Here's a somewhat more convoluted example:
The first image is what came out of the initial prompt. Then I tried to get it to improve things and to add animations; the second image is the end result (and yes, the sun is animated as described). Here's the complete chat log:
https://gist.github.com/int19h/18bb9d96b4d8418da2531513970d8...
It's obviously far from perfect, but the fact that it can do this much already - meaning that it understands broadly how the objects would relate to each other spatially for something like this - is already extremely impressive IMO, given that its training and inputs are entirely textual.