Handling image and other visual entities is hard. While going from text (and extracting some understanding of that text) to an image is much easier than the reverse.