Large-Scale Training of Generative Adversarial Networks

Abstract

Generative models have transformed various fields, particularly image synthesis, by providing a data-centric approach to visual content creation. This thesis investigates generative adversarial networks (GANs), a prominent class of generative models recognized for their fast inference and control over synthesized outputs. Despite their advantages, GANs encounter difficulties when scaling to large, diverse datasets, such as training instability and mode collapse. To tackle these issues, we devise novel techniques and network architectures for scaling GANs. Our contributions encompass Projected GANs, which achieve state-of-the-art performance on 22 benchmark datasets and train up to 40 times faster; StyleGAN-XL, a GAN trained on ImageNet, constituting the new state-of-the-art; and StyleGAN-T, a large-scale text-to-image synthesis model, the first GAN surpassing one billion parameters in model-size. By progressing towards developing the first GAN foundation model, we aspire to further enhance the field of generative artificial intelligence, expanding its impact on image synthesis and unlocking new applications, such as real-time image editing, personalized content generation, and interactive virtual experiences.