SEEK-CAD: A SELF-REFINEDGENERATIVEMODEL- ING FOR3D PARAMETRICCAD USINGLOCALINFER- ENCE VIADEEPSEEK

The advent of Computer-Aided Design (CAD) generative modeling will significantly transform the design of industrial products. This study investigates the incorporation of Chain-of-Thought (CoT) feedback within a self-refinement mechanism for generating 3D parametric CAD models using a training-free methodology with the locally deployed inference LLM DeepSeek-R1. We begin by rendering an initial generated parametric CAD model into a sequence of step-wise perspective images, processed by a Vision Language Model (VLM) along with corresponding CoTs, to evaluate the CAD model generation. The gathered feedback is then utilized by DeepSeek-R1 to refine the initial model for subsequent generations. Furthermore, we introduce a 3D CAD model dataset structured around the SSR (Sketch, Sketch-based feature, and Refinements) triple design paradigm, encompassing a variety of CAD commands tailored to meet industrial application needs and supporting LLM generation. Extensive experiments validate the effectiveness of the Seek-CAD approach across various metrics, demonstrating its potential to enhance CAD parametric modeling efficiency and performance.