Artistic Color Style Transformation in Paintings using Generative Adversarial Networks

Color style transfer in paintings is a style-constrained image transformation technique in digital painting that supports the mapping of color features from a source image to a target image without losing the original content and structure. Traditional methods—like histogram matching and rule-based techniques—are inflexible and tend to lose the semantic and stylistic richness that is intrinsic to artworks. Neural style transfer methods provide better performance but usually need large amounts of paired data and heavy computational power, reducing their applicability to a wide range of styles. For overcoming these issues, this paper introduces a new approach named SADA-GAN (Spatial and Channel Attention-based Dual-branch GAN), which incorporates spatial and channel attention mechanisms in the framework of a dual-branch encoder-decoder model. This model encodes content and style attributes separately and combines them adaptively for improved global consistency and localized detail in stylized results. The realism and perceptual quality is further enhanced with a multi-loss adversarial training approach. The experimental results indicate that the proposed model realizes an FID score of 8.1 and 96% user satisfaction, superior to conventional and baseline GAN models such as CycleGAN, AGGAN, and NST. These results confirm the effectiveness and usability of the model in creative digital uses, providing a robust solution to high-quality artistic color conversion.