I'm surprised Stable Diffusion and others haven't cracked legible text by now. OCR has been solved for decades now, surely they could include some OCR term in the loss function while training it?