Most of these stem from the way DALLE-2 is designed - in particular the composition, counting, negation and concept blending examples. In the paper they explain

Most of these stem from the way DALLE-2 is designed - in particular the composition, counting, negation and concept blending examples. In the paper they explain that these aspects were deliberately sacrificed in order to improve image quality and diversity, and for the ability to generate variations of an existing image. These aspects were much better in the prior version (GLIDE), which conditioned on text tokens instead of just a fixed length vector.