Waiting weeks for a scanlation group to pick up the next chapter is the normal manga experience. This post walks through the practical workflow for translating a whole chapter yourself — what to prepare, where AI helps, and where it still needs a human check.
1. Start from a clean chapter file
The easiest container is a CBZ: it is just a ZIP archive with image files inside. Rename it to .zip and you can inspect every page. Keep JPG, PNG or WebP, one file per page.
Two rules save a lot of pain later:
- Check the page order. Some scanners export 001, 002, ... 010, 011 — fine. Others use 1, 10, 11, 2, which sorts wrong. Natural sort is the first thing any translator should do.
- Keep double-page spreads as one image. Splitting a spread splits speech bubbles across the gutter and produces nonsense.
If you want the longer explanation of the format, I wrote a short guide here: what a CBZ file is.
2. Choose between overlay and whole-page translation
There are two real approaches:
Overlay / typeset. OCR detects bubbles, translates the text, then draws the new text over cleaned bubbles. The art stays identical. The cost is infrastructure: font matching, bubble masking, line breaking, and a glossary so names and honorifics stay consistent. It also struggles with text drawn into the artwork, like sound effects and signs.
Whole-page image translation. An image model gets the page plus instructions and returns the page with translated lettering. It handles bubbles, signs and sound effects in one pass, and keeps panel order. The trade-off: image models may slightly redraw line art or screentones. For a quick personal read that is often acceptable; for archival-quality scans it is not.
3. Translate the whole chapter, not page by page
Chapter-level context matters. A character's speech style, an ongoing joke or a name introduced on page 3 should carry to page 40. Page-by-page translation drifts.
A practical pipeline is: unpack the archive, sort pages naturally, translate each page with shared context, then repack. Long chapters (30–60 pages) also expose failures faster — one bad page in a batch is easy to spot, ten inconsistent ones are not.
4. Always preview before paying or shipping
AI output is not deterministic. Before committing to a full-resolution export, read the free/low-res preview end to end. Look specifically at:
- small text inside panels and margins
- hand-lettered sound effects
- pages with heavy screentone (models like to "clean" it)
- honorifics and names across the chapter
- right-to-left reading order
This is the step most people skip, and it is the one that decides whether the result is usable.
5. Keep speed effects and page shape
A translated page should still feel like the original: same trim size, same panel borders, same spread shape. If the output changes aspect ratio or crops the page, the reading experience breaks even when the translation is correct.
Where a tool like LingoManga fits
I built LingoManga around this workflow. You upload a ZIP or CBZ chapter, choose an output language, and read the whole chapter as a free low-resolution preview with a watermark. If it looks right, one payment of $4.90 unlocks the clean full-resolution ZIP. No credits, no per-page math.
- Output languages: English, Spanish, Brazilian Portuguese, French, Italian, German
- Source language is auto-detected; non-Japanese pages may be less reliable, so check the preview
- 3–60 pages per chapter, double-page spreads stay one page
- No account needed to preview
- A real public-domain sample chapter: lingomanga.com/sample
What still needs a human
Machine translation is a first pass, not a substitute for a scanlator. Slang, layered jokes and cultural references still read better when a person rewrites them. Treat AI as a way to read a chapter now, and as a draft when you care about publication quality.
If you have a chapter that is stuck in raw Japanese, the fastest path is to check a free preview first and only pay for the pages that actually read well.
