Abstract: Architectural visualization typically relies on renderings and animations as its two primary modes of presentation. Conventional production workflows, however, suffer from low modeling efficiency, slow rendering speeds, and inadequate material and lighting fidelity. This study investigates the application of the Stable Diffusion generative model for producing architectural renderings and animations from three perspectives: model comprehension and training, image generation and refinement, and case application and analysis. The technical principles and parameter control details of LoRA model training, ControlNet-based conditioning, text-to-image (T2I) generation, and image-to-image (I2I) translation are examined in depth. Furthermore, by benchmarking against the current workflow of architectural scheme design, a comparative analysis between AI-generated and manually produced visualizations is conducted. The results reveal that AI image generation offers high efficiency, automation, and fine-grained detail optimization, enabling rapid style diversification and iterative design exploration. Consequently, high-quality renderings can be produced in significantly reduced timeframes, thereby shortening the overall design cycle.