How to translate raw manga in your browser (2026 guide)

You’re reading raws because waiting six weeks for the official English chapter is not a personality trait you possess. The problem: actually understanding the raws. Here’s the state of the art in 2026 — and why it’s finally good enough that you can do all your translation directly in the browser.

Why classic OCR translators break on manga

Translating manga sounds like it should be a solved problem. We have Google Translate. We have DeepL. We have a thousand desktop OCR utilities. So why does every solution feel awful?

Because manga isn’t paragraph text — it’s typography inside images. The text lives as pixels, not characters. Three things make it especially hard:

• Bubbles are bounded, not flowing. A speech bubble is a contained unit of meaning. Treating the page as a giant text block and dumping the OCR output into one big translation throws away who’s speaking to whom.
• Reading order varies. Right-to-left for Japanese manga. Top-to-bottom for Korean webtoons. Mixed inside the same panel for sound effects vs. dialogue. Linear OCR gets the order wrong, and bad order means bad translation.
• Sound effects need translation too. ドン! has a specific weight — “THUD!” or “BOOM!” depending on context. OCR alone gives you transliteration; you need someone (or something) to read the scene.

The old method, and why we’re past it

For years the workflow was: download the raw chapter as a zip, extract images, feed each page to a desktop OCR tool (Mokuro, Capture2Text), copy the extracted text into a translator (DeepL, ChatGPT), paste the result somewhere you could read it alongside the image. Five tools, twenty minutes per chapter, mediocre output.

Browser-based attempts existed (Yomu, ImageTranslate plugins) but they almost all overlayed translated text on top of the artwork, or dumped a wall of romaji in a side panel. Neither felt like reading manga.

The 2026 approach: image-edit AI

The pivot is models that don’t just see images but re-renderthem. Google’s Gemini image-edit family (released through 2025–26) can take a manga panel, identify each speech bubble, understand the language and tone, and produce a new version of the same panel with English typed inside the original bubble — same hand, same layout, art untouched. That’s a categorical change from OCR-plus-translate pipelines, which could only ever overlay text or dump it in a sidebar.

When done right, the workflow becomes:

• You hover a panel you can’t read.
• Click the translate button that appears on the image.
• The panel gets re-rendered with English in the bubbles, in the same hand.
• The page stays a manga page. The artwork stays the artwork.

The full browser-only setup

Whether you use MochiTranslate or another tool, the essential components are the same. Here’s what you need:

• A modern Chromium browser (Chrome, Brave, Edge, or Arc). Translation extensions ship as Chromium-first because of the better extension API.
• A Gemini API key. Open a Google AI Studio account, generate a key, enable billing. Each panel costs a few cents at current Gemini image-edit pricing — a $5 top-up gets you a long way if you stick to spot-translating the panels you can’t read rather than batch-translating whole chapters. (BYO means you pay Google directly, no markup, and your translation cache survives.)
• A bubble-aware, image-edit translation extension. The bubble-aware part is critical — that’s the difference between “Google Translate covering the page” and “the chapter reads like a fan translation.” And image-edit (re-rendering, not just transcribing) is what keeps the English inside the original bubble instead of pasted on top of the artwork.
• A raw manga source. Many readers use the official publishers’ sites (MangaPlus is a popular choice), MangaDex, or Comick, all of which carry untranslated chapters. Translation is for personal reading; please don’t republish or sell translated chapters.

Tips that make translations actually good

• Use Gemini’s image-edit model. It’s the only model right now that reliably re-renders manga bubbles back into the panel (rather than just transcribing them to a sidebar). MochiTranslate’s popup recommends the right model when you set it up.
• Set the tone preference. A samurai action manga and a slice-of-life rom-com need different voice choices in English. Most good translators let you pick.
• Translate panel-by-panel, every time. Spot translation is the cost-aware default — tap the bubble that blocked you, not the whole chapter. Spread-page action with overlapping speech and SFX benefits from this tighter scope anyway: you get a cleaner re-render when the model isn’t juggling twelve bubbles at once.
• Build a glossary. For long-running series, capture the recurring proper nouns and attack names so the translation stays consistent across chapters.

Day-one English, panel by panel

The translation problem is good enough in 2026 to read raws comfortably — provided you reach for spot translation instead of batch. Parse what you can on your own, tap the bubble you can’t. No desktop pipeline, no batch waiting, no costing yourself $5 per chapter.

If you want the bubble-aware, in-panel rendering handled by an extension that runs on your own API key, MochiTranslate is the one we make. Same studio, same one-time-pricing, no-tracking philosophy as MochiDim.