Show HN: Instantly Translate Manga – TranslateManga

[kadeus]

https://translatemanga.net/en

Since I was young, I've loved anime, and over the years, manga has brought me joy, given me courage, and sparked excitement in my heart.

However, as I read more, I realized that many of these manga weren't translated at all. I also came across some AI-based translation tools, but the results often fell short.

So, I decided to create a tool that allows manga fans to read and enjoy their favorite manga, no matter the language or whether a translation team is available.

This product has just been launched, and there are certainly areas that can be improved. However, with time, I'm confident it will only get better.

You're welcome to try it out and share your valuable feedback!

[hmm37]

Any hopes of this or any other software being able to translate games on the fly. Would be neat to be able to play old games that were never translated.

[chocolatkey]

You have much competition in this space with for example https://ichigoreader.com https://ismanga.com and many more (just search the chrome extension store for examples). What is your plan to stand out and remain competitive?

On a side note, if anyone’s been watching the fan translation community, it’s clear this is the end game.

[chucky_z → chocolatkey]

I don't actually think it's 'the end game.'

A vast majority of AI translations are usually enough to get the gist of the story but tend to miss a lot of really important with-context story clues. What I've found is that the best case is that you can get a 'speed translation' with AI, and then have a translator come in behind and clean it up. Much like using AI to generate the base of an image and then painting over top of it to fix any artistic choices it's made that the artist doesn't agree with.

[chocolatkey → chucky_z]

I agree that AI translations are not close to the standards many people would expect, but there is a significant, growing population of readers (many with EN as a second language) that are fine with reading MTL using extensions such as this or fan-translated MTL with human assistance, and it is what many new fan translators are opting for as old groups retire. The “speed translators” using MTL also demotivate the slower, better translators, as many people will no longer be as interested in their release. I don’t think human translation will cease for a long time, but in the meantime, MTL as a percentage of new fan translations will continue to grow

[Karrot_Kream → chocolatkey]

The quality vs speed translating war has been ongoing forever now. Before I retired (a long time ago now) I had transitioned to quality groups because the speed oriented groups were producing such bad translations. MTL is just accelerating this.

[VincentEvans → chocolatkey]

To provide a counterpoint to your hypothesis that “people are fine with easily accessible but subpar results” I often switch from dubbed anime english voiceover to original japanese that I don’t understand and prefer reading subtitles because they convey the meaning and humor much better than the americanized voiceover, and the voice actors are better.

I hope that AI generated simulacra will eventually take its rightful place next to microwave dinners in consumption hall of fame once the novelty wears off.

[gamacodre → VincentEvans]

> they convey the meaning and humor much better than the americanized voiceover

I'm glad this isn't just me. I occasionally watch anime and the English dubs are seemingly universally terrible compared to the originals. Subtitles are annoying but nothing like listening to someone mumble their way though a script.

[BizarroLand → chucky_z]

I've read a few that were mangled horribly, people who didn't read other versions so they lost context or made really bad choices in the translation.

It will be a while before not only can the languages be translated but also the intent communicated in context.

[Karrot_Kream → chocolatkey]

It's been the end game of fan translation forever. I used to be a manga, LN, and anime scanslator/fansubber in another life.

For popular manga the "meta" today is roughly:

1. Wait for the chapter to be released in China or Taiwan where copyright isn't as broadly enforced.

2. Upload the Chinese raws.

(Steps 1-2 usually happen in a different ecosystem than translation.)

3. Have a Chinese/English bilingual person read the manga and turn it into translated text. These folks are usually pre-teens or teenagers. If this is a popular series, they're under intense speed pressure (the group that drops first wins all the imaginary internet points, or basically all the page views and torrent peers.) The quality of these translations are usually anywhere between average to bad. The kids read barely enough to understand what's going on (given the intense time pressure) and that they're already starting from a translated source (JP -> ZH.)

4. Have your typesetter clean the manga and prepare it for English. This is generally done by also a pre-teen. If they're under time pressure they'll do the bare minimum here to get it out quickly.

5. If you have enough staff and the series is popular, an editor will then edit the text to make sure it makes sense and perhaps rewrite bits to make it reflow into the text bubbles. If you don't, they're often the same person as the TL or the typesetter/cleaner.

6. Upload onto the sites/drop the torrents.

If you can read Japanese you'll realize that a lot of these translations are speed oriented and the quality is quite bad. Sometimes quality-oriented groups will go ahead and revise an already released manga but more often than not once some scanslations are out, others will consider the series "translated" and bad fan scanslations will live on.

