- Images
-
According to OP themselves:
Do you remember Baka? It was a PoC character which people liked her, but she was based on an online service that got taken down, so she got dropped. So I thought I'd make a remake, but this time with Kohaku!
This time, the AI is local and much more is included: Emotions, actual responses, customization and much more!
There's a server instance running locally on your machine with no internet connection and nothing is sent anywhere.
A chatbox window appears where you can type something. This is sent to the model and Kohaku answers in character. There are no scripted lines, an actual model generates actual responses.
She's modelled on the character Kohaku from Tsukihime: a cheerful maid on the surface with a sharper side underneath.
Her replies come out in the style of a visual novel with little action cues woven in between the dialogue.
Her emotions update in real time too, read straight out of the shared memory into the CNS, so the sprite reacts to her mood in each response.
She also SORTA remembers the conversation, at least for a while (it depends on the context size, 16384 should be fine?).
Every message reloads a rolling window of recent exchanges into the model, so she's not answering each line completely blind.
However, that window is capped, so older stuff eventually falls off to prevent the context from becoming excessive, so don't expect her to recall everything!
In terms of performance, I didn't want this to be a "works great on my machine, breaks everywhere else" situation, so she auto-tunes herself: the context size (i.e. how much she remembers and how long she can ramble) scales off your free RAM automatically, unless you override it in OPTION.ini. The thread count scales off your core count independently of that setting.
If you set something that your system can't handle, it will warn you and let you decide whether to send it anyway or fall back to something safer instead of just crashing.
The choice of model is also up to you. Smaller 1B models keep things fast and light, while bigger 3B+ models offer a more coherent experience if you have RAM to spare, albeit at the cost of slower replies. Everything user-facing lives in OPTION.ini, same as any other character option, no digging through code required!
Always remember that your first reply will be a little slower than usual, but your subsequent replies to Kohaku will be faster.
On my website, in the “Tools” section, there will be two more models where all you need to do is change the template name to "Kohaku" and place it in the "Model" folder within the System folder.
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Quick Questions:
- The loading animation froze?! Did MUGEN froze?!
Neither the engine nor the character has frozen, it's just loading the model! Wait a bit!
- It's all black, nothing appears.
Restart MUGEN and try again.
- Why is the text so poorly positioned?
That's because this character was designed using MUGEN's default font, not the custom fonts from other screenpacks.
In any case, it shouldn't be too much of a problem.
- Sometimes Kohaku's replies include strange symbols...
I can't fix this! Because it's the AI itself that's doing it! The responses aren't scripted but are generated on the fly,
and since the AI models were trained on a lot of different factors, this is what happens!
Do you remember Baka? It was a PoC character which people liked her, but she was based on an online service that got taken down, so she got dropped. So I thought I'd make a remake, but this time with Kohaku!
This time, the AI is local and much more is included: Emotions, actual responses, customization and much more!
There's a server instance running locally on your machine with no internet connection and nothing is sent anywhere.
A chatbox window appears where you can type something. This is sent to the model and Kohaku answers in character. There are no scripted lines, an actual model generates actual responses.
She's modelled on the character Kohaku from Tsukihime: a cheerful maid on the surface with a sharper side underneath.
Her replies come out in the style of a visual novel with little action cues woven in between the dialogue.
Her emotions update in real time too, read straight out of the shared memory into the CNS, so the sprite reacts to her mood in each response.
She also SORTA remembers the conversation, at least for a while (it depends on the context size, 16384 should be fine?).
Every message reloads a rolling window of recent exchanges into the model, so she's not answering each line completely blind.
However, that window is capped, so older stuff eventually falls off to prevent the context from becoming excessive, so don't expect her to recall everything!
In terms of performance, I didn't want this to be a "works great on my machine, breaks everywhere else" situation, so she auto-tunes herself: the context size (i.e. how much she remembers and how long she can ramble) scales off your free RAM automatically, unless you override it in OPTION.ini. The thread count scales off your core count independently of that setting.
If you set something that your system can't handle, it will warn you and let you decide whether to send it anyway or fall back to something safer instead of just crashing.
The choice of model is also up to you. Smaller 1B models keep things fast and light, while bigger 3B+ models offer a more coherent experience if you have RAM to spare, albeit at the cost of slower replies. Everything user-facing lives in OPTION.ini, same as any other character option, no digging through code required!
Always remember that your first reply will be a little slower than usual, but your subsequent replies to Kohaku will be faster.
On my website, in the “Tools” section, there will be two more models where all you need to do is change the template name to "Kohaku" and place it in the "Model" folder within the System folder.
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Quick Questions:
- The loading animation froze?! Did MUGEN froze?!
Neither the engine nor the character has frozen, it's just loading the model! Wait a bit!
- It's all black, nothing appears.
Restart MUGEN and try again.
- Why is the text so poorly positioned?
That's because this character was designed using MUGEN's default font, not the custom fonts from other screenpacks.
In any case, it shouldn't be too much of a problem.
- Sometimes Kohaku's replies include strange symbols...
I can't fix this! Because it's the AI itself that's doing it! The responses aren't scripted but are generated on the fly,
and since the AI models were trained on a lot of different factors, this is what happens!
