SuGaR: Surface-Aligned Gaussian Splatting for Efficient 3D Mesh Reconstruction and High-Quality Mesh Rendering

Guédon and Lepetit tackle a practical gap in 3D GaussianSplatting: while 3DGS renders well, its optimised primitive cloud is unstructured and extracting a mesh from it is not straightforward. Marching Cubes, the default for Neural SDFs, does not apply directly because Gaussians do not implicitly define a level set. SuGaR closes this gap with a regularisation term that encourages Gaussians to align with the underlying scene surface and a downstream mesh extraction procedure tailored to that alignment.

The regulariser penalises Gaussians whose elongation deviates from the local tangent plane suggested by their neighbours, biasing the optimisation toward thin, surface-conforming primitives. Once trained, the authors run Poisson reconstruction on the oriented Gaussian centres to produce a clean, watertight triangle mesh in minutes. Poisson reconstruction scales gracefully and preserves detail better than Marching Cubes applied to noisy implicit fields, which is the standard fallback for NeRF-style methods.

A final, optional refinement stage binds Gaussians to the surface of the extracted mesh, parameterising each primitive by a barycentric coordinate on a face. The mesh and the bound Gaussians are then jointly optimised through the splatting renderer, so the mesh provides edit handles while the Gaussians provide rendering quality. This enables sculpting, rigging, animation, and relighting through conventional graphics tools acting on the mesh, with the Gaussians automatically following along.

SuGaR sits alongside 2D Gaussian Splatting as one of the two main lines of work on extracting accurate geometry from a Gaussian field, and the released code is a common reference for editable Gaussian assets.