In order to solve the problems of insufficient local image details and lack of global sample diversity caused by training imbalance and data scarcity in architectural design generation, a hybrid generation model combining GAN (Generative Adversarial Network) and Stable Diffusion algorithm is constructed. GAN adversarial training is used to improve local details, and Stable Diffusion is used to improve global style consistency, aiming to achieve intelligent architectural design with high stability and practicality. Using data such as building floor plans, 3D models and renderings, a dataset covering various architectural styles is constructed. The consistency and accuracy of data input are ensured through image normalization, resolution adjustment and style annotation. Based on the Stable Diffusion model, domain-specific conditional control is introduced to achieve precise regulation of architectural style. Combined with the adversarial training mechanism of GAN, local details are optimized to improve the overall image refinement. A mature open source model is selected as the initial model. By adjusting hyperparameters such as learning rate and diffusion steps, the model is gradually converged. Incremental training methods are introduced to ensure the balance between local details and global style during training. The experimental results show that the SSIM (Structural Similarity Index Measure) value of the Stable Diffusion + GAN hybrid model is 0.912, with an error range of ± 0.021. It not only has the best generation quality but also the best stability. In terms of building functional layout, the core accessibility index of the Stable Diffusion + GAN hybrid model is 0.87 ± 0.03, the visibility of public areas is 68.3% ± 2.1%, the compliance rate with regulations is 92.7% ± 1.2%, and the expert score is 4.3 ± 0.4, all of which meet the industry standard reference values. The experimental results prove the effectiveness of this paper’s research on architectural model analysis.