I wonder if it's possible to train to read text encoded as one colored pixel per letter, or even per token.