A Style-Based Generator Architecture for Generative Adversarial Networks

CVPR 2019  ·  Tero Karras, Samuli Laine, Timo Aila ·

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.

PDF Abstract CVPR 2019 PDF CVPR 2019 Abstract

Datasets


Introduced in the Paper:

FFHQ

Used in the Paper:

CelebA-HQ LSUN

Results from the Paper


Task Dataset Model Metric Name Metric Value Global Rank Result Benchmark
Image Generation CelebA-HQ 1024x1024 StyleGAN FID 5.06 # 2
Image Generation FFHQ 1024 x 1024 StyleGAN FID 4.4 # 12
Image Generation LSUN Bedroom StyleGAN FID-50k 2.65 # 1
Image Generation LSUN Cat 256 x 256 StyleGAN Clean-FID (trainfull) 8.72 ± 0.03 # 3
Image Generation LSUN Churches 256 x 256 StyleGAN FID 4.21 # 14
Clean-FID (trainfull) 4.75 ± 0.01 # 3
Image Generation LSUN Horse 256 x 256 StyleGAN Clean-FID (trainfull) 4.78 ± 0.03 # 3

Methods