Text-to-Image Generation via Semi-Supervised Training

Synthesizing images from text is an important problem and has various applications. Most of the existing studies of text-to-image generation utilize supervised methods and rely on a fully-labeled dataset, but detailed and accurate descriptions of images are onerous to obtain. In this paper, we introduce a simple but effective semi-supervised approach that considers the feature of unlabeled images as "Pseudo Text Feature". Therefore, the unlabeled data can participate in the following training process. To achieve this, we design a Modality-invariant Semantic- consistent Module which aims to make the image feature and the text feature indistinguishable and maintain their semantic information. Extensive qualitative and quantitative experiments on MNIST and Oxford-102 flower datasets demonstrate the effectiveness of our semi-supervised method in comparison to supervised ones. We also show that the proposed method can be easily plugged into other visual generation models such as image translation and performs well.