ViGoR-Bench - Rule Obey: leaderboard

Metric: Rule-obey rate (%): share of final outputs that follow the instruction's explicit constraints (for example no wall crossing in a maze), over ViGoR-Bench's 918 generative visual-reasoning tasks (physical, knowledge and symbolic reasoning: sorting, assembly, Sudoku, mazes, jigsaws, equations, function plots and more), judged by Gemini-2.5-Pro against the human-verified reference image or answer; image-to-image models without chain of thought; higher is better. Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 18 models tracked.

Top models

#ModelScoreOverall rank
1Nano Banana Pro62
2Nano Banana30.9
3GPT-image-129.3
4Seedream 4.028.7
5Qwen-Image-Edit-250914.3
6Lumina-DiMOO13.2
7FLUX.1-Kontext-dev13.1
8FLUX.2-dev11.5
9LongCat-Image-Edit11.4
10Qwen-Image-Edit-251111.1

Interactive version: theaggregate.ai/benchmark?slug=vigor-bench-rule-obey · How It Works · Data refreshed daily, snapshot 2026-10-11.