CRD-CGAN: category-consistent and relativistic constraints for diverse text-to-image generation