Exploring Optimized Generation Methods for Post-War Cityscapes Restoration Based on Stable Diffusion Model

Nowadays, frequent local wars have inflicted severe damage on urban built environments, presenting substantial challenges for post-war restoration. Moreover, the scarcity of architectural imagery further exacerbates these challenges. In this context, virtual restoration techniques have shown significant advantages in speed and accuracy over traditional experience-based methods. This paper aims to explore the potential of artificial intelligence in the restoration of architectural ruins and the generation of visual predictions. Specifically, we compared the performance of pix2pix GAN and Stable Diffusion Models in architectural restoration, then further applied Stable Diffusion Models based on a modern style to the entire post-war restoration process spanning time. Notably, the optimization of its U-NET module through rule-enhanced learning and the precise mapping of image features through ControlNet improved the accuracy and coherence of restoration. Experimental findings indicate that Stable Diffusion Model surpasses traditional machine learning approaches in preserving architectural characteristics and styles, effectively addressing the issues of paired training data scarcity and minor facade feature dissipation, while astutely retaining selective elements indicative of war-induced architectural damage and aging. © The Author(s) 2025