CaricatureGS: Exaggerating 3D Gaussian Splatting Faces With Gaussian Curvature
Eldad Matmon, Amit Bracha, Noam Rotstein, Ron Kimmel
2026-01-12
Summary
This paper presents a new method for creating realistic and customizable 3D caricatures of faces, meaning it takes a normal face and exaggerates features like the nose or eyes in a controllable way.
What's the problem?
Existing methods for exaggerating 3D faces often result in blurry or unrealistic images. Simply stretching a 3D model doesn't look good, and current techniques struggle to balance exaggeration with maintaining a photorealistic appearance. The initial attempt to directly deform a modern 3D representation called 'Gaussian Splatting' also didn't produce good results.
What's the solution?
The researchers developed a system that first exaggerates the underlying shape of the face based on its curves. Then, instead of directly manipulating the 3D Gaussian Splatting data, they cleverly create fake training images by warping the original face to match the exaggerated shape. They then train the system using both real face images and these fake, exaggerated images, allowing it to learn how to create both normal and caricatured faces from a single 3D model. Finally, they added a way to smoothly transition between the original and exaggerated forms for real-time control and editing.
Why it matters?
This work is important because it allows for the creation of high-quality, controllable 3D caricatures that look very realistic. This has potential applications in areas like creating personalized avatars for games or virtual reality, special effects in movies, or even just fun filters for social media. It improves upon previous methods by producing better results and offering more control over the caricature process.
Abstract
A photorealistic and controllable 3D caricaturization framework for faces is introduced. We start with an intrinsic Gaussian curvature-based surface exaggeration technique, which, when coupled with texture, tends to produce over-smoothed renders. To address this, we resort to 3D Gaussian Splatting (3DGS), which has recently been shown to produce realistic free-viewpoint avatars. Given a multiview sequence, we extract a FLAME mesh, solve a curvature-weighted Poisson equation, and obtain its exaggerated form. However, directly deforming the Gaussians yields poor results, necessitating the synthesis of pseudo-ground-truth caricature images by warping each frame to its exaggerated 2D representation using local affine transformations. We then devise a training scheme that alternates real and synthesized supervision, enabling a single Gaussian collection to represent both natural and exaggerated avatars. This scheme improves fidelity, supports local edits, and allows continuous control over the intensity of the caricature. In order to achieve real-time deformations, an efficient interpolation between the original and exaggerated surfaces is introduced. We further analyze and show that it has a bounded deviation from closed-form solutions. In both quantitative and qualitative evaluations, our results outperform prior work, delivering photorealistic, geometry-controlled caricature avatars.