DiffCap-Bench - Composition: leaderboard

DiffCap-Bench (2026) image difference captioning restricted to composition differences, including cropping. Macro-averaged F1* in percent, combining the share of true differences described and the share of described differences that are correct, as judged by Gemini 2.5 Pro; higher is better.

Metric: Macro-averaged F1* (%) on composition differences, including cropping: harmonic mean of recall* (share of key differences described) and precision* (share of described differences that are correct) on the 1,075 image pairs of DiffCap-Bench (6,713 human-verified difference items), one fixed captioning prompt, open models at temperature 0.1 and proprietary models at their defaults, a Gemini 2.5 Pro judge matching each described difference against the key difference list; higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 14 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1GPT-5.291.5#190
2Gemini 3 Pro88#146
3Qwen 3 VL Plus79.3#514
4Kimi K2.5Non-reasoning77.5#272
5Step3 VL 10B77.2#1058
6Grok 4.2075.8#307
7MiMo-VL-7B-RL-250875.2#829
8Seed1.6-VLThinking73.6#389
9Qwen 3 VL 32B Instruct72.3#607
10Qwen 3 VL 8B Instruct69.2#1043
11Qwen 3 VL 32B (Thinking)69.2#577
12Qwen 3 VL 8B (Thinking)66.7#924
13InternVL3.5-8B62.9#1618
14InternVL3.5-38B59.2#1026

No result here: #2 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=diffcap-bench-composition · How It Works · Data refreshed daily, snapshot 2026-11-22.