Manga OCR
Manga OCR is how the text gets out of a manga page: the dialogue in its bubbles, the captions, the sound effects drawn into the art. Serifu reads the page with a vision model that sees all of it at once, shows what it read line by line, and translates it in the same run.
What is manga OCR?
OCR is optical character recognition: turning the text in an image into text a computer can use. Manga OCR is that job on a comic page, and it is harder than it sounds. The dialogue is drawn into speech bubbles, the narration sits in captions, the sound effects are part of the art, and Japanese and Chinese are set in vertical columns that run right to left. A translator that gets a bubble wrong gets its translation wrong too.
How Serifu reads a manga page
Four steps, on every page of a volume.
It finds every text region
A detector on your computer finds the bubbles, the captions and the sound effects; a segmentation step outlines each balloon, so one region never spans two of them. Nothing is typed in by hand and no box has to be drawn first.
It reads the page whole
The page and a crop of every region go to the vision model together, with the two pages before it, your glossary and what you told Serifu about the story. That context is what keeps a name, a pronoun or a number right from one page to the next, and what tells a sound effect from a line of dialogue.
It shows you what it read
Select a bubble and its source text is beside the translation, line by line. If a character was read wrong, correct the source and Re-translate this line; if a name should stay the same, add it to the glossary. The reading is not a black box you have to accept.
When it misses text
A sign in the background, a sound effect over the art, a bubble the detector did not find: Box, read and translate reads that one region and letters it in place. Mask tells Serifu to treat an area as blank, and Read and translate this page again reads the page from its image once more.
Serifu vs the manga-ocr model
manga-ocr is the open-source model most people find first when they search for manga OCR. We gave it and Serifu's default reader the same 24 crops from a Japanese manga, Shuho Sato’s Give My Regards to Black Jack vol. 1, and checked each reading by hand. A reading counts only if every character is right. Text drawn on the art is where they differ most. It is a small test on one volume, so it shows the kind of mistake each makes, not an error rate.
| Reader | All 24 crops | In bubbles (8) | On the art (16) |
|---|---|---|---|
| Serifu's default reader (gemini-3.8-flash) | 24 / 24 | 8 / 8 | 16 / 16 |
| manga-ocr (open source) | 17 / 24 | 7 / 8 | 10 / 16 |
Manga OCR in five languages
Serifu reads Japanese, Korean, Simplified Chinese, Traditional Chinese and English comics. There is no OCR model to choose and nothing to install: one reader, chosen and measured by us, handles all five, vertical columns included.
How to read a manga page with Serifu
- 01
Import the volume
A folder of pages, a CBZ, ZIP or PDF, or images; pick the pages you need. Serifu sorts them and shows the page order and the reading direction before anything runs.
- 02
Read and translate
Choose the translation model and start. Each page is read from its image and translated in one run, and the price is shown first: from 1 credit a page. The reading is part of that price, not a separate charge.
- 03
Check the lines that matter
Open a page and select a bubble: its source and translation are there. Fix a misread character and Re-translate this line, or leave the translation and keep reading. A line Serifu missed is a Box, read and translate away.
What manga OCR does not do
It does not export a page's text as a file: the source and the translation stay with the project, line by line. It does not read a page without translating it, and there is no OCR model to pick per language.
Manga typesettingManga OCR cost
Reading comes with Read and translate a page: 1 credit with gpt-6-luna, 1.5 with gemini-3.8-flash and 3 with claude-opus-5-5, the reading and the first translation together. Box, read and translate is 0.2 credit for the read plus the model's price for the line, asked first. 5 credits are free when you sign up, and credit packs never expire.
Manga OCR FAQ
Does Serifu do OCR on its own?
Reading is part of Read and translate: a page is read from its image and translated in one run, at the model's page price. There is no separate charge for the reading alone.
Which languages does the manga OCR read?
Japanese, Korean, Simplified Chinese, Traditional Chinese and English.
Does it read vertical Japanese?
Yes. Vertical Japanese and Chinese columns are read as columns, right to left, as they are drawn.
Can I export the text?
Not yet. The source text and its translation are in the editor, line by line, and stay with the project. Exporting them as a text file is not supported.
What if it misses a bubble or misreads a character?
Box, read and translate reads a region the detector missed, and letters it in place. A character read wrong is one edit away: correct the Source, then Re-translate this line. If the whole page was read badly, Read and translate this page again.
Is my page uploaded to Serifu?
The detector runs on your computer. A page is sent to be read and translated and is not kept on serifu.net; the volume stays on your computer.
Does it need a graphics card?
No. Reading runs on Serifu's servers: no model to download and no graphics card to own. The app is for macOS and Windows.
Can I do a whole volume?
Yes: Read and translate N pages runs the volume, and each page is charged when it starts.