Gaussian Eigen Models for Human Heads

Figure 1 . We propose a method that represents 3D Gaussian head avatars in a network-free form as ensembles of eigenbases (GEM). Only a linear combination of these bases is needed to generate new primitives, which can be splatted using 3D Gaussian Splatting. We demonstrate that the necessary coefficients for a specific expression can be regressed from single images, enabling real-time facial animation and crossreenactment. The simplicity of GEM results in highly efficient storage and rendering times.