The agentic tooling in some of the latest models is getting pretty crazy too - somebody on Reddit recently shared an example of using GPT-6 Astra to generate sprites with Aseprite, an open-source, popular pixel art editor and it worked shockingly well.
Added support for GPT 2.5
The key is making them one frame at a time, rather than asking for the whole sheet. Adherence frame-to-frame with a reference image is really good, so just prompt with the previous + direction. I finally got Nano Banana to make fairly decent fluid animations that way.
I have abandoned the work I was doing, but if you provide the image constraints along with some examples and descriptions of those examples using the terminology of your constraints, you should get something iterable.
That said for single images the old model was already okayish for prototypes