Abstract
This paper presents ImagineNet, a tool that uses a novel neural styletransfer model to enable end-users and app developers to restyle GUIs using animage of their choice. Former neural style transfer techniques are inadequatefor this application because they produce GUIs that are illegible and hencenonfunctional. We propose a neural solution by adding a new loss term to theoriginal formulation, which minimizes the squared error in the uncenteredcross-covariance of features from different levels in a CNN between the styleand output images. ImagineNet retains the details of GUIs, while transferringthe colors and textures of the art. We presented GUIs restyled with ImagineNetas well as other style transfer techniques to 50 evaluators and all preferredthose of ImagineNet. We show how ImagineNet can be used to restyle (1) thegraphical assets of an app, (2) an app with user-supplied content, and (3) anapp with dynamically generated GUIs.