Good morning, afternoon, or evening, everyone! This is different from my original content, that being the two articles I have out at the time of writing (lol).
Recently, I’ve picked up a new habit, and that is translation of Japanese and other manga, you can find the list of the current manga’s I’m working on below. Along with using OpenAI’s gpt-4o for translations, though sometimes using gpt-oss-20b and gemma-12b depending on how hard the language is, so for the most part, costs are cheap, being ~0.80 USD to translate a 200 page volume, with the rest just the time it takes to run and perform the machine translation, redrawing, and other things on my PC, then if I have the time, touching up any weird placements or translations with GIMP.
For now, don’t expect great things. I’ll work on these as I have time, but I do want to at least finish “Ma O Eru!” and get up to speed with “Kono Inu to Kitara Mago no Koto shika Kangaeteinai” once the new volumes come out.
If anyone is interested in helping, just reach out on Discord or via email. I pay for all of these out of my pocket, so any help is appreciated. I’m also more than open to dropping MTL, assuming someone comes along to assist.
The Pipeline
The workflow for scanlation is deceptively complex. What looks like “just translate and slap text on it” actually involves:
- Raw acquisition — getting clean, high-resolution scans
- Translation — the actual language work (or in my case, throwing tokens at gpt-4o)
- Cleaning — removing the original Japanese text from the artwork
- Redrawing — reconstructing the artwork behind where text was
- Typesetting — placing the English text with appropriate fonts, sizing, and placement
- Quality check — proofreading, checking for consistency, touching things up in GIMP
Machine translation gets you about 60% of the way there. The remaining 40% is where all the nuance, humor, and cultural context lives. MTL without editing is how you end up with manga that reads like it was written by a particularly confident but confused alien.
Why MTL Is Both Amazing and Terrible
Modern machine translation (particularly the newer LLM-based approaches) has genuinely gotten impressive for Japanese -> English. For straightforward dialogue, it’s often serviceable. But manga dialogue is rarely straightforward:
Original: ちょっと待ってよ！そんなの聞いてないんだけど！MTL: "Wait a moment! I haven't heard of such a thing!"Better: "Hold on, nobody told me about this!"The MTL version is technically correct but reads like a textbook. Natural English manga dialogue needs to feel like people actually talk, with contractions, sentence fragments, and the occasional creative liberty.
Swapping between gpt-4o, gpt-oss-20b, and gemma-12b depending on the complexity of the source material helps keep costs down while still getting decent output. The harder stuff — slang, cultural references, wordplay — that’s where you either pay for a better model or spend time manually fixing things.
If you’re getting into scanlation, start with a series that has simple dialogue and minimal SFX. Shounen action manga is a good starting point — the dialogue tends to be direct and the SFX are usually just impact sounds.
Manga To Be Translated
- Ma O Eru!
- Kono Inu to Kitara Mago no Koto shika Kangaeteinai
Kensei Maou no Yarinokoshi
Was It Worth It?
Honestly? It’s been more fun than I expected. There’s something satisfying about taking a raw page and turning it into something that reads naturally in English. Plus I’ve picked up way more drawing skills from the redrawing process than I ever expected. It’s also given me a new appreciation for the work that goes into official translations — they have entire teams of translators, editors, and artists working together to make it look seamless.