Generative Modeling of B-Reps

• Generative Modeling of B-Reps is a process of synthesizing complete, editable CAD models with watertight topology using neural and probabilistic methods.
• It employs unified representations such as Indexed B-Reps, chain complexes, and holistic token sequences to capture both geometric and topological dependencies.
• This approach overcomes traditional pipeline limitations by integrating structure-aware neural architectures, yielding higher validity and efficiency in CAD synthesis.

A boundary representation (B-Rep) is the canonical geometric and topological format for representing solid objects in computer-aided design (CAD) and manufacturing. Generative modeling of B-Reps seeks to synthesize complete solids with watertight, manifold topology and tractable, editable parametric geometry, directly from neural or probabilistic models. This field targets the automated creation, reconstruction, and manipulation of CAD models at the semantic level of faces, edges, and vertices, in contrast to mesh or occupancy-based approaches. Recent advances have focused on overcoming the rigidity and error-prone multistage pipelines inherent in traditional bottom-up B-Rep synthesis, devising unified representations and generative frameworks that integrate both geometry and topology with high fidelity and scalability.

1. Formalization and Definitions of B-Rep Generative Modeling

Generative modeling of B-Reps is defined as learning a conditional distribution over tuples $\mathcal{M} = (F, E, V, \partial, \mathcal{P})$, where faces $F$ are parametric surface patches, edges $E$ are curves, vertices $V$ are points, $\partial$ encodes explicit incidence/topological relations (e.g., which edges bound each face, which vertices lie at endpoints of edges), and $\mathcal{P}$ stores the geometric parameters of each primitive. A valid B-Rep is both watertight and manifold: each edge has precisely two incident faces and two endpoints, all faces form closed loops, and the assembly admits explicit algebraic or spline-based geometric parametrizations for exact representation and downstream editing (Liu et al., 2024). Typically, generative tasks address unconditional synthesis, point cloud/model reconstruction, conditional generation (class, image, text), and autocompletion or inpainting of partial models.

2. Unified Representations and Their Structural Implications

Several paradigms exist for encoding the B-Rep structure for generative modeling:

• Indexed Boundary Representation (IBR): Converts the B-Rep graph into ordered sequences of vertices, edges, and faces, enabling autoregressive pointer-based generation of each element with explicit references between them (Jayaraman et al., 2022).
• Chain Complexes: Treats B-Reps as algebraic chain complexes, representing each order (vertex, edge, face) as a commutative group and topology as boundary maps, enforcing manifoldness and closed-loop constraints via linear algebra (Guo et al., 2022).
• Holistic Token Sequences: Packs geometry tokens, position tokens, and topology index tokens into a single, hierarchically ordered stream, supporting autoregressive next-token models in a unified vocabulary (Li et al., 23 Jan 2026, Xu et al., 2 Dec 2025).
• Voronoi and Particle-Based Flattening: Partitions the input space by the classical Voronoi diagram of GT primitive centers, recovering cells via a learned binary boundary mask and fitting primitives per cell, or flattens the cell complex into unordered sets of compositional k-cell particles with shared latents at interfaces (Liu et al., 2024, Lu et al., 25 Jan 2026).
• Topological Decoupling: Explicitly splits topology generation from geometry, often modeling edge-face and edge-vertex adjacency relationships independently with Transformers or Graph diffusion models, then sequentially generating geometric attributes conditioned on sampled topology (Li et al., 17 Mar 2025, Lai et al., 7 Jul 2025).
• Volumetric Distance Functions (BR-DF): Encodes surface and per-face topology into signed and unsigned distance functions, reconstructing faceted B-Reps by robust Marching Cubes-and-Triangles decoding (Zhang et al., 18 Nov 2025).
• Direct NURBS Parameterization: Learns to encode and decode surfaces using normalized NURBS parameters, drastically reducing representation size and improving face quality (Fan et al., 2024).

These representations define the possible workflows for joint or decoupled geometry/topology generation and directly determine the efficiency, scalability, and editability of the resulting generative models.

