GenCAD: Image-conditioned Computer-Aided Design Generation with Transformer-Based Contrastive Representation and Diffusion Priors
Summary
GenCAD presents an image-conditioned generative model for parametric CAD, delivering both 3D CAD output and the full CAD command history. It uses an autoregressive transformer encoder, contrastive cross-modal learning, and a latent diffusion model to synthesize CAD programs conditioned on images, enabling precise, modifiable designs and CAD retrieval of thousands of programs. The work is accompanied by a GitHub codebase, highlighting practical AI tooling for automated CAD design.