edger477 wrote: Wed Jun 14, 2023 8:36 pm
They are normally putting out "xPT model" (x pre-trained model). For our use case, we would need to take source code and build model from scratch by feeding it different kind of files to analyze. It is quite a lot of work.
Indeed, pre-trained models won't work at all. We would need to build a library of annotated stim file parts (morphones) for the AI to learn (i.e. what feels good, what feels like a sharp sting, what is a slow/fast/steady/... ramp-up etc.).
To generate a good model, we need lots of data but since the audio is enough, we probably have a good set of data, even if it is still a relatively small set compared to others (still Dynascore also was trained with only about 1000 pieces from different genres).
gynephilos wrote: Wed Jun 14, 2023 9:10 pm
The thing I'd like to have with an AI thing is biomedical feedback from the cock hero player. This would enable the AI program to customize the experience and edging levels for the individual player.
That would be amazing, but will take even longer. There is work on edging detection and some progress on various Discord channels though. Basically the goal would be to prolong the plateau phase nearly indefinitely. But to do this live will be a real challenge.
diglet wrote: Wed Jun 14, 2023 9:34 pm
This methodology is really not suited for generation of stimfiles. Stimfiles have very specific characteristics with (in practice) very limited degrees of freedom. These AI's are intended to be used in applications where those characteristics aren't known or their dimensionality is too large to properly encode it.
The strength of AI is exactly to detect correlations that are unknown. Then they can be used to make a new piece of data that preserves this correlation but creates a new sensation (in case of audio file) that works just as well but maybe feels completely different. Or (more likely) feels like a variation of the input.
The goal with annotated training is not to create a stim file from the video contents directly (that's one step further and eliminates a lot of freedom for the creator) but rather to describe the stim experience for parts of the video (e.g. "from 1:20 to 2:24 generate a stim file like ~some anntoated stim file name~ with a steady ramp up, no breaks, no shocks, base frequency 700 Hz, triphase, stroking").
