DiffCap-Bench - Action and Pose: leaderboard

DiffCap-Bench (2026) image difference captioning restricted to changes in actions and body poses. 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 changes in actions and body poses: 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
1Qwen 3 VL Plus75.8#512
2Gemini 3 Pro75.5#145
3Qwen 3 VL 32B (Thinking)68#573
4Seed1.6-VLThinking67.5#385
5Qwen 3 VL 8B Instruct66.9#1029
6Grok 4.2066.8#301
7Kimi K2.5Thinking66.8#266
8Qwen 3 VL 32B Instruct65.2#606
9MiMo-VL-7B-RL-250863.5#823
10Qwen 3 VL 8B (Thinking)63.3#910
11GPT-5.262.9#185
12InternVL3.5-38B62.7#1021
13InternVL3.5-8B59.9#1604
14Step3 VL 10B57.7#1080

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

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