Learn, Imagine and Create: Text-to-Image Generation from Prior Knowledge