DiffCap-Bench - Expression: leaderboard

DiffCap-Bench (2026) image difference captioning restricted to facial expression changes. 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 facial expression changes: 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 April 2027. 14 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1Qwen 3 VL Plus79.1#512
2Gemini 3 Pro72.1#145
3Grok 4.2071.6#301
4Qwen 3 VL 32B Instruct66.2#606
5InternVL3.5-38B63.8#1021
6Qwen 3 VL 8B Instruct63.7#1029
7Kimi K2.5Thinking60.6#266
8InternVL3.5-8B58.2#1604
9Seed1.6-VLThinking55.9#385
10Qwen 3 VL 32B (Thinking)54.2#573
11Qwen 3 VL 8B (Thinking)48.3#910
12MiMo-VL-7B-RL-250845#823
13Step3 VL 10B37.8#1080
14GPT-5.230#185

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-expression · How It Works · Data refreshed daily, snapshot 2026-11-21.