Improved Consistency Regularization for GANs

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