Next3D: Generative Neural Texture Rasterization for 3D-Aware Head Avatars

Figure 1 . Our 3D GAN synthesizes generative, high-quality, and 3D-consistent facial avatars from unstructured 2D images. Unlike current animatable 3D GANs that only modify yaw-pitch head poses and facial expressions, our approach enables fine-grained control over fullhead rotations, facial expressions, eye blinks, and gaze directions with strict 3D consistency and a high level of photorealism. Our approach also provides strong 3D priors for downstream tasks such as 3D-aware stylization.