I imagine it to work with a combination of text location detection, traditional OCR and LLM based translation where each translated piece of text gets summarized and added to a running summary that is prepended to each new piece of text.

Interactive would mean that the user can edit and insert info about which character the text belongs to or whether it is just a general description, or give additional context or ask questions about the translation, alternative translations, to explain ambiguities, alter the tone and style etc.