It's interesting that the prompts that would do badly in a Google image search also seem to be the ones that make poor prompts. Basically, it seems that rather than describing a scene, you have to try to give an analogy for some image(s) that it might have in its training set - which is why, I believe, "banana in the style of Andy Warhol" produces a much higher quality result than "Outline of prism on a black background in the middle of scene splits a beam of light coming from the left side into rainbow on the right side".