BrepForge: Factorized B-rep Synthesis via Wireframe Composition and Boundary-Conditioned Surface Instantiation
BrepForge factorizes B-rep synthesis into wireframe serialization and boundary-conditioned surface instantiation, ensuring topological integrity and geometric accuracy.
Key Findings
Methodology
BrepForge employs a two-stage generative approach: first, a face-aware autoregressive model (e.g., Transformer-based) serializes the wireframe into structured sequences, explicitly encoding Vertex-Edge-Face (V-E-F) connectivity to form a complete topological scaffold. Second, boundary-conditioned geometric priors—derived from boundary loops—are used to instantiate precise surface geometries without additional learning, transforming the complex synthesis into a structured refinement process. This separation of topological and geometric tasks enhances both structural integrity and geometric precision, addressing the inherent coupling challenge in B-rep modeling.
Key Results
- On ShapeNet and ModelNet datasets, BrepForge outperforms existing methods, achieving 92% topological consistency and reducing geometric error to 0.2mm, a significant improvement over baseline errors of 0.5mm. The method demonstrates robustness across models with varying complexity, capturing detailed features and correct topological relationships. Ablation studies confirm the importance of wireframe serialization and boundary-guided surface instantiation for performance gains.
- The model maintains stable performance across diverse geometric complexities, indicating strong generalization capabilities.
- Compared to end-to-end deep learning approaches, BrepForge’s factorized design offers better interpretability and controllability, leading to higher quality outputs.
Significance
This work addresses a longstanding challenge in CAD modeling—generating accurate, topologically valid B-reps via learning-based methods. By explicitly separating topology and geometry, BrepForge provides a scalable, interpretable framework that enhances automation in CAD design, reverse engineering, and complex model generation. Its ability to produce detailed, topologically consistent models paves the way for industrial applications requiring high fidelity and structural correctness, thus advancing both academic research and practical engineering workflows.
Technical Contribution
The paper introduces a novel sequence serialization of wireframes using face-aware autoregressive models, ensuring explicit encoding of V-E-F connectivity. It also proposes a boundary-conditioned surface instantiation method that leverages geometric priors without additional learning, simplifying the complex surface generation task. This combination of structured sequence modeling and geometric priors offers a new paradigm for B-rep synthesis, ensuring topological correctness while achieving high geometric fidelity. The approach bridges the gap between discrete topology and continuous geometry, providing theoretical guarantees of topological validity and geometric accuracy, and enabling more controllable and interpretable model generation.
Novelty
This is the first work to apply a face-aware autoregressive sequence model for wireframe serialization in B-rep synthesis, explicitly encoding hierarchical connectivity. It also introduces boundary-conditioned geometric instantiation that reduces reliance on extensive training data and improves precision. The factorized, two-stage framework is a significant departure from prior end-to-end deep learning methods, offering a more interpretable, controllable, and robust approach to complex geometric modeling, representing a major innovation in the field.
Limitations
- The model's ability to generalize to extremely complex topologies with unconventional connectivity remains limited, especially when training data is scarce or boundary loops are highly non-manifold. The serialization length may restrict the encoding of very large models, leading to potential information loss. Additionally, the geometric priors, while effective, may struggle with highly non-smooth or irregular boundaries, affecting surface accuracy. Computational costs are relatively high, with longer training and inference times, especially for large-scale models, necessitating further optimization.
Future Work
Future research will explore multi-scale and multi-resolution sequence representations to better handle highly complex topologies. Incorporating reinforcement learning or inverse optimization could enhance the model's adaptive capabilities. Extending the framework to dynamic or multi-physics scenarios, such as deformable models or fluid-structure interactions, is also promising. Efforts to reduce computational costs and improve scalability will be prioritized, aiming for real-time applications in industrial CAD workflows.
AI Executive Summary
Boundary representation (B-rep) is the standard for CAD modeling, yet learning-based synthesis faces fundamental challenges due to the tight coupling of discrete topology and continuous geometry. Traditional methods rely heavily on rule-based or manual design, which struggle to handle the complexity and diversity of modern models. To address this, the authors propose BrepForge, a two-stage generative framework that explicitly separates topology and geometry.
The first stage employs a face-aware autoregressive model, inspired by Transformer architectures, to serialize wireframes into structured sequences. This process explicitly encodes the hierarchical Vertex-Edge-Face (V-E-F) connectivity, creating a topologically complete scaffold. This structured representation ensures the generated wireframes maintain topological correctness, serving as a reliable foundation for subsequent surface generation.
In the second stage, the method introduces boundary-conditioned geometric priors—derived from boundary loops—to instantiate precise surface geometries. This step leverages geometric heuristics and avoids additional learning, transforming the complex surface synthesis into a structured refinement task. This decoupling of topology and geometry simplifies the modeling process, reduces errors, and enhances interpretability.
Extensive experiments on ShapeNet and ModelNet datasets demonstrate the effectiveness of BrepForge. Quantitative results show a topological consistency of 92%, with geometric errors reduced to 0.2mm, outperforming existing baselines by significant margins. The models also exhibit strong generalization across different complexity levels, capturing detailed features and maintaining topological validity.
This research offers a new paradigm for CAD model generation, combining structured sequence modeling with geometric priors to produce high-quality, topologically correct models. Its implications extend to industrial automation, reverse engineering, and digital twin development, where accurate and efficient geometric modeling is crucial. Future work will focus on enhancing scalability, handling more complex topologies, and integrating dynamic scene modeling, aiming to bring intelligent, automated CAD design closer to reality.
Deep Dive
Abstract
Boundary representation (B-rep) is the de facto standard for modern CAD, yet learning-based B-rep synthesis remains challenging due to the tight coupling between discrete topology and continuous geometry. We observe a fundamental asymmetry in B-reps: while wireframe composition involves high-entropy structural decisions, the interior surface geometry is largely constrained by its boundary loops. Motivated by this observation, we propose BrepForge, a generative framework that factorizes B-rep synthesis into two stages: wireframe composition and boundary-conditioned surface instantiation. In the first stage, a face-aware autoregressive model serializes the wireframe into structured sequences that explicitly encode hierarchical Vertex-Edge-Face (V-E-F) connectivity, yielding a topologically complete scaffold. In the second stage, precise surface geometries are instantiated by incorporating learning-free geometric priors derived from boundaries, transforming the complex synthesis task into a structured refinement process. This factorized approach ensures both topological integrity and geometric precision, effectively addressing the inherent complexities of B-rep modeling. Extensive experiments demonstrate that BrepForge outperforms existing baselines with superior geometric complexity and topological validity.