We target the problem of zero-shot anomaly detection, in which a model is pre-trained on a
set of seen classes and expected to detect anomalies in other unseen classes at test time …