Diffusion-CAD: Controllable Diffusion Model for Generating Computer-Aided Design Models

Generative methods for creating computer-aided design (CAD) models have gained significant attention over the past two years. However, existing methods lack fine-grained control over the generated CAD models, making it difficult to manage details such as model dimensions and the relative structure of components. To address these limitations, this study introduces Diffusion-CAD, a diffusion-based generative approach that outputs CAD construction sequences. Diffusion-CAD iteratively denoises Gaussian noise into continuous CAD vectors, which are then transformed into discrete CAD sequences. We designed classifier-free and classifier-guided methods to control the distribution of Gaussian noise, CAD sequences, and noisy CAD vectors separately, thereby achieving a variety of fine-grained control tasks. Extensive experiments demonstrated the superior performance and novel capabilities of the proposed method for conditional generation tasks.