Surrogate-based optimization is widely used for aerodynamic shape optimization, and its effectiveness depends on representative sampling of the design space. However, traditional sampling methods are hard-pressed to effectively sample high-dimensional design spaces. This paper introduces DiffAirfoil, a newairfoil sampling method based on diffusion in an automatically learned latent space. DiffAirfoil is highly data-efficient and requires significantly fewer training geometries than Generative Adversarial Networks. It ensures the validity of sampled airfoils through an automatic parameterization. We demonstrate DiffAirfoil’s capability to generate diverse and valid 2D airfoil shapes, while also facilitating conditional generation without the need for adaptation or retraining. Comprehensive benchmarks show that our method offers significant advantages in data efficiency, ease of implementation, and adherence to conditions. Therefore, DiffAirfoil presents a promising approach to enhancing sampling efficiency for surrogate models in aerodynamic shape optimization tasks.