BrepGPT: Autoregressive B-rep Generation with Voronoi Half-Patch

TL;DR

BrepGPT introduces a decoder-only autoregressive model for full B-rep CAD generation by unifying geometry and topology through the Voronoi Half-Patch representation. It leverages dual VQ-VAEs to produce compact vertex-based tokens and trains a GPT-style Transformer to sequentially generate B-rep information, followed by detokenization into complete models. The approach achieves state-of-the-art performance in unconditional B-rep generation and demonstrates strong capabilities in conditional generation (class, point cloud, image, text), autocompletion, and interpolation. This framework offers a scalable, single-stage alternative to hierarchical and diffusion-based CAD generation methods, with potential to improve geometric fidelity and topological coherence in industrial design workflows.

Abstract

Boundary representation (B-rep) is the de facto standard for CAD model representation in modern industrial design. The intricate coupling between geometric and topological elements in B-rep structures has forced existing generative methods to rely on cascaded multi-stage networks, resulting in error accumulation and computational inefficiency. We present BrepGPT, a single-stage autoregressive framework for B-rep generation. Our key innovation lies in the Voronoi Half-Patch (VHP) representation, which decomposes B-reps into unified local units by assigning geometry to nearest half-edges and sampling their next pointers. Unlike hierarchical representations that require multiple distinct encodings for different structural levels, our VHP representation facilitates unifying geometric attributes and topological relations in a single, coherent format. We further leverage dual VQ-VAEs to encode both vertex topology and Voronoi Half-Patches into vertex-based tokens, achieving a more compact sequential encoding. A decoder-only Transformer is then trained to autoregressively predict these tokens, which are subsequently mapped to vertex-based features and decoded into complete B-rep models. Experiments demonstrate that BrepGPT achieves state-of-the-art performance in unconditional B-rep generation. The framework also exhibits versatility in various applications, including conditional generation from category labels, point clouds, text descriptions, and images, as well as B-rep autocompletion and interpolation.

Figures (16)

• Figure 1: Geometric sampling strategy for Voronoi Half-Patches (VHP). Left: Input B-rep model. Middle: Voronoi partitioning in parametric space. Right: Geometric sampling within Voronoi regions, showing vertices (purple), curve samples (green), and surface samples (blue) with sampling paths.
• Figure 2: Half-edge next pointer sampling illustration. Left: Vertex-edge graph structure. Middle: Decomposition into directed half-edges. Right: Next pointer establishment through strategic sampling, where terminal samples from adjacent half-edges are stored in the current VHP.
• Figure 3: Distribution of topological elements in 2,000 B-rep models from the DeepCAD dataset. For models with equivalent face counts, the number of edges consistently exceeds the vertex count, while half-edges show the highest frequency. As the face count surpasses 100, the half-edge count approaches approximately five times the vertex count.
• Figure 4: Pipeline overview of BrepGPT. The Connect VQ-VAE encodes topological relationships between B-rep vertices through pairwise connectivity classification. The VHP VQ-VAE encodes geometric information within Voronoi Half-Patches through MSE regression, taking feature vectors from the start and end vertices of each directed half-edge. The GPT-style Transformer performs autoregressive generation by concatenating vertex coordinates, connectivity, and VHP tokens in sequence.
• Figure 5: Left: All pairwise connections between B-rep vertices (green). Center: Pairwise connections between vertices within connected components (different colors represent distinct subgraphs). Right: Comparison of required vertex pair computations with and without connected component tokenization, demonstrating the reduction in computational complexity.