Compositional Text-to-Image Generation with Dense Blob Representations