Progressive Compositionality In Text-to-Image Generative Models