Spectral Normalization for Generative Adversarial Networks

Addressing the well-known instability of generative adversarial network training, the paper proposes spectral normalization, a new weight normalization technique applied to the discriminator. The method is designed to be computationally light and easy to incorporate into existing GAN implementations, making it a low-cost modification for practitioners seeking more stable training.

The authors test spectral normalization on the CIFAR-10, STL-10, and ILSVRC2012 (ImageNet) datasets and experimentally confirm that spectrally normalized GANs, or SN-GANs, can generate images of better or equal quality relative to previous training stabilization techniques. The approach became a widely adopted tool for stabilizing GAN discriminators.