DistortBench - Compression: 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 two compression types (JPEG, JPEG2000), in percent; higher is better.

Metric: Joint accuracy (%) on the two compression types (JPEG, JPEG2000), 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 2031. 14 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1Qwen 3.5 4BNon-reasoning56.4#1241
2InternVL3.5-38B56.4#1021
3Qwen 3.5 9BNon-reasoning55.1#828
4Qwen 3.5 35B A3BNon-reasoning55#492
5Qwen 3.5 27BNon-reasoning54.9#375
6Kimi-VL-A3B-Instruct50#2253
7Gemma 3 27B48.1#1663
8InternVL3.5-14B46.8#1504
9InternVL3.5-8B46.7#1604
10Gemma 3 4B44.6#5339
11GPT-5.443.7#135
12Kimi-VL-A3B (Thinking)41.9
13Gemma 3 12B41.2#2059
14Qwen 2 VL 7B40.4#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-compression · How It Works · Data refreshed daily, snapshot 2026-11-21.