DistortBench - Noise: 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 on the five noise distortions (white, colored white, impulse, multiplicative, denoising), in percent; higher is better.

Metric: Joint accuracy (%) on the five noise distortions (white, colored white, impulse, multiplicative, denoising), a category of 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 March 2027. 14 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1Qwen 3.5 27BNon-reasoning65.7#375
2Qwen 3.5 35B A3BNon-reasoning60.5#492
3Qwen 3.5 9B60.1#828
4Qwen 3.5 4BNon-reasoning59.7#1241
5GPT-5.451.8#135
6Kimi-VL-A3B-Instruct51.3#2253
7InternVL3.5-38B46.6#1021
8Gemma 3 12B44.8#2059
9Kimi-VL-A3B (Thinking)44.2
10Gemma 3 27B43.4#1663
11InternVL3.5-8B42.2#1604
12Gemma 3 4B41.6#5339
13InternVL3.5-14B37.8#1504
14Qwen 2 VL 7B34#2275

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

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