DiffCap-Bench - Background: leaderboard

DiffCap-Bench (2026) image difference captioning restricted to background 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 background 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: Estimated already saturated. 14 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1Kimi K2.5Non-reasoning90#266
2Gemini 3 Pro89.4#145
3GPT-5.289#185
4Qwen 3 VL Plus87.9#512
5Seed1.6-VLThinking87.8#385
6Qwen 3 VL 32B (Thinking)85.7#573
7Step3 VL 10B84.1#1080
8MiMo-VL-7B-RL-250883#823
9Grok 4.2082.7#301
10Qwen 3 VL 8B (Thinking)81.4#910
11Qwen 3 VL 32B Instruct81.3#606
12Qwen 3 VL 8B Instruct77.3#1029
13InternVL3.5-38B73.8#1021
14InternVL3.5-8B71.9#1604

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