Many recent approaches to learn submodular functions suffer from limited expressiveness.

Explanations forthisphenomenon arecontroversial: Whilemostworksattribute the artifacts to the generator, other works point to the discriminator. We take a sober look at those explanations and provide insights on what makes proposed measures against high-frequency artifacts effective.

Withthetwo-branch architecture, itnaturally incorporates pre-trained moving statistics in BN layers during fine-tuning, exploiting more priorknowledge ofpre-trained networks.

Assuming that small spatial structures are more-or-less consistent across these measurements, these methods project to a local low-rank approximation of the data [37,31].