Optimization strategies for generative AI via visual-feature clustering and aesthetic semantic modeling of responsive architectural facades

This study focuses on responsive architectural facades as the core research object. It examines how generative Artificial Intelligence (AI) can better understand and generate their complex visual and aesthetic characteristics. Responsive architectural facades are highly dynamic, adaptive, and visually complex. However, current generative AI systems still face limitations in semantic understanding, structural control, and aesthetic consistency when handling this type of architectural design task. These limitations often lead to semantic discontinuity, visual drift, and weak design controllability. Therefore, a clear research gap exists in developing a structured semantic modelling pathway that can connect facade visual features with promptbased generative control in architectural design. The study aims to build a multi-layer framework integrating K-means clustering and prompt optimization. It aims to establish a mapping from visual features to semantic labels and prompt strategies. Based on that, the framework proposes clustering-driven prompts to achieve better semantic control and human–AI collaboration compared to experience-based prompting. A mixed-method approach was adopted. A responsive architectural façade dataset was built through keyword extraction, image collection, and preprocessing. Kmeans clustering was then applied to derive visual prototypes and semantic categories. These outputs were translated into a structured prompt system and tested through multi-round generation experiments, followed by algorithmic evaluation and expert interviews. The results show that clustering-driven prompts improve semantic stability, style controllability, and generative consistency. Compared with experienc based prompts, the proposed method better represents dynamic facade characteristics and environment–interaction relationships. The framework also proved to improve cognitive feedback efficiency in human–AI collaboration. In conclusion, this study fills an important semantic gap in generative AI-enabled facade design and provides practical support for intelligent architectural design.