How are machine translated romhacks made?

Beavis

New Challenger
Level 0
16%
Joined
Oct 22, 2024
Messages
4
Level up in
21 posts
Reaction score
6
Points
1,052
I know the use of AI is frowned upon but I'm curious about how it's done? Are there even any advantages to using ai for a translation romhack compared to doing it the traditional way?
 
As with a lot of things, it depends on how you use AI itself for the project.

You could ask for it to make the patch directly, script included. Entirely possible in some cases depending on how the game files are stored internally, but the quality will be sorely lacking if no proof-reading is done while comparing it to a trustworthy script.

Ideally, you'd want AI to assist you with making the tools that you'd use to insert the game text yourself and then repacking it into the game proper. That way you'd avoid years of learning to code while still guaranteeing that the final script would be something you approved.
 
As far as i know, the hacking part (like extracting and re-inserting the texts of the game) is still done the traditional way... though AI might be used to help creating the necessary tools.
(I'm not sure how useful it might be in that area)

The main part that is done by AI is translating the text.

Actually it's nothing new... using tools such as Google Translate has been pretty common in the translation scene.

In the past, romhackers had to find a human translator to take care of translation, while the romhacker would focus on the hacking parts...
However, it wasn't uncommon that either the translator or romhacker would stop working on it, and the project would become dead...

What has changed now is that AI can produce a level of acceptable quality of translation... while the older tools wasn't able to do so.

Thanks to it, the bottleneck of coupling a human translator and a romhacker is lifted.
Romhackers can prepare the translation tools, and provide a bare-bones translation using AI, and leave the quality work and polishing to anyone with translation skills who is able to do so.
 
it can speed up the process, but in my opinion, it should be used by someone who understands the original language so they can double-check the translation before releasing it.
 
it can speed up the process, but in my opinion, it should be used by someone who understands the original language so they can double-check the translation before releasing it.
yes this makes sense I don't think any quality translation could be made if someone doesn't proofread it.
 
As @JustLooking123 and @Atma said, it varies a lot, and AI can definitely speed up the process.
I recently completed a sort of translation mod for an indie game called Late Homework, although I never released it because of a couple of Unity-related technical issues that, to be honest, I never cared enough to overcome.

What counts as "the traditional way" also depends on what you mean by it. Fan translations, official translations, and localizations (which aren't the same thing) all vary greatly in both workflow and quality.

First of all, CAT (Computer-Assisted Translation) tools aren't new, and neither is machine translation. Long before LLMs, translators used CAT tools with translation memories, terminology databases, glossaries, and dictionaries to keep terminology and phrasing consistent across a project. Nowadays many CAT tools can integrate AI providers like Anthropic, Google, OpenAI, DeepL, or even services like Linguee through APIs, allowing you to get contextual suggestions based on your own prompts (e.g. "Translate this 1990s RPG while preserving a slightly archaic tone."). They're powerful tools, but they only take you so far if you haven't decided your translation approach.

That's really the bigger issue. Machine translation has existed for years in video games, fansubs, editorial work, and the translation industry in general, but you still have to decide how you want to translate.

For example, Sweet Home (the Famicom horror RPG that inspired Resident Evil) has multiple English translations. One follows the Japanese script as closely as possible, while another deliberately adapts parts of the script to match the movie more closely and make certain scenes less obtuse for English-speaking players. The latter project even documented the reasoning behind those decisions. Those are examples of different translation philosophies—roughly speaking, equivalence-oriented translation versus Skopos theory (and of course, these aren't the only approaches, nor are they mutually exclusive). One prioritizes preserving the source text, while the other prioritizes communicating effectively with a specific audience.

The important point is that no tool—a dictionary, a CAT program, or an AI model—can make those decisions for you.

You can see the same thing in older Nintendo of America localizations. Many games from the '80s and '90s replaced Japanese jokes or references with American pop culture or puns. Some people find those charming, while others think they've aged poorly outside their original cultural context. An AI assistant can help explain those references or suggest alternatives, but if you lack cultural and linguistic knowledge of both the source and target languages, the result can still feel awkward or forced. LLM quality also varies by language; most models are strongest in English and are still noticeably weaker when handling the nuances of many other languages.

That's also where translation and localization diverge. Translation focuses on rendering the text into another language, while localization adapts it to the target audience's culture, expectations, and experiences.

On the technical side, AI can also be genuinely useful. My own project stalled because I never re-enabled the Unity scene needed to switch languages, not because translating the script was the difficult part. I could probably throw together a quick proof of concept today with a mod loader and ChatGPT or Gemini, but handling the technical side still requires some understanding of the game and a willingness to read documentation. These tools won't do all the work for you, but they can point you in the right direction and save a lot of time.

