Photorealistic facial exaggeration by coupling curvature-weighted mesh deformation with anisotropic 3D Gaussian Splatting. Real-time view synthesis, controllable exaggeration, and faithful identity preservation.
We present CaricatureGS, a curvature-aware 3D Gaussian Splatting framework for photorealistic facial caricature. Our approach couples curvature-weighted Poisson deformation on a fitted FLAME mesh with an anisotropic 3D Gaussian representation. We introduce a Local Affine Transformation (LAT) module to synthesize pseudo ground-truth caricature supervision and propose a joint optimization scheme that alternates training over original and exaggerated frames. The result is a single 3D Gaussian representation that renders identity-preserving faces with continuous exaggeration control and real-time novel-view synthesis.
Given an input videos we fit per-frame FLAME mesh and estimate Gaussian curvature K on it. We exaggerate the shape geometry by solving a Poisson equation weighted by K.
Local Affine Transfer (LAT) warps original frames to the exaggerated projections, yielding supervision for the Gaussians optimization.
Jointly optimize one set of anisotropic 3D Gaussians on original + caricatured frames for consistent geometry & radiance. Alternating between GT and GT* images allows for smooth interpolation between representations.
@misc{matmon2026caricaturegs,
title = {CaricatureGS: Exaggerating 3D Gaussian Splatting Faces With Gaussian Curvature},
author = {Eldad Matmon and Amit Bracha and Noam Rotstein and Ron Kimmel},
year = {2026},
eprint = {2601.03319},
archivePrefix = {arXiv},
primaryClass = {cs.GR},
url = {https://arxiv.org/abs/2601.03319}
note = {Accepted to 3DV 2026}
}