Exaggerating Gaussian Splatting Faces With Gaussian Curvature

We explore the exaggeration and caricaturization of human faces. Creating compelling 3D caricatures requires significant deformation of facial geometry while preserving identity, visual realism, and consistency across viewpoints. Achieving these goals calls for an effective combination of geometry-based and appearance-based approaches, enabling smooth and controllable shape deformation while maintaining high-quality, 3D-consistent renderings. We investigate the intersection of geometric facial deformations with the recent 3D Gaussian Splatting (3DGS) representation and show that a naive combination of the two does not produce satisfactory results. To address this challenge, we introduce a dedicated training scheme that adapts the appearance representation to the deformed geometry, enabling a broader range of facial exaggerations while preserving rendering quality. The proposed formulation further provides explicit geometric control and naturally supports localized facial editing. Further analysis using a face-recognition model indicates increased inter-subject separation in identity embedding space, suggesting that caricaturization enhances identity-distinctive facial features.

M.Sc. student under the supervision of Prof.Ron Kimmel.