How to interpret different ELBO values when checking anomaly detection possibilities of VAE model on different "testing" datasets? The higher the ELBO value of the model when testing it on different datasets, the better our model in detection of anomaly which follows "testing dataset distribution"?
the better our model in detection of anomaly
VAE is a generative model, not a discriminative one, so I'm not sure what you mean here.
More generally, higher mean ELBO over the test dataset doesn't mean that you have a better model of the test dataset. If you train a VAE to model the standard gaussian, and then test it on a dataset with just 0's, the model is going to assign higher mean ELBO rather than if you'd tested with a standard gaussian dataset.