When I was active the steps were different; they've changed as they've responded to time pressures. It's already a huge race to the bottom. Most of the big series are getting translated by pre-teens and teens. Some more niche series are being tackled by college students. Doing AI assisted translation is just the next step in an ongoing war to speed up translations.

[prussia → Karrot_Kream]

The manga scanlation world is pretty big, but that's not been my experience and understanding at all.

1. For truly popular mangas, no serious group is going to wait for a chapter to be released in China or Taiwan[0]. Someone that lives in Japan will go out and buy it. Or wait a few days for the it to posted online for sale by the publisher (though sometimes most recent chapter will be free for a few days). It's all illegal so why would it matter if the raws were obtained in China/Taiwan vs. Japan/online anyways?

3-5. For many of the popular series, more "groups" than you would expect are really just one person. And perhaps irregularly a cleaner and/or a typesetter (eg KireiCake[1]). Some more "professional"/commercial groups have much larger teams, and even paid translators, but they tend to do porn (especially erotic BL), because that's where the money is at. I've never heard of pre-teens being part of these groups, but I guess no one would really advertise that they are a pre-teen. Shitty translations (or shitty MTLs) discouraging others is a real problem, especially for more niche series.

[0] Unless the Chinese language release is at the same time of the Japanese one? At least for moderately popular manga, I know the English release is usually months (or more commonly, years) behind. [1] https://www.reddit.com/r/manga/comments/1e1bkw9/a_statement_...

[Karrot_Kream → prussia]

Yeah the manga scanslation world is really big so I have no doubt there are differences in each of the points, except:

> I've never heard of pre-teens being part of these groups, but I guess no one would really advertise that they are a pre-teen

I was a pre-teen when I started scanslating. I did this for ages and did all the roles that were available back then. I started editing/managing and head TLing multiple projects as a teen. I "retired" in college. I don't want to dox myself too much but the pipeline from scanslater to low-pay intro commercial translation work was very common back when I did it, since it's easy to hit N2 if you were TLing from Japanese and many of the better TLs were N1 or better. While this isn't the case for me, many of the TLs were heritage speakers. From the Discords I'm in these days I'd say the age distribution remains very similar, with a mean around 16. I mean it makes sense, who wants to do this much work with no pay?

[novok → Karrot_Kream]

A halfway tool that would let you take the text and auto-typeset it in would probably be a big thing. All you would need is the equivalent of a subtitles file at that point. The tool will identify regions to change and how to change them, the original text and the machine translation text. Then, all you would need is somebody who can do translations properly afterward to give a better translation, and with less friction, it will happen a lot more. And you could have a user rating of which translation is better.

I could also imagine a tool that would take a page and describe per panel. Then, the translation files with LLMs could have a better idea of the context and probably make a better translation versus a speech bubble by speech bubble type translation since they would only interact with the intermediate 'subtitles' file. It would also be a good accessibility thing for blind people to read a comic.

[allenu → chocolatkey]

Do you know if there's a camera-based app that translates on the fly? These all seem to use uploaded images, but I'd love to be able to flip through an actual physical book and translate via an app.

[spacebruce → allenu]

The current version of Google Translate is close to managing this, it's reasonably good at picking out separate text bubbles, even in handwritten script, but you might have to play with camera positioning and lighting.

it's best when faced with long stretches of dialogue, the translation does start to fail with short snappy "action manga" japanese with short context-less sentences and exclamations.

[theogravity]

This might be useful for just checking the general content of a chapter you're interested in if it hasn't been translated yet, but it's not clear if it handles things like varying fonts / sizes used to convey the emotion of spoken dialog, does consistent translation (eg does it remember stylistic choices it has made chapters before it), or handles tricky items that might be difficult to localize.

Also, how does it work, what is the tech behind it? Are you doing any of the training yourself?

[spondylosaurus → theogravity]

Another thing I'm not sure machine translation can really "nail" is cultural context, or even little linguistic cues and other tidbits. I like when translators explain in the margins that one character is speaking in a certain register for XYZ reason, or that there's been a shift in a certain relationship signaled by a change in how they address each other, etc.

That said, I did just read a great series last night whose human translation ended right before the final two volumes, and hasn't been updated in nearly 8 years... so I may need to try some machine translation on those last two volumes just to see how things end.

