Dealing with the business side ended up being too much hassle, so it’s open source now. You can get it here: GitHub - Funscript-Flow/Funscript-Flow: A tool that uses computer vision to generate .funscript files for video files
From this:
To this:
(Unaltered output shown in OFS)
Features
- Completely automatic motion-tracked funscript generation for any video.
- Easy to use — Just download it and point it at your 3-terabyte “Taxes” folder.
- Runs on your computer without uploading anything anywhere.
- Generates scripts for entire libraries, fast (usually faster than watching it, in my environment).
- No GPU required (but the more CPU cores, the better).
- For scripters, there’s an option to disable keyframe reduction and export the raw motion data, so you can fine-tune it yourself
Technical Details:
This uses a purely mathematical approach with no machine learning, and can make a good funscript for pretty much any source video. It’s performed well on:
- Blurry 2005 cell phone footage from that time in the walk-in.
- The stylings of directors with shaky hands, and unhealthy fondness for Dutch angles.
- The full Hentai beastiary, with its wide variety of appendages.
- Good, wholesome VR with steady cameras and simple motions.
- Beach Volleyball (Don’t judge)
It doesn’t need any specific body part to be in frame, because, much like the common velociraptor, it can only see motion (though it does correct for camera movement and figure out orientation using math).
Known Limitations:
It’s good enough most of the time (I’ve been enjoying its output almost exclusively for a few weeks), but it’s no substitute for an expert scripter.
It can’t tell if whatever’s on screen is bouncing because of a hearty laugh or a vigorous pounding, so it scripts all motion as the latter (you may notice that there are no idle periods in the heatmap). In testing, this has been an advantage.
I’m communicating by edits here because anti-spam is preventing me from joining the conversation below:
The niche this tool satisfies is videos that don’t work well with object detection (because there aren’t a lot of good landmarks on screen, the participants aren’t human, the camera isn’t consistent enough, or any number of other reasons).
It’s not going to fare well in head-to-head comparisons with AI tools on clean POV video (VR or not), because of inherent limitations of the “motion-only” approach it uses.
That’s all to say that it’s not working for you in its current form, it probably won’t in the future. Planned improvements are more likely to help with speed than quality (though an invert is definitely doable, and I can add it soon).
That said, I have a large library of varied content, and have really enjoyed being able to use the Handy with it, even if the sync isn’t exact. I hope you enjoy that too.
Sample Scripts (unaltered output running with default settings, with VR checked for VR videos, and keyframe reduction enabled for the Jada Stevens video):
Valentina Nappi, Sheila Ortega (The Heat Wave) 3584p T69.funscript (715.7 KB)
shoplyfter_kylie_kingston_and_natalie_knight_full_low.funscript (863.2 KB)
The_Devil_In_Disguise_samsung_180_180x180_3dh.funscript (620.3 KB)
JC Wilds (A Ride On The Wild Side) 2700p T69.funscript (518.4 KB)
Mae Milano (Rent Free) 3584p T69.funscript (473.4 KB)
Pov Threesome - Jada Stevens.funscript (2.5 MB)
BadoinkVR_Eyes_Wide_Slut_(GearVR).funscript (506.1 KB)
Flawless Busty Babes Octavia Red & Hailey Rose Bring the Holiday Party to the Bedroom - S19;E9_NF Busty_480p.funscript (501.1 KB)
Bee.Movie.2007.funscript (878.6 KB)
