GAN-based Manchu paper-cutting style transfer with augmented reality interactive visualization

Manchu paper-cutting, a distinctive intangible cultural heritage of Northeast China, is at increasing risk of decline as skilled artisans age and traditional transmission channels erode. This paper proposes an integrated framework that couples an improved Generative Adversarial Network (GAN) with an augmented reality (AR) interactive visualization system for Manchu paper-cutting style transfer and cultural dissemination. A dedicated dataset of 2,136 annotated images spanning six motif categories was constructed, with copyright provenance recorded and cultural authenticity verified by Manchu paper-cutting inheritors. The generator combines multi-scale residual modules, a self-attention layer and a symmetry-consistency loss to jointly model fine incisions and long-range bilateral motifs, while a spectrally normalized multi-scale discriminator enforces texture authenticity. A multi-objective loss function that unites adversarial, perceptual, style-matching and edge-preservation terms sharpens the crisp contours central to paper-cutting aesthetics. Against seven baselines that now include CycleGAN, Pix2Pix, AdaIN, MUNIT, a diffusion-based transfer (InST), a transformer-based transfer (StyTr 2 ) and the paper-cutting-specific CutGAN, the proposed model attains an FID of 54.37 and the best FSIM, LPIPS and SSIM scores. A seven-expert cultural-authenticity panel further rates our outputs highest across five heritage-specific dimensions. The AR system, built on Unity, ARCore and MediaPipe with gesture-based interaction, is evaluated against a non-AR baseline in a within-subject study (n = 40) and significantly outperforms it on immersion, cultural perception and two-week knowledge retention (all p < 0.001, Wilcoxon signed-rank). Code, pretrained weights and evaluation scripts are deposited in a public archive to support reproducibility.