Manga OCR: Setup, Models & Translation

Manga OCR is the job of reading the text inside manga images. Most people searching for it mean one project in particular: manga-ocr, a free Python library by kha-white that reads Japanese from a cropped speech bubble and returns plain text. It does not find bubbles, translate lines, or put text back on the page.

So the right manga OCR setup depends on what you want at the end. For a Japanese sentence to look up, install manga-ocr. For a page you can read in English, you need OCR plus translation and lettering, which is what an OCR manga translator does in one step.

What is manga OCR?

OCR stands for optical character recognition: turning pixels into text you can copy. Manga OCR applies that to comic pages, where the text sits in bubbles, runs vertically, and overlaps artwork.

The names get confusing because one project appears under several spellings:

• manga-ocr: the GitHub repository and the package you install from PyPI.
• manga_ocr: the command you run in a terminal and the Python module you import. The class is MangaOcr.
• manga-ocr-base: the trained model the library downloads from Hugging Face, published as kha-white/manga-ocr-base.
• mangaocr: the one-word label many apps and searches use for the same engine.

When a tool's settings list "Manga OCR" as an engine, it usually means this model or a converted copy of it. Generic "manga OCR" on a marketing page may mean something else entirely, so check which engine a product actually names.

Why manga needs its own OCR

Document OCR expects tidy horizontal lines on a plain background. Manga rarely offers that. The manga-ocr project was built around five problems that show up on nearly every page:

• Vertical and horizontal text, often on the same page.
• Furigana, the tiny reading aids printed beside kanji.
• Text overlaid on images, not on clean white space.
• A wide variety of fonts and styles, from book type to shouted lettering.
• Low-quality images, such as old scans and compressed screenshots.

It also reads multi-line text in a single pass, so a whole bubble can go in at once without being split into lines. It works as a general printed Japanese OCR too, and the README suggests it should do a decent job on novels and video games. That makes it a useful answer to "image to text Japanese" when the image is a small, tight crop.

How the manga-ocr-base model works

The manga-ocr-base model is an end-to-end vision encoder-decoder. A vision transformer looks at the image, and a BERT-style decoder writes out Japanese text token by token. It was trained on the Manga109-s dataset plus synthetic data.

The input size explains the most common manga OCR mistake: passing a full page. Here is the same page resized two ways, using the model's own preprocessing settings.

Full page at model input size: dialogue turns to blur

Bubble crop at the same size: every character survives

That is why the library has no "upload a page" mode. Something has to find the text regions first, whether that is your hand with a screenshot tool or a detector such as comic-text-detector.

How to install and run manga-ocr

This manga OCR tutorial follows the official README. You need Python 3.9 or newer. The very newest Python release may not work yet, because PyTorch often takes time to catch up.

1. Install the package

pip install manga-ocr

If you use uv, run uv add manga-ocr, or try it without adding it to a project:

uvx manga_ocr

For manga-ocr GPU use, install PyTorch from its official instructions first. The library picks CUDA or Apple MPS when available and otherwise falls back to the CPU.

2. Read images in Python

from manga_ocr import MangaOcr

mocr = MangaOcr()
text = mocr('/path/to/bubble.png')
print(text)

You can pass a file path or a PIL image. The first run downloads the model, which may take a few minutes. After that, recognition happens on your own machine.

3. Or run it in the background while reading

This is the setup many Japanese learners use. Start the clipboard watcher:

manga_ocr

Capture a region with a tool like ShareX or Flameshot. manga-ocr reads the image from your clipboard and writes the text back, where a dictionary such as Yomitan can pick it up. Wait for the OCR ready message in the log before you start.

To watch a screenshot folder instead:

manga_ocr "/path/to/sharex/screenshot/folder"

Clipboard mode replaces any image you copy with recognized text, so use folder mode if you also need to paste images normally.

Common install problems

• DLL load failed while importing fugashi: this might come from the Microsoft Store build of Python. Try installing from python.org.
• manga_ocr command not found: try python -m manga_ocr.
• mecab-python3 fails on ARM: see the workaround linked in the repository's troubleshooting notes.
• Clipboard mode on Linux: install wl-copy for Wayland or xclip for X11.

