Latent Style: multi-style image transfer via latent style coding and skip connection

Unsupervised multi-style image translation is an important and challenging study in the task of image translation. The translation relations between interrelated images should be analyzed from multiple angles as these relations are not merely unidirectional or based on a single factor. Multi-style image translation algorithms have recently emerged to establish a multifaceted relationship between coupled images and interpret their features, which can fully express the content and semantic information of these images. One key algorithm, the multimodal unsupervised image-to-image translation (MUNIT), achieves reasonable unsupervised translation, but its image style representation is random noise, which leads to suboptimal multi-style representation. In order to achieve better multi-style image translation, we propose an improved MUNIT scheme equipped with style coding, skip connection, and a self-attention mechanism. The proposed scheme pays more attention to image style coding as well as the global and detailed image information. Through extensive experimental comparisons with state-of-the-art methods on various image translation tasks, the advantages of this scheme are demonstrated qualitatively and quantitatively. The code and tutorials have already released at https://github.com/huawang123/LatentStyle.