Abstract: We consider the problem of zero-shot anomaly detection in which a model is pre-trained to detect anomalies in images belonging to seen classes, and expected to detect anomalies from unseen ...