DistortBench: leaderboard

DistortBench (2026) shows a vision-language model a distorted photo without its original and asks it to name the distortion type and severity among four options (chance 25 percent). Joint accuracy over all 27 distortion types and five severity levels, in percent; higher is better.

Metric: Joint accuracy (%) on DistortBench's 13,500 four-choice questions (27 distortion types at five KADID-10k calibrated severity levels on 100 KADIS-700k photos, no reference image), the answer naming both type and severity, answered-only accuracy (unparseable responses, at most 3.4 percent, leave the denominator), chance 25; higher is better. Source: arxiv.org. Saturation forecast: Around July 2027. 14 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1Qwen 3.5 27BNon-reasoning61.9#375
2Qwen 3.5 35B A3BNon-reasoning56.8#492
3Qwen 3.5 4BNon-reasoning56.1#1241
4Qwen 3.5 9B55#828
5InternVL3.5-38B52.6#1021
6Kimi-VL-A3B-Instruct52.4#2253
7GPT-5.451.6#135
8Gemma 3 27B47.5#1663
9Gemma 3 12B46.6#2059
10InternVL3.5-8B44.4#1604
11Kimi-VL-A3B (Thinking)44
12InternVL3.5-14B43.4#1504
13Qwen 2 VL 7B36.7#2275
14Gemma 3 4B33.7#5339

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

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