GeoR-Bench: leaderboard

Strict accuracy in percent: the share of edited images that pass the reasoning, consistency and image quality checks at once, averaged over 440 geoscience image-editing samples in six categories (GeoR-Bench, 2026), judged by Gemini 3 Flash with task-specific rubrics. Higher is better.

Metric: Strict accuracy (%): share of samples whose output passes the reasoning, consistency and image-quality checks at once (mean of the six category accuracies), 440 reasoning-informed geoscience image-editing samples in six categories (geomorphology, hydrology, atmosphere and ocean, cryosphere, GIS and spatial geometry, crustal science), one generated output per sample, Gemini 3 Flash judge with task-specific rubrics; higher is better. Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 19 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1GPT Image 242.7
2Nano Banana 239.2
3Nano Banana Pro35
4GPT Image 1.526.1
5Seedream 5.019.6
6Nano Banana14.6
7Seedream 4.511.7
8FLUX 2 Max10.3
9Qwen-Image-Edit-251110.3
10FLUX 2 Pro5.63
11FLUX.2-dev5.07
12Seedream 4.04.99
13Step1X-Edit3.45
14DreamOmni3.15
15Step1X-Edit (Thinking + Reflection)2.67

Interactive version: theaggregate.ai/benchmark?slug=geor-bench · How It Works · Data refreshed daily, snapshot 2026-11-21.