SagaGAN: Style Applied using Gram matrix Attribution based on StarGAN v2

Abstract

Image-to-image translation aims to convert an image from one domain to another while preserving its content. AdaIN (Adaptive Instance Normalization) is a widely used style application method, but it may not fully capture the fine-grained visual characteristics of complex styles. We propose SagaGAN, a novel approach that combines the gram matrix with AdaIN to better capture and transfer style information. We introduce two loss functions: G1 loss and G2 loss, which focus on the differences between gram matrices of the style, generated, and input images. These losses enable SagaGAN to learn richer style information. Additionally, we incorporate a perceptual loss alongside the cycle consistency loss to maintain a balance between style application and content preservation. Experimental results demonstrate that SagaGAN effectively applies style information, leading to improved image generation performance compared to existing models. By leveraging the gram matrix to capture complex style characteristics while preserving content, SagaGAN enhances the style transfer capabilities of models like StarGAN v2.