[Tools] SLR Translator 3 - Translation Engines for RPG Maker VX, VX Ace, MV, MZ, and Pictures

[Shisaye]

Overview:
SLR Translator is a tool I've been developing over the past few years to translate RPGM games at a high standard.
(This will actually work on a fairly old laptop as long as it doesn't overheat.)
Also includes the DSLR Translation engine to use whatever AI (LLM) you want as long as its api supports 'v1' or 'v4' requests.
(Pre-configured for a free LM Studio setup using Qwen.)
As well as the SEP picture translation system to be able to translate (horizontal) picture based text offline in bulk.

Spoiler: Regarding Premium Features (Is SLR Translator really free?)

SLR Translator and all its features are completely free. There are absolutely no premium features, or deals with any company or person. You can choose to use DSLR with a non-free endpoints, but that has nothing to do with me.
I'm not a "Chinese Shill". I wish I was and they would pay me, but no they do not even know, or care that I exist. I just haven't found a western LLM, yet, that can beat Chinese ones in cost/performance ratio.
You can get better quality from other ones, but they are way too expensive to be worth it.

I am also not part of the premium plan of Krakenfiles, I get nothing if you buy premium from them.

I have a SubscribeStar page because someone told me to make one at some point, but it does not get you anything extra and nobody donates. I'm not even close to the threshold of ever being able to withdraw money.
I'm really just dumb enough to do this as an unpaid hobby since 4 years.

Important Notes:
Press F1 with the tool open for tutorials on how to use this.
The SLR translation option will only work if you have a working SugoiV4 server running locally on the default port 14366.
You can visit Ming Shiba's Patreon, or the F95 Sugoi thread to find out more about Sugoi.
The V9-V9.4 Toolkits are not compatible with SLR, but 9.5 works again.
The DSLR translation option does not require SugoiV4, and provides a vastly superior translation, but requires setup of a local LLM, or needs an external provider, read the manual.
SEP works on its own.
RTFM.​

Developer: Shisaye
Version: 3.015
System Requirements: At least 8GB of RAM.
For free locale use of DSLR at least 6Gb of VRAM on a CUDA/VULKAN capable GPU.
For free locale use of the DSLR Glossary system at least 8Gb of VRAM on a CUDA/VULKAN capable GPU.
Cover Art: FatKE

Installation:

Spoiler

Extract the tool and launch it using the SLR Translator executable.
If you have problems extracting, please update your 7zip to the latest version.
After launching the tool, please press F1 to get to the help page explaining all basic stuff.

Changelog:
See attached "Changelog.txt".

DOWNLOAD:

SLR Translator 3.015:
(Includes DSLR and SEP.)

You must give the following reaction to unlock:

You must give the following reaction to unlock:

SHA-256 of the archive:
38355f5179c343dd49316f05adb25ae2fe56186544c13d69d9189179bbe4969d

SLR Translator Linux 3.015:
(Includes DSLR and SEP.)

You must give the following reaction to unlock:

You must give the following reaction to unlock:

SHA-256 of the archive:
ebd0cd1b32f7996ee31aaa201cf9b348432e8f5b3b337d4b1e5b24bf38849147

Additional Tools:
(Extractors, Editors, etc. Also includes a minimalist Sugoi Repack.)
(I did not make any of these.)

You must give the following reaction to unlock:

You must give the following reaction to unlock:

SHA-256 of the archive:
61bfb54dc700ecd875a0d28fd10d032058db4b901bc191ddbd2257008c926f3b

Spoiler: Old T++ based version

SLR Translator 2.107:
(Includes DSLR.)

You must give the following reaction to unlock:

You must give the following reaction to unlock:

SHA-256 of the archive:
20affbd87223cb6053b59ebcfb601a6394e0462105726322302b88a2a8058daf

SLR Translator Linux 2.107:
(Includes DSLR. Is incomplete)​

You must give the following reaction to unlock:

You must give the following reaction to unlock:

SHA-256 of the archive:
7dd325bc89b999a51f9f91e38f891dce6188218c7bc1fa531b7d4398e4ee0d60

SEP Addon:
(Image translation. Currently doesn't work with Linux.)​

You must give the following reaction to unlock:

You must give the following reaction to unlock:

SHA-256 of the archive:
dd686de45f5824a8ef06add91ceb353a61e41cda3c0be442edc5e91c18820ae5
​

 

[Entai2965]

Best mascot ever.

