A Comparative Analysis of Deep Learning Models for Text-to-CAD Generation

Generating precise 3D Computer-Aided Design (CAD) models from natural language text is a significant challenge in engineering. In this seminar, we investigate this problem by comparing two deep learning architectures: the foundational DeepCAD model and the state-of-the-art Text2CAD transformer. First, we implemented DeepCAD and identified a critical cuboid bias, where the model fails to capture complex geometries. Second, we implemented the Text2CAD framework, which utilizes a BERT encoder and cross-attention decoder. Our comparative analysis (implementation and testing) demonstrates that the Text2CAD architecture achieves a 10x reduction in invalid models and superior parametric precision.