Abstract: Aerodynamic shape design is a critical aspect of aircraft development. To enhance design efficiency, overcome the topological constraints of parametric methods in early conceptual design, and assist designers in rapidly generating initial candidates, this study investigated the problem of intelligent 3D aerodynamic shape design. A generative model-based framework was proposed, which employs point clouds as a unified geometric representation. First, a geometric foundation model based on an autoencoder architecture was constructed and pre-trained in a self-supervised manner to extract low-dimensional latent vectors of 3D aerodynamic shapes. Then, a conditional diffusion model guided by aerodynamic performance was built and trained in the latent space to enable controllable generation of shape representations, which were subsequently decoded into point clouds. Additionally, for aerodynamic coefficient prediction, a mapping network was appended to the decoder of the foundation model to enable fast performance estimation. Preliminary validation of the framework was conducted. The results show that the framework can efficiently generate a large number of point-cloud-based aerodynamic shapes in an end-to-end manner according to target performance, and the generated shapes can be further screened by the aerodynamic performance prediction model. This framework demonstrates strong generalization capability and is applicable to generative design of 3D aerodynamic configurations for various architectures.