SAM-GAN: A Single-Image GAN Model Guided by Synergistic Attention

In real life, there is often an insufficient number of images available for the model to learn from in image generation tasks. Previous work shows that learning the distribution of patches within a single image can possibly gain powerful prior. But training a single-image generative model is still a tough mission since the features of an image patch are considerably less detailed and abstract compared to a distribution of images. That's why the existing methods are relatively rare, such as SinGAN, ConSinGAN, and CCASinGAN. This paper proposes a single-image GAN model (SAM-GAN) guided by a synergistic attention mechanism which consists of low-dimensional and high-dimensional blocks focusing on both global structural and detailed texture features in the meantime. We conduct numerous experiments to evaluate the model's performance qualitatively and quantitatively. The outcomes demonstrate SAM-GAN's superiority over other state-of-the-art models, showcasing its potential to generate more intricate and realistic images.