DiffCap-Bench - Spatial: leaderboard

DiffCap-Bench (2026) image difference captioning restricted to spatial rearrangements of objects. 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 spatial rearrangements of objects: 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
1Gemini 3 Pro83.5#145
2GPT-5.282.7#185
3Qwen 3 VL Plus78.6#512
4Kimi K2.5Thinking78#266
5Seed1.6-VLThinking75.6#385
6Qwen 3 VL 32B (Thinking)75#573
7Step3 VL 10B73.9#1080
8MiMo-VL-7B-RL-250873.9#823
9Grok 4.2073.8#301
10Qwen 3 VL 32B Instruct71.5#606
11Qwen 3 VL 8B (Thinking)71.1#910
12Qwen 3 VL 8B Instruct67#1029
13InternVL3.5-38B63.1#1021
14InternVL3.5-8B63#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-spatial · How It Works · Data refreshed daily, snapshot 2026-11-21.