Playground v3: Improving Text-to-Image Alignment with Deep-Fusion Large Language Models