Reproduce “CADMorph: Geometry-Driven Parametric CAD Editing via a Plan-Generate-Verify Loop” on OpenResearch

About this paper

Weijian Ma, Shizhao Sun, Ruiyu Wang, Jiang Bian

CADMorph introduces an iterative framework for geometry-driven parametric CAD editing, integrating a Parameter-to-Shape latent diffusion model and a Masked-Parameter-Prediction large language model within a plan-generate-verify loop. The system achieves superior geometric adherence, structural preservation, and generation fidelity, outperforming baselines with a 0.687 IoU, 0.009 mean Chamfer Distance, and a 3.1% invalid rate.

Key insights

• An iterative plan-generate-verify loop effectively leverages pretrained domain-specific foundation models to perform complex CAD editing without requiring scarce triplet datasets for training.
• Cross-attention maps from a Parameter-to-Shape model can precisely identify which segments of a parametric sequence contribute to specific geometric parts, enabling targeted modifications and promoting structural preservation.
• The combination of diverse candidate generation from an LLM and shape similarity verification with a latent diffusion model, along with a priority queue for selection, enhances both shape fidelity and the robustness of the editing process.

Summary by alphaXiv.