Visual anomaly detection, the industrial gold standard.

MVTec AD is the reference benchmark for unsupervised visual anomaly detection in industrial inspection. It judges systems on Image AUROC — how cleanly the detector separates defective from pristine parts on a held-out test set of real factory imagery.

Image AUROC, near the ceiling.

Image AUROC measures the area under the Receiver Operating Characteristic curve for the binary defective-versus-pristine decision at the image level. A perfect detector scores 100; the current SOTA sits at 99.6, which is why the leaderboard is now effectively saturated and the interesting research has moved to pixel-level AUROC and PRO.

Because scores cluster in the 97–99 range, small differences are not noise. A detector that gives up a tenth of a point is giving up detections on thousands of real inspections per shift.

Real factory images, real defects.

MVTec AD was released by MVTec Software GmbH. The dataset contains real industrial-inspection images across 15 object and texture categories — bottles, cables, capsules, carpet, grid, hazelnut, leather, metal nut, pill, screw, tile, toothbrush, transistor, wood, zipper — with pixel-accurate annotations of defect regions.

Models train on defect-free examples only; the test split contains both normal and defective samples, and the score is the mean over the 15 categories.

Reported, then reproduced.

Each row above is a reported Image AUROC from the submitting paper or repository. Values here are preserved verbatim — where a paper reports different numbers under different inference settings (e.g. with or without test-time augmentation), the row reflects the best reported figure the authors stand behind.

Full policy: /methodology.