Abstract
Disentangling content and style information of an image has played animportant role in recent success in image translation. In this setting, how toinject given style into an input image containing its own content is animportant issue, but existing methods followed relatively simple approaches,leaving room for improvement especially when incorporating significant stylechanges. In response, we propose an advanced normalization technique based onadaptive convolution (AdaCoN), in order to properly impose style informationinto the content of an input image. In detail, after locally standardizing thecontent representation in a channel-wise manner, AdaCoN performs adaptiveconvolution where the convolution filter weights are dynamically estimatedusing the encoded style representation. The flexibility of AdaCoN can handlecomplicated image translation tasks involving significant style changes. Ourqualitative and quantitative experiments demonstrate the superiority of ourproposed method against various existing approaches that inject the style intothe content.