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

SIGGRAPH Asia 2025  ·  ACM Transactions on Graphics, Vol. 44, No. 6, Article 226

Pu Li1,2    Wenhao Zhang1    Weize Quan1    Biao Zhang3    Peter Wonka3    Dong-Ming Yan1,2

1MAIS, Institute of Automation, Chinese Academy of Sciences    2University of Chinese Academy of Sciences    3KAUST

Paper arXiv Code
BrepGPT teaser

Top: B-rep models generated by BrepGPT. Bottom: Visualization of corresponding Voronoi Half-Patches, where distinct colors represent regions in the parametric space of each face geometrically closest to their boundary curves.

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 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, VHP unifies geometric attributes and topological relations in a single, coherent format. We further leverage dual VQ-VAEs to encode both vertex topology and VHP geometry into compact vertex-based tokens. A decoder-only Transformer then autoregressively predicts these token sequences, which are decoded into complete B-rep models.

BrepGPT achieves state-of-the-art performance in unconditional B-rep generation and supports a range of applications including conditional generation from category labels, point clouds, text, and images, as well as B-rep autocompletion and shape interpolation.

Fast-forward video

Method

BrepGPT pipeline
Pipeline overview. 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. The GPT-style Transformer autoregressively generates vertex token sequences that are decoded into complete B-rep models.
VHP sampling
Voronoi Half-Patch sampling. Left: input B-rep. Middle: Voronoi partitioning in parametric space. Right: geometric sampling with vertices (purple), curve samples (green), and surface samples (blue).

Results

DeepCAD comparison
Qualitative comparison on the DeepCAD dataset, ordered by increasing vertex complexity.
ABC comparison
Qualitative comparison on the ABC dataset.

Citation

@article{li2025brepgpt,
  title     = {BrepGPT: Autoregressive B-rep Generation with Voronoi Half-Patch},
  author    = {Li, Pu and Zhang, Wenhao and Quan, Weize and Zhang, Biao and Wonka, Peter and Yan, Dong-Ming},
  journal   = {ACM Transactions on Graphics},
  volume    = {44},
  number    = {6},
  pages     = {226:1--226:18},
  year      = {2025},
  publisher = {ACM},
  doi       = {10.1145/3763323}
}

© 2025 Pu Li et al.  ·  Page template inspired by Nerfies.