Abstract

Manually creating realistic, digital human heads is a difficult and time-consuming task forartists. While 3D scanners and photogrammetry allow for quick and automatic reconstruction of heads, finding an actor who fits specific character appearance descriptions can bedifficult. Moreover, modern open-world video games feature several thousands of characters that cannot realistically all be cast and scanned. Therefore, researchers are investigating generative models to create 3D head shapes which fit a specific character appearance description. While current methods can generate believable head shapes quite well, generating a corresponding high-resolution, high-quality, production level texture which respects the character’s appearance description has been a major challenge.This work presents methods which combine to generate synthetic face textures at the video game production quality level. Production quality face textures have the following constraints: (i) game artists control the generative process by providing precise appearance attributes, the face shape, and the character’s age and gender, (ii) the texture must be at least 4096×4096resolution, and (iii) the texture must be in bidirectional reflectance distribution function parameters (diffuse albedo, roughness, specular albedo) for raytraced rendering under different lighting and viewing conditions.The proposed methods build upon earlier deep learning approaches addressing similar problems. Several key additions to these methods have been created to be able to use them in this context, specifically for providing easy artist controls and to work with limited training data. The implementation results show that in spite of training with a limited amount of training data, just over 100 samples, the machine learning model produces realistic textures which comply to a diverse range of skin, hair, lip and iris colors specified through a novel,yet intuitive, description format and augmentation thereof.