An Analysis of Text-to-Image Synthesis

Abstract

The technique of generating an image from the given text description is called Text-to-image synthesis. Human beings can visualize and conceptualize images described by natural language elucidations. Which appears to be a simple task for the human brain is one of the foremost challenging problems in Natural Language Processing and Computer Vision. The digital world is overflooded with various types of images. But still getting an image of the required resolution, scale or mode is a distant dream. Capturing advanced medical images or satellite images with needed quality is very challenging due to the complexity involved. Text-to-image synthesis is a blessing for such fields. Some other fields like fashion designing, interior designing, etc. are the application areas of text-to-image synthesis. It is considered a hard problem compared to the reverse problem of captioning of images due to its multimodal nature and complexity involved in problem-solving. With the introduction of Generative Adversarial Networks (GANs), text-to-image synthesis has gained new heights of success and received much attention. Many image synthesis models have been proposed in the literature, where GANs play a pivotal role. This paper aims to give a glimpse of the revolutionary models in the field of text to image synthesis. We also discuss the evaluation metrics and datasets used and the various issues and the likely directions of future research.