Attribute-Guided Face Generation Using Conditional CycleGAN

摘要: We are interested in attribute-guided face generation: given a low-res input image, an attribute vector that can be extracted from high-res image (attribute image), our new method generates for the satisfies attributes. To address this problem, we condition CycleGAN and propose conditional CycleGAN, which is designed to (1) handle unpaired training data because low/high-res images may not necessarily align with each other, (2) allow easy control of appearance generated via demonstrate high-quality results on synthesize realistic easily controlled by user-supplied attributes (e.g., gender, makeup, hair color, eyeglasses). Using as identity produce corresponding incorporating verification network, network becomes identity-guided produces interesting transfer. three applications CycleGAN: identity-preserving superresolution, swapping, frontal generation, consistently show advantage method.