Abstract
Recent work has increased the performance of Generative Adversarial Networks(GANs) by enforcing a consistency cost on the discriminator. We improve on thistechnique in several ways. We first show that consistency regularization canintroduce artifacts into the GAN samples and explain how to fix this issue. Wethen propose several modifications to the consistency regularization proceduredesigned to improve its performance. We carry out extensive experimentsquantifying the benefit of our improvements. For unconditional image synthesison CIFAR-10 and CelebA, our modifications yield the best known FID scores onvarious GAN architectures. For conditional image synthesis on CIFAR-10, weimprove the state-of-the-art FID score from 11.48 to 9.21. Finally, onImageNet-2012, we apply our technique to the original BigGAN model and improvethe FID from 6.66 to 5.38, which is the best score at that model size.