I also read somewhere that this system has special logic added to it that judges how humans would aesthetically judge the final image, so in a way the impressive aesthetic qualities of these images isn't totally coincidental.
This makes sense, OpenAI has been experimenting with human-in-the-loop type refinement systems for GPT-3 (see InstructGPT https://openai.com/blog/instruction-following/). It would make sense that they would use something similar for Dall-E 2 as well.