GenScale - Common-Object Scale: leaderboard
GenScale (2026) evaluates whether image generators render familiar objects at physically plausible relative sizes, scoring each anchor-target object pair on a five-point ordinal scale with a Gemini 3.1 Pro Preview judge calibrated against nine human raters. This board is the share of common-object pairs judged plausible in text-to-image generations, pooling natural-depth and same-plane scenes, in percent, higher is better. Pairs in images that fail a visual-quality prefilter or miss an object are excluded, so each model is scored only on the pairs it rendered validly.
Metric: Plausible pairs (%; share of anchor-target common-object pairs in text-to-image generations judged plausible in relative size, score 3 of 5, natural-depth and same-plane splits pooled, Gemini 3.1 Pro Preview judge calibrated against nine human raters, five samples per pair, majority vote; pairs in images failing a visual-quality prefilter or missing an object are excluded, 943 to 1,136 scored pairs per model). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Nano Banana 2 | 56.5 | |
| 2 | GPT Image 2 | 51.9 | |
| 3 | Grok Imagine | 46.7 | |
| 4 | Z-Image-Turbo | 46.1 | |
| 5 | Qwen-Image-2512 | 40 | |
| 6 | FLUX.2 (GenScale variant unspecified) | 33.9 | |
| 7 | Stable Diffusion 3.5 Large | 31.1 |
Interactive version: theaggregate.ai/benchmark?slug=genscale-common-object-scale · How It Works · Data refreshed daily, snapshot 2026-11-12.