VoroLight: Learning Voronoi Surface Meshes via Sphere Intersection

VoroLight is a differentiable framework that achieves smooth and regular 3D shape reconstruction by associating Voronoi surface vertices with trainable spheres and optimizing a sphere-intersection loss to promote controlled degeneracy while preserving the watertight and convex properties of Voronoi diagrams.

The Gist

Imagine you are trying to build a 3D model of a complex object, like a heart or a bunny, using only a bag of marbles. In the world of computer graphics, these marbles are called generators. If you arrange them so that every point in space is closest to one specific marble, you create a Voronoi diagram.

Think of a Voronoi diagram like a honeycomb. Each cell in the honeycomb belongs to one "queen" (marble). The walls between the cells are perfectly flat planes. This is great for making strong, watertight structures (like a 3D-printable lamp), but there's a catch: standard honeycombs are made of flat, angular faces. If you try to make a smooth curve, like the curve of a heart, a standard honeycomb looks jagged and bumpy, like a low-resolution video game character.

Enter VoroLight.

The researchers behind this paper asked: "How can we make these honeycomb walls curve smoothly without breaking the rules of the honeycomb?"

The Magic Trick: The "Sphere Intersection"

Usually, to make a honeycomb wall, you just need two marbles. But to make a smooth curve, you need a bit of magic. The team realized that if you force three or more marbles to meet at a single point in a very specific way, the sharp corners disappear, and the surface becomes smooth.

They call this "Voronoi Degeneracy." It sounds scary, but think of it like this:

• Normal Honeycomb: Three bees meet at a corner. It's a sharp point.
• VoroLight Honeycomb: We force the bees to meet in a way that creates a perfect, smooth curve, like the intersection of three soap bubbles.

To achieve this, VoroLight doesn't just move the marbles (generators). It attaches a virtual, trainable sphere to every single corner (vertex) of the honeycomb.

Imagine every corner of your 3D model has a tiny, invisible balloon attached to it. The computer's job is to blow up or shrink these balloons until they all touch each other at exactly the same two points. When these "balloons" (spheres) intersect perfectly, the honeycomb walls between them naturally smooth out into a beautiful, curved surface.

Why is this a big deal?

1. It's "Watertight" and Strong: Unlike other methods that might leave tiny holes or self-intersecting messes (like a tangled ball of yarn), VoroLight guarantees a perfect, solid shell. It's like building a house where every brick fits perfectly, and there are no gaps for rain to get in.
2. It Works with Anything: You can feed VoroLight a messy cloud of points, a blurry photo, a 3D scan, or even a mathematical formula. It figures out how to arrange its "marbles" and "balloons" to match the shape, no matter how the data is given.
3. It's Ready for 3D Printing: Because the result is a solid, smooth, and mathematically perfect structure, you can send it straight to a 3D printer. The paper even shows off some cool Voronoi lamps that look like glowing, organic honeycombs.

The Process in Simple Steps

1. The Rough Draft: The computer starts with a rough, bumpy honeycomb shape that vaguely looks like the target object (like a heart).
2. The Balloon Training: The computer attaches those invisible "balloons" (spheres) to the corners. It then plays a game of "tug-of-war," adjusting the size and position of the balloons and the marbles until the balloons all intersect perfectly.
3. The Smooth Result: As the balloons align, the jagged honeycomb walls melt into a smooth, curved surface that still looks like a honeycomb but feels like a smooth sculpture.
4. Filling the Gaps: Once the outside shell is perfect, the computer fills the inside with more marbles to create a solid 3D structure, ensuring the inside matches the outside perfectly.

The Analogy: Sculpting with Soap Bubbles

Imagine you are a sculptor.

• Old methods were like trying to carve a smooth statue out of a block of wood using a chisel. You get close, but you always end up with little chisel marks and rough edges.
• VoroLight is like blowing soap bubbles. You arrange the bubbles so they naturally press against each other. Where they touch, they form smooth, curved walls. VoroLight is the "mathematical wind" that arranges these bubbles (spheres) so perfectly that they form a solid, 3D-printable object that looks like it was grown, not carved.

In a Nutshell

VoroLight is a new tool that teaches computers how to build 3D shapes that are both structurally perfect (like a honeycomb) and visually smooth (like a sculpture). It does this by using a clever trick of intersecting spheres to smooth out the sharp corners of a mathematical grid, turning jagged data into beautiful, printable 3D objects.

Technical Summary

1. Problem Statement

Reconstructing 3D shapes from diverse data sources (point clouds, images, implicit fields) is a fundamental challenge in computer vision. Existing methods face a trade-off between geometric guarantees and flexibility:

• Implicit Methods: (e.g., NeRF, DeepSDF) handle complex geometry well but require non-differentiable extraction steps (like Marching Cubes) to produce meshes, often resulting in non-manifold or non-watertight surfaces.
• Explicit Mesh Methods: (e.g., direct prediction) offer direct access to mesh elements but often rely on predefined grids or templates, limiting geometric flexibility and failing to guarantee watertightness or self-intersection freedom.
• Differentiable Voronoi Methods: (e.g., VoroMesh) produce watertight, convex, and topologically consistent meshes by optimizing generator positions. However, standard Voronoi configurations rely on stable vertices (defined by exactly 4 generators in 3D), which results in locally uneven, "bumpy" surfaces lacking smooth curvature. Furthermore, previous differentiable Voronoi approaches are often restricted to point-cloud supervision and lack mechanisms to enforce higher-order geometric regularity.

The Core Challenge: How to learn a smooth, watertight, and topologically consistent Voronoi surface mesh that can be trained via gradient descent on multimodal data, while overcoming the inherent "jaggedness" of standard Voronoi diagrams.

