The lazy way: one prompt
Give one page to an AI that can edit images (ChatGPT, Gemini and the like) with this sentence. Good enough to get the gist for yourself.
Translate all the Japanese text on this manga page into English. Keep the original lettering style, colors and layout, leave everything else unchanged, and keep the original aspect ratio.
Mia's tips
- It redraws the whole page rather than editing the text. Zoom in and the line art shifts slightly; the proportions can stretch a little too.
- Tiny text outside the bubbles is where it slips most: it's small, blurry, and nothing points at who said it, so it guesses.
- Every page starts from zero, so a catchphrase can come out differently on each page. Paste in the character sheet below if you want it consistent.
The good way: split it up
Scanlation groups already work as a relay: translator, proofreader, cleaner, typesetter. Split the job the same way, use the best tool for each step, fix only the step that went wrong, and never touch the original art.
- Find the text: comic-text-detector finds the text blocks in comics, with a mask for every letter. It sometimes misses ruled narration boxes; the model picks those up in step three.
- Read it (OCR): manga-ocr is made for manga Japanese and handles vertical and hand-lettered text. No GPU needed.
- Read the whole chapter, write a character sheet, translate: give the model every page plus the OCR text, with the prompt below.
- Clean: inside bubbles, just paint over the letter strokes with the bubble color; for text drawn over art, LaMa repaints what was behind it.
- Letter: horizontal, all caps, comic-style. For hand-drawn sound effects, keep the original and add a small translation next to it.
Don't want to wire it up yourself? BallonsTranslator and manga-image-translator are ready-made open-source tools doing the same steps, and you can plug your own model and the prompt below into the translation step.
You are a scanlation group of one: translator and proofreader. This folder holds one complete chapter of a Japanese manga:
- pages/p001.png … the original pages, in order. Japanese manga reads right to left, top to bottom.
- boxes/p001.jpg … the same page, with every text block the program found marked by a red box and a blue number.
- ocr.json … for every block on every page: its number, position in pixels, and the Japanese the OCR read. The OCR misreads characters, drops furigana, merges two neighboring bubbles into one block, and misses some text.
Work in this order. Actually open the images at every step; don't rely on the OCR text alone.
1. Read from the first page to the last. Understand the plot, who's who, who says each line, and in what tone.
2. Write bible.md (character sheet + glossary):
- For each character: Japanese name, English name, role, and how they talk (what they call themselves and others, catchphrases, tone, how old they sound).
- Proper nouns, code names, organizations, places, recurring slogans: one translation each.
- Words with furigana (ruby): what the kanji say, how the ruby reads, which two meanings the author wants at once, and how to handle it in translation.
3. Proofread page by page, block by block: fix OCR errors; label each block's type (dialogue / thought / narration / sfx / sign: text in the art / aside: small text next to a bubble); name the speaker. Where the OCR merged two bubbles into one block, keep it as one block and separate the two lines with " / ".
4. Translate into English (en):
- Read like a published translation: natural, not word for word.
- Every character talks the way the character sheet says; a child should sound like a child (simple words, maybe clumsy sentences) but stay readable.
- Deliberate mistakes, mispronunciations, wordplay and puns: find something with the same effect in English. Don't "correct" the character.
- Sound effects become English sound effects, the way American comics letter them.
- Every name, term and catchphrase stays consistent across the whole chapter.
5. Any line you're unsure about and a human should look at: set flag to true and say why in note (two possible readings, a pun, depends on later chapters…).
Output translation.json:
{"pages":[{"page":"p001","blocks":[{"id":1,"type":"dialogue","speaker":"…","ja":"proofread Japanese","en":"…","flag":false,"note":""}],"missing":[{"where":"where the missed text is","ja":"…","en":"…"}]}]}
Every block in ocr.json gets exactly one entry (matching id): don't skip, merge or add blocks. Text the boxes missed goes into that page's missing list.
Character sheet template
The prompt above has the AI write one itself. You can also fill it in first; change one line and the whole chapter follows (say, keeping a catchphrase the way fans say it).
# Characters | Japanese | English | Role | How they talk (what they call themselves and others, catchphrases, tone, age) | |---|---|---|---| | アーニャ | Anya | girl, reads minds | calls herself "Anya," her dad "Papa"; short sentences, a little clumsy, always readable | | | | | | # Glossary | Japanese | English | Notes | |---|---|---| | わくわく | Waku waku! | catchphrase, same everywhere | | | | | # Furigana and wordplay | Where | Kanji | Reading | The two meanings | How to translate | |---|---|---|---|---| | | | | | |
Mia's tips
- I used OpenAI's GPT-6.1 Sol at xhigh reasoning, in Codex: the pages and OCR sit in one folder and it opens them page by page. Use a model that can see images. Feed it OCR text alone and you lose who's speaking and how.
- The most valuable output is the handful of lines it flags: a name said wrong on purpose, whether there's a pun. Checking those beats proofreading everything from scratch.
- It can get clever: on its first read-through it mistook a plain line for a pun and planned to invent one. Have it go back over the character sheet and the translation at the end; it caught itself.
- Furigana is the hard part of manga: the kanji say one thing, the reading says another. Give it its own section in the character sheet so the translation tries to keep both.
- Use manga you bought, or chapters a publisher shares for free. Read the result yourself; don't post it.