Edit:
SLR Translator has a lot of optimizations based around Japanese to English translation + the Sugoi Offline Translator v4.0 NMT model. The Sugoi Offline Translator v4 NMT model is included in Sugoi Toolkit v6-v9+.

Using SLR Translator in its intended configuration requires:
1) The Sugoi Offline Translator v4.0 model in either PyTorch/Fairseq format or in CTranslate2 format.
2) Software that can run that model on either the CPU or GPU.

The PyTorch/Fairseq/original version of the model and a small flask server to run it are available from the Official Toolkit (~6.5+ GB), but FOSS alternatives are also available here for convenience, reduced size, better performance, improved compatibility, added cache, added UI, transparency, stability, etc.

I went ahead and bundled everything together in a Sugoi-Repackage to bootstrap the Sugoi Offline Translator v4.0 NMT model + SLR Translator as fast, as clearly, and as reliably as possible. The Sugoi-Repackage is already included under AdditionalTools.7z\apps\Sugoi-Repackage-CPU in the OP. Just double click on the .bat to set up the software and then add the CTranslate2 format model to the correct folder to begin using SLR Translator right away.

Spoiler: Sugoi-Repackage Installation

1. Most users should use the Sugoi-Repackage-CPU (32MB -> ~900 MB after setup) if just starting out.
- This is already included under AdditionalTools.7z\apps\Sugoi-Repackage-CPU in the OP.
- CUDA users should download one of the CUDA versions.
- The CUDA version supports both CPU and CUDA. CUDA is faster, but it requires an Nvidia GPU and takes a lot more space (~5 GB+).
2. Double click on the .bat to setup the local server software.
- It will download, install, and configure CTranslate2/PyTorch into an isolated Python environment.
3. CPU users should download the model in the CTranslate2 format (+1 GB).
- CUDA users should download the model in PyTorch/Fairseq format (+1 GB).
4. If using CTranslate2, copy the CTranslate2 model to Sugoi-Repackage\sugoi_v4model_ctranslate2.
- If using the PyTorch/fairseq model, copy the model to Sugoi-Repackage\sugoi_v4model_fairseq.

For more detailed notes on the Sugoi-Repackage, see the main thread and expand the small summary below.

Spoiler: Background information

Disclaimer: I did not write most of the information. I just copied it for everyone else's documentation and bundled the work of other people together to make their software easier to use.

Summary: The default software that Sugoi Offline Translator v4.0 uses to run the model is an old version of Facebook/Meta's fairseq project + a small flask server. Issues:

- This old unupdated version of Fairseq used is very slow.
- It does not release memory properly after use which means the system will eventually run out of memory while using it.
- Due to the platform specific nature of the toolkit, this software does not reliably work on other operating systems due to configuration limitations especially when using Nvidia CUDA on Windows 7, 8, Linux, Mac OSX. Even though PyTorch and CTranslate2, the software that runs the model, supports these configurations, the toolkit does not.
- It is complicated to switch between the different PyTorch and CTranslate2 models and configurations (CPU, CUDA, DirectML).
- Translations are not cached.
- Early versions of Sugoi Toolkit v9 dropped support for T++/SLR Translator (unofficially). The developer, MingShiba, may re-add support at some point.

Basically, the Sugoi Toolkit is the worst option to run the Sugoi Offline Translator v4 NMT model because it has the least compatibility, the slowest performance, and the most bugs. The software is just that bad, especially the non-existent technical documentation. The Sugoi-Repackage fixes all of the above problems and uses py3translationServer as the server software since they had the best documentation, cache, and it also includes a UI at http://localhost:8501 for those who want to use the model directly.

This thread is to discuss SLR Translator. To discuss the repackage, please use the repackage main thread.

 

[Shisaye]

To everyone using SLR translator, just to say it again in a more clear fashion.

If you've made a translation, regardless if you think it's bad, regardless if you think it's too incomplete, heck even if it has minor crashing issues:
Please actually publish the translation.

