According to Nvidia, its GAN is built around a concept called “style transfer.” Rather than trying to copy and paste elements of different faces into a frankenperson, the system analyzes three basic styles — coarse, middle, and fine styles — and merges them transparently into something completely new.
Coarse styles include parameters such as pose, the face’s shape, or the hair style. Middle styles include facial features, like the shape of the nose, cheeks, or mouth. Finally, fine styles affect the color of the face’s features like skin and hair.
According to the scientists, the generator is “capable of separating inconsequential variation from high-level attributes” too, in order to eliminate noise that is irrelevant for the new synthetic face.
Join the conversation as a VIP Member