🤖 FunGen - AI-Powered Funscript Generation - VR & 2D POV - Join the Discord

[k00gar]

I’ve always been open to criticism, viewing it as a path to progress.

However, I believe there may be a misunderstanding about the AI component of this project and its potential for significant near-future enhancements.

The AI layer primarily involves Computer Vision, which identifies body parts in images. This component won’t determine, for example, whether a hand detected near the penis area is interacting or not. That assessment is made by the second layer of logic.

This second layer handles tasks such as filling detection gaps, managing flicker and false positives, assessing distance, prioritizing, cleaning the signal, amplifying, normalizing and more.

This layer is purely algorithmic, not AI-driven, and consists of thousands of lines of code addressing hundreds of edge cases. In the current version under development, the AI component represents only the first of ten processing stages.

I haven’t viewed this project through rose-tinted glasses for quite some time, I can tell you

I / We’ve offered this for free with no asking back, and I’ve increasingly considered stepping back from it.

Just trying to stay motivated, even when some simply complain about the installation process being a pain in the B, or when the results are erratic.

Both complaints I fully understand, but I gladly accept help on both, from whomever is happy working for free on this aside from their daily work and other real life responsibilities

[Rriik]

I think the most reasonable way to go about it is put big ass disclaimers everywhere that this is a WIP and done for free as a passion project. You cannot change people complaining or giving criticism (be it constructive or not), but you can excuse yourself for only being human. Draw motivation from what you do and from the progress you do make, rather than letting the voices of others pull you down. And otherwise just tell people to read the disclaimers.

People should also ask themselves the following: would you rather prefer this technology be made available to you all for free, decentralized and democratized, or would you instead prefer it be abused by money sucking patent-ridden anti-competitive anti-consumer sextech corporations? I know what i would prefer. And if you do too, just have patience and support those who can make this possible. Rome wasn’t built in a day, but it did survive the testament of time.

[justsoicanfav]

My intention wasn’t to demotivate anyone. I was simply stating where it’s at now and it was meant for people who come to this thread and might want to give it a try. It’s best for all involved that people know what they can expect from the tool. I should think everyone would agree with this sentiment. All the optimistic hype from some users here does is get people funneled to try to use it with unrealistic expectations and then you’re going get bombarded with the kind of criticism that is hard to take as constructive.

It is impressive what it can do with some parts and it’s loads better than other solutions we’ve had for it. Especially if you’re finding solutions to convert the YOLO model results into accurate funscripts without using AI, but creating clever algorithms. I just don’t see it as something I can use for scripting YET. Big emphasis on the word YET. I feel the sentiment is entirely fair and not harsh at all and I do apologize if it was taken that way.

[k00gar]

To be crystal clear @justsoicanfav : no offense taken, no worries.

Just as a reminder, the tool was initially designed for non-scripters users, to simply get a fast script for whatever video they might want.

Then, we started getting some interest from scripters, and we are really flattered about it, but it comes with a whole next level of expectations (quality of endpoints timings, amplitude, etc.) and we are trying to take it to the next level also from feedback collected from this non-initially expected expert user base.

I will try and release an alpha sneak peek branch today or tomorrow based on the PROD branch, only for feedback, with no support at all:

• sex position isolation + discarding of unwanted body parts
• sex position bookmarks in the funscript (cuepoints)
• cleaner signal
• amplification of signal, within wanted boundaries (100 - 40 for blowjob, 100 - 0 for most others)
• adjusted penis locking logic
• smooth transition system between body parts switch during blowjob

This could be a way to offer some progress while we are working on the fully rehauled version, which will take much more time to come out.

[roa]

please replace me - scripting takes too much time

[golfball]

I have written scripts for SLR so i know what i am doing, but i am afriad i am not being succesful at installing this software, is there a walkthru, i cannot follow the instructions on Github

[justsoicanfav]

If you’re comfortable with Discord, I would suggest joining their server. There are experienced people there who are really helpful. It can be confusing to get it set up if you aren’t used to python or using command lines and terminals. Making the install easy for ordinary folks would probably halt the development of the actual features for a long time. @golfball

[justsoicanfav]

That’s fair if it wasn’t made for that use originally. And I can also understand that not everyone needs/wants the level of accuracy you get from an experienced manual scripter. I would obviously also prefer being able to generate something and not have to waste time doing manual work.

[k00gar]

Really appreciate the mention @Alexus , again, thanks a lot https://discuss.eroscripts.com/t/swallowbay-slr-savannah-bond-blowjob-savannahs-peaches-and-meringue-multi-axis/246944

Also, the crash-dummies alpha test version mentioned below is now live. Advanced technical users only, no support provided, use with caution, not a final product, etc.