SLR Translator was never meant for individual throwaway translations.
The point was to enable even someone with no coding skills whatsoever to keep the translation train rolling and get rid of the last bit of untranslated games, no matter how niche and hopeless they seem.

While I'm at a good pace, having made over 400 translations over the past 4 years, games still release every day, and there's still hundreds of untranslated ones out there that could at least use "some" translation.

If you wouldn't know where to post them or it seems like a really stupid process. Just dump them in the "mini releases" section of the translation release section. I do not enforce any formatting rules there.
Just make a thread, name it after the game, and put your translation patch in there. You can obviously add more information, but you do not have to.

Also tell me about failed translations, because then there's obviously an issue with the tool that I need to fix. Don't just delete it and move on.

 

[Shisaye]

1.105 fixed a really annoying error that was causing quite a few games running on VX Ace to crash during certain actions. I recommend updating to that one in case you had crashing issues before.
(I obviously always recommend updating, but I know people don't really want to.)

 

[Shisaye]

In case you were also annoyed by the "POTENTIALFILENAME" error code spam, I've completely removed it in v1.112 now.
I've improved escaping and detection of that stuff to the point that it "should" no longer be necessary.

 

[Shisaye]

Concerning v1.119.
This one will feel like a massive downgrade on some games, but it's "working as intended".
Basically it will spam lost placeholder errors, but that doesn't actually mean those cells looked any better beforehand, it just actually tries, and reports failure...

The big issue is and always was custom colors.
When Japanese text suddenly has a word in a different color.
In the actual text box that looks like this:

TEXTTEXT\C[12]COLOREDTEXT\C[0]TEXTTEXT

And the big issue is, how the fuck do you feed that to Sugoi?
If you just feed TEXTTEXT, COLOREDTEXT, TEXTTEXT separately while escaping the color commands, the result is horrible because they need the context of one another to make sense.

If you feed it TEXTTEXTCOLOREDTEXTTEXTTEXT, the resulting translation will be decent, but different length and you have no way of knowing which word or words are actually supposed to be in the color box.
Because Japanese is almost always shorter than English, it will almost always be wrong.

So unless you just remove all colors from all text, your only option is to isolate the color block, translate it on its own, and then feed Sugoi TEXTTEXTPLACEHOLDERTEXTTEXT.
But the only way for that to give the desired result is if Sugoi on one hand leaves enough of the placeholder intact so the color block can be inserted back in, but on the other hand doesn't ignore its existence and actually treats it as a "thing" to integrate in its output.
And that's super inconsistent, hence all those previously hidden bad translations.

 

[mrericsan]

Hi there Shisaye, question. Is there a global text replacement function in the translator. The translator did its job by flagging it as potential file names. But its a 2k+ lines edit, just wondering whether there is a easy way to fix this.

 

[Timurama]

MTL is lowest of the low, just learn japanese

 

[Timurama]

And by the way we have had textractor for ages

 

[Shisaye]

Hi there Shisaye, question. Is there a global text replacement function in the translator. The translator did its job by flagging it as potential file names. But its a 2k+ lines edit, just wondering whether there is a easy way to fix this.

What version are you using? How is it being flagged?
The POTENTIALFILENAME error code was retired in 1.112 and shouldn't actually exist anymore.

 

[mrericsan]

I am using 1.102. But that the flagging is not the problem. it was actually an error that can be fix by
un-translating the file names (I tested it). Issue is those few sound files are all over the place (2000+ times lol) and i have to overwrite it back to the original file name in Japanese. Thus i wonder there exist a global text replacing functions.
P.S. i was also wondering if i can just change the file name to the translated one, will that break the game in any way.

Thx

 

[Shisaye]

I am using 1.102. But that the flagging is not the problem. it was actually an error that can be fix by
un-translating the file names (I tested it). Issue is those few sound files are all over the place (2000+ times lol) and i have to overwrite it back to the original file name in Japanese. Thus i wonder there exist a global text replacing functions.
P.S. i was also wondering if i can just change the file name to the translated one, will that break the game in any way.

