I'm attempting to design an AD application, but my dataset has quite a lot of intraclass variance, in that there is quite a lot of variance amongst "normal" images. I've been using PaDiM, which I understand has low complexity, and this has caused the model to have a very high false positive rate (up to 40%). I've then placed a classifier upstream of the model, which separates the data into a small number (3-4) of subclasses of normal images and hoped that a separate PaDiM model for each subclass would perform well, but even then it's struggling to avoid FPs.

I've decided that PaDiM probably doesn't have sufficient complexity to encapsulate the variance in the normal images and improve the FP rate, but now I'm stuck on what AD algorithm to try next. Is there some kind of available ranking in terms of complexity of models available in Anomalib? Or, can someone who knows the algos well provide me with some suggestions? I'll also take any suggestions about approaches to AD problems with high variance in non-anomalous images. Thanks everyone!