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

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.

• The paper introduces a novel Transformer-based model that autoregressively generates B-Reps by unifying topology and geometry tokenization.
• It employs a breadth-first traversal of the face adjacency graph to ensure precise, watertight representations of complex geometries.
• The model outperforms prior approaches by improving generation speed and achieving over 50% validity on models with up to 100 faces.

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

Introduction

The paper introduces AutoBrep, a Transformer-based model designed for autoregressive generation of Boundary Representations (B-Reps) in Computer-Aided Design (CAD). B-Reps are integral to precise 3D modeling, yet generating them with exact geometry and watertight topologies remains a challenge. AutoBrep directly addresses this by employing a unified tokenization scheme that dynamically encodes both geometric primitives and topological relationships into discrete tokens, facilitating improved generation quality and speed while maintaining scalability to complex geometries.

Model Architecture and Features

AutoBrep utilizes a Transformer model that processes sequences of discrete tokens representing B-Rep faces, edges, and their adjacency (Figure 1). This framework synthesizes B-Reps through a sequence of tokens, which are generated based on a breadth-first traversal (BFT) of the face adjacency graph. This method contrasts with previous iterative or multi-stage approaches by combining geometric and topological data into a single autoregressive sequence.

Figure 1: Sequence ordering follows a breadth-first traversal (BFT) of the B-Rep face adjacency graph.

The model distinguishes itself with a local topological reference system. By assigning dynamic tokens within a local context corresponding to the traversal level, AutoBrep efficiently manages vertex-face-edge relationships without relying on global coordinate ordering. This local context is crucial for maintaining watertightness and enhancing model performance, as shown in experiments where global referencing decreases quality metrics significantly.

Results and Evaluation

Experiments conducted using the extensive ABC dataset demonstrate that AutoBrep exceeds prior models like BrepGen and Holistic Latent (HoLa) representations, particularly in terms of generation speed and validity of complex solids. Quantitative metrics like Coverage (COV), Minimum Matching Distance (MMD), and Jensen-Shannon Divergence (JSD) validate its superior alignment with real-world data.

Figure 2: Validity of generated B-Reps as a function of the face count.

Results indicate that AutoBrep maintains over 50% validity for models with up to 100 faces, compared to much lower rates for other models as complexity increases. The use of a holistic Transformer model, alongside a novel tokenization approach, promotes both efficiency and robustness against error accumulation, typical in multi-stage generation pipelines.

Autocompletion Capabilities

Beyond unconditional generation, AutoBrep supports B-Rep autocompletion by seamlessly integrating specific user-provided constraints. This feature addresses scenarios where partial assemblies require completion within specific geometric bounds, which is pivotal for practical CAD applications in manufacturing and design.

Figure 3: Autocompletions from faces representing the assembly interfaces (orange).

The model's unified token framework allows for exact preservation of user-provided geometry and ensures watertight completions, significantly enhancing user control and practical utility in assembly-driven workflows.

Conclusion

AutoBrep represents a significant advancement in CAD modeling, providing a scalable, efficient solution for generating complex B-Rep geometry. Through its unified topology and geometry tokenization, the model not only surpasses existing benchmarks in quality and validity but also introduces practical tools for controlled autocompletion. Future work will likely explore enhanced robustness against intricate geometries and extend the model's adaptability to broader CAD applications, including parametric and procedural design tasks.