Text-Adaptive Generative Adversarial Networks: Manipulating Images with Natural Language

This paper addresses the problem of manipulating images using natural description. Our task aims to semantically modify visual attributes of object in an image according to the text describing the new visual. Although existing methods synthesize images having new attributes, do not fully preserve text-irrelevant contents of the original image. In paper, we propose the text-adaptive generative adversarial network (TAGAN) generate semantically manipulated images while preserving text-irrelevant. The key to our method is the text-adaptive discriminator that creates-level local discriminators according to input text to classify-grained attributes independently. With this discriminator, the generator to generate images where only regions that correspond to the given text modified. Experimental results show that our method outperforms existing on CUB and Oxford-102 datasets, and our results were mostly preferred a user study. Extensive analysis shows that our method is able to disentangle visual attributes and produce pleasing outputs.