AI-Based Pre-Renewal Design for Historic Building Facades: An AIGC–LoRA Framework with Collaborative Assessment

Historic conservation areas face the challenge of balancing heritage preservation with modern adaptation, often resulting in irreversible risks. AIGC technology offers an effective solution to mitigate these renewal risks. Current methods struggle with three bottlenecks: a lack of high-quality datasets, difficulty integrating expert and public preferences, and generating diverse proposals under complex preservation rules. This study proposes an AI-driven pre-renewal framework for building facades, which involves (1) virtual pre-renewal design using a large language model (LLM) to generate facade proposals based on “non-change,” “permissible,” and “prohibited” rules; (2) a multi-stakeholder evaluation system integrating expert and public judgments via the Bradley–Terry model; and (3) LoRA fine-tuning of Stable Diffusion XL to optimize facade generation. In the case study of the Shangxijie and Xiaxijie Historic Conservation Area of Jincheng Ancient Town, the framework was implemented in three stages. First, LLM-generated facades addressed data scarcity by adhering to preservation constraints. Second, an online platform integrated expert and public evaluations to refine the training dataset. Finally, LoRA fine-tuning improved the model’s contextual fidelity and stylistic coherence. Quantitative analysis showed that LoRA models outperformed the base model in authenticity and fidelity. Historic models achieved the highest fidelity (FID = 23.4, SSIM = 0.918, CLIPScore = 0.842), Style-Coordinated models performed stably (composite score = 0.82 ± 0.05, SSIM = 0.884), and Style-Incompatible models showed greater variability (mean = 0.78, SD = 0.09). The expert–public collaborative mechanism validated the iterative “generate–evaluate–refine” workflow as a sustainable approach for heritage facade renewal.