DiffCap-Bench - Text and Symbol: leaderboard

DiffCap-Bench (2026) image difference captioning restricted to added, removed or modified text and symbols. 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 added, removed or modified text and symbols: 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 Plus90.5#512
2Qwen 3 VL 32B (Thinking)88.1#573
3Gemini 3 Pro88#145
4Kimi K2.5Thinking88#266
5GPT-5.286.3#185
6Seed1.6-VLThinking86.3#385
7MiMo-VL-7B-RL-250884.3#823
8Qwen 3 VL 8B (Thinking)83.7#910
9Step3 VL 10B80.9#1080
10InternVL3.5-8B78.9#1604
11InternVL3.5-38B78.5#1021
12Qwen 3 VL 8B Instruct77.3#1029
13Grok 4.2076.5#301
14Qwen 3 VL 32B Instruct76.2#606

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-text-and-symbol · How It Works · Data refreshed daily, snapshot 2026-11-21.