Aircraft design involves optimization within extremely high-dimensional geometric design spaces, posing a fundamental challenge for aerodynamic shape optimization. Although geometric filtering has improved design-space compactness compared to conventional parameterization methods, a compelling need remains to further shrink the design space without sacrificing design quality. This study introduces a physics-regularized generative parameterization that integrates data-driven aerodynamic modeling with generative learning to achieve a compact space for aerodynamic shape optimization. The approach begins by constructing a modal design space via deep-learning-based geometric filtering. A self-evolving aerodynamic modeling framework with adaptive sampling is then developed to efficiently introduce physical regularization. An optimization-based sampling method is proposed to generate a diverse set of feasible aerodynamic shapes under both geometric and performance constraints. A generative model trained on this physics-regularized database provides a low-dimensional latent space for shape optimization. Validation on blended-wing-body configurations demonstrates that the proposed method maintains geometric completeness while reducing design variables by [Formula: see text] relative to the state-of-the-art geometric filtering. Consequently, the surrogate-based optimization cost decreases by over 55%, and the drag prediction error in fast data-driven optimization drops from 32.5 to 3.2%. These results confirm that the proposed physics-regularized generative parameterization enables a compact geometric space to effectively meet various aircraft design requirements.