Reproduce “Towards Valid B-Rep Generation: Training-Free Wireframe Anomaly Detection and Repair” 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
Jingyu Wu, Youcheng Cai, Tengyu Luo, Ligang Liu
Wireframe Detection and Repair (WDR) is introduced as a training-free framework that integrates into multi-stage B-Rep generation pipelines to detect and rectify geometric and topological anomalies in intermediate wireframes. The WDR framework consistently improved the kernel-checked validity of generated CAD models by 10.9 to 26.9 percentage points across various datasets and models, while largely preserving generative diversity and quality.
Key insights
- Combining multimodal risk signals—visual, geometric, and topological—provides a more comprehensive and accurate prediction of downstream B-Rep invalidity, as demonstrated by the GTAD's F1-score of 81.97%.
- Energy-guided local candidate reranking during autoregressive decoding effectively steers the generation towards lower-risk wireframes, achieving superior validity compared to parallel "Best-of-N" reranking.
- The framework's plug-and-play nature allows for substantial improvements in B-Rep validity without the need to retrain or fine-tune large pre-existing generative models, making it a computationally efficient solution.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.