Huh, very interesting, thank you. In my (limited) tests with GPT-4 vision (before 4o), it did really well with just the image, but it's good to know that the OCR can help.