Abstract
3D head animation has seen major quality and runtime improvements over thelast few years, particularly empowered by the advances in differentiablerendering and neural radiance fields. Real-time rendering is a highly desirablegoal for real-world applications. We propose HeadGaS, the first model to use 3DGaussian Splats (3DGS) for 3D head reconstruction and animation. In this paperwe introduce a hybrid model that extends the explicit representation from 3DGSwith a base of learnable latent features, which can be linearly blended withlow-dimensional parameters from parametric head models to obtainexpression-dependent final color and opacity values. We demonstrate thatHeadGaS delivers state-of-the-art results in real-time inference frame rates,which surpasses baselines by up to ~2dB, while accelerating rendering speed byover x10.