I will try and release an alpha sneak peek branch today or tomorrow based on the PROD branch, only for feedback, with no support at all:

• sex position isolation + discarding of unwanted body parts
• sex position bookmarks in the funscript (cuepoints)
• cleaner signal
• amplification of signal, within wanted boundaries (100 - 40 for blowjob, 100 - 0 for most others)
• adjusted penis locking logic
• smooth transition system between body parts switch during blowjob

This could be a way to offer some progress while we are working on the fully rehauled version, which will take much more time to come out.

[hectorgh]

Thanks for sharing!
My 5090 is ready to tes this tool

[StillHorizon]

I hope this will help those who perhaps prefer a more visually guided approach to installing this software,

Step 1

Step 2

Step 3

step 4, join the discord to gain access to the model to be used (assuming you just want to get it up and working without any optimization, just get the n size model ending in .pt, place them in the “models” folder)

Step 5
Click view → settings check the Yolo model is pointing to the .pt file correctly

Step 6 (last step)
Restart by closing the GUI, and run the python FSGenerator.py command again in the terminal!

The above follow closely to the provided instruction in the github page. The commands ran in the image also assumes you have a cuda capable GPU (e.g. NVIDIA). Hope it helps out if anyone is interested.

Might be outdated if the newer version is out, however it should be working as of now. If you face any error, if you search your error/question in the discord it should likely already be answered.

[Narru]

This tools great for missionary, doggy and also cowgirl when it’s not just grinding. It really struggles with blowjobs and handjobs though. Hopefully future updates make these parts better. At the moment you really need to re-work a lot of the script in OFS after.

[tanner]

Project update?

[k00gar]

Sure.

We fully rewrote the major part of the code base in an alternate version. We have solid signal, good performance, but are still testing and experimenting with pose models and optical flow, with JAV VR in focus.

Still a lot to do though, and very less time as our jobs got quite overwhelming for the past few months leaving us less time for this project than we would like.

Anyway, I also released an experimental crash-dummies version few weeks ago and posted a few teasing pics about the above mentioned experiments.

Also, got a few feature requests in the Discord that I intend to fill in as soon as I can.

I understand this is disappointing, but we are taking it slow to avoid losing interest or simply getting a bit burnt out by this project.

It will be ready when it’s ready

Stay tuned.

[cyberbloney]

if a pov scene was occluding the members, aka a closeup where the camera doesnt see the genitals, im guessing it wont tack anything either? Or would. Itninterpolate from the rest of the torso as well?

[k00gar]

We could either track the boobs or use optical flow to get a global movement as an additional approach (which we are currently experimenting to also deal with occlusions).

On a secondary note, the previous pipeline allowed me to run 8k VR footage at 40 fps on my Mac. New experimentation today is showing 150fps. 2D 4k at close to 250/300 fps. Video reading + yolo detections.

Stay tuned.

[k00gar]

This is getting a lil bit out of hand haha

Here is a sneak peek at the version in development. Allows you to generate and edit the funscript right from the app, load an existing funscript and inject inside it, etc.

[ggboy]

[cyberbloney]

is it gonna be an exe file or is it gonna be another github release?

[_toast]

I’ve just found this thread and it has been fascinating! Might I ask I question regarding your missionary / close-up logic?

I’ve created a Yolo11m Model based on currently 5500 hand-picked and annotated frames that detects 7 body parts (it took forever, but learning to work with cvat, yolo, runpod etc. was fun). I’ve also written a small tool for the video inference/label-data generation.

I’m currently planning the broader logic and edge cases, but I’m undecided on how to approach the ‘penis is 90%+ outside the visible area’ scenes, like close-up missionary or standing missionary where the focus is on the face/upper body.

My first idea was to pre-process the video with ffmpeg in order to move the viewpoint a bit down, but that would only fix one or two missionary scenes and not the general ‘estimate where the penis is and how far it penetrates/is covered’ problem. Most missionary scenes I’ve used are clear enough in the lower half to detect the penis/vagina without pre-processing, but many are just too blurry on the bottom for even the human eye to accurately say ‘yea, thats a dick’.

So my current approach would be to estimate the position/penetration based on a not yet specified amount of time before the penis left the viewpoint and continue that pattern (with some variation) until the penis re-enters the viewpoint.

Could you give me a hint on how to best proceed here? I’m sure you’ve tried a fair amount of solutions already.