MATGAN: Unified GANs for Multimodal Attribute Transfer by Coarse-to-Fine Disentangling Representations

Abstract

Image attribute transfer aims to change an image to a target one with desired attributes. There are mainly two challenges for this task: multi-domain transfer and attribute-level multimodality. The first means editing multiple attributes using a single model and the second means diverse appearances for the target attribute. Existing methods cannot address the two problems simultaneously. Moreover, many works focus on image-level multimodality rather than attribute-level. In this paper, we propose a novel coarse-to-fine disentangling representation framework MATGAN to achieve Multimodal Attribute Transfer. In the coarse disentangling stage, we propose to embed images onto a content space and an attribute space for image-level multimodality. In the fine disentangling stage, we further disentangle the attribute space to bind with each attribute for attribute-level multimodal and multi-domain transfer. Extensive experiments demonstrate the effectiveness of our approach.