Multi-view Consistent 3D Gaussian Head Avatars 'without' Multi-view Generation

MVCHead generates consistent 3D Gaussian head avatars from a single 2D image, without multi-view capture or 3D supervision.

It combines hierarchical state space blocks, a hierarchical bi-directional scan, and an SE(3) multi-view critic to improve geometry and texture consistency.

The method learns both conditional and unconditional head models directly from 2D images, making single-shot avatar generation more practical.

The paper also introduces FaceGS-10K, a large-scale dataset of ready-to-use 3D Gaussian head assets for research and training.