Research map / Graph learning / Graph self-supervised learning
BrepGiff: Lightweight Generation of Complex B-rep with 3D GAT Diffusion
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 BrepGiff, a lightweight diffusion model that encodes B-rep CAD models as 3D face-centroid graphs and denoises them with a degree-guided GAT, enabling generation of complex B-rep models (up to 100+ faces) on two RTX 4090 GPUs.
Problem
Direct generation of complex B-rep CAD models is hard because B-rep data mixes parametric geometry with topology; prior direct methods (e.g., BrepGen) handle only about 50 faces and need heavy compute.
Approach
Represent a B-rep as a 3D graph whose nodes are face centroids (MST-reduced, with UV-sampled face/edge geometry as attributes), and run a DDPM-style diffusion that removes/adds edges guided by node degrees, with a cross-attention GAT denoiser that predicts edges only for active nodes.
Key results
- Unconditional generation on DeepCAD-matched data: COV 75.52%, MMD 0.98, JSD 1.17, Valid 67.7%, beating BrepGen (73.87% COV, 62.9% valid).
- Model has 7.80e7 parameters versus 4.56e8 for BrepGen, training in 4 days on a single RTX 4090 (BrepGen: 7 days on two RTX 4090s).
- With max 100 faces, validity is 62.9%; removing GAT drops validity for >100-face models from 67.7% to 31.8%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use BrepGiff instead of BrepGen | BrepGiff · ABC unconditional generation | COV (%): 73.87 -> 75.52 | Helped |
| Use BrepGiff instead of BrepGen | BrepGiff · ABC unconditional generation | Valid (%): 62.9 -> 67.7 | Helped |
| Remove 1D CNN edge feature extraction (ablation) | BrepGiff · B-rep models up to 120 faces (50K models) | Valid (%): 58.8 -> 50.3 | Hurt |
| Remove 2D CNN face feature extraction (ablation) | BrepGiff · B-rep models up to 120 faces (50K models) | Valid (%): 58.8 -> 22.4 | Hurt |
| Remove GAT denoising (ablation) | BrepGiff · B-rep models up to 120 faces (50K models) | Valid (%): 58.8 -> 22.9 | Hurt |
| Use 100 diffusion steps instead of 1000 (ablation) | BrepGiff · B-rep models up to 50 faces | COV (%): 68.55 -> 35.22 | Hurt |
| Use 400 diffusion steps instead of 1000 (ablation) | BrepGiff · B-rep models up to 50 faces | COV (%): 68.55 -> 62.15 | Hurt |
| Use 700 diffusion steps instead of 1000 (ablation) | BrepGiff · B-rep models up to 50 faces | COV (%): 68.55 -> 68.45 | Hurt |
Surprising findings
- Validity for complex models with >100 faces falls sharply without GAT (67.7% to 31.8%), indicating the GAT denoiser is critical for topology consistency.
Limitations
- Failure cases include missing or broken faces, sensitive to the number of diffusion steps (under- or over-processing).
- Baseline BrepGen could not be trained on >50-face models with two RTX 4090s, so no direct comparison on complex models.
- Reconstruction relies on B-spline fitting and a GNN-predicted adjacency, and the training set excludes NURBS and other complex surfaces.
Experimental setup
Datasets: ABC dataset DeepCAD data split
Benchmarks: ABC dataset
Metrics: COV MMD JSD Novel Unique Valid
Baselines: DeepCAD SolidGen BrepGen
Builds on: EDGE DDPM BrepGen Graph Attention Network
Closest papers by meaning
- DTGBrepGen: A Novel B-rep Generative Model through Decoupling Topology and Geometry (CVPR 2025 · similarity 0.91)
- Mesh BDF: Barycentric Dominance Field for 3D Native Mesh Generation (NeurIPS 2026 · similarity 0.85)
- SpelsNet: Surface Primitive Elements Segmentation by B-Rep Graph Structure Supervision (NeurIPS 2024 · similarity 0.84)
- CraftsMan3D: High-fidelity Mesh Generation with 3D Native Diffusion and Interactive Geometry Refiner (CVPR 2025 · similarity 0.83)
- Hyperbolic Geometric Latent Diffusion Model for Graph Generation (ICML 2024 · similarity 0.83)
- CADDreamer: CAD Object Generation from Single-view Images (CVPR 2025 · similarity 0.83)
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.