The article presents GAvatar, a method for generating high-fidelity 3D avatars from text prompts. The method utilizes a new 3D representation, 3D Gaussian Splatting, to create realistic, animatable avatars. The paper discusses the limitations of mesh or NeRF-based representations and how GAvatar addresses these issues. With the help of Gaussian splatting, GAvatar overcomes learning instability and captures fine avatar geometries. The method enables large-scale generation of diverse avatars, surpassing existing methods in appearance and geometry quality. It also achieves fast rendering at 1K resolution.
Publication date: 18 Dec 2023
Project Page: https://nvlabs.github.io/GAvatar
Paper: https://arxiv.org/pdf/2312.11461
