[EN] MTL Scanlation Guide

[kn4ll]

Ah yes, what a very intelligent response. If a completely normal message convinces you that I'm "psychotic" or "on drugs" then clearly you're a waste of my time and everyone else's.

I think we might have different definitions of completely normal then.

 

[kn4ll]

I've gotten some questions about panelcleaner and how it works, so I've just made some screenshots here over how to use it.



First you'll just load the files or folders. Default is cleaning (Panel Cleaner is a program that uses OCR to clean panels after all), so you'll have to switch to OCR. For OCR, using default settings is fine. You might need to tweak it a little to get cleaning to work if you want to use it for that.



After OCR has been finished, you'll be able to manually go through and correct mistakes in ordering and writing. Here it's really helpful to have some hiragana, katakana, and general kanji knowledge. Panel Cleaner which uses Manga OCR is a bit worse on handwritten text than google lens, so you might want to double check any text that's handwritten.



Easiest and best way to translate the finished text file is usually uploading it to an LLM like chatgpt or gemini, I prefer gpt. If you log into chatgpt you get a somewhat better model which is worth it in my opinion, you only get like 10 free uses a day but it's enough for a few chapters. You might also want to ask chatgpt to break down sentences if you can't figure out the sentence (not seldom it has used the wrong meaning of a kanji, or like for handwritten text it has gotten a kanji wrong).

 

[GeographicalNonsense]

If you're MTLing, use the latest models like Deepseek, it is free after all, and it can think the context of the Japanese text, but not as good as humans

 

[deleted-user-7d07f52915]

Imo, a lot of us are ESL, and couldn't tell the difference between edited MTL and actual translation. The problem is some "scanlation" groups just use mildly-edited MTL and didn't check if the translation makes sense. From I can tell, all manhwa "scanlators" are doing it this way. A bunch of scumbags. I wonder why they have so many defenders. A simple google search or keeping a list of names and terms would improve their "translation" significantly.

Things are especially bad with CJK novels. Shamelessly garbage MTL for most of them.

 

[kn4ll]

Imo, a lot of us are ESL, and couldn't tell the difference between edited MTL and actual translation. The problem is some "scanlation" groups just use mildly-edited MTL and didn't check if the translation makes sense. From I can tell, all manhwa "scanlators" are doing it this way. A bunch of scumbags. I wonder why they have so many defenders. A simple google search or keeping a list of names and terms would improve their "translation" significantly.

Things are especially bad with CJK novels. Shamelessly garbage MTL for most of them.

I think the benefits outweighs the risks and issues greatly. With the access to MTL a significantly larger amount of people will be able to translate manga. It's not gatekept by translators anymore, and people can basically translate anything they want. Since anyone can do it, you've already gotten a pretty big increase in the amount of scanlated manga. I think that people can greatly increase the quality of their translations by doing some really simple things, like learning hiragana, katakana, some kanji and basic Japanese grammar to avoid a lot of the stupid mistakes a lot of people do. And also translate handwritten text, although that has become a bit more difficult lately as google lens and google translate on pc don't provide the OCR anymore (at least it doesn't for me), so you'll have to use a tablet or phone to get the google OCR. A little bit of a pain, but it's worth the effort if there's a lot of handwritten text, as manga ocr isn't particularly great at picking it up.

Even if MTL opens up to more people using scummy behavior, that's on people that have reduced morals, not on MTL honestly. No different from someone using a gun for hunting and recreation, and someone using it to murder people. I think it's important to make a distinction between using MTL and having scummy behavior, sadly a lot of people don't seem to be able to make this obvious distinction. Would it be any less scummy if they translated without MTL, but with shitty editing, paywalls, and sniping? No it wouldn't, groups like galaxy degen scans, soseki scans, and arang scans have been doing this shitty behavior for years.

Like to get started, you'd only need to download photoshop, either legally or cc15 or cs6 from your favorite torrent site, the scanlator school font pack under resources, panel cleaner and google lens on your phone or tablet, hakuneko, calibre/dededrm or other tools to retrieve raws, and stuff like IOPaint for AI redrawing (although content aware fill mostly do the job in most of cases), and a free chatgpt/gemini account or similar. Everything you can get for free, and now scanlation is something that can be done by anyone that only cares to put in a little work and effort.

 

[kn4ll]

Tried to install IOPaint on a new computer, man that program is hard to install. Seems like it doesn't work with anything else than python 3.12.10. Don't install pytorch and torchvision first, it will download it by itself, and it seems like installing the wrong version botches the IOPaint install. Managed to get it to work after a lot of fumbling.
Tried installing panel cleaner through terminal afterwards, and it used some of the same programs/dependacies (pillow I think) as IOPaint, and couldn't install it since it needed an older version, so I ended up just using the windows dowload.
And even though hakuneko has failure message after failure message, somehow it manages to work completely fine after I've installed it.
Man all this computer stuff is too hard for me.

Edit: Tried to install IOPaint on another pc. Looks like it doesn't work anymore. Even tried using miniconda and stuff like that. Oh well.
Editedit: And now it works again! Yay! Guess the fault wasn't at my end.