Thx

Well it depends on the context/content of the cells.
If they all include something unique to them like a special word or used character you could just use regex in the Ctrl+F menu.

For example if they all include the capital characters "SE" you could clear them all using:

[\s\S]*SE[\s\S]*

If you check the "match case" and "regexp" options, and replace it with nothing.

If they do not include anything identifiable like that, you need to look at their context entries and clear that using a custom script, but that's a bit more annoying.

Which game are we talking about? I would need to take a look at that myself.

 

[Shisaye]

I am using 1.102. But that the flagging is not the problem. it was actually an error that can be fix by
un-translating the file names (I tested it). Issue is those few sound files are all over the place (2000+ times lol) and i have to overwrite it back to the original file name in Japanese. Thus i wonder there exist a global text replacing functions.
P.S. i was also wondering if i can just change the file name to the translated one, will that break the game in any way.

Thx

I've found and fixed your issue in v1.123.
It was just a broken escaper pattern and a complicated replace script is not necessary.
But if you notice something like that again please just report it so It's fixed for everyone instead of just fixing it yourself.

Here's a proof of concept translation using the fix:
https://www.anime-sharing.com/threads/english-slr-mtl-breeding-presser-rj01260762.1599611/

 

[Amniotic]

I have found that using MTools text extraction nearly never results in any breakages. I then just run it through some scripts and batch it to sugoi. Once done MTool reinjects it back into the game.

Instructions:
Load a game via Mtool then navigate to the translation tab. Click export offline tranlation file. It will be save in the games folder. Copy the json to the ordered action scripts folder and then execute the scripts name something like "1. run first". You will now see a bunch of text files in the folder. Cut the text files 0, 1, 2, 3 into the translation folder (leave the temp ones alone). Start up Sugoi then execute the translate.js script. Once it has completed cut the translated files into the ordered actioin folder again and run the script "2. run second". Now you have a translated json to import back into MTool.

Here are the scripts.

Perhaps you could use mtool to extract the text, to then import it into your translator++ fork perhaps resulting in no breakages.

 

[Shisaye]

I have found that using MTools text extraction nearly never results in any breakages. I then just run it through some scripts and batch it to sugoi. Once done MTool reinjects it back into the game.

Instructions:
Load a game via Mtool then navigate to the translation tab. Click export offline tranlation file. It will be save in the games folder. Copy the json to the ordered action scripts folder and then execute the scripts name something like "1. run first". You will now see a bunch of text files in the folder. Cut the text files 0, 1, 2, 3 into the translation folder (leave the temp ones alone). Start up Sugoi then execute the translate.js script. Once it has completed cut the translated files into the ordered actioin folder again and run the script "2. run second". Now you have a translated json to import back into MTool.

Here are the scripts.

Perhaps you could use mtool to extract the text, to then import it into your translator++ fork perhaps resulting in no breakages.

SLR Translator already has support for MTool files.
If you load a game into MTool, you can switch to the translate tab, and then click on "Export the original text needs to be translate". That will create a file called "ManualTransFile.json".
You can load this file into SLR Translator by using the "Json Parser" option.
Once translated you can load it back into MTool using the "Load translation file" button.

This is useful for translating something like Wolf RPG that is currently not supported by SLR.

For actually supported RPGM engines I do not really see the point, because translation quality is just going to be worse than just going through the normal SLR process.
I am not aware of any crashing issues with my Tool unless you've actually made a mistake.

 

[Ikamusumechan]

Yamiyashiki Trial Ver Hello!Sorry for bothering but how can we fix the text overlap seen like this?

 

[Ikamusumechan]

also thanks a lot for this tool and other explanations!

 

[Shisaye]

Yamiyashiki Trial Ver Hello!Sorry for bothering but how can we fix the text overlap seen like this?
View attachment 62737
View attachment 62738

This game is using an incredibly inefficient custom text box. Especially during special cutscenes like the intro.
Unless you want to remake the plugin you can pretty much just wrap it even tighter and/or change the font and font size.
For example during a quick test I wrapped the project to "size", changed the font to stint and the font size to 20.

