OpenVLM A-Bench - Aesthetic Quality: leaderboard
A-Bench Aesthetic Quality accuracy from the OpenCompass OpenVLM Leaderboard (VLMEvalKit): multiple-choice questions on AI-generated images, judging aesthetic quality such as composition, colour and lighting. A dimension of A-Bench-P2 (low-level quality perception). Accuracy in percent, higher is better.
Metric: Accuracy (%). Source: huggingface.co. Saturation forecast: Estimated already saturated. 160 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 1.5 Pro | 70.8 | #454 |
| 2 | InternVL3-78B | 69.9 | #361 |
| 3 | Moonshot v1 8K | 68.5 | #622 |
| 4 | Qwen 2 VL 7B | 68 | #743 |
| 5 | Qwen 2 VL 72B | 66.7 | #450 |
| 6 | Grok 2 (1212) | 66.2 | #502 |
| 7 | InternVL3-14B | 65.8 | #506 |
| 8 | InternVL3-38B | 65.3 | #413 |
| 9 | InternVL2-8B | 65.3 | #813 |
| 10 | GPT-4.1 (2025-04-14) | 63.9 | #257 |
| 11 | Gemini 1.5 Flash | 63.9 | #620 |
| 12 | Qwen 2.5 VL 7B | 63.5 | #641 |
| 13 | GPT-4.1 Nano | 63 | #730 |
| 14 | MiniCPM-V-2.6 | 62.6 | #827 |
| 15 | InternVL3-8B | 62.1 | #624 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=openvlm-a-bench-aesthetic-quality · How It Works · Data refreshed daily, snapshot 2026-11-05.