Well, it's technically PETSCII art, but the model can be trained on any bitmap charset. So you can make ATARIART or SPECCYART or whatever.

It's not correct to say that most of the information is in the color, though.

I can't remember if it's explained in the readme or not, but what makes this different from other converters is that it matches 1bit 8x8 blocks from the source image to the best matching character, increasing the perceived resolution of the output image.

It's a dumb idea I wanted to implement to amuse myself, and it's not better or worse than any similar tools, but AFAIK it uses a novel approach: fast by using a custom trained NN and uniquely looking output, similar to what a human artist might create by hand.

For no particular reason I will share with you my favorite by an unknown artist.

____\o/______/\__

It's called Help shark.