DiffCap-Bench - Object: leaderboard

DiffCap-Bench (2026) image difference captioning restricted to object additions, deletions and replacements. 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 object additions, deletions and replacements: 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
1Gemini 3 Pro87.7#145
2Kimi K2.5Thinking86.6#266
3Seed1.6-VLThinking86.2#385
4Qwen 3 VL Plus85.3#512
5Qwen 3 VL 32B (Thinking)85#573
6GPT-5.283#185
7MiMo-VL-7B-RL-250881.4#823
8Step3 VL 10B79.7#1080
9Qwen 3 VL 8B (Thinking)79.4#910
10Grok 4.2077.4#301
11Qwen 3 VL 8B Instruct72.4#1029
12Qwen 3 VL 32B Instruct71.9#606
13InternVL3.5-38B70.8#1021
14InternVL3.5-8B69.3#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-object · How It Works · Data refreshed daily, snapshot 2026-11-21.