CVPR2020

MSG-GAN: Multi-Scale Gradients for Generative Adversarial Networks

Animesh Karnewar, Oliver Wang

Abstract

Figure 1: Results of our proposed MSG-GAN technique where the generator synthesizes images at all resolutions simultaneously and gradients flow directly to all levels from a single discriminator. The first column has a resolution of 4x4 which increases towards the right reaching the final output resolution of 1024x1024. Best viewed zoomed in on screen.