dataset: 300 images at 768x768

recaptioned with qwen3.6 vl (still occasionally confused about the gender with Hal Fosters drawings)

will very rarely produce unprompted gibberish text, but can reproduce original comic book lettering on the plus side