Typological Transcoding Through LoRA and Diffusion Models: A Methodological Framework for Stylistic Emulation of Eclectic Facades in Krakow
Year: 2025
Type: article
Abstract: Stylistic emulation of historical building facades, especially for complex and data - scarce styles such as Krakow's Eclecticism, poses significant challenges for AI. This study develops a methodological framework for a deeper, typologically - informed "transcoding" of style beyond mere visual mimicry, which is crucial for heritage preservation, urban renewal, and digital heritage. We integrate architectural typology with Low - Rank Adaptation (LoRA) fine - tuning of a Stable Diffusion model, specifically the FLUX architecture. The process includes typology - guided image dataset preparation (selection and hierarchical labeling) and precise control of LoRA training parameters like learning rate and loss value. The typologically - guided LoRA - tuned model significantly outperforms baseline models in quantitative metrics—FID, LPIPS, and PSNR—and in expert qualitative evaluations regarding realism and stylistic accuracy. LoRA's loss value and application weights effectively balance creative variation and faithful style emulation. This research validates the synergy between architectural typology and AI for sophisticated stylistic emulation, positioning AI as an intelligent creative partner for nuanced reinterpretation. However, challenges in achieving deeper semantic understanding and 3D modeling persist.
Source: Buildings
Authors: Zequn Chen, Nan Zhang, Chaoran Xu, Zhiyu Xu, Songjiang Han, Lishan Jiang
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $3,137
OpenAlex ID: W4411782460