Classes of GANs

Based on the discrepancy measure, GANs fall into two broad classes.

1. Basic GANs use an f-divergence: vanilla GANs, CGANs, DCGANs. They mainly rely on the relatively unstable KL and JS divergences. KL shoots to infinity for non-overlapping distributions; JS becomes log 2 for any two non-overlapping distributions, so it is more stable than KL.

2. GANs using IPMs: Wasserstein GANs (WGAN, WGAN-GP) and MMD-GANs, whose discriminator implements an IPM such as the Wasserstein distance or Maximum Mean Discrepancy.