MMR-AD - VisA Localization: leaderboard

Zero-shot industrial anomaly localization on the VisA images of MMR-AD (2026): given a normal reference and a test image, the model draws a box around the anomaly, counted correct at an IoU of at least 0.1 with an annotated region. Accuracy in percent, averaged over product classes; higher is better.

Metric: Anomaly localization accuracy (%) on VisA test images (from MMR-AD): a predicted anomaly box counts as correct at IoU of at least 0.1 with an annotated anomalous region, zero-shot, the model sees a spatially aligned normal reference image and the test image with the MMR-AD instruction template; dataset-level average over the subdataset's product classes; higher is better. Source: arxiv.org. Saturation forecast: Around April 2027. 13 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1GPT-518.5#178
2O4 Mini12.3#368
3Qwen 2.5 VL 72B11.1#781
4Gemini 2.5 Pro10.5#276
5Llama 4 Maverick6.7#1110
6QVQ-Max4.8#949
7GPT-4o3.7#751
8InternVL3-8B2.9#1782
9InternVL3-38B2.2#939
10Qwen 2.5 VL 7B2.2#1862
11InternVL3.5-8B2.2#1604
12InternVL3.5-38B1.8#1021
13Gemma 3 27B0.5#1663

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=mmr-ad-visa-localization · How It Works · Data refreshed daily, snapshot 2026-11-21.