Abstract: Recent breakthroughs in text-to-image synthesis have been driven by diffusion models trained on billions of image-text pairs.