Can Visual Clusters Support Semantic Steering? A Cross-Scale Study of Responsive Façade Images

Abstract

Active responsive façades are commonly classified by function or technology, while their visual organization across component and whole-façade scales remains underexplored. This study asks whether image features can organize cross-scale differences and whether cluster-derived semantics can improve generative steering. It introduces a traceable dual-scale workflow linking visual features, clustering, architectural semantics, and image generation. The dataset comprises 75 matched pairs of component close-ups (Set A) and whole-façade views (Set B). Each image was encoded by 43 features: SSIM, mean RGB values, a 32-bin gradient orientation histogram, and seven Hu moments. In two-dimensional PCA space, K-means yielded six scale-specific clusters, interpreted through feature evidence, metadata, and researcher-reviewed GPT-assisted descriptions. Cross-scale agreement was weak but nonrandom (ARI = 0.033; NMI = 0.194; p = 0.0065); however, this association became nonsignificant after feature-group equal weighting (p = 0.829). Tectonic type was associated with clustering at both scales. In a 144-image experiment, semantic prompts increased visual–prototype similarity by 0.0339 (p = 0.0215) and the own-prototype–strongest-competitor margin by 0.0306 (p = 0.0093). Top-1 identification rose from 13.9% to 44.4%, although the difference was nonsignificant (p = 0.0586). The contribution is a traceable visual–semantic process for exploratory generative steering, rather than a fixed taxonomy or stable class-specific control.