Abstract
Neural style transfer (NST), where an input image is rendered in the style ofanother image, has been a topic of considerable progress in recent years.Research over that time has been dominated by transferring aspects of color andtexture, yet these factors are only one component of style. Other factors ofstyle include composition, the projection system used, and the way in whichartists warp and bend objects. Our contribution is to introduce a neuralarchitecture that supports transfer of geometric style. Unlike recent work inthis area, we are unique in being general in that we are not restricted bysemantic content. This new architecture runs prior to a network that transferstexture style, enabling us to transfer texture to a warped image. This form ofnetwork supports a second novelty: we extend the NST input paradigm. Users caninput content/style pair as is common, or they can chose to input acontent/texture-style/geometry-style triple. This three image input paradigmdivides style into two parts and so provides significantly greater versatilityto the output we can produce. We provide user studies that show the quality ofour output, and quantify the importance of geometric style transfer to stylerecognition by humans.