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

TL;DR

MSG-GAN introduces multi-scale gradient flow to improve stability and performance in high-resolution image synthesis, offering an effective alternative to progressive growing techniques in GAN training.

Contribution

The paper proposes MSG-GAN, a novel method enabling multi-scale gradient flow from discriminator to generator, enhancing stability and versatility across datasets and architectures.

Findings

Stable convergence across various datasets and resolutions

Matches or exceeds state-of-the-art GAN performance

Operates with fixed hyperparameters across different settings

Abstract

While Generative Adversarial Networks (GANs) have seen huge successes in image synthesis tasks, they are notoriously difficult to adapt to different datasets, in part due to instability during training and sensitivity to hyperparameters. One commonly accepted reason for this instability is that gradients passing from the discriminator to the generator become uninformative when there isn't enough overlap in the supports of the real and fake distributions. In this work, we propose the Multi-Scale Gradient Generative Adversarial Network (MSG-GAN), a simple but effective technique for addressing this by allowing the flow of gradients from the discriminator to the generator at multiple scales. This technique provides a stable approach for high resolution image synthesis, and serves as an alternative to the commonly used progressive growing technique. We show that MSG-GAN converges stably on a variety of image datasets of different sizes, resolutions and domains, as well as different types of loss functions and architectures, all with the same set of fixed hyperparameters. When compared to state-of-the-art GANs, our approach matches or exceeds the performance in most of the cases we tried.