DiffCap-Bench - Image Property: leaderboard

DiffCap-Bench (2026) image difference captioning restricted to global image property changes such as resolution or saturation. 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 global image property changes such as resolution or saturation: 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: Around February 2027. 14 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1GPT-5.289.1#190
2Gemini 3 Pro87.8#146
3Kimi K2.5Non-reasoning87.2#272
4Qwen 3 VL Plus85.9#514
5Seed1.6-VLThinking83.6#389
6Qwen 3 VL 32B Instruct82.9#607
7InternVL3.5-38B78.9#1026
8InternVL3.5-8B77.2#1618
9Qwen 3 VL 8B Instruct76.2#1043
10Qwen 3 VL 32B (Thinking)75.9#577
11MiMo-VL-7B-RL-250875.8#829
12Step3 VL 10B75.7#1058
13Qwen 3 VL 8B (Thinking)74.1#924
14Grok 4.2072.7#307

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-image-property · How It Works · Data refreshed daily, snapshot 2026-11-22.