This workflow is a one-shot comic translator, built to make fast, rough translations of comics that haven't been released by human translators yet.

It uses vision LLM via API to build a summary/translation of the page for context, then a YOLOv8 model to isolate each speech bubble, then a smaller LLM to script the individual bubbles, and finally Flux2 Klein to replace the text. The modified text is masked back to the original page after an upscale to make it nice and clean.

It caches intermediate results to text/image files, both so that it can be restarted mid-run without too much penalty, and so that you can correct egregious mistranslations or retry garbled text.

Of course, even when working one speech bubble at a time, Flux2 still produces garbled text sometimes, and the translation produced by LLMs is really not on the level of a skilled human translator. But I find personal use of this script to be a nice way to enjoy comics that are too obscure to see even a fan translation uploaded.

Other language pairs besides Japanese-to-English may work by modifying/translating the prompts, but I haven't tried it.