Title: Structural Preservation in Diffusion-Based Architectural Style Transfer: A Quantitative Study of the Structure-Style Trade-off
Authors: Ho Tran Chau Thao, Nguyen Minh Duc
Volume: 10
Issue: 8
Pages: 257-265
Publication Date: 2026/08/28
Abstract:
Architectural style transfer must alter appearance while preserving geometry, yet generative models often bend columns and misalign window grids, producing outputs unusable for design work. This paper presents ArchStyleAI, a Stable Diffusion v1.5 pipeline combining ControlNet with M-LSD line maps and a LoRA style adapter an approach already used by Kuang et al. (2023), but previously evaluated only through visual inspection or realism metrics (CLIP Score, FID) that do not measure geometric fidelity to the source image. The contribution is a measurement: a line-map SSIM computed between M-LSD representations of source and output. Sweeping Img2Img Strength (S) at 0.60, 0.75, and 0.90 on 50 Gothic facades, CLIP Score rose from 29.30 to 30.05 and LPIPS from 0.4256 to 0.5149, while line-map SSIM fell only from 0.7562 to 0.7479 - a narrow range attributable to the sparsity of M-LSD maps, which lets LPIPS track visible degradation more sensitively. S = 0.75 is therefore recommended as a working default rather than an established optimum, since the differences between settings were not tested for significance. All configurations are reported for replication; because evaluation images overlapped with the training set, findings describe parameter sensitivity rather than generalisation to unseen buildings.