Download #1: Mega
This is an unofficial, non-commercial translation patch. This file contains only modified text strings and does not include the original game engine, executable files, images, music, or any other copyrighted game assets. Users must purchase and install a legitimate, authorized copy of the original visual novel to apply this patch.
How To Install : 1. Extract everything in the RAR into your game installation folder
(the one containing SIK_DL.EXE).
2. Double-click install.bat there.
AI Model #1: DeepSeek V4.1 Flash 1017
I've been looking at my VNDB wish list to see what to translate, I added this to my list in 2012 and figure I would work on this.
the ai model took about 20 minutes to dump everything, but it's refusing to translate.
Even though no age is ever explicitly said, just because they present as students it immediately assumes they're underage
Reasoning with it didn't work, so I guess I'll have to try my local model and see how long it takes to translate, which I also don't know how the quality differs between local and cloud.
A look at my gpu while running the model locally, running the rtx 3080 10gb at 90% power limit. Not doing much with the PC while it's translating, randomly looking at tiktok, horny from going through the previous visual novel I released and wondering what I'm going to goon to, I guess if I can hold back and hope this translation completes soon I'll goon to that.
So it took around 1 hour and 20 minutes to translate 12,000 lines locally using a rtx 3080, I'm guessing the cloud model probably would have taken 20 minutes or less, or so I thought.
Turns out it only extracted 100 lines out of 12,000, typed it out which I don't understand why, an hour and 20 minutes wasted.
After about an hour getting no where, I'm going to try a model that might have built in translating.
Attempted to use this model, but even after trying to optimize it to fit in 10gb of vram, it was way to slow, but now I'm trying the same model but a 5.77GB sized version, which after optimizing is working full speed.
Went to sleep, tried other programs and models, but due to the limitation of my rtx 3080, it wasn't going so well, after some AI discussion, I might have found out how to get some openrouter models to translate which I'm trying now
after main failed attempts, I found another method that might work, but it takes around 2 hours to translate
that also failed, but I think I found a method that finally works, I can see the text file being updated with english that deepseek refused to translate.
Here's how openrouter looks running this model which is Qwen2.5 72B Instruct, it has translated around 600/11524, in about 20 mins which cost 1 cent. I'm not sure completely how this works, for the most part I run everything through fcc-claude but this is using a python script using a openrouter api key, I'm curious if I could have just used fcc.
Okay, after lots of trial and error, I finally found a model on openrouter that works.
So I'm running Qwen3.6 Plus, I think I just googled needing a uncensored model on openrouter and google recommended it, ran it through ffc-claude, asked it to translate the file, I mentioned that there would be sexual content that needs translating, it then recommended a few different models
It chose Llama 4 Maverick by default which seems to be doing all the translation currently, I still need to figure out what's best after the translation is done, is Qwen3.6 Plus good enough to repack the file and get the visual novel working, or do I switch back to DeepSeek V4.1 Flash 1017 and let that finish the project and hope it doesn't reject the translation due to sexual content.
A look at pricing, so far DeepSeek has been the quickest and cheapest, but it has limitations on sexual content, I guess I don't really know what Qwen has done so far, and Maverick has the ability to translate sexual content but is much slower than DeepSeek.
This VN was a nightmare to work, the three previous NVs took around $2 each to finish, this took almost $10, lots of issues with word wrapping, text not appearing, text clipping into each other, but I think a big issue was DeepSeek needing to go through errors in the translation file that the other model did.
Things we learned, DeepSeek V4.1 Flash 1017 won't translate extreme sexual content and we'll have to switch to another model which I know how to do now.