Making scripts is incredibly time consuming, and very repetitive.
In fact, it seems like something AI should be well suited for: the input is lots of pixels with big swaying movements of input values, and the output is one (or a few for OSR2) values that basically mirror those input movements.
Seems like it should be possible to identify “cocks” (or hips, or tits, or dildos, or hands) in video, and match movement to any cocks or hips, or the difference between them, and produce an output value relatively easily.
It has been tried by a number of people/groups with little success. Two raw/needed to much touch up afterwards. If someone can make it work great, I think we are a ways off still from the individuals I have talked to.
The tricky part is that all training has been made seems to recognize faces or full body only. You need to train your AI yourself first and that is extremely time consuming and requires lots of prepared raw material to train on. You often don’t see the full bodies and they are in all kinds of positions and angles. All that you need to train your AI for. In case of VR you have the additional problem with a distorted raw video.
There are projects trying to render a skeleton on people in videos (e.g. OpenCV together with OpenPose), but those are not detailed enough to get something that is suitable to generate a detailed script.
I actually have a prediction that eventually, there will be enough scripts made that it will be possible to train a ML model to create scripts based on existing scripts.
You could categorize the ‘type’ of action that a video segment is “handjob, blowjob, titfuck, vaginal, anal, footjob” (there’s already an AI that has done basically this), and then have a series of models that you’ve trained on funscripts matching video segments of that type to estimate what the funscript position value should be for that frame.
Combined with some filtering and whatnot, I reckon you could get very acceptable funscripts with that kind of method! There might even be enough data to do something like that now. It would be so much work though
Deep fakes are the closest i’ve seen so far in model training (face train) and tested myself. The only thing I could think of working would be to train a mask of how much the dick is showing and then base the stroke for this.
Do you have something of this you’d like to share to improve or give feedback if it’s still in progress?
It can eventually work but you need a dedicated group to get it right and probably a lot of investment while only a few companies will buy the end product. And even then it will still need to be checked for mismatches. Certainly in VR where it is sometimes imposible to see the cock or the lenght of the movement, as a scripter you already have to guess, computers will have a lot of trouble with those parts.
