Abstract
Recall that most of the current image style transfer methods require the user to give an image of a particular style and then extract that styling feature and texture to generate the style of an image, but there are still some problems: the user may not have a reference style image, or it may be difficult to summarise the desired style in mind with just one image. The recently proposed CLIPstyler has solved this problem, which is able to perform style transfer based only on the provided description of the style image. Although CLIPstyler can achieve good performance when landscapes or portraits appear alone, it can blur the people and lose the original semantics when people and landscapes coexist. Based on these issues, we demonstrate a novel framework that uses a pre-trained CLIP text-image embedding model and guides image style transfer through an FCN semantic segmentation network. Specifically, we solve the portrait over-styling problem for both selfies and real-world landscape with human subjects photos, enhance the contrast between the effect of style transfer in portrait and landscape, and make the degree of image style transfer in different semantic parts fully controllable. Our Generative Artisan resolve the failure case of CLIPstyler and yield both qualitative and quantitative methods to prove ours have much better results than CLIPstyler in both selfies and real-world landscape with human subjects photos. This improvement makes it possible to commercialize our framework for business scenarios such as retouching graphics software.
Abstract (translated)
URL
https://arxiv.org/abs/2207.11598
