<jats:p>Historic district preservation faces major challenges in reconstructing destroyed architectural buildings while maintaining environmental and artistic balance, especially with incomplete records. This study introduces a computer vision–based generative adversarial network (CV-GAN) approach for reconstructing historic district artistic styles. The proposed method employs an enhanced StyleGAN2 framework with direct feature extraction modules and environmental coherence loss to ensure stylistic consistency. A pre-processing pipeline extracts multi-scale architectural features via convolutional neural network-based structures, followed by adversarial synthesis. Experiments conducted on a dataset of 11 763 images across five architectural categories demonstrate significant improvements over existing methods. CV-GAN achieves a Composite Reconstruction Score of 82.4%, outperforming StyleGAN2 (79.2%) and Pix2Pix (73.4%), while reducing the Fréchet inception distance score to 40.82 from 42.85. Whole-block reconstruction success reaches 73%, exceeding baseline approaches by 32%. Style-specific performance includes Neoclassical (81.7%), Victorian (86.3%), Gothic Revival (61.2%), and Georgian (48.2%) styles. The computation cost averages 5.8 min per building with 8.7GB memory, or 42.6GB at district level. The results indicate that degraded facades can be realistically recovered from limited photographic records, providing a feasible solution to the ‘reconstruction truth against aesthetic truth’ dilemma and offering cultural heritage practitioners an effective tool for architectural restoration in urban contexts.</jats:p>