The result looked like this:

 

[Ikamusumechan]

This game is using an incredibly inefficient custom text box. Especially during special cutscenes like the intro.
Unless you want to remake the plugin you can pretty much just wrap it even tighter and/or change the font and font size.
For example during a quick test I wrapped the project to "size", changed the font to stint and the font size to 20.
View attachment 62752View attachment 62753View attachment 62754View attachment 62755View attachment 62756
The result looked like this:
View attachment 62757View attachment 62758

Thank you!

 

[Shisaye]

Just to have it said.

When I started working on the SLR project the MTL options were really bad and there was even a short moment in time in which my Sugoi/dictionary hybrid was arguably the best option out there.

Now "AI" has evolved to the point that this project feels very outdated.
But I cannot really add support for more modern models because I neither have the necessary hardware, nor the funds to get that hardware. I can barely pay my bills.
Technically someone could provide me a remote server for testing to outsource the processing, but I just do not think that's worth it. Considering this is just for RPG Maker games this project is probably not for someone with a high end gaming rig anyway.

I still think SLR has a reason to exist because not everyone has a high end rig capable of running large models, and so SLR Translator is still the only free, as open source as possible, offline alternative for someone to get something "decent enough" to be able to finish a game without knowing Japanese, but I do not think I will develop it further besides fixing bugs/crashes reported to me. (And please do actually report them to me.)

I had plans to rework Scripts parsing for VX/Ace to make proper automation for it, as well as implementing automatic font changes and other QOL stuff, and at some point before my GPU died I was even working on an OCR system to automate the translation of picture based text, but given the hardware constraints and the complete lack of funding, these changes would ultimately just be a lot of unpaid work for incremental improvements to a project, that already gave up on being "good".

But if someone else has any ambitions for this project or finds parts of it useful, you can obviously change and redistribute any part of SLR however you like as long as you keep it open source. (GPLv3 license and all that.)

 

[anioshi200us]

thx

 

[Amniotic]

The current local LLM in my experience are not as good Sugoi. They get the genders and idioms right if prompted well but they're all over the place contextually. I haven't tried one since summer but that was 14B with a Japanese Qlora at 8 bit quant.

Look into Intels OpenVino framework, it is very fast for OCR on CPUs.

Anyway thanks for the translation software, especially the additional software. I can run sugoi on ROCm now.

Check out CUDO compute, they offer RTX Quadro 4000 for like 0.36 cents an hour.

 

[Amniotic]

Just to have it said.

When I started working on the SLR project the MTL options were really bad and there was even a short moment in time in which my Sugoi/dictionary hybrid was arguably the best option out there.

Now "AI" has evolved to the point that this project feels very outdated.
But I cannot really add support for more modern models because I neither have the necessary hardware, nor the funds to get that hardware. I can barely pay my bills.
Technically someone could provide me a remote server for testing to outsource the processing, but I just do not think that's worth it. Considering this is just for RPG Maker games this project is probably not for someone with a high end gaming rig anyway.

I still think SLR has a reason to exist because not everyone has a high end rig capable of running large models, and so SLR Translator is still the only free, as open source as possible, offline alternative for someone to get something "decent enough" to be able to finish a game without knowing Japanese, but I do not think I will develop it further besides fixing bugs/crashes reported to me. (And please do actually report them to me.)

I had plans to rework Scripts parsing for VX/Ace to make proper automation for it, as well as implementing automatic font changes and other QOL stuff, and at some point before my GPU died I was even working on an OCR system to automate the translation of picture based text, but given the hardware constraints and the complete lack of funding, these changes would ultimately just be a lot of unpaid work for incremental improvements to a project, that already gave up on being "good".

But if someone else has any ambitions for this project or finds parts of it useful, you can obviously change and redistribute any part of SLR however you like as long as you keep it open source. (GPLv3 license and all that.)

Also I have an old AMD RX 590 8GB that definitely works and an AMD Vega 56 8GB which are supported by ROCm 5.7 and DirectML (AMD r9 290 too but that's ancient and I think it's dead) sat in a draw doing nothing. The Vega 56 needs the thermal paste and thermal pads changing but it functions but has more compute and memory bandwidth than the 590. I'd happily give either to you for free (if you cover shipping) if you want one, but understand if you rather keep your anonymity or if shipping is expensive.

