Style-Specific Diffusion for Architectural Classification: Enhancing Performance Through Class-Adapted LoRA Augmentation

• Gustavo Henrique Maia Dias UFPB
• Thaís Gaudencio do Rêgo UFPB
• Yuri de Almeida Malheiros Barbosa UFPB

Resumo

This paper addresses the challenge of limited training data in architectural style classification by proposing a novel data augmentation framework based on latent diffusion models with class-specific Low-Rank Adapters (LoRA). We focus on Brazilian architectural styles as a case study of a broader class of visual classification problems that share key characteristics: subjective categorization with expert disagreement, ambiguous class boundaries, continuous rather than discrete feature spaces, and historically evolving categories. Our approach generates synthetic images preserving style-defining features while introducing meaningful variations to enhance classifier performance. Unlike traditional augmentation methods that operate at the pixel level, our diffusion-based approach captures underlying semantic distributions, making it particularly valuable for domains where class distinctions depend on subtle feature combinations rather than isolated elements. We evaluate multiple neural network architectures and augmentation strategies across five architectural styles (Baroque, Neoclassical, Eclectic, Neo-Gothic, and Modernist). Results demonstrate significant performance improvements, with combined generative and traditional augmentation yielding an average accuracy gain of 5.8%. Notably, single-layer architectures outperformed multi-layer counterparts when using augmented data, suggesting reduced overfitting. Style-specific improvements ranged from 0% for Modernist (already at 100% accuracy) to 10.5% for Baroque style. Our methodology provides a versatile framework applicable to numerous domains sharing these characteristics, including art style classification, historical artifact categorization, fashion trend analysis, and other fields where categories exist on continuums rather than as discrete entities.