CGANs Based User Preferred Photorealistic Re-stylization of Social Image

In social networks, it is important to re-stylize a randomly taken photo to exhibit a unique individual character. Previous stylization methods either respect to a motivation of improving perceptual quality or artistic style transfer, are neither personalized nor photorealistic. Besides, a strong constraint on scene consistency of reference image is always required, which is not easy to meet for a customized application. In this paper, we propose a customized photorealistic re-stylization method referred to a group of user favorite images with loose scene consistency. To better express user preferred style, reference images are selected from the perspective of photographer where image content and composition are jointly considered and weighed by user preference of light and color. To achieve high perceptual quality, we map image pixels and styles based on Conditional Generative Adversarial Networks. Comprehensive experiments verify our method could improve user preferred photo re-stylization and bring in less artificiality.