Abstract
We propose a rejection sampling scheme using the discriminator of a GAN toapproximately correct errors in the GAN generator distribution. We show thatunder quite strict assumptions, this will allow us to recover the datadistribution exactly. We then examine where those strict assumptions break downand design a practical algorithm - called Discriminator Rejection Sampling(DRS) - that can be used on real data-sets. Finally, we demonstrate theefficacy of DRS on a mixture of Gaussians and on the state of the art SAGANmodel. On ImageNet, we train an improved baseline that increases the bestpublished Inception Score from 52.52 to 62.36 and reduces the Frechet InceptionDistance from 18.65 to 14.79. We then use DRS to further improve on thisbaseline, improving the Inception Score to 76.08 and the FID to 13.75.