QuadGPT: Native Quadrilateral Mesh Generation with Autoregressive Models

TL;DR

QuadGPT is introduced, the first autoregressive framework for generating quadrilateral meshes in an end-to-end manner and significantly surpasses previous triangle-to-quad conversion pipelines in both geometric accuracy and topological quality.

Abstract

The generation of quadrilateral-dominant meshes is a cornerstone of professional 3D content creation. However, existing generative models generate quad meshes by first generating triangle meshes and then merging triangles into quadrilaterals with some specific rules, which typically produces quad meshes with poor topology. In this paper, we introduce QuadGPT, the first autoregressive framework for generating quadrilateral meshes in an end-to-end manner. QuadGPT formulates this as a sequence prediction paradigm, distinguished by two key innovations: a unified tokenization method to handle mixed topologies of triangles and quadrilaterals, and a specialized Reinforcement Learning fine-tuning method tDPO for better generation quality. Extensive experiments demonstrate that QuadGPT significantly surpasses previous triangle-to-quad conversion pipelines in both geometric accuracy and topological quality. Our work establishes a new benchmark for native quad-mesh generation and showcases the power of combining large-scale autoregressive models with topology-aware RL refinement for creating structured 3D assets.

Figures (17)

• Figure 1: QuadGPT can generate diverse, high-quality quad meshes conditioned on point clouds.
• Figure 2: Comparison of topological quality across different pipelines. Iso-surfacing methods produce dense, unstructured triangular meshes. Autoregressive triangle generation followed by heuristic conversion fails to create coherent structure. Our QuadGPT directly generates native quadrilateral meshes with clean, artist-friendly edge flow.
• Figure 3: Pipeline of QuadGPT. First, an autoregressive Hourglass Transformer is pre-trained to generate mesh sequences conditioned on an input point cloud. Subsequently, the model is fine-tuned using Truncated Direct Preference Optimization (tDPO), where a preference dataset is automatically constructed by comparing truncated sequences via a novel topological reward.
• Figure 4: Qualitative Comparison against Indirect Autoregressive Pipelines. The top four rows show results on out-of-distribution dense meshes generated by Hunyuan3D lai2025hunyuan3d, while the bottom four rows showcase performance on artist-designed meshes. Baseline methods followed by tri-to-quad conversion often produce topological artifacts and lose geometric detail. QuadGPT consistently generates meshes with superior topological coherence and fidelity across both domains.
• Figure 5: Qualitative Comparison against a Field-Guided Method. Field-guided methods like QuadriFlow can be unstable on meshes with complex topology or sharp features.