A Style-Based Generator Architecture for Generative Adversarial Networks

This paper introduces a new generator architecture for generative adversarial networks (GANs) that allows for unprecedented control over image synthesis. By automatically learning and separating high-level attributes (e.g., pose and identity when trained on human faces) and stochastic variations (e.g., freckles, hair), the style-based generator enables intuitive, scale-specific manipulations. This leads to better understanding of how GANs work and creates opportunities for improved image editing and generation.