I highly recommend Hades V1 for pixel art, Plant Milk/Minij for usual generations. Ive linked them below.

I have been experimenting behind curtains. Some bad models, some private trainings. One of experiments was to train a LoRA directly with pixel art model(Hades V1). Now, my DAKU dataset is only 18 preprocessed images of x8 pixel size and has something like 400x400px portraits. Either that or the pixel checkpoint resulted in this interesting thing - the images are not actually pixel art, but with x1 LoRA strength mostly seem like a totally different style, only partially looking like initial dataset.
The results kinda neat, so why not release it. I will be experimenting with that more and try researching, but all of this is buzz-consuming. Will take some time.