GGAvatar: Dynamic Facial Geometric Adjustment for Gaussian Head Avatar
GGAvatar overviewAbstract
Reconstructing animatable 3D head avatars from target subject videos is challenging because identities and facial deformations vary substantially. GGAvatar introduces a coarse-to-fine Gaussian representation with a Neutral Gaussian Initialization Module and a Geometry Morph Adjuster. Gaussian primitives are paired with a deformable triangular mesh for robust geometric initialization, while global deformation bases provide a fine-grained, low-dimensional model of facial motion.
Type
Publication
Pacific Graphics Conference Papers and Posters
GGAvatar reconstructs high-fidelity animatable Gaussian head avatars from monocular RGB video. Its geometry-aware coarse-to-fine design improves both novel-view rendering and expression animation.
