Abstract: Multidisciplinary design optimization is a primary approach for advanced aircraft to balance aerodynamic and stealth performance. In recent years, generative adversarial network-based design methods have been developed, yet they suffer from performance-awareness deficiency and scattered performance distribution of generated configurations. To address these issues, this paper introduces a regression model as a performance predictor into the conditional Wasserstein generative adversarial network with gradient penalty (CWGAN-GP), and proposes a generative design framework guided by performance regression, termed guided CWGAN-GP (GCWGAN-GP). This framework establishes a guidance loss based on the deviation between predicted performance and target conditions, thereby enabling performance feedback. A dual-constraint mechanism is constructed, comprising a regression-model-based performance constraint and a discriminator-based geometric constraint. Taking lift and radar cross section (RCS) as target performance metrics, a dataset is built through parametric modeling, the vortex lattice method, and the physical optics method, and aerodynamic/stealth design of a flying-wing configuration with 28-dimensional shape design variables is conducted. Results demonstrate that, through the optimization of multiple training strategies, the GCWGAN-GP framework achieves favorable convergence. Compared with the baseline CWGAN-GP model, the GCWGAN-GP model can stably generate layout schemes with clear and consistent geometric features, while achieving highly concentrated target performance and effectively balancing aerodynamic and stealth performance. The framework realizes performance-oriented data learning and high-performance shape generation, validating its feasibility and superiority in multidisciplinary design optimization of aircraft.