3. Generative Modeling Frameworks: Pipelines and Algorithms

Generative B-Rep modeling pipelines can be categorized as follows:

• Autoregressive Transformer-Based Pipelines: Sequentially sample vertices, then edges as pointers into the vertex pool, and faces as edge sets or surface tokens, typically with causal masking and pointer-softmax architectures to guarantee valid reference orderings. Unified tokenization schemes, as in AutoBrep and BrepARG, enable single-stack Transformers to capture geometric and topological dependencies in a breadth-first or DFS traversal order (Li et al., 23 Jan 2026, Xu et al., 2 Dec 2025).
• Diffusion Models: Employ DDPM frameworks operating over node features in hierarchical trees (BrepGen), chain complexes (ComplexGen), or graph structure (GraphBrep). Unified denoising or flow-matching is used for both position and latent geometry, with topology implicitly enforced by deduplication, merging, or explicit adjacency diffusion (Xu et al., 2024, Lai et al., 7 Jul 2025, Lu et al., 25 Jan 2026). Geometry is often handled via VAE-latent or spline parameter diffusion, with attention-derived conditioning from prior levels.
• Top-Down Partitioning and Structure-Aware Generation: The Split-and-Fit pipeline leverages a neural Voronoi diagram—learned with NVD-Net via voxelwise binary classification in 3D U-Nets—to split the volume into structure-adaptive cells. Each cell undergoes primitive fitting by least-squares minimization over the analytic surface family, followed by intersection curve and vertex recovery; adjacency is inferred from touching cell boundaries (Liu et al., 2024).
• Holistic Latent Space Models: Surface-embedding–only latent spaces, where curves and vertices are decoded via neural intersection networks, recast topological learning as a regular geometric decoding problem, supporting multimodal conditioning and reducing pipeline ambiguity (Liu et al., 19 Apr 2025).
• Volumetric Generative Pipelines: Latent diffusion among SDF/UDF fields, decoded via Marching Cubes and Triangles, ensure watertightness and unconditional topology validity. These pipelines permit batch-wise synthesis of solid collections, with multi-branch 3D U-Net cross-attention handling face–face and face–surface dependencies (Zhang et al., 18 Nov 2025).

A sample pipeline, based on Split-and-Fit (Liu et al., 2024):

1. Convert input (point cloud, mesh, UDF) to 256³ voxel grid storing unsigned distance and gradients.
2. Apply 3D U-Net (NVD-Net) for binary boundary classification with focal loss to handle class imbalance.
3. Recover Voronoi cells by region-growing; fit analytic primitive per cell by minimizing parametric residuals.
4. Analytically intersect adjacent cells for curves and infer vertices at curve intersections.
5. Assemble full B-Rep from recovered faces, curves, vertices, with explicit incidence relations.

4. Evaluation Metrics and Benchmarks

Standard evaluations of generative B-Rep models adopt:

• Point-cloud-based distribution metrics:
  • Coverage (COV): proportion of reference test shapes covered by generated samples (thresholded Chamfer).
  • Minimum Matching Distance (MMD): mean nearest-neighbor Chamfer between generated and test sets.
  • Jensen–Shannon Divergence (JSD): histogram-based comparison of voxelized occupancies.
  • Light-Field Distance (LFD): evaluates overall visual fidelity in rendered view space.
• CAD-specific metrics:
  • Valid ratio: percentage of generated models that are watertight and topologically consistent.
  • Primitive-level F1: detection of ground-truth vertices, curves, or surfaces within a threshold.
  • Topological consistency: F1 of face–edge and edge–vertex incidence recovery.
  • Novelty/Uniqueness: fraction of generated shapes not in training or that are distinct within the sample batch.
  • Cyclomatic complexity: loop count in the wireframe graph.

Relevant benchmarks include ABC (parametric CAD), DeepCAD (sketch-and-extrude), Furniture (category-conditioned), and 1M+ unique STEP solids in ABC-1M (Xu et al., 2 Dec 2025). Models are quantitatively compared across these metrics and datasets, often showing AutoBrep, BrepARG, and BrepGPT outperform prior frameworks, achieving higher validity and fidelity with significantly reduced training and inference times.

Method	COV (%)	MMD×10	JSD×10	Valid (%)	Reference
BrepARG	75.45	0.89	1.02	87.60	(Li et al., 23 Jan 2026)
DTGBrepGen	74.52	1.07	1.02	79.80	(Li et al., 17 Mar 2025)
Split-and-Fit	–	–	–	–	(Liu et al., 2024)
AutoBrep	71.49	1.45	0.97	70.8	(Xu et al., 2 Dec 2025)
BrepGPT	79.3	0.960	0.840	83.9	(Li et al., 27 Nov 2025)

Qualitative assessments (e.g., user studies) confirm that top-down, structure-aware methods (Split-and-Fit) yield cleaner B-Reps with fewer over-segmentations and higher plausibility of primitive adjacencies.

