How AI manga translation works

From raw pages to translated manga in seconds.

Step 1

Upload your pages

Drag and drop PNG, JPG, or WebP images -- or upload a ZIP archive. No file size surprises: we accept up to 20 MB per page.

Upload individual pages or a full chapter as a ZIP archive. The translator accepts PNG, JPG, and WebP formats at any resolution. For best OCR results, use the original scan quality -- avoid heavily compressed JPGs. ZIP uploads process all pages in order, maintaining chapter continuity for the glossary.

Step 2

OCR & text detection

Our GPU-powered detector (CTD) scans every page and identifies text bubbles, sound effects, and narration boxes with sub-pixel accuracy.

The Comic Text Detector (CTD) is a deep learning model trained specifically on manga, manhwa, and manhua layouts. It identifies speech bubbles, narration boxes, free-floating text, and sound effects -- including vertical Japanese text, open-background Korean text, and dense Chinese dialogue. Each detected region is passed to the OCR engine, which reads the source characters with support for Japanese, simplified Chinese, and traditional Chinese.

Step 3

Translate & inpaint

A large language model translates the OCR text into your target language. Meanwhile, LaMa Large inpaints (erases) the original text so the page is clean.

Translation uses a premium LLM (currently Google Gemini 2.5 Flash) that receives all text regions on the page at once. This full-page context approach produces natural dialogue that respects speaker relationships, honorifics, and narrative flow. Simultaneously, LaMa Large neural inpainting removes the original text and reconstructs the underlying art -- even in detailed backgrounds and patterned screentone. If you have a per-series glossary, your locked terms are injected into the LLM prompt to ensure consistent names and terminology.

Step 4

Typeset & download

The translation is rendered onto the page with matched fonts and bubble sizing. Download your finished pages individually or as a ZIP.

The typesetting engine places translated text onto the clean inpainted canvas with automatic font sizing to fit each bubble. The renderer uses a specialized manga-to-English engine that respects bubble boundaries and adjusts line breaks for readability. Download pages individually as PNG or get the full chapter as a ZIP. On Creator and Pro plans, you can review every bubble in the built-in editor before downloading -- adjusting wording, font size, and placement.

Performance and specs

Processing speed

Under a minute per page

The GPU pipeline translates and typesets a page in seconds, then a quality check re-checks every bubble, so most pages are ready in under a minute. Pages of a chapter are processed in parallel, so a typical 20-page chapter is usually done in a few minutes.

GPU infrastructure

NVIDIA A40 / RTX 4090

The pipeline runs on high-end NVIDIA GPUs (A40 or RTX 4090) for fast, consistent processing. GPU instances scale automatically to handle demand, so there is no queue during normal usage.

Supported formats

PNG, JPG, WebP, ZIP

Upload individual images in PNG, JPG, or WebP format (up to 20 MB per page), or a ZIP archive containing an entire chapter. Output is always PNG for maximum quality. ZIP uploads return a ZIP of all translated pages.

Supported languages

Source: Japanese, Chinese. Target languages below.

LanguageCodeStatus
EnglishENGLaunch
SpanishESPLaunch
FrenchFRALaunch
Portuguese (BR)PTBLaunch
GermanDEULaunch
RussianRUSLaunch
TurkishTRKLaunch
IndonesianINDLaunch
VietnameseVINLaunch
Chinese (Simplified)CHSLaunch
Chinese (Traditional)CHTLaunch
KoreanKORLaunch
ArabicARABeta

Who uses MangaGloss

Watch the pipeline in action

AI Manga Translator: Translate a Whole Raw Chapter in 1 Minute · Chapters & transcript

Manga OCR translator: how AI reads every bubble
Why AI translation still needs a human edit

All MangaGloss videos →

Common questions

How fast is the translation?
Most pages are ready in under a minute: the GPU pipeline (detection, translation, inpainting, typesetting) takes a few seconds, and a quality check then re-checks every bubble for overflowing text, missing lines and untranslated words. Pages of a chapter run in parallel, so a typical 20-page chapter is usually done in a few minutes.
Does the AI handle sound effects?
Speech bubble text and bubble-contained sound effects are detected, translated, and typeset automatically. Art-integrated SFX -- onomatopoeia drawn directly into the illustration -- are a known limitation across every manga translation tool and are on our development roadmap. The text detector can identify most common SFX, but heavily stylized or art-blended effects may be missed.
Can I edit the AI's output?
Yes. On Creator and Pro plans, the built-in editor lets you click any translated text bubble, edit the wording, adjust font size and placement, and re-render the page instantly -- all in your browser. This is the key difference between MangaGloss and other tools: you get an AI first draft plus a professional editing workflow, not a take-it-or-leave-it output.
What source languages are supported?
Japanese and Chinese (both simplified and traditional) are supported at launch. Korean source text is on our roadmap. The pipeline handles vertical and horizontal text, tight bubbles, open backgrounds, and long-strip webtoon formats across all source languages.
Is the free tier lower quality?
No. Every tier, including the free pages, uses the exact same GPU pipeline, premium LLM engine and neural inpainting. The only differences between tiers are page volume, the watermark on pages translated without an account, the built-in editor and the per-series glossary.

Ready to try it?

Upload your first pages and see the results in seconds.

Start translating
How AI Manga Translation Works | MangaGloss