DiffusionAD is more like a Anomaly generator rather than Anomaly Detector

[scteam1994]

Anomaly detection is a challenging task because it's hard to get anomalous data, and often there is none available at all. Adding anomalous data to the training process and calculating the loss using mask labels is completely disconnected from real-world scenarios. No one working on anomaly detection would use such methods to improve model performance. The only value of this paper, in my opinion, lies in the generator part of the model, which could potentially be used to synthesize more anomalous images when there is already a certain amount of anomalous data available. However, this is not within the scope of anomaly detection (AD).