The generative framework FaceLit is capable of generating 3D faces that can be rendered under various user-defined lighting conditions and views, fully learned from 2D images without any manual annotations.

The model learns to generate the shape and material properties of the face to produce realistic facial images with multi-view 3D and lighting consistency when rendered from the natural statistics of pose and lighting.

The framework provides methods to generate realistic faces with explicit lighting and view control on multiple datasets (FFHQ, MetFaces, and CelebA-HQ). Demonstrate state-of-the-art photorealism among 3D perception GANs on the FFHQ dataset, achieving a FID score of 3.5.