2. Methodology: VoroLight

VoroLight introduces a differentiable framework that explicitly learns higher-order Voronoi degeneracy to control surface curvature. The method consists of four main stages:

A. Mesh Initialization

• Boundary-Reflection Strategy: An initial rough Voronoi surface is generated by placing generators in a voxel grid. Exterior generators are placed in a thin band outside the shape, and interior generators are created by reflecting exterior ones across the estimated surface normal.
• Differentiable Formulation: Uses a closed-form, differentiable expression for Voronoi vertices (based on generator positions) to perform initial shape fitting and regularization.

B. Sphere Intersection Training (The Core Innovation)

Inspired by the non-differentiable VoroCrust algorithm, VoroLight lifts the concept of sphere-based construction into a trainable framework.

• Trainable Spheres: Each Voronoi surface vertex $v$ is associated with a trainable sphere $S_v = (c_v, r_v)$ (center and radius).
• Degeneracy via Intersection: To achieve smooth curvature, the method enforces higher-order degeneracy. For every polygonal face $f$, a pair of target intersection points $(p_f^{(1)}, p_f^{(2)})$ is defined along the face normal.
• Sphere-Intersection Loss ($L_{sphere}$): The optimization forces all spheres incident to a specific face to pass through these same two target points.
  • Intersection Loss ($L_{int}$): Minimizes the power distance of incident spheres to the target points.
  • Exclusion Loss ($L_{excl}$): Ensures non-incident spheres do not enclose the target points, preventing geometric inconsistencies.
• Effect: This constraint forces the generators to align in a way that creates degenerate vertices (shared by 5+ generators), which couples neighboring face orientations and produces smooth, well-regularized surfaces.

C. Multimodal Supervision

Because the loss functions are defined directly on surface vertices, VoroLight supports supervision from various input modalities:

• Implicit Fields: Minimizes squared level-set error.
• Point Clouds/Meshes: Minimizes bidirectional Chamfer distance.
• Multi-view Images: Uses differentiable rasterization (nvdiffrast) to minimize silhouette, normal map, and neural shading losses.

D. Volumetric Extension

The framework naturally extends to volumetric tessellations:

• Interior Generators: Additional generators are placed deep inside the shape.
• Voronoi-Verified Sampling: An iterative process ensures these interior points do not alter the optimized surface boundary.
• Centroidal Voronoi Tessellation (CVT): Interior generators are optimized under a CVT loss to create a uniform, high-quality volumetric mesh with globally consistent surface-interior topology.

3. Key Contributions

1. VoroLight Framework: The first differentiable Voronoi formulation that explicitly learns higher-order degeneracy via trainable vertex spheres and a sphere-intersection loss, enabling smooth surface reconstruction.
2. Multimodal Supervision: A unified framework capable of training on implicit fields, point clouds, meshes, and multi-view images without changing the core geometric representation.
3. Volumetric Consistency: A natural extension to volumetric meshes that guarantees watertightness and consistent topology between the surface and the interior, demonstrated via 3D-printable lamp designs.
4. Robustness & Regularity: Demonstrates superior surface regularity (smoother curvature) and robustness to noise compared to prior differentiable Voronoi methods (VoroMesh).

4. Experimental Results

The authors evaluated VoroLight on 22 objects using Thingi10K and Objaverse datasets across three input modalities:

• Implicit Fields: On analytic shapes (Taubin heart, cube with cavity), sphere-intersection training significantly reduced curvature metrics (Mean $|H|$, Median $|H|$, P95 $|H|$) and SDF loss, confirming the generation of smoother surfaces.
• Point Clouds & Meshes:
  • Compared to VoroMesh, VoroLight produces smoother surfaces with fewer high-curvature outliers and more aligned Voronoi faces.
  • While VoroMesh preserves slightly finer high-frequency details (lower Chamfer distance), VoroLight offers better geometric regularity and stability under noisy inputs.
  • Noise Robustness: Under increasing Gaussian noise (up to 2%), VoroLight maintains coherent cell structures, whereas VoroMesh fragments into irregular, disconnected cells.
• Multi-view Images:
  • Compared to TetSphere Splatting, VoroLight achieves lower Chamfer distance and higher F1 scores on both 300 random views and sparse 6-view setups.
  • Crucially, VoroLight produces globally consistent, watertight volumetric meshes, whereas TetSphere often results in overlapping elements or gaps between local tetrahedral clusters.
• Ablation Studies:
  • Triangle-face relaxation: Subdividing polygon faces into triangles (making the system locally consistent with exactly 3 spheres per face) improves reconstruction fidelity (lower CD2) while maintaining similar smoothness.
  • Initialization: The method is robust to coarse initial grid resolutions.

5. Significance and Impact

• Bridging Geometry and Learning: VoroLight successfully bridges the gap between the strict geometric guarantees of Voronoi diagrams (watertightness, convexity) and the flexibility of deep learning.
• Simulation-Ready Outputs: By producing watertight, manifold, and self-intersection-free meshes with consistent interior topology, the output is immediately suitable for downstream physics simulations (fluid dynamics, solid mechanics) and 3D printing, areas where implicit-to-mesh extraction often fails.
• Controlled Degeneracy: The paper provides a novel theoretical and practical approach to "learning" degeneracy in geometric structures, a concept usually avoided in standard computational geometry, to achieve smoothness in discrete representations.
• Practical Application: The demonstration of 3D-printable Voronoi lamps highlights the method's ability to generate physically valid, artistic, and structurally sound objects directly from data.

In summary, VoroLight represents a significant step forward in differentiable geometry processing, offering a robust, smooth, and topologically guaranteed alternative to both implicit reconstruction and standard explicit mesh prediction.