A Transformer Guided Generative Adversarial Network (TG-GAN) for Style Transfer in Artistic and Natural Scene Images

The artistic style transfer is a technique to produce aesthetically pleasing images by merging the semantic content of one domain with the stylistic attributes of another.Most attentionbased GAN structures have been unable to achieve structural consistency and style fidelity across heterogeneous datasets.This paper presents a Transformer-Guided Generative Adversarial Network (TG-GAN) that incorporates multi-head self-attention into the generator to improve cross-domain feature alignment while ensuring perceptual realism.The role of the transformer is to flexibly align content-style relations using a novel adaptive token fusion approach, guided by a perceptual-adversarial optimization process.Results based on qualitative and quantitative evaluations on MS-COCO to WikiArt and MS-COCO to Flickr Landscapes demonstrate that TG-GAN achieves superior results over both StyTr² and DualStyleGAN in terms of structural integrity and stylization quality.The models proposed achieved an SSIM of 0.803, an FID of 24.5, and a Style Classification Accuracy of 91.3%, which is better than existing transformer-based GAN frameworks.The framework provides a promising pathway for scalable cross-domain and multimodal style transfer while also offering additional perspectives for integrating transformer architectures with generative adversarial learning.