Enhance Image Style Transfer with Depth Spatial Modulation

Yang Peng, Peng Lu, Wang Yin

Abstract: Style transfer technology is an image editing technique that creates new images by blending the content features of content reference images and the style features of style reference images. In past style transfer algorithms, the transfer results often failed to strictly follow the guidelines given by the content and style references, resulting in discrepancies between the contours of the generated objects and the content references, as well as instances where the colors within the objects bled out. In this work, we attempt to augment traditional style transfer algorithms with a structural branch, filling in structural information using spatial modulation. As this approach does not disrupt the existing architecture of the generative network, it maintains a degree of universality. Experimental results suggest that our proposed method can enhance the transfer effectiveness of existing style transfer networks in terms of authenticity, content, and style.