Funscript AI Generation - VR (& 2D POV now?) - Join the Discord :)

[k00gar]

FYI, @spatialflux and I are planning to release the new version of the AI model this week-end.

Expect important quality of detection improvement!

Also, we are working on a full rehaul of the “tracking” logic, this will take more time, but in the end should bring faster processing and major funscript output improvements.

I will also make a proper announcement, but currently as a beta test, we have a bot with us in the Discord!

You can use it to upload, search, download, request funscripts etc.

Be in touch with it via PM / direct messages only, starting by !help for the list of commands

ENJOY!

The bot just started logging downloads btw:

[k00gar]

New YOLO Model Update – More Reliable, More Precise!

@spatialflux and I are thrilled to announce a major upgrade to our YOLO detection model, bringing better stability, and more reliable tracking across all classes!

What’s New?
More Stable Training – We extended training with more epochs and a lower learning rate, ensuring the model learns deeper patterns without overfitting.
Enhanced Object Recognition – We tweaked multiple hyperparameters and switched to a new optimizer, making the model much better at detecting less common objects.
Smarter Fine-Tuning – We ran a dedicated fine-tuning phase without mosaic or scale augmentation, helping the model handle real-world cases more effectively.

Key Improvements:
Flicker Ratio Reduced – Smoother, more consistent tracking with fewer ID switches.
Better Track Stability – Fewer unique track IDs = more reliable object persistence across frames.
Increased Detection Performance – Higher accuracy in recognizing faces, hands, and key body parts.

This update is a huge step forward in model reliability and brings us closer to seamless, flicker-free tracking!

In the attached screenshot, left side being the new model, you can expect improvements also on BJ scenes with better face detection and tracking!

As usual, download the models from the #resources from the discord, place them in your “models” folder, and don’t forget to change the model you are using via the Settings Panel !

Introducing AI Funscript Bot!
Upload your AI generated scripts to share with the community.
Request new scripts and let the AI generate them for you.
Browse and download from an ever-growing collection!

Exciting times to be alive!

Give it a spin and let us know your feedback!

[k00gar]

Our Masterplan Roadmap is Live!

We’ve just shared our “Masterplan” roadmap over in ⁠Discord.

If you’re interested in contributing or have ideas to share, don’t hesitate to poke @spatialflux and I — we’d love to have you on board!

[Alexus]

You are doing an amazing job – this is the evolution and future of this site. I hope you achieve great success, and (jokingly) that you won’t be stopped by people making money from complex frame-by-frame manually scripts.

[Rriik]

Honestly i don’t believe this will cause an uproar in the community, IF done right. And that means to keep this tool as open as possible, involve the community in the big-picture direction as much as possible, and always make it clear that the humans always have the power (and full rights, in case doublevr reads this, lmao) to use and change whatever the AI generates. AI is best used as a tool to speed up workflows, not as a replacement for human intuitiveness and creativity.

So far the creators are definitely following that path, and I think it will continue to grow as more people find out about this. Open-source projects like these have 100% the possibility to overcome challenges and become greater than the closed-source scams that normie sex-tech companies usually like to promote on these forums. The community should benefit from this the most, we still have the power to not let this niche keel over to big tech monopolies.

[k00gar]

Thanks, @Alexus! This has now become a (very small - we’re two now) collaborative effort, with incredible support from the community.

We’re making the most of our free time between work and family commitments to give back to the community.

The response from scripters and users has been very benevolent, with a warm welcome and plenty of constructive feedback.

There are still many questions to answer and challenges to tackle, but we’re excited to keep pushing forward!

[k00gar]

Committed this fix in the main branch :

→ Fixed state.frame_end ghosting in the GUI, which was leading to early cut of processing of videos after a shorter one was already processed in the same GUI instance.

[k00gar]

Latest comparison graphs on a passthrough video.

• In red = man made posted on eroscript today (awesome, bumdude as usual, thanks!)
• In blue = our pipeline.

And we are currently fully rehauling the logic, we’ll make it even better

[Clava01]

Hy easyest way to install possible create exe file?

[cprn.1337]

Thank you for this, I just started generating my first funscript.

I have a ton of VR vids waiting to get scripted.

[k00gar]

Haha, nice.

My advice would be: don’t script them all at once

We are working on an even better version, yet to be released.

Might need some couple + weeks though.

[cprn.1337]

I played around with it today, and it still needs some work, with time it will probably get really good, below you can find a comparison for a paid script vs generated, in certain positions it works great, in others not so much. Still I enjoyed playing around with it.

[GoNikky]

Hey @k00gar, have you considered combining your approach with @Zalunda’s Motion Vectors Plugin (How to use FunscriptToolBox MotionVectors Plugin in OpenFunscripter - #92 by Zalunda)? It might work better than the tracking boxes alone?

[k00gar]

Not yet, and the latest version of the AI body part detection model is quite strong and reliable already.

Short misses on frames can be filled by Kalman filters, as well as short obstructions.

But longer obstructions might need a different approach with a flow analysis.

We will definitely look into that direction once we will have pushed our current approach as far as we can, and see what is left to be filled.

[k00gar]

For sure the latest section is way too shallow in terms of strokes, but I do not see many strokes missing

Absolutely agree on some sections being hit or miss though.

If you have any of those examples (video reference, timestamp, issue), please share in the Discord, as we are trying to build a benchmark library so as to evaluate the next big update in the making.

[Zalunda]

I’ve been thinking about using NVidia Optical Flow to get motion vectors but I haven’t found the time yet. It seems that it can be combined with an object tracker AI.

NVOFA Tracker - NVIDIA Docs

My understanding is that it can improve the object tracking frame-by-frame by using the motion vectors to guess where an object will be in the next frame (i.e. if the motion vectors “under” the object is moving 10 pixels to the left, the object is probably 10 pixel to the left on the next frame). But I’m not sure if its only to speed up the process or if it does a better job overall.

I don’t know, maybe at some point we’ll get reliable multi-modal data that could be fed to a machine-learning algorithm:

• Motion vectors
• Object tracking
• Pose tracking
• Depth tracking
• Audio tracking (i.e. track ‘fap-fap-fap-fap…’ sound during sex, ‘pop’ sound during blowjobs, etc).
• Reference funscript

But I agree with K00gar that it’s probably better to fine-tune each approach and see where it can get us. It’s already challenging enough.

[Dimava]

Is tgere a way to integrate this with Blender?

As OFS is now a legacy software I think it’s time to consider how a future one should look like