 

[nickyturner666]

honestly, as a beginner myself, i do shamefully use a bit of mtl now and then, although i do make sure to double check sentences through anything like jisho or search anything online to see how it is interpreted by others to see if what im translating is going to be more accurate, so it won't end up complete slop. mtl can tend to miss stuff like references and slang, so it's always good to ask for help online with proper resources and information available, or from someone else who is experienced. either way, i do appreciate the tutorial you're giving for others and i figured i could give a bit of my two cents for anyone else here

 

[kn4ll]

honestly, as a beginner myself, i do shamefully use a bit of mtl now and then, although i do make sure to double check sentences through anything like jisho or search anything online to see how it is interpreted by others to see if what im translating is going to be more accurate, so it won't end up complete slop. mtl can tend to miss stuff like references and slang, so it's always good to ask for help online with proper resources and information available, or from someone else who is experienced. either way, i do appreciate the tutorial you're giving for others and i figured i could give a bit of my two cents for anyone else here

I think it's better showing people how it's possible doing MTL most effectively and precise instead of doing what has been done until now, which is just trying to put any guide or help down and say don't use MTL. Then people are just going to use it anyways, but poorly. Stuff like knowing basic Japanese, hiragana, and katakana helps immensely and don't require much time to learn, and actually translating handwritten text instead of just skipping it makes a scanlation much better.

Right now I'm thinking of making a program similar to ichigoreader and cotrans that automate the entire process, but more for scanlation purposes than reading, and that makes a psd that you'll only need to proofread and edit. You already have the tools from panel cleaner to ocr and clean, AI generative fill like lama that can redraw the OCR'd text that can't be cleaned, chatgpt that is far better at translating if you translate an entire chapter (like using .csv) instead of page by page like ichigoreader and cotrans, and automatic typesetting similar to ichigoreader and cotrans. All of the tools to do it should be readily available, it's just about putting them together. Though I've only started learning programming two weeks ago, but according to a friend of mine it shouldn't be too hard to make.

Though I suspect that there are several groups that already have made similar tools like this, they're just not public. Especially the new coin based pay wall groups that appeared lately and put out a ton of chapters in a short time. Though three of them have already been removed because of the new addition to the 1.5 rule.

Edit: If you install cotrans on your machine, you can apparently save as psd to use like this. Haven't tried it out myself though, but my guess would be that it's what the coin based paywall groups use. And this one, though I'm not sure if it has psd support. Gotta check them out when I have time. None of them seem to translate as chapter script though and works page by page, which isn't great. I guess AI is moving so fast that they're getting a bit obsolete already.

 

[kn4ll]

So there's still evolution in the field. I suspect that some of these newer tools are used by the coin based paywall groups that have been banned on mangadex. One example is Koharu, which pretty much automates the entire scanlation process and export it as a psd for editing. We've come a long way from cotrans and ichigoreader, that's for sure.

 

[happypurple]

AI still has the same fundamental flaws it's always had.

You can tell text is AI when it's made from the statistical averages of human speech, instead of "thoughts" "experiences" or even "emotions" (yucky flesh puppet words.) The reason AI doesn't say the same thing every time you give it the same input, probabilistic generation, also means that hallucinations are inevitable. Advancements till this point haven't made hallucinations impossible, just less likely and easier to miss when they occur. In a way it's almost worse for translation if you care about accuracy, since the errors are more likely to go ignored or unnoticed.

Believe it or not, translation is more than binary errors you can catch with some proofreading! There's actual human-level judgement needed for the best quality product. AI has mediocre writing quality since it's literally the average of human writing. When corpos save a dime by getting the AI to do a first pass before handing the manga off to their 60 cents/page translators, the end product is limited by the AI generated base which revisions get made from.

The same fundamental principal behind your phone's keyboard trying to guess the next word in your sentence is now on the front lines for artistic expression at some low-cost b2b manga localization companies. It's pretty telling that the AI startups trying to translate magna/etc into English with AI are based outside of native English speaking countries. Can you imagine an American company taking marvelslop comics to Japan with chatGPT generated dialogue? They'd get laughed out of the pacific ocean. AI is a copy of the shape of human writing, rather than actual intelligence. The AI has no mind, experiences, memories, consciousness, or understanding of language as we humans understand it, not even just in the sense that there's nothing within the AI capable of understanding something.

I can't get mad at an inanimate object like ai, or a toaster, but I CAN get mad at the wonderful person dumping 1000 broken toasters in my driveway, while expecting me to thank him for the free toasters. I just love it when the guy using MTL cannot even be bothered to double-check that the AI got the pronouns of the characters right, so everyone in their upload becomes male. Clearly our generation's Shakespeare.

Can everyone please switch to Chinese already? It would genuinely mean so much to me if you could do what you're doing to Chinese instead. I can't blame the people who speak English as a second language for not being the one of the 1-in-4 EN speakers who are native, but can we at least be honest with ourselves? Can someone give me a reason to humor these MTLers and treat them as my equals, when they have nothing to offer except the time they're wasting getting the AI to do the part of scanlation that needs human judgement the most?

Much love
XOXO