Reproduce “AutoBrep: Autoregressive B-Rep Generation with Unified Topology and Geometry” 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
Xiang Xu, Pradeep Kumar Jayaraman, Joseph G. Lambourne, Yilin Liu, Durvesh Malpure, Pete Meltzer
AutoBrep, developed by Autodesk Research, introduces an autoregressive Transformer model that unifies the generation of B-Rep geometry and topology by tokenizing both into a single sequence. The system demonstrates improved B-Rep validity rates of 70.8% and significantly faster inference speeds, alongside enabling robust B-Rep autocompletion.
Key insights
- Unified tokenization of geometry and topology, processed by a single autoregressive Transformer, mitigates error accumulation and simplifies the generation pipeline compared to multi-stage approaches.
- The use of Finite Scalar Quantization (FSQ) effectively discretizes latent geometric embeddings, providing a compact and precise representation while achieving 100% codebook usage.
- Breadth-first traversal (BFT) with local topological references is crucial for robust B-Rep generation, outperforming global or coordinate-based ordering by focusing on relevant local structures and improving validity.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.