DiNGGA: Dirichlet and Noise-Guided Gaussian Avatar for Fine-grained Facial Detail

초록

Recent Gaussian Splatting–based head avatar reconstruction methods achieve photorealistic rendering by binding Gaussian primitives to 3D Morphable Models (3DMMs), enabling geometrically consistent deformation and animation control. However, the optimization process lacks explicit surface opacity constraints, causing inherently opaque facial regions to converge to semi-transparent states. This phenomenon leads to Gaussians from the face and the back of the head being rendered and optimized together, which hinders the model from capturing intricate geometric details such as expression-dependent wrinkles and non-rigid deformations. To address these issues, we propose DiNGGA (Dirichlet and Noise-Guided Gaussian Avatars). First, we fill the interior of the 3DMM head volume with noise 3D Gaussians, effectively mitigating the semi-transparency artifacts. Second, we introduce a 3D Dirichlet PDF-based height field defined over each mesh triangle of the face region, enabling explicit modeling of expression-dependent wrinkles and non-rigid surface deformations. Extensive experiments show that DiNGGA effectively mitigates surface transparency artifacts and significantly enhances fine-grained facial detail reconstruction compared to prior methods.

키워드