We present VoroLight, a differentiable framework for 3D shape reconstruction based on Voronoi meshing. Our approach generates smooth, watertight surfaces and topologically consistent volumetric meshes directly from diverse inputs, including images, implicit shape level-set fields, point clouds and meshes. VoroLight operates in three stages: it first initializes a surface using a differentiable Voronoi formulation, then refines surface quality through a polygon-face sphere training stage, and finally reuses the differentiable Voronoi formulation for volumetric optimization with additional interior generator points.
Stage I: We initialize a watertight Voronoi surface mesh from the input using a differentiable Voronoi formulation. The surface is defined as Voronoi faces between oppositely labeled generators.
Stage II: We refine the surface quality by associating each face vertex with a trainable sphere and each face with a trainable offset distance. Through optimization, this sphere-based representation enables the surface to self-regularize into smooth, uniform Voronoi geometry. We investigate two refinement schemes: polygon-face constraints form a globally over-constrained system with small residuals, while triangle-face constraints are locally consistent and achieve near-perfect boundary alignment.
Stage III: We extend the refined surface to a volumetric mesh by adding interior generator points. The differentiable Voronoi formulation is reused to optimize the complete volumetric structure, producing high-quality, topologically consistent 3D meshes with convex polyhedral cells.
All results shown below use the triangle-face sphere refinement scheme.
Volumetric mesh generation from implicit shape representations.
Mesh inputs are converted to point cloud reconstruction problems by densely sampling points on the target surface. Target mesh input reconstruction, compared to VoroMesh.
300-camera view inputs reconstruction, compared to TetSphere Splatting.
From an input front-view RGB image, we use Huanyuan3D to generate five additional RGB views (left, right, top, bottom, and back), as well as the corresponding normal fields of these total six views. VoroLight then reconstructs the 3D shape by optimizing the Voronoi surface to match six-view silhouette masks and normal fields via differentiable rasterization. After obtaining this refined geometry, the geometry is kept fixed, and a neural deferred shader (NDS) is trained on the RGB views to reproduce the input appearance on the Voronoi mesh.
@article{lu2025vorolight,
title={VoroLight: Learning Quality Volumetric Voronoi Meshes from General Inputs},
author={Lu, Jiayin and Jiang, Ying and Yang, Yin and Jiang, Chenfanfu},
journal={arXiv preprint arXiv:2512.12984},
year={2025}
}Jiayin Lu is partially supported by the UCLA Institute for Digital Research and Education (IDRE) Postdoctoral Fellowship.