Reproduce “TG-Diff: Coupling Discrete Topology Diffusion and Topology-conditioned Geometry Diffusions for B-Rep 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
MingZe Sun, Haiyong Jiang, Bingchen Yang, Haoxuan Song, Yidi Li, Jun Xiao, Peter Wonka
Researchers at the University of Chinese Academy of Sciences introduced TG-Diff, a two-stage diffusion framework for B-rep generation that decouples topology and geometry using a surface-centric representation. This approach yielded a validity of 85.6% on DeepCAD-0-30 and nearly doubled structural diversity compared to prior methods, while reducing computational overhead.
Key insights
- Decoupling topology and geometry generation, coupled with a simplified surface-centric representation, inherently reduces geometric-topological inconsistencies and simplifies the generation process.
- Using discrete diffusion for topology ensures a high success rate and diversity in structural generation, which then effectively conditions subsequent geometry creation.
- FiLM-style modulated attention in the geometry diffusion phase effectively incorporates topological information, enabling consistent and accurate surface generation aligned with the structural plan.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.