> Current AI image models can't understand grid-based pixel art.

sounds like a good use case to fix this problem from the model layer. an image gen model that is trained to make pixel perfect art.