Title: QuadLink: Autoregressive Quad-Dominant Mesh Generation via Point-Relation Learning

Abstract:

The creation of production-ready, quad-dominant meshes is fundamental to contemporary 3D content pipelines. However, deriving anisotropic, quad-dominant structures from point clouds remains a significant hurdle, as current techniques are generally restricted to outputting either exclusively triangular meshes or isotropic quadrilateral meshes. To address this, we introduce QuadLink, a comprehensive three-stage framework that generates quad-dominant meshes by establishing structured connections between points.

QuadLink approaches polygonal mesh creation through a hybrid centroid-conditioned vertex linking model. The process begins by predicting a unified collection of anchors, which includes both vertices and face centroids. Subsequently, the model learns links conditioned on these centroids to associate vertices with their respective face centers. Finally, it constructs polygonal faces using a quad-first assembly strategy, underpinned by robust geometric verification protocols. This link-centric approach facilitates the efficient production of sparse, anisotropic quad-dominant meshes characterized by coherent edge flow, while simultaneously accommodating hybrid polygonal topologies.

To generate training data for this architecture, we propose a Tri-to-Quad Operator. This tool transforms artistic triangle meshes into quad-dominant datasets by employing global merge selection techniques. Our extensive experimental results demonstrate that QuadLink successfully generates production-ready quad-dominant meshes directly from point clouds, surpassing previous baselines in both geometric fidelity and topological quality. Furthermore, the method natively supports hybrid polygonal topology, allowing it to generalize to arbitrary n-gon meshes without requiring any modifications to its architecture.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC