S-Avatar: Diffusion-Guided Gaussian Head Avatars from a Single Image

Researchers introduced S-Avatar, a single-image method for creating photorealistic 3D head avatars.

It combines diffusion guidance with 3D Gaussian splatting to produce high-resolution avatars, then aligns them with the FLAME head model.

A binding template enables real-time expression animation and controllable avatar motion.

The method outperforms prior approaches on public datasets for novel-view synthesis and expression quality.

The work improves realism, view consistency, and real-time animatability for VR and AR avatars.