On the other hand now I have Ctranslate2 working with ROCm, I'd happily do the translation and then send you the files to do the fixes and word wrap. I wouldn't be reliable as I go through phases where I don't have any interest in playing games for a couple of months, then for a few weeks I'll play loads.

Alternatively I'll gift some sheckles so you can use cloud computing. It'd be in a cryptocurrency though, so you'd have to convert it dollars/sterling/roubles/yen/euros or whatever.

Just let me know.

Apologies if anything in this post is breaking any site rules.

Btw Google cloud gives you like $200 free credit when you sign up for the first time. They have both NMT and LLM finetuned translation APIs.

 

[Shisaye]

Also I have an old AMD RX 590 8GB that definitely works and an AMD Vega 56 8GB which are supported by ROCm 5.7 and DirectML (AMD r9 290 too but that's ancient and I think it's dead) sat in a draw doing nothing. The Vega 56 needs the thermal paste and thermal pads changing but it functions but has more compute and memory bandwidth than the 590. I'd happily give either to you for free (if you cover shipping) if you want one, but understand if you rather keep your anonymity or if shipping is expensive.

On the other hand now I have Ctranslate2 working with ROCm, I'd happily do the translation and then send you the files to do the fixes and word wrap. I wouldn't be reliable as I go through phases where I don't have any interest in playing games for a couple of months, then for a few weeks I'll play loads.

Alternatively I'll gift some sheckles so you can use cloud computing. It'd be in a cryptocurrency though, so you'd have to convert it dollars/sterling/roubles/yen/euros or whatever.

Just let me know.

Apologies if anything in this post is breaking any site rules.

Thank you for the offer, but I really don't think that's worth the trouble.

To reasonably run the type of LLMs I would consider superior to Sugoi, I would need something comparable to an RTX 3090 and then I would need to completely remake large parts of the translator because the way text is currently fed to the model doesn't give it enough context. (Sugoi works really well with short isolated inputs, but those "AI"s really do not.)

I have no idea how long that would take, or if I would even have the skill to make it work at all and so I really wouldn't want someone else to pay for the time it takes me to throw shit at all wall hoping something sticks.

It's also not like this translator is something world changing here anyway and from what I've seen in the last 3 years demand is also pretty low.
People used me to get quick free low effort translations when I was taking requests on F95, but interest in the actual translator was always very low.
The few donations I've gotten until now did not even cover the power consumption during processing, let alone the 2 GPU's I've killed while working on over 400 different translations during that time.

I should probably do the responsible thing and instead of continuing to sink all my free time into this, get a second job and actually get some savings so I can pay for "hobbies" like this myself.

 

[Amniotic]

Thank you for the offer, but I really don't think that's worth the trouble.

To reasonably run the type of LLMs I would consider superior to Sugoi, I would need something comparable to an RTX 3090 and then I would need to completely remake large parts of the translator because the way text is currently fed to the model doesn't give it enough context. (Sugoi works really well with short isolated inputs, but those "AI"s really do not.)

I have no idea how long that would take, or if I would even have the skill to make it work at all and so I really wouldn't want someone else to pay for the time it takes me to throw shit at all wall hoping something sticks.

It's also not like this translator is something world changing here anyway and from what I've seen in the last 3 years demand is also pretty low.
People used me to get quick free low effort translations when I was taking requests on F95, but interest in the actual translator was always very low.
The few donations I've gotten until now did not even cover the power consumption during processing, let alone the 2 GPU's I've killed while working on over 400 different translations during that time.

I should probably do the responsible thing and instead of continuing to sink all my free time into this, get a second job and actually get some savings so I can pay for "hobbies" like this myself.

Offers always there but yeah seems to much effort.

But why would you need a 3090? The vega managed 40 t/s with vulkan.

All the best translation llm models at the moment are under 10b. A 7/8b model (like bigwebdatas ja-en Gemma on huggingface) in 4 bit gguf is around 4.5GB. Add to that your dictionary using embedding a temp of 0 and some form of chunking for context you'd easily fit that in 8GB of vram.

I know nothing of parsing text or regex or whatever else you have to spend your time on. So yeah probably not worth the hassle.