VDiff-Bench: leaderboard
Fine-grained image-difference identification: 1,756 four-choice questions over 1,543 image pairs covering position, motion, colour, appearance, noise, texture, size, OCR and illumination changes; invalid outputs count as wrong, random is 25%. Higher is better. Source: arxiv.org/abs/2609.06245.
Metric: Choice Accuracy (%). Source: arxiv.org. Saturation forecast: Around February 2027. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 89.6 | #64 |
| 2 | Kimi K2.5 | 89.2 | #152 |
| 3 | Gemini 3.5 Flash | 88.8 | #66 |
| 4 | GPT-5.4 | 84.4 | #83 |
| 5 | Kimi K3 | 84.2 | #35 |
| 6 | Grok 4.3 | 67.3 | #204 |
| 7 | Gemini 2.5 Flash | 65.9 | #255 |
| 8 | Qwen 3 VL 8B (Thinking) | 58 | #423 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=vdiff-bench · How It Works · Data refreshed daily, snapshot 2026-11-05.