AutoBrep: Autoregressive B-Rep Generation with Unified Topology and Geometry

TL;DR

AutoBrep tackles the challenge of end-to-end B-Rep generation by unifying geometry and topology into a single discrete token stream processed by a GPT-style Transformer. It introduces a novel tokenization that encodes both latent geometry (C,G) and local topology (T) within a breadth-first traversal, along with level-based dropout to train on long CAD sequences. The approach achieves superior quality, watertightness, and inference speed compared to multi-stage baselines, and naturally supports B-Rep autocompletion with exact preservation of user-provided faces. This work enables scalable, controllable CAD generation and sets the stage for more robust interaction with design workflows. It also provides two new datasets, ABC-1M and ABC-Constraint, to support unconditional generation and constrained autocompletion tasks.

Abstract

The boundary representation (B-Rep) is the standard data structure used in Computer-Aided Design (CAD) for defining solid models. Despite recent progress, directly generating B-Reps end-to-end with precise geometry and watertight topology remains a challenge. This paper presents AutoBrep, a novel Transformer model that autoregressively generates B-Reps with high quality and validity. AutoBrep employs a unified tokenization scheme that encodes both geometric and topological characteristics of a B-Rep model as a sequence of discrete tokens. Geometric primitives (i.e., surfaces and curves) are encoded as latent geometry tokens, and their structural relationships are defined as special topological reference tokens. Sequence order in AutoBrep naturally follows a breadth first traversal of the B-Rep face adjacency graph. At inference time, neighboring faces and edges along with their topological structure are progressively generated. Extensive experiments demonstrate the advantages of our unified representation when coupled with next-token prediction for B-Rep generation. AutoBrep outperforms baselines with better quality and watertightness. It is also highly scalable to complex solids with good fidelity and inference speed. We further show that autocompleting B-Reps is natively supported through our unified tokenization, enabling user-controllable CAD generation with minimal changes. Code is available at https://github.com/AutodeskAILab/AutoBrep.

Figures (8)

• Figure 1: Geometric representation learning. Each face is represented by a 32$\times$32 point grid sampled uniformly in the parameter domain of the underlying surface, while ensuring the $(u_\text{min},v_\text{min})$ point is the lowest lexicographically. Deep compression dcae is used to compress the point grid into low dimensional latent embeddings that are discretized using finite scalar quantization fsq into integer tokens.
• Figure 2: Top row: Sequence ordering follows a breadth-first traversal (BFT) of the B-Rep face adjacency graph (top-left). A new face (orange) and the connecting edge(s) are added incrementally during traversal. Previously visited faces are colored in gray. Bottom row: Unified discrete tokens include face geometry $F$, edge geometry $E$, and topology reference token $T$ for the face-edge incidence. $T$ is dynamically assigned to every face within a local context window for unique face references. This local window corresponds to the BFT level and is updated when traversal proceeds to the next window (from green to blue). * indicates the two vertical edges connecting the same set of semi-cylindrical faces.
• Figure 3: Randomly sampled unconditional results from (a) BrepGen xu2024brepgen, (b) HoLa HolaBRep25, and (c) our method AutoBrep. Last two rows highlight the generated hard complexity solids (blue color). Our method generates high-quality B-Reps with better complexity, diversity and watertightness.
• Figure 4: Validity of generated B-Reps as a function of the face count.
• Figure 5: Autocompletions from faces representing the assembly interfaces (orange), showing that a wide variety of shapes and styles can be generated.