A Style-Based Generator Architecture for Generative Adversarial Networks

by 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.

Associated video at https://www.youtube.com/watch?v=kSLJr...

12 pages, ebook

Published December 12, 2018

Reviews

[Keith · Rating 3 out of 5]

For the non-specialist, this would seem easier to follow in combination with the video demonstration of the results, which makes it clear that "the camera doesn't lie" is no longer true. "It's time to start RUNNING!" as the game-show host said to the Butcher of Bakersfield.

The gist, as far as I understand it, is that the team has developed modifications to a standard artificial intelligence system ("generative adversarial networks," basically multiple AI engines working against each other to produce more plausible results) that, once trained on a sufficiently large database of photographs, allow for the production of novel images while providing relatively granular control of "styles," which are essentially clusters of parameters such as pose, camera angle, and face-shape at the coarse level; facial features, hair style, and furniture at the middle level; and texture, color, and microfeatures at the fine level. This, again, is easier to illustrate than to describe, so check the video or the illustrations in the paper.