Expressive Head Avatar Modeling From Monocular Video of Neutral Expression

We study the reconstruction of high-quality 3D head avatars. Our goal is to reduce the reliance on dense capture data for most existing approaches, which limits their practicality. Recent advances address this using single or few input images by either training a prior model or fine-tuning multi-view diffusion models to generate pseudo training points. While they fall short in producing multi-view-consistent, high-fidelity results aligned with the input data. This motivates us to explore a more practical and user-friendly input setting. Modern smartphones such as Apple's Face ID already guide users to slowly rotate their heads in front of a single camera, enabling the capture of facial data across varying viewpoints with minimal effort. This simple and intuitive scanning motion has become a widely accepted user habit and provides sufficient geometric information-highlighting a natural opportunity for 3D head avatar creation from monocular videos of neutral expression. Under this settinng, we introduce R$^{2}$2 Avatar, a lightweight and user-friendly framework for generating expressive 3D head avatars, which adopts a Reconstruction-by-Restoration strategy avoiding large-scale model pretraining while achieving high-quality animatable avatars. Specifically, a geometry-guided warping module first synthesizes coarse expression variaHons from the neutral input. Then, a restoration module refines the warped results by recovering highfrequency facial details, including mouth interior, with the help of a data-driven 2D animation prior. These restored images serve as supervision targets to optimize the final avatar. Experiments demonstrate that our method produces realistic avatars with improved expression diversity and view consistency compared to baseline approaches.