | FLUX.2-klein-9B | 3.2 | |
| 10 | HunyuanImage-3.0 | 0.3 |
Interactive version: theaggregate.ai/benchmark?slug=paintbench-symbolic-reasoning · How It Works · Data refreshed daily, snapshot 2026-11-12.