Which manga OCR tool fits your job?

The right tool depends less on the engine and more on what you do with the output. There is no official hosted manga OCR online service from the project, and Hugging Face listed no inference provider for the model when we checked.

mokuro is the natural next step for readers. It combines comic-text-detector with manga-ocr, then lets you read in a browser with selectable text. If a forum post recommends an OCR manga reader for study, it is often this.

For manga OCR ONNX builds, community exports exist on Hugging Face for runtimes without PyTorch. Check the model card before you rely on one: some exports build on a retrained manga-ocr-base-2025, whose own card was still an empty template when we looked.

What OCR model does MangaTranslator use?

If you searched for the MangaTranslator OCR model, you probably mean one of several open-source apps. Most of them no longer depend on a single manga OCR engine, as this snapshot from their public READMEs shows.

The takeaway for choosing a manga OCR model: use manga-ocr for tight Japanese crops, and look at PaddleOCR-VL or a multi-language engine when you need several languages or whole-page reading. These projects change quickly, so confirm against the current README before committing.

From manga OCR to translation

OCR is one stage in manga OCR translation. Getting from a raw page to a readable one involves several jobs, and manga-ocr covers exactly one of them.

You can assemble this yourself: OCR each crop, paste the text into a translator, then erase and re-letter in an image editor. That gives you full control, and it is worthwhile for a release you plan to polish. It is slow if all you want is to read the page.

An OCR manga translator runs those stages for you. Upload a complete page to our single-page manga translator, choose your language, and it detects the text, translates it, cleans the original lettering, and returns the page as an image. It accepts PNG, JPG, JPEG or WebP up to 15 MB, with Japanese, Korean or Chinese source pages.

Source page: Japanese dialogue

Reading copy: English dialogue

Pick the workflow by source: use batch translation for up to 20 page images, the CBZ translator for a CBZ or ZIP chapter, or the browser extension for images on a page you are reading. For raw Japanese specifically, the Japanese manga translator guide covers furigana, honorifics and dropped subjects.

Manga OCR limits and how to check results

The most important manga OCR limit is easy to miss. The README warns that the model always tries to recognize something, even on an image with no text, and may "dream up" realistic sentences. Plausible Japanese is not proof of a correct read.

One independent test shows where it breaks. A Rust port of manga-ocr published its results on a real page: 4 of 9 bubbles passed. The failures were a tiny caption, slanted action text, shouted lettering, a neighboring line leaking in, and one wrong character. That is a single page on a port, not a benchmark, and the longest bubble on that page passed, so length alone does not predict failure. The README adds its own cautions: longer text raises the chance of errors, and handwriting probably won't work.

Be careful with published accuracy tables as well. Some manga OCR articles quote figures above 99%, usually with no sample or method. Scores depend on which pages you test, so try your own pages instead.

When you check manga OCR output, look at:

• Crop edges: tighten the crop if text from a neighboring bubble appears.
• Long bubbles: split them and run each part again.
• Similar characters: compare katakana such as ン and ア, or シ and ツ, against the image.
• Names: check readings in context, since the model sees no surrounding pages.
• Empty regions: treat any output from a blank or artwork-only crop as invented.
• Sound effects: read them yourself, as they are often drawn as art rather than lettered.

Translated pages need the same care. A smooth sentence can still come from a misread character. Our guide to translating Japanese manga images lists what to compare before downloading.

Is manga OCR free?

Yes, for the open-source route. manga-ocr is free under Apache-2.0, and once the model is downloaded, recognition runs on your computer, so your pages stay local. What you pay is setup time, disk space and patience on slower hardware.

A free OCR manga translator is a different matter, because translation and image cleanup cost money to run. On AI Manga Translator, starting a translation requires sign-in and uses one credit per page, and failed tasks return their credit. Any trial credits on your account let you test a page; downloading results requires a purchased credit pack. See Pricing for current packs.

Uploading a page to any online service sends it to that service, so read the privacy policy first, and only process pages you own or have permission to translate.

Frequently asked questions

Choose your next step

If you want the Japanese text, install manga-ocr and crop tightly. If you want to read the page, skip the setup: start with one dialogue-heavy page and compare the result with the original.

Translate one manga page →