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

Software
61

And here for this one, the blowjob/handjob section seems messy, looks like the viewing angle and the kalman filtering logic calls for some more tuning and a different logic.

Also the stroke length estimation logic needs to be reworked, as part of it relies on a rolling window, which in case of non movement can induce a downward or upward (unwanted) slope.

https://discuss.eroscripts.com/t/milfvr-jizzlejuice-jizzlejuice-daisy-fuentes/205280

2400×1600 307 KB

image
image1920×418 79.7 KB

milfvr-daisy fuentes-jizzlejuice-jizzlejuice-3600p-180_180x180_3dh_LR_reencvw_15.0.funscript (234.0 KB)

2 Likes
62

Same remarks for this one, more food for thought:

https://discuss.eroscripts.com/t/naughtyamericavr-katrina-jade-damon-dice-in-american-daydreams/218929

image
image1920×419 76.3 KB

Katrina Jade [ADD] Oct 16, 2024R_6kvr265_reencvw_15.0.funscript (145.4 KB)

63

I understand your point. I program with Arduino for fun. Some time ago, I wrote a long program to upload code to memory using the I2C protocol. It was my first real experience.

Feeling embarrassed and “scared” about what others might say about my code, I gathered my courage and published it on a forum. It turned out to be a positive experience because some people gave me valuable suggestions on how I could optimize it, and overall, it was a success.

In the end, my code (just like yours) does its job.
Could it be optimized and refined? Absolutely! But it still works, and that deserves respect.

3 Likes
64

And last one for today, now I will need a few days to dive into the scripts, isolate segments showing discrepancies and various errors so as to understand what kind of combinations drives the detections and or algorithm out of the road.

https://discuss.eroscripts.com/t/ar-porn-sasha-tatcha-fit-and-fired-up-passthrough/194093

image
image1920×420 78.6 KB

ARPorn_Sasha Tatcha_Fit and Fired Up_4000p_8K_original_FISHEYE190_alphavw_15.0.funscript (166.7 KB)

65

The more I think of it, the more I am getting convinced the pose model would be an awesome complement to the approach I am following now. Happy to talk with you if you’d be interested to join initiatives.
In any case, still working on the code, might be able to share it by end of january, in case it could be of any interest for personal use.
In the meantime, happy to generate funscripts in trade of constructive feedback so as to tweak the algorithm. Thank you!

1 Like
66

Ha ha, thanks again for sharing experiences and confidence building.

I will share it sooner than later and can try and give you a preview if you would like.

It’s just I would like to collect as much feedback as possible beforehand so as to tweak and adjust before making it available, so I could explain what are the possibilities and limits so users are not frustrated and disappointed
 (and I already know my blowjob/handjob logic got broken while tuning other setups
 requiring fixing)

67

Out of curiosity, have you tried testing this on a furry VR video? Just to see how well a model trained on humans works on anthros.

68

I have not even tried JAV-VR yet :sweat_smile:

Please share a link and I will be happy to see and share what it comes up with :slight_smile:

Though, I am not expecting great results, as the training happened purely on “real human” images, part of which jav-vr (let’s say 10% or so of mosaic related stuff).

My guess is that we would need to fine-tune the model for your specific needs, which should be totally doable. Might not be my top priority as I would like to fix the broken parts of the algo first, second, tune the so-so sections, and then the rest of suggestions :slight_smile:

69

I’m really glad JAV-VR is also taken into consideration with this project. I’ve been watching this topic ever since you first posted and I’m really impressed at how fast it has progressed. Thank you very much for working on this and I look forward to trying out when it’s ready.

1 Like
70

Thank you @Ufungus !

My guess is JAV-VR will need far more annotated training pictures than I have trained the model with, and might need some logic adaptation. Currently in the process of testing some with @Emmi, we will try and see what we can make out of it :slight_smile:

1 Like
71

Quick update from the day, while experimenting with three videos provided by @Emmi, 2 of them were JAV and results were not good. Actually, part of the algorithm relies on detecting the glans for transition, ending penetration sections, etc. And a mosaic’ed glans is not helping with the detection, but I will look into it later.

So I decided to focus on the third one, from BadoinkVR.

I noticed an unwanted gap in the generated funscript, had to write a debugger to understand precisely what happened at that very point. The model was actually kind of “hallucinating” thus triggering a mess in the algorithm.

So I started rewriting the tracking logic (again
), with what I would consider better results, but still in the process of refining other more difficult sections and transitions.

In parallel, I am re-running a full detection on another video, with full scale instead of a downscaled version I was using for speed sake, along with a yolo 11m model instead of the n version I was also using for speed sake. Takes much more time that way, but I want to compare results and see if the potential quality gain is worth it or not.

Also noticed the Visvalingam–Whyatt I was using to simplify the signal and lessening the number of positions in the funscript was set with with a factor too strong, thus losing some details. Needs tuning also.

In the videos below, the green line is the original signal, when the blue one is interpolated based on the generated funscript positions.

