Lightweight Generative Adversarial Networks for Text-Guided Image Manipulation

We propose a novel lightweight generative adversarial network for efficient manipulation using natural language descriptions. To achieve this, a new-level discriminator is proposed, which provides the generator with-grained training feedback at word-level, to facilitate training a generator that has a small number of parameters, but can still focus on specific visual attributes of an image, and then edit them affecting other contents that are not described in the text. , thanks to the explicit training signal related to each word, the can also be simplified to have a lightweight structure. Compared the state of the art, our method has a much smaller number of parameters, still achieves a competitive manipulation performance. Extensive results demonstrate that our method can better disentangle visual attributes, then correctly map them to corresponding semantic, and thus achieve a more accurate image modification using natural descriptions.