GenScale - Scale Correction: leaderboard
GenScale (2026) evaluates whether image generators and editors keep familiar objects at physically plausible relative sizes, scoring object pairs with a Gemini 3.1 Pro Preview judge calibrated against nine human raters. This board is the share of implausible object pairs from failed generations that are judged plausible after a general-purpose image editor corrects the image, pooling an autonomous setting and one with explicit resize instructions, in percent, higher is better. Before editing, 20.4 percent of these pairs are plausible.
Metric: Plausible pairs after editing (%; share of implausible pairs from failed generations judged plausible after a general-purpose editor corrects the image, score 3 of 5, autonomous-discovery and explicit-instruction settings pooled, Gemini 3.1 Pro Preview judge calibrated against nine human raters, five samples per pair, majority vote; 20.4% before editing, 190 to 197 scored pairs per editor). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT Image 2 | 53.3 | |
| 2 | Nano Banana 2 | 32.5 | |
| 3 | Seedream 4.5 | 31.6 | |
| 4 | FLUX.1 Kontext-dev | 28.5 | |
| 5 | Qwen-Image-Edit-2511 | 28.4 |
Interactive version: theaggregate.ai/benchmark?slug=genscale-scale-correction · How It Works · Data refreshed daily, snapshot 2026-11-12.