Compositional Visual Generation With Enhanced Language Guidance