[X] My Assistant

This manga has everything I want to find in a good body swap manga:
https://e-hentai.org/g/2186781/daee634118/

But I wanted to read it in English. I also think that others would enjoy it in the common tongue (yeah, English isn't my first language either).
How are things nowadays? Do we have to manually clean the balloons and then fill them with manually translated text? Are there some neural network to speed up the job? At least cleaning the balloons I'm fairly certain (since I work with these) that it can be done automatically, just need some good/huge database of original+cleaned images (assuming that the community preserves cleaned and original versions out there...).

On a related note, I remember that there used to be folks who took the challenge themselves and did the job... how to contact them? How much they charge (or they do it for free?) ?

Requests and bounties isn't really for discussion, so moved here. Probably worth skimming threads here in case it's already been discussed.

[mangadex.org] https://mangadex.org/chapter/d7c64a71-b97f-...-6e2f5eb97746/1

[capture2text.sourceforge.net] http://capture2text.sourceforge.net/

Combined with Textractor + DeepL Extension with clipboard monitoring or the DeepL App and you've got an adequate means to read hentai without sucking off somebody for a translation. Can also be used for playing Japanese games. It's far from perfect but so are the kind of stories you get from hentai anyway.

I'm thinking that if I have both original and clean version, theoretically, I could get text boundaries by getting absolute difference of both images, then applying textract (or another OCR) for jpn in vertical, finally for each bbox I could invoke Google API to translate that sentence. That would give me the translation.

Using morphological filters, floodfill and findcountours algorithm, I could detect balloons' contours in clean image. Assuming that I know from which balloon come each sentence, I would need a good strategy to find the right font size to write translated sentences in an aesthetically pleasant way in clean image.

TL;DR That workflow would have original and cleaned image as input, and output translated image. Could be done online in Google Colab, no need to download and install any tool.

Also, like I said before, given enough examples (and patience (IMG:[invalid] style_emoticons/default/tongue.gif) ), a DNN (UNET? Pix2Pix? ) could be trained to map original images in clean images.

Too complex, I know... also have some limitations, like all the text needs to be contained inside balloons, cleaning must be consistent, and so on.
But, afaik, doable and in the end a tool for lazy ones like myself.



@marien
Could you elaborate more? I didn't understand your workflow.

This post has been edited by barack obrahma: Apr 11 2022, 23:21

@marien
Could you elaborate more? I didn't understand your workflow.

Capture2Text is a powerful OCR tool you can use to read Japanese text on images. Said text is then copied to your clipboard. Textractor monitors your clipboard and can automatically translate whats on it with your chosen plugin and settings. Google, DeepL, Systran, etc. Since Capture2Text already has Google Translate built in, you can use DeepL in Textractor for an alternate translation. I won't go into all the details but the only thing you really need is Capture2Text. It's simple and easy to use while Textractor takes some getting used to.

It sounds like you're planning to use this tech for editing purposes. I don't recommend that since you're not a native English speaker. I was only suggesting you use these to read hentai on your own.

I'm thinking that if I have both original and clean version, theoretically, I could get text boundaries by getting absolute difference of both images, then applying textract (or another OCR) for jpn in vertical, finally for each bbox I could invoke Google API to translate that sentence. That would give me the translation.

You seem to be missing the all-important step of actually translating the text. Maybe you should include that instead of pointlessly feeding the Japanese script into Google's gibberish-generator.
Even if you're content with MTL for the sake of reading a doujin you want, this approach of trying to do everything automatically is nonsense. A proper MTL, as much as it ever can be "proper", requires every output to be curated and checked manually, or you WILL get garbage. Not just sentence-by-sentence, but phrase-by-phrase, and checking both the meaning and how the pieces all connect together grammatically. And if the grammar and the vocabulary both need to be done manually, it's basically just a manual TL at that point.

It sounds like you're planning to use this tech for editing purposes. I don't recommend that since you're not a native English speaker. I was only suggesting you use these to read hentai on your own.

Thnx for the info, I'll look into that.
But I'm not planning on editing/publishing translations, just the QnD to MTL something, easily reproducible by others with low tech skills. (IMG:[invalid] style_emoticons/default/cool.gif)

You seem to be missing the all-important step of actually translating the text. Maybe you should include that instead of pointlessly feeding the Japanese script into Google's gibberish-generator.
Even if you're content with MTL for the sake of reading a doujin you want, this approach of trying to do everything automatically is nonsense. A proper MTL, as much as it ever can be "proper", requires every output to be curated and checked manually, or you WILL get garbage. Not just sentence-by-sentence, but phrase-by-phrase, and checking both the meaning and how the pieces all connect together grammatically. And if the grammar and the vocabulary both need to be done manually, it's basically just a manual TL at that point.

Good points sir, but I'm not advocating MTL over tedious, boring and time consuming HTL. It still has a place... like, as an example, traditional knife makers still do quality grade crafts. But some people just want to slice their bread, GTD style, and not have a deep emotional connection with their knife, like others want. Some people just want to have a general idea of whats happening in a manga, and not have a commercial ready-to-publish material at hands. Maybe they do too, but a MVP (minimum viable product) might be enough for that person to realize that they dont like that much that story, and no need to pursue a better translation. OTOH, the MVP might enhance ones senses and make it look for a better translation.
So, a quick, fast and not reliable translation isn't a competition for a slow, tedious and excellent translation. They both have space... that said, I don't think this post should be about it. I'm just looking for tools, pipelines, workflows to automatize MTL. The benefits of HTL over MTL, is the scope of many other posts, afaik, no need to have one more here. (IMG:[invalid] style_emoticons/default/wink.gif)

Papago exist and it is good enough for an EOP to have a basic understanding of what's happening in a basic doujin.

Machine translation is the future of translating. I do believe that. You can't beat AI and it will only get better and better. The main problem with Japanese is that it is context heavy and even the most complex AI translation platform struggle with this. Japanese MTL will be the ultimate test for a MTL.ẞ


For basic hentai, MTL is actually good enough especially Vanilla doujins. MTL will struggle even with the basic manga. It's unable to retain any context from whatever it is translatinf

This post has been edited by shakuganaexa: Apr 12 2022, 16:17

Why? People who are fine with MTL can just use their google apps and translate it that way. Or save the pages and use Google Keep. Or use an OCR software like mentioned above. I don’t see a need to have any pipeline, if the goal is to just read MTL on your own time.

Anyone who wanna read MTL already have the tools needed, just gotta not be lazy.