StyleGAN: a style-controlled face generator
On 12 December 2018 NVIDIA described StyleGAN, the generator of an adversarial network in which the latent code controls style at every level through AdaIN while separate noise adds random detail. On the new FFHQ face set its FID was 4.40 against 8.04 for Progressive GAN.
Why it matters
Photorealistic faces at 1024x1024 came to be generated with high-level attributes (pose, identity) separated from stochastic detail (freckles, hair), each controllable apart. The FFHQ set came with the paper.
FFHQ is 70,000 face images at 1024x1024 from Flickr, more varied in age, ethnicity and background than CelebA-HQ. Training took a week on an NVIDIA DGX-1 with eight Tesla V100s. The first version only plans to release the set; the record does not claim it was released that day.