Article: Design of an art style transfer teaching platform based on generative adversarial networks Journal: International Journal of Arts and Technology (IJART) 2026 Vol.16 No.7 pp.30 - 49 Abstract: This paper proposed generative adversarial network (GAN) with a shared latent space (sLS-GAN) to improve controllability and cultural adaptability in art style transfer. By integrating a variational autoencoder (VAE) with adversarial learning, it constructs a shared latent space that enables high-quality bidirectional translation between greyscale and colour image domains. Latent-space alignment improves realism and semantic coherence, while cycle-consistency regularises forward-backward mappings. A dual up-sampling/dual down-sampling design enhances structural stability across domains. In addition, a residual saliency network strengthens salient-region modelling and improves efficiency, reducing reliance on explicit content-preservation constraints. Experiments on the WikiArt Paintings and SemArt datasets show that sLS-GAN achieves an FID of 106.45 on WikiArt and outperforms representative baselines in Inception Score and PSNR, indicating improved semantic consistency, diversity, and perceptual quality. In greyscale colorisation, sLS-GAN reduces parameters by 89.5% and FLOPs by 87.8% versus conventional models, delivering substantial computational savings. Inderscience Publishers - linking academia, business and industry through research

Title: Design of an art style transfer teaching platform based on generative adversarial networks

Authors: Henan Zeng; Yu Shu

Addresses: College of Art and Design, Maanshan University, No. 8 Huangchi Road, Dangtu County, Maanshan, 243100, Anhui, China ' College of Electronic Commerce, Anhui Business College, Wuhu, 241003, Anhui, China

Abstract: This paper proposed generative adversarial network (GAN) with a shared latent space (sLS-GAN) to improve controllability and cultural adaptability in art style transfer. By integrating a variational autoencoder (VAE) with adversarial learning, it constructs a shared latent space that enables high-quality bidirectional translation between greyscale and colour image domains. Latent-space alignment improves realism and semantic coherence, while cycle-consistency regularises forward-backward mappings. A dual up-sampling/dual down-sampling design enhances structural stability across domains. In addition, a residual saliency network strengthens salient-region modelling and improves efficiency, reducing reliance on explicit content-preservation constraints. Experiments on the WikiArt Paintings and SemArt datasets show that sLS-GAN achieves an FID of 106.45 on WikiArt and outperforms representative baselines in Inception Score and PSNR, indicating improved semantic consistency, diversity, and perceptual quality. In greyscale colorisation, sLS-GAN reduces parameters by 89.5% and FLOPs by 87.8% versus conventional models, delivering substantial computational savings.

Keywords: generative adversarial network; GAN; greyscale-colour image translation; bidirectional variational autoencoder; cycle consistency; residual saliency network.

DOI: 10.1504/IJART.2026.154950

International Journal of Arts and Technology, 2026 Vol.16 No.7, pp.30 - 49

Received: 24 Dec 2025
Accepted: 07 Mar 2026
Published online: 20 Jul 2026 *

Title: Design of an art style transfer teaching platform based on generative adversarial networks

Keep up-to-date