When humans look at a 3D model of a coffee mug, we intuitively understand its global structure: it is a solid object with a continuous surface and exactly one handle creating a single hole. However, standard deep learning models view 3D shapes quite differently. Whether predicting raw vertex coordinates, voxel grids, or point clouds, neural networks excel at matching visual appearances but frequently fail to preserve the underlying mathematical rules of geometry—its topology.
If a network predicts a mesh by outputting a "triangle soup" or independent surface patches, as noted by Angela Dai et al. in Scan2Mesh: From Unstructured Range Scans to 3D Meshes, there is no explicit mechanism to ensure that edges align perfectly. This leads to severe topological defects: non-manifold geometry (where an impossible number of faces share a single edge), self-intersections, floating disconnected components, or microscopic gaps that destroy watertightness.
These are not just cosmetic flaws. In downstream engineering applications like Finite Element Analysis (FEA) or fluid dynamics, a non-watertight mesh causes physics simulations to fail completely. In medical imaging, generative models that accidentally introduce unintended connections between anatomical structures render digital twins useless for simulation, a vulnerability explored by Farhad R. Nezami et al. in Probing the limits and capabilities of diffusion models for the anatomic editing of digital twins. For manufacturing, fragmented meshes cannot be 3D printed without extensive manual cleanup, as highlighted by Kristen M. Edwards and Brandon Man in Sketch2Prototype: rapid conceptual design exploration and prototyping with generative AI.
To solve this, a major research direction focuses on embedding topological constraints directly into the generative process. Three primary strategies have emerged to guarantee structural integrity: diffeomorphic flows, space partitioning, and structural connectivity grammars.
Diffeomorphic Flows: Deforming Space Without Tearing
One elegant way to ensure topological consistency is to avoid changing the topology at all. Instead of generating a mesh from scratch, a model can start with a known, valid template shape (such as a sphere) and deform it into the target geometry.
To guarantee that this deformation never introduces self-intersections, tears, or holes, the mapping must be a diffeomorphism—a continuous, invertible function that maps one topological space to another.
[ Spherical Template ] ---> ( Smooth, Invertible Flow ) ---> [ Target Geometry ]
(Perfect Topology) (No tearing or gluing) (Preserved Topology)
In Hybrid Neural Diffeomorphic Flow for Shape Representation and Generation via Triplane, Kun Yeun Han et al. leverage this concept by encoding deformations into high-resolution triplane features. Because the underlying neural flow is mathematically constrained to be invertible, the network can stretch and bend the template to match complex target shapes (like highly detailed organs) while guaranteeing that the final surface remains a closed, single-component manifold.
Similarly, Alexandre Bône et al. introduce specialized exponentiation layers in Auto-Encoding Meshes of any Topology with the Current-Splatting and Exponentiation Layers. By computing velocity fields and integrating them over time, the network generates smooth, reversible warpings. This prevents the "topological noise" and self-intersections that typically plague models predicting raw vertex positions.
Space Partitioning: Watertightness by Construction
Template deformation works beautifully when the target shape shares the same topology as the template. But what if the target domain contains shapes with an arbitrary or unknown number of holes, such as chairs or buildings?
Instead of moving vertices on a surface, we can achieve guaranteed watertightness by carving up the underlying 3D space using implicit primitives.
In BSP-Net: Generating Compact Meshes via Binary Space Partitioning, Zhiqin Chen et al. design a network that outputs sets of planes (half-spaces). By taking the mathematical intersection of these half-spaces, the network forms convex primitives.
Planes (Half-spaces) ---> Intersection ---> Convex Primitive ---> Union ---> Final Watertight Mesh
Why does this guarantee a watertight mesh? A convex polytope formed by intersecting planes is, by definition, a closed solid volume with no gaps. By taking the union of multiple convex primitives, the network can assemble complex, sharp-edged geometries. Because the building blocks themselves are mathematically closed, the resulting output is guaranteed to be a solid manifold, entirely free of surface holes or overlapping patches.
Structural Grammars: Learning the Language of Connectivity
A third approach treats mesh generation not as a continuous geometric problem, but as a discrete sequence guided by structural rules—akin to a language grammar.
Traditional autoregressive models often fail because they predict vertex coordinates without understanding how those vertices should connect. To fix this, Yawar Siddiqui et al. introduce MeshGPT: Generating Triangle Meshes with Decoder-Only Transformers. They quantize mesh faces into a vocabulary of geometric "words." A decoder-only transformer then learns the sequence of faces, effectively capturing the underlying "grammar" of clean, artist-constructed topology. By predicting faces in a coherent sequence, the model avoids the chaotic, over-tessellated outputs of unconstrained coordinate predictors.
This structural approach is especially vital for Boundary Representations (B-reps) in Computer-Aided Design (CAD), where faces, edges, and vertices must align with zero numerical tolerance. In BrepGen: A B-rep Generative Diffusion Model with Structured Latent Geometry, Xiang Xu and Xiang Xu 0008 represent CAD models as hierarchical trees. By using node duplication to explicitly encode topology within a diffusion framework, they prevent the tiny numerical discrepancies that usually cause adjacent faces to separate, ensuring solid-body integrity.
Handling Complex Junctions
While maintaining a simple manifold is ideal for isolated objects, the real world often requires representing complex, intersecting structures—such as geological layers or internal architectural walls.
To address this, Mohammad Moulaeifard et al. step outside traditional manifold constraints in Subdivide and Conquer: Adapting Non-Manifold Subdivision Surfaces to Surface-Based Representation and Reconstruction of Complex Geological Structures. They implement non-manifold subdivision algorithms that explicitly model "singular" edges where three or more faces meet. By mathematically defining how these junctions refine during subdivision, they ensure that complex internal intersections remain perfectly sealed and connected, preserving the physical validity required for downstream fluid flow and heat transfer simulations.