Abstract: Accurate and flexible geometric parameterization is a cornerstone of modern aerodynamic design and optimization. While recent advances in generative modelling have improved airfoil design workflows by learning latent representations, most existing approaches still rely on traditional parametric functions, which inherently constrain the design space and limit geometric diversity. To overcome these limitations, we propose Airfoil-DDPM, a non-parametric airfoil generation framework based on Denoising Diffusion Probabilistic Models (DDPM). Our method directly learns the geometric distribution of airfoil shapes from data and generates high-fidelity designs without any predefined parameterization. By integrating a UNet backbone with attention mechanisms, Airfoil-DDPM captures fine-scale geometric features and ensures smoothness and curvature continuity. Experimental results demonstrate that Airfoil-DDPM accurately reproduces both the geometric and aerodynamic characteristics of standard airfoils such as NACA0012 and RAE2822. Furthermore, qualitative visualizations confirm its superior ability to learn the UIUC library’s data distribution and generate diverse, physically plausible airfoil geometries. Our approach offers a flexible and efficient alternative to traditional parameterization-based methods, paving the way for nex_t-generation aerodynamic design.