SpatialGen-Bench (Visual Answering): leaderboard
The generative-pixels protocol of SpatialGen-Bench (ProVisE): image-generation and editing models answer the same spatial tasks by producing pixels rather than text; the unweighted mean of all 14 integrated task accuracies, including the three expansion tasks. Only source-complete runs receive an overall.
Metric: Overall Accuracy (%). Source: zju-omniai.github.io. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 Image 2 | 55.44 | |
| 2 | Nano Banana 2 | 49.35 | |
| 3 | GPT-5 Image Mini | 40.97 | |
| 4 | Seedream 4.5 | 38.84 | |
| 5 | JoyAI-Image | 36 | |
| 6 | FLUX.2 [klein] 9B | 35.25 | |
| 7 | SenseNova-Vision-7B-MoT | 32.22 | |
| 8 | FLUX.2 [klein] 4B | 29.33 | |
| 9 | OmniGen-v1 | 27.44 | |
| 10 | BAGEL-7B-MoT | 27.1 |
Interactive version: theaggregate.ai/benchmark?slug=spatialgen-bench-visual-answering · How It Works · Data refreshed daily, snapshot 2026-11-05.