We propose a generative framework, FaceLit, capable of generating a 3D face that can be rendered in various lighting conditions and user-defined views, learned solely from 2D images in the wild without any manual annotation. Unlike existing jobs that require careful capture setup or human labor, we rely on out-of-the-box lighting and pose estimators. With these estimates, we incorporated the Phong reflectance model into the neural volume rendering framework. Our model learns to generate the shape and material properties of a face in such a way that, when rendered according to natural posing and lighting statistics, it produces photorealistic facial images with multi-view 3D and lighting consistency. Our method enables the photorealistic generation of faces with explicit lighting and view controls on multiple data sets: FFHQ, MetFaces, and CelebA-HQ. We show state-of-the-art photorealism among 3D-aware GANs in the FFHQ dataset achieving a FID score of 3.5.
