In automobile aerodynamic design, balancing design features and aerodynamic performance is essential. Therefore, shapes that improve aerodynamic performance while maintaining the features of an already decided design are required, but developing such shapes quickly to meet these requirements is not easy. This study proposes an AI model that quickly modifies the shape, based on the customer’s needs and the initially decided design, to improve aerodynamic performance while maintaining the automobile's defining design features. We target the aerodynamic drag coefficient (Cd) as the performance metric. The proposed model represents each automobile as a multidimensional vector extracted by a neural network and enables continuous shape editing by manipulating components of this vector. Implemented as an encoder–decoder architecture, the system encodes 3D geometry into a latent space, conditions the latent code on a user-specified target Cd reduction, and decodes the modified code back into a 3D geometry. The network is trained to produce plausible, aerodynamically effective shapes while retaining visual design cues, following the approaches of Miao et al. [1] and Vatani et al. [2]. In this study, we evaluate the proposed model by using original, high-fidelity shape data generated through morphing based on production car SUV designs, under conditions that closely resemble actual design practice. The method for creating shape data is based on Fu et al. [3]. This approach deliberately avoids using traditional research benchmark datasets (e.g., DrivAer, ShapeNet) in order to more deeply assess the practical usefulness of the method in real-world design settings. By applying the proposed model to production car designs and evaluating the quantitative change in Cd and the qualitative assessment of shape, we demonstrate the extent to which this method can contribute to practical design work. ACKNOWLEDGEMENT We gratefully acknowledge Mazda Motor Corporation for providing the production SUV shape dataset used in this study. REFERENCES [1] J. Miao, T. Ikeda, B. Raytchev, et al., Fine-grained 3D vehicle shape manipulation via latent space editing. Machine Vision and Applications, vol. 36, no. 117, pp. 1-15, 2025. [2] P. Vatani, M. Elrefaie, F. Nazarpour, et al., TripOptimizer: Generative 3D Shape Optimization and Drag Prediction using Triplane VAE Networks, 5 Sep 2025. [3] H. Fu, Q. Wang, T. Nakashima, et al., Gradient-Enhanced Kriging-Based Parallel Efficient Global