AnomalyVFM -- Transforming Vision Foundation Models into Zero-Shot Anomaly Detectors

Researchers introduced AnomalyVFM, a framework that turns vision foundation models into stronger zero-shot anomaly detectors.

It generates more diverse synthetic anomalies and adapts models with low-rank feature adapters plus a confidence-weighted pixel loss.

Using RADIO as the backbone, it achieved 94.1% average image-level AUROC across nine datasets, beating prior methods by 3.3 points.

The result suggests vision-only foundation models can outperform earlier anomaly detectors with better synthetic training data and efficient adaptation.