Research map / Image generation / Autoregressive and latent generation
DTGBrepGen: A Novel B-rep Generative Model through Decoupling Topology and Geometry
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DTGBrepGen, which decouples B-rep generation into a two-stage Transformer topology generator (edge-face then edge-vertex adjacency) followed by Transformer diffusion models for B-spline geometry.
Problem
Automatically generating valid and high-quality B-rep CAD models is hard because topology and geometry are tightly interdependent, and prior methods neglect topological constraints, hurting validity and geometric accuracy.
Approach
Decouple topology from geometry: generate edge-face adjacency (as a shared-edge count matrix) and then edge-vertex adjacency with a Transformer and pointer network, then generate face bounding boxes, vertex coordinates, edge and face B-spline control points with separate Transformer diffusion models, and sew the result with OpenCascade.
Key results
- On DeepCAD, validity rises from 68.23% (BrepGen) to 79.80%, with COV 74.52% and JSD 1.02 versus BrepGen's 72.38% and 1.29.
- On ABC, validity rises from 47.11% (BrepGen) to 62.08%.
- Class-conditioned Furniture validity is about 10 points higher than BrepGen across categories (e.g., Table 49.82% vs 28.50% Bathtub).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DTGBrepGen instead of BrepGen | B-rep generation · DeepCAD | Valid (%): 68.23 -> 79.80 | Helped |
| Use DTGBrepGen instead of BrepGen | B-rep generation · ABC | Valid (%): 47.11 -> 62.08 | Helped |
| Use DTGBrepGen instead of BrepGen | Class-conditioned B-rep generation · Furniture: Table, Chair, Bed, Sofa | COV (%): 65.08 -> 70.20; 61.24 -> 69.48; 59.30 -> 71.11; 67.82 -> 73.46 | Helped |
| Use DTGBrepGen instead of BrepGen | Class-conditioned B-rep generation · Furniture categories | Valid (%) | Helped |
| Replace B-spline representations with point-cloud-based latent encoding (ablation) | DTGBrepGen · DeepCAD | Valid (%): 79.80 -> 71.77 | Hurt |
| Replace B-spline representations with point-cloud-based latent encoding (ablation) | DTGBrepGen · ABC | Valid (%): 62.08 -> 56.63 | Hurt |
Surprising findings
- Using ground-truth topology gives only 90.24% geometry validity on DeepCAD, indicating valid topologies are not always suitable for geometry generation.
Limitations
- Overall validity still leaves room for improvement due to gap between learned and true topology distributions.
- Struggles with complex models having intricate geometry and complicated structure.
- Non-watertight solids and self-intersecting geometries appear in failure cases.
Experimental setup
Datasets: DeepCAD ABC Furniture
Benchmarks: DeepCAD ABC Furniture
Metrics: COV MMD JSD Novel Unique Valid
Baselines: DeepCAD BrepGen SolidGen
Builds on: BrepGen SolidGen DeepCAD PolyGen
Closest papers by meaning
- BrepGiff: Lightweight Generation of Complex B-rep with 3D GAT Diffusion (CVPR 2025 · similarity 0.91)
- Enhancing Implicit Shape Generators Using Topological Regularizations (ICML 2024 · similarity 0.85)
- CADDreamer: CAD Object Generation from Single-view Images (CVPR 2025 · similarity 0.85)
- CraftsMan3D: High-fidelity Mesh Generation with 3D Native Diffusion and Interactive Geometry Refiner (CVPR 2025 · similarity 0.85)
- Geometry-Informed Neural Networks (ICML 2025 · similarity 0.85)
- SILSA: Sliding-Window Slice Latents for Topology-Preserving High-Resolution 3D Generation (NeurIPS 2026 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.