A Style-Based Generator Architecture for Generative Adversarial Networks
This work addresses the challenge of intuitive and high-quality image generation for applications like computer vision and graphics, representing a novel architectural advancement rather than an incremental improvement.
The paper tackles the problem of controlling image generation in GANs by proposing a style-based generator architecture that automatically separates high-level attributes and stochastic details, enabling scale-specific synthesis control. The result is improved state-of-the-art performance in distribution quality metrics, better interpolation, and enhanced disentanglement of latent factors, with new automated methods introduced for evaluation.
We propose an alternative generator architecture for generative adversarial networks, borrowing from style transfer literature. The new architecture leads to an automatically learned, unsupervised separation of high-level attributes (e.g., pose and identity when trained on human faces) and stochastic variation in the generated images (e.g., freckles, hair), and it enables intuitive, scale-specific control of the synthesis. The new generator improves the state-of-the-art in terms of traditional distribution quality metrics, leads to demonstrably better interpolation properties, and also better disentangles the latent factors of variation. To quantify interpolation quality and disentanglement, we propose two new, automated methods that are applicable to any generator architecture. Finally, we introduce a new, highly varied and high-quality dataset of human faces.