The MVTec AD 2 Dataset: Advanced Scenarios for Unsupervised Anomaly Detection
Paper • 2503.21622 • Published
Industrial defect inspection: demo assets
Release bundles (weights, thresholds, reference statistics, metrics) and gallery images for the interactive demo of https://github.com/yassineerraji/Industrial-Visual-Defect-Detection-Anomaly-Segmentation.
| Bundle | Model | Supervision |
|---|---|---|
releases/patchcore |
patchcore | normal images only |
releases/autoencoder |
autoencoder | normal images only |
releases/unet |
unet | pixel masks (grouped 5-fold CV + final all-data model) |
Every metric in these bundles comes from tracked experiments; see the GitHub README for the protocol and results. Anomaly scores are unnormalised and are not probabilities.
Licence. Trained on MVTec AD 2 (MVTec Software GmbH), licensed CC BY-NC-SA 4.0. Model weights are distributed under the same terms, for non-commercial use. The gallery images are taken from the MVTec AD 2 public test set. Dataset: Heckler-Kram et al., The MVTec AD 2 Dataset: Advanced Scenarios for Unsupervised Anomaly Detection, arXiv:2503.21622, 2025.