Reproduce “GraphBrep: Learning B-Rep in Graph Structure for Efficient CAD Generation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Weilin Lai, Tie Xu, Hu Wang
GraphBrep, developed by researchers from Hunan University, Beijing Institute of Technology, and SAIC GM Wuling, introduces a graph-based diffusion model for generating Boundary Representation (B-Rep) CAD models more efficiently by explicitly learning topology. This method reduces training time by up to 31.3% and inference time by up to 56.3% compared to prior art, while achieving higher validity rates of 66.62% on the DeepCAD dataset.
Key insights
- Explicitly modeling B-Rep topology as a weighted undirected graph with an adjacency matrix significantly reduces data redundancy and computational complexity compared to implicit or tree-based representations.
- Decomposing the complex joint distribution of geometry and topology into a sequence of conditional diffusion models allows for more manageable and efficient learning of each component.
- Conditioning the generation of geometric edges on the pre-generated surfaces and their explicit topological connections (adjacency matrix) ensures geometric plausibility and avoids fixed, redundant edge counts per surface.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.