OpenVLM A-Bench - Generative Distortion: leaderboard
A-Bench Generative Distortion accuracy from the OpenCompass OpenVLM Leaderboard (VLMEvalKit): multiple-choice questions on AI-generated images, judging generation artefacts such as malformed hands or distorted objects. 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 | GPT-4.1 (2025-04-14) | 70.4 | #257 |
| 2 | Gemma 3 27B | 67.2 | #606 |
| 3 | GPT-4.1 Mini | 65.2 | #363 |
| 4 | Qwen 2 VL 72B | 63.6 | #450 |
| 5 | Aquila-VL-2B | 62.4 | #1216 |
| 6 | InternVL3-8B | 61.6 | #624 |
| 7 | InternVL3-78B | 60.4 | #361 |
| 8 | Gemma 3 12B | 59.6 | #674 |
| 9 | InternVL3-38B | 59.2 | #413 |
| 10 | Qwen 2.5 VL 7B | 58 | #641 |
| 11 | Qwen 2 VL 7B | 57.6 | #743 |
| 12 | Grok 2 (1212) | 57.2 | #502 |
| 13 | GPT-4.1 Nano | 55.6 | #730 |
| 14 | Gemini 1.5 Pro | 55.6 | #454 |
| 15 | Llama 3.2 11B Instruct | 55.6 | #1108 |
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-generative-distortion · How It Works · Data refreshed daily, snapshot 2026-11-05.