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.
|Task||Dataset||Model||Metric name||Metric value||Global rank||Compare|
|Image Generation||CelebA-HQ 1024x1024||StyleGAN||FID||5.06||# 1|
|Image Generation||FFHQ||StyleGAN||FID||4.40||# 1|