Domain-adaptive generative data augmentation for building façade classification: Linking synthetic data to performance across urban contexts

The limited availability of labeled data remains a major barrier to deploying computer vision models for building façade material classification. This study proposes a domain-adaptive generative augmentation framework that integrates architecture-aware prompt design with LoRA-based fine-tuning of a diffusion model. The framework generates synthetic façade images that preserve material-specific textures and region-dependent visual patterns. Unlike prior studies that mainly assess visual realism, this study links generative augmentation strategies directly to downstream classification performance. Experiments examined the effects of generation hyperparameters and LoRA training data volume, and the framework was validated using geographically distinct datasets from Korea and the England. Compared with conventional augmentation, the proposed method improved classification accuracy from 85.00% to 93.67% in Korea and from 71.67% to 83.67% in England. The results demonstrate that domain-adaptive generative augmentation reduces the synthetic-real data gap and improves practical classification robustness under data-scarce conditions for urban-scale built environment analysis.