It seems like a good pixel art model should be trained from scratch to emit high-quality low-resolution outputs, rather than trying to jerry-rig a generalist high-resolution model to unnaturally output tiny images (or even worse, a regular grid of exactly NxN color blocks).