Here is an example of a properly working section:

When the beginning of this one is a TOTAL mess (the good news being it is only a matter of algorithm and tuning, this can be fixed):

Broken BJ:

Next short term steps:

  • Fix the “grinding” scenes
  • Smoothen the transition between positions
  • Rework the handbjob/blowjob algo
  • 


But now that I have this debugger to make use of log files of 250MB+ per video, this is of great help to assess, fix, tune, troubleshoot, etc. :slight_smile:

2 Likes
72

I tried one of your v15.0 scripts posted above and yes the BJ/handjob is rly broken :smiley:. Looking forward for your next version of generator to test it out :+1:

73

Thank you so much, I really appreciate the feedback.
I know the BJ/HJ is VERY broken at this stage, I have been struggling with what to focus on when a mouth is either sucking the penis or just licking it, and when one or two hands are touching it and jerking it while mouth is there, what should the algo focus on, 
 terrible to automatize.

Not focusing on this part right now, but still trying to think about it and come up with something ok-ish.

Same with grinding / rubbing parts


How about the other parts of the scripts ? Was anything acceptable ?

Been rewriting again some parts today, but not satisfactory yet. Might need a few days out of the project before coming back with impulse and energy.

Thank you again for trying and providing feedback. Feel free to share a video that we could work on if you’d like.

74

I didnt watch the whole video, just checked some parts for 30-40s but yes, some parts were more watchable and not distracting as other.
For example:
doggy ± 15:49m
missionary ± 21:50m and 30m (but the one around 19m was messy)
cowgirl 36m (but no cowgirl 35m when shes leanign towards the camera)
rev cowgirl 38m/41m

1 Like
75

First, thanks again!

Based on the timestamps you provided, I guess this is the Katrina Jade ADD video that you tested.

The messy missionary part seems to be when the penis & pussy are in the very lowest part of the video, thus inducing strong distortion, image recognition is no good, hence messing with the tracking. The plan is to fallback to tracking the boobs, with either a time phasing or a reverse position (boobs at the lowest when penetration at the deepest, and vice versa. Works with bigger boobs, less true with smaller boobs with less inertia though). It was there in a previous version of the algo, but I have not yet reimplemented that when rewriting from scratch the algo a couple days ago.

When she leans toward the camera, same, we lose tracking and details. Need to come up with something smart. And same with grinding that I need to implement, I might need to rely on a pose estimation model and track the movement of the estimated hips. Or do something with the navel when visible.

Anyway, your feedback is truly valuable, thank you again.

Last but not least I am still debating with two approaches:

  • Estimate the penetration depth based on the recent history of movements of the tracked body part distance towards the base of penis (highest is 100, lowest is 0, and every position in between is proratized, while keeping only a history of 200 last movements)
  • Estimate the penetration depth based on the visible part of the penis vs the size detected when penis and glans were last seen. Could be more accurate, yet for extremely well endowed stunt cocks, penetration might not go balls deep
 Hence ruining the approach. Instead, I might need to go with an in between approach, where the highest point is the estimated or known tip of glans, and lowest is the lowest known position of penetrated body part (i.e. pussy, butt, mouth). Handjob calls for something different.

This second approach, before enhancement, gives more nuanced funscripts like this:

image
image3084×192 89.5 KB

Instead of something like this with the first approach:

image
image3102×188 100 KB

Which can also be nuanced depending on the depth history of the body part positions we want to consider for max / min / proratization process


76

Oh, yes. I forgot to mention it was Katrina. Sorry :grin:

The messy missionary part seems to be when the penis & pussy are in the very lowest part of the video, thus inducing strong distortion.

Maybe stupid idea but what about tracking hips/crotch here too like u mention on griding ? When they are on each other its 0 and the rest can be decided by penis-pussy range.
Navel or boobs could work too i guess, but lets say in 21:15m the penis area wouldnt be visible at all, then the boobs position stays almost the same aka very small movement there. So in some cases it would be hard to get data.

Just quick thinking about it, dont take me too serious. I have no experties in this :slightly_smiling_face:

77

No, this is not stupid, but the the pose estimation model might not be able to determine their position either, for the same distortion reason.

However, I just came up with this, the white track being a combination of detected left and right hip.

Here at 19min:

Here at 21min15s:

But here at 35min, it is a mess:

Boy, this is getting painful :sweat_smile:

78

I pictured it somehow like this. But the picture is taken from OFS with VR video mode, so i dont know if it will look the same in your helper.

Pic

Bez nĂĄzvu
Bez názvu877×625 62 KB

Is the cowgirl messy because her hands are in way to see her hips ?

79

No, I mean the case you present seems pretty straightforward and either image recognition and tracking or pose estimation tracking should work ok-ish.

I was referring to the example of the third video of my previous post, kind of grinding.
Image recognition should work, kalman filtering should help with the occlusion, though there is no major movement between the base of penis and base of vagina that the algo can grab. And the pose estimation does not work at all the green and red dots being the estimation of the position of left and right hip as per yolo 11 pose base model. Totally unusable.

Hence, for grinding or whatever we could call this type of interaction, we would first need to know if penetration is ongoing or not. And find a reliable anchor to estimate the movement intensity /depth. Sometimes could be the breast area, sometimes the hips, sometimes the navel, sometimes a combination of them.

Kinda struggling right now.

80

You mean the cowgirl from 35min mark ? Looks like I deleted the log file, I need to regenerate it so we can debug. Let me try and get back to you on this.