It is a challenge for these models to generate images of counterintuitive or unusual situations that aren't depicted in the training set. For example, if you ask for a small cube sitting on top of a large cube, you'll likely get the correct result on the first attempt. Ask for a large cube on a small cube and you'll probably get an image of them side-by-side or with the small cube on top instead. The models can generalize in impressive ways, but it's still limited.