Ultimately, I think it's less about the tool and more about how you organize your project. If your terminology stays consistent, your character voices remain coherent, you have a clear translation philosophy, and you know who your audience is, then the tool comes second. Looking into localization kits (lockits), terminology management, or even how professional CAT workflows are organized will probably teach you more about modern translation than debating whether AI is "good" or "bad." Even a well-organized spreadsheet can be enough if you're disciplined.

P.S. The recent wave of Metal Gear Solid (PSX) modding was initially helped by Claude. People used it to assist with reverse engineering and tool development, and now they've reached the point where they're building their own editors and utilities. I think that's a good example of AI acting as an accelerator rather than a replacement.
 
What game / console do you have in mind?
Ask ChatGPT about your project, it will guide you and help you step by step.
 
I know the use of AI is frowned upon but I'm curious about how it's done? Are there even any advantages to using ai for a translation romhack compared to doing it the traditional way?
It would be easier and quicker to ask an AI this question, it would give you a coherent reply as long as you're specific about the question you're asking it.
People are bad at explaining and you got smart-asses saying stuff like "in order to make bread from scratch, you must first create the universe". People complain about AI slop, but if you think about it, it is people who are slop.
 
I personally just call them LLMs. The algorithm is really good at finding patterns, which is exactly what you want when translating something.

What I’ve been doing is scanning the RAM while debugging the game, then creating a log that gets passed to an LLM to find patterns. After that, I manually create a replacement. With today’s models, you can even have the LLM write a replacement tool for you, but I prefer doing it manually because I learn a lot from the process as a game developer.

When it comes to translation, the LLm follows your system prompt, meaning the quality of your prompt directly dictates the quality of the translation. Currently, AI chatbots like Gemini or ChatGPT are widely used in Japan, so their Japanese language interpretation and grammar accuracy are exceptionally strong.

This makes LLM translation completely different from traditional tools like Google Translate. LLMs understand the broader task based on contextual patterns, so the main limitation now becomes how the game itself is programmed.

And that brings us to the core issue, Japanese characters (Katakana, Hiragana, etc.) are stored as bytes in memory. When you try to translate text into English and patch it back in, you often hit a character limit because the English translation is much longer than the Japanese original, which can cause the game to crash.

To solve this, you have to build another system to generate shorter translations. But this creates a new problem, sometimes your only option is to use a shortened phrase that sounds robotic. It isn't something a human translator could fix either ,sometimes your hands are simply tied by the memory limits.
 
What game / console do you have in mind?
Ask ChatGPT about your project, it will guide you and help you step by step.
Of all the different AI options to use, ChatGPT is by far the worst for something like this. Asking it a very generic question like "how do I translate X game" is going to give you something so vague and generic it will most likely be useless.

The reality here is that translation patches still require quite a bit of human involvement if you want them done well. Every game is different in how it works, so you will need to do some initial snooping around and reverse engineering on your own. Even if you're using AI to help you'll need to do this first step to help guide the AI so it makes good decisions and correct assumptions. If you don't do this, you're going to get a hot mess.

The Grandia Digital Museum AI patch is a prime example of this. The AI took old tools of mine and tried to work with them, but because the person driving it didn't know what they were doing the AI made a lot of bad assumptions resulting in truncated cave speak, wrong fonts being used, menu text running outside the box, bloated script sizes, etc. The things it tried to do on it's own like the graphics decompression resulted in a mess of python scripts that no one else can use as they rely on having a specific memory dump from an emulator.

AI is a tool, it's only as good as the person using it. If you know what you're doing it can be very useful in reverse engineering and making simple tools. I've used it a few times in creating tools for certain custom compression schemes I've ran into in some games. When you don't know what you're doing however, it's going to create slop.
 

Users who are viewing this thread

Connect with us

Support this Site

RGT relies on you to stay afloat. Help covering the site costs and get some pretty Level 7 perks too.

Featured Video

Latest Threads

Smash Remix + Extra released

New characters and New stages in a new released update.

Read more

Does anyone remember the Pink Panther movie, the first one?

When I was a kid, this movie was playing on "Sessão da Tarde," which is a Brazilian movie...
Read more

How much do you care about digital id in your game services?

This is rarely talked about, but independently of backlash, we all feel digital id is coming...
Read more

Forum Hangman!

Oh, yes... another Sammy game xD

Everyone knows how to play Hangman, right?

You have unlimited...
Read more

Hey! Max Payne just turned 25!

I'm a day late but it was released 25 July of 2001 and was quite a groundbreaking game for...
Read more

Online statistics

Members online
262
Guests online
3,832
Total visitors
4,094

Forum statistics

Threads
21,771
Messages
549,926
Members
1,015,754
Latest member
daniel101010

Today's birthdays

Advertisers

Back
Top