[parsimo2010]

Similar to other posts- you have competition in this space. Can you show some examples of how your results are better than competitors?

Also, some clarity on what counts toward the limit of "20 free translations" would probably help. Is it 20 speech bubbles? 20 pages? 20 images?

I think seeing these things would motivate people that might be interested but aren't sure if it is worth the attempt.

[vunderba]

In addition to lots of commercial competition, there's also a halfway decent open source comic translation app that uses automatic OCR + LLMs (byok).

https://github.com/ogkalu2/comic-translate

[m4rtink]

This reminds me of something that has been bugging me for quite a while - are there some extensions/services that would do video translation ?

What I mean is that you have a video with text in it in a foreign language (labels on stuff, signboards, even hardcoded subtitles) in a foreign language & the tool would translate those to a language you can read (like English) in place.

While I'm aware of all the various edge cases and challenges for this, its not a theoretically thing - I can already pause the video & use Google translate on my phone via the camera feature to translate whats on the TV/monitor quite well.

I would just want something a bit more integrated so that I don't need to fumble with my phone every time I want to translate something on the screen. :P

Even automated still frame translation would be enough - just pause the video & the tool would analyze the still frame & translate any text it can find in place.

Like, this is evidently possible nowadays, but I have yet to see it implemented in in an integrated manner so far.

[qq99]

Nice idea :) I just realized (thanks to your post) that implementing a solution to this is pretty tricky.

e.g., the bottom right panel, 1 sentence is split up into 2 speech bubbles

「この帽子を」＋「お前に預ける」meaning approximately, "I'll leave this hat in your care"

Any solution here has to kinda grab the words for a whole panel yet still know how to segment it (and read it RTL, top-to-bottom), and then ideally render something like:

"I'll entrust this hat..." + "...to you." or something similar for some sensible split into 2 word bubbles.

In the top-left, 「おれの大切な帽子だ」, the translation of "It's my hat" is quite bad. "It's my precious hat." would be much better. Maybe it's an OCR bug? Because LLM should get that nuance pretty easily.

[bdhcuidbebe]

> I also came across some AI-based translation tools, but the results often fell short.

> Professional manga and comics translation powered by AI.

How are your AI based tool any different?

[dankwizard → bdhcuidbebe]

Because I've added preamble and instructions "Always make sure you're translating correctly". This keeps mine ahead of the competition. Please sign up

[gherard5555]

>Professional manga and comics translation HAHAHAHAAHAHAHAHAHA

[TheAceOfHearts]

Ideally, fans only have to translate a manga once and then they can publish the English version for everyone to access. Probably not the best model from your business' perspective, but there's no reason to keep repeating work when you can easily save and share the results. This is a case where community and business interests are unlikely to become fully aligned.

[Klaster_1]

Come to think of this, doesn't manga represent a perfect dataset? Why not train a model on original/translated page pairs? This way, you achieve a full E2E pipeline where you don't have to worry about cleaning, joining pages, translation, editing. Or is this too computationally prohibitive at the moment?

[asukachikaru]

There seems to be an obvious mistake in the landing page example. The word `ガキ`(kid / brat) is translated as oyster. I guess the AI mistook the katakana `gaki` into hiragana `kaki (かき)`. You may want to take that into consideration.

[fuzzythinker]

What is a "translation"? It needs to be spelled out. 1 page, 1 chapter? 20 pages free or 1200 pages/mo is horrendous. If chapter, a chapter's length differs greatly among mangas.

[novok]

I would put more demo pages and make them larger / zoomable. It's hard to read the current one. I would make the demo pages have tricky things beyond clear speech bubbles too, but are still essential for understanding, like text not in a bubble.

[hiccuphippo]

How does it deal with wordplay and double entendre? As I understand, a lot of Japanese comedy is based on that.

[Karuma → hiccuphippo]

You're worried about something that only comes up once every few thousands of sentences...

I'd be more worried because even the simplest and shortest sentences are horribly translated in those example pages.

[rererereferred → Karuma]

Once every thousand sentences? You probably don't read Kumeta.

[yborg]

Looks like this got slashdotted, uploading an image and clicking 'translate' just spins forever.

[okkdev]

I really can't see any of these tools actually succeed in good translations instead of surface level slop...

[hooverd]

I think MTL is a great tool but there will always be a place for proper scanlations.

[ge96 → hooverd]

Sees Lain's face