Image Synthesis with Aesthetics-Aware Generative Adversarial Network.

With the advance of Generative Adversarial Networks (GANs), image generation has achieved rapid development. Nevertheless, the synthetic images produced by the existing GANs are still not visually plausible in terms of semantics and aesthetics. To address this issue, we propose a novel GAN model that is both aware of visual aesthetics and content semantics. Specifically, we add two types of loss functions. The first one is the aesthetics loss function, which tries to maximize the visual aesthetics of an image. The second one is the visual content loss function, which minimizes the similarity between the generated images and real images in terms of high-level visual contents. In experiments, we validate our method on two standard benchmark datasets. Qualitative and quantitative results demonstrate the effectiveness of the two loss functions.