5. Scalability, Multimodal Conditioning, and Efficiency

Generative B-Rep models are increasingly capable of handling complex solids with large face/edge counts and supporting multimodal input conditioning:

• Scalability: Unified token streams, autoregressive sequencing, and particle set flow-matching architectures enable reliable scaling to 100+ faces per solid (valid ratio remains >50–80% at maximum complexity in AutoBrep, BrepARG, and Flatten The Complex) (Li et al., 23 Jan 2026, Xu et al., 2 Dec 2025, Lu et al., 25 Jan 2026).
• Multimodal inputs: Conditioning pipelines ingest point clouds (via PointNet (Liu et al., 19 Apr 2025) or Michelangelo (Li et al., 27 Nov 2025)), images (DINOv2, Sonata), sketches, or text semantics (CLIP), yielding direct B-Rep reconstructions for arbitrary input modalities.
• Training/inference speed: Unified representations markedly reduce computational cost. BrepARG training/inference times are 1.2 days/1.5 s, compared to 3.0 days/3.6 s for DTGBrepGen and 7.5 days/8.4 s for BrepGen, showing 2–5× speedups (Li et al., 23 Jan 2026).
• Watertightness and Robustness: BR-DF's volumetric approach yields provable 100% validity via robust topological recovery (Zhang et al., 18 Nov 2025). Particle set representations support non-manifold synthesis and local inpainting (Lu et al., 25 Jan 2026).

6. Limitations and Open Challenges

While recent methods have advanced B-Rep generative modeling, several limitations persist:

• Surface representational challenges: In neural or quantized UV-grid representations, inconsistent geometry can produce invalid trims and stitching failures; methods like NeuroNURBS alleviate undulating artifacts but still require preprocessing and are limited to untrimmed NURBS faces (Fan et al., 2024).
• Topology–geometry coupling: Cascaded multistage pipelines may accumulate errors or propagate ambiguities; particle-based and holistic latent methods show promise in reducing these effects, but joint differentiable graph reconstruction remains an active area (Xu et al., 2024, Liu et al., 19 Apr 2025, Lu et al., 25 Jan 2026).
• Expression of complex primitives: Support for free-form, non-uniform B-splines, T-splines, and subdivision schemes is limited; extensions to parameterize control grids natively are under investigation (Fan et al., 2024, Xu et al., 2024).
• Metric limitations: Coverage, MMD, and JSD capture only partial aspects of manufacturing validity; new metrics for manifoldness, curvature continuity, and assembly/fit constraints are needed (Lai et al., 7 Jul 2025).
• Multiple solid bodies and assemblies: Most current models generate only single bodies; extension to hierarchical/assembly-level generation is ongoing (Xu et al., 2024, Lu et al., 25 Jan 2026).

7. Prospects and Emerging Directions

The future of B-Rep generative modeling involves the integration of several advances:

• Hierarchical and hybrid generative sequencing to manage large-scale, highly detailed solids.
• Direct, end-to-end latent modeling on raw parametric surface representations (NURBS, T-spline), potentially without autoencoding or preprocessing stages (Fan et al., 2024).
• Generalized flow-matching and particle set models that can handle surface, curve, vertex entities, with global context awareness and flexible combinatorial structures, supporting both manifold and non-manifold output (Lu et al., 25 Jan 2026).
• Enhanced conditioning and autocompletion pipelines enabling designer-in-the-loop workflows, controllable assemblies, and robust inpainting from partial B-Reps (Xu et al., 2 Dec 2025, Li et al., 27 Nov 2025).
• Topology-aware loss functions and evaluation metrics reflecting industrial requirements for manufacturability and structural correctness (Lai et al., 7 Jul 2025).

Generative modeling of B-Reps now benefits from a spectrum of approaches—structure-aware partitioning, holistic tokenization, chain complexes, decoupled topology-generation, and robust volumetric field models—each contributing to increased accuracy, validity, and practical tractability for AI-driven CAD synthesis. Methods such as Split-and-Fit (Liu et al., 2024), BrepARG (Li et al., 23 Jan 2026), BrepGen (Xu et al., 2024), DTGBrepGen (Li et al., 17 Mar 2025), AutoBrep (Xu et al., 2 Dec 2025), and particle/flow-matching schemes (Lu et al., 25 Jan 2026) set a new standard for representation and generation of boundary representations in computational design.