Make It Count: Text-to-Image Generation with an Accurate Number of Objects