Transforming Architectural Visualizations for Generative xR Renderings using Fine-tuned AI Models

This paper explores the transformative potential of extended reality (xR) and generative artificial intelligence (AI) in architecture visualization. For heightened application, the authors mainly focus on generating monoscopic 360 images that are transferable into xR devices. Investigating monoscopic 360 architectural renderings through generative AI powered by few-shot learning methods, the authors adapt different styles while reflecting on real-life streetscapes. This research presents a methodical strategy that incorporates supplementary training through the stages of data collection and preparation, fine-tuning of hyper-parameters, and training processes. The effectiveness of few-shot learning heavily depends on the availability of high-quality training data with consistent representation, emphasizing the importance of the Data Collection and Preparation stage. The authors employ the Low-rank Adaptation (LoRA) approach to facilitate few-shot learning for architectural visualization. The research underscores the considerable impact achievable through the integration of extended reality with Generative AI into the architecture and planning process, presenting a valuable and transformative tool for designing buildings that seamlessly incorporate existing physical settings with prompted aesthetic visualizations.