FunAutoScripter: a fast, motion based, fully unattended funscript generator (free and open source)

[AtumaPingas (CoolUsernameGottaStayAnonymous)]

Context

I’ve been interested in automatic funscript generators, but I kept running into the same problems. Some are too slow, some sometimes detect nothing at all, some need too many steps, some are paid, some hallucinate actions, and some need constant human supervision.

I’m studying for an engineering degree while working, so I have no time to script or build scripting apps. Instead, I designed an algorithm that is optimized to make the best script possible with the requirements in mind, then vibe coded it until I had a working app. It’s free and open source of course.

Requirements

• High processing speed, which means no neural networks or ML models
• Works with different kinds of scenes (insertion, HJ, BJ, fingering, dance)
• Runs completely unattended and produces a usable script
• Available as a command line tool and as a GUI
• CPU only, no GPU needed. The only dependencies are numpy, scipy, OpenCV and ffmpeg, plus DearPyGui for the GUI
• It can’t rely on object detection, so it works on every kind of video that has oscillating motion like a cooking video. It detects movement and it scripts it, the question is if it’s the right thing to track or not

What I’m willing to give up on

Every approach has downsides, and this one has some.

• Orientation accuracy: sometimes the script comes out inverted. To keep it consistent, the 100 position points up by default, and the 100 of every stroke in a chapter follows that guide. You can select a chapter and press i to invert it.
• Positional accuracy: velocity matching is the priority, position comes from integrating the velocity so its not perfect.
• Unrelated motion: it can track movement that isn’t the activity, especially periodic motion stronger than the action you want. When the main motion pauses, the automatic region may follow whatever moves next.
• Camera motion: it isn’t compensated.
• Single tracking point: It can’t track 2 bodies at the same time.

How it works

1. Decode and downscale. ffmpeg outputs greyscale frames at low resolution (192 px wide, 30 fps by default).
2. Optical flow. Dense Farneback flow is computed on every frame pair, then averaged into a grid of cells (12 px by default). This is the only slow step, so it’s cached to disk.
3. Scene cut detection. Cuts are detected from frame difference spikes. Region choice, direction and normalisation all reset at each cut.
4. Region of interest. In a sliding window (3 s by default), candidate regions are ranked by motion intensity and periodicity (the autocorrelation peak of their motion). The winner only replaces the current region if it beats it by a margin, which avoids jitter.
5. Direction and velocity. The dominant motion axis comes from a magnitude-weighted PCA on the flow vectors. The velocity is measured along that axis.
6. Integration and filtering. Velocity is integrated into position and then passed through a zero-phase Butterworth bandpass (0.05 to 6 Hz by default) to remove drift and noise.
7. Normalisation and enhancement. The signal is scaled to 0 to 100 with a rolling percentile window (30 s by default). Optional enhancements like smoothing, gain, boost, offset and a second pass for small fast strokes are available.
8. Keyframes. Peaks and valleys are extracted with minimum amplitude and interval constraints, and a speed limit (400 units/s by default) protects the device.

TL;DR: the computer chooses the best place to track, then tracks it.

Features

• Automatic chapter detection (hard cuts, fades, axis changes and zone changes, all optional)
• Chapters are saved next to the video in a .chapters.json file and restored automatically
• Per-chapter manual ROI, zoom and invert
• Hysteresis switching of the region of interest
• Configurable low-pass and high-pass filters, enhancement and normalisation
• Zoom option that crops the motion zone before downsampling, which helps with subtle motion.
• Second-pass enhancement for small fast strokes hidden inside a larger slow oscillation.
• 3D support for SBS and TB videos, plus a VR mode that zooms in to remove edge distortion (proper reprojection would be better)
• Optional GPU acceleration (OpenCL), multithreading by splitting the video into chunks, and batch processing of several files at once
• Resumable batches: videos that already have a script are skipped

Everything is configurable through command parameters or GUI sliders. Because optical flow is cached, you can change any parameter that doesn’t affect the flow and fine-tune the funscript instantly, without recalculating. This also works in bulk: after a batch, changing something like gain or smoothing regenerates every cached video almost instantly.

Performance

It processes 200+ fps on my PC, sometimes 500, parallel extraction makes it double the speed for me. 500 fps processing speed is insane and with the video at 30fps imagine the time it takes for a video to be scripted,and the quality is pretty good at least in my experience, it’s not perfect but the best approach I have seen to date.

Tools used

• Video processing: ffmpeg, OpenCV
• Numerical processing: SciPy, NumPy
• Graphical interface: DearPyGui

Demo

All demos were made completely unattended. The results are exactly as the program produced them, with no human intervention. I also still don’t know what are the best parameters

GUI

The interface is split into windows, each with its own purpose:

• Control: process videos or trigger actions.
• Video: the video with overlays for optical flow, region of interest, candidates, window averages, magnitude, frame magnitude, periodicity, and the measurement axis with live funscript arrows.
• Video Control: toggle overlays, set manual zoom or manual region of interest.
• Parameters: optical flow processing and post-processing settings.
• Overview: the timeline, color coded by direction, position, pace, velocity, amplitude, etc…
• Signals: the optical flow signal projected onto the measurement vector, and the funscript itself.
• Folder Videos: see which videos are already scripted and select videos to process.
• Status: detailed processing progress.

It uses DearPyGui, so the look is similar to OpenFunScripter. There are 17 color themes.

Video examples

Download

Builds for Windows and Linux are on the GitHub releases page. You can also run it from source or use Docker.

Repo:

GitHub - pingasAtumas/FunAutoScripter: Fast funscript generator for interactive...

Releases page:

Releases · pingasAtumas/FunAutoScripter

Download the funautoscripter_gui for the graphical user interface

It’s 100% vibe coded. The AI also wrote a document explaining all the math, its in the repository, and a guide with placeholder images, which I’ll replace eventually.

This software is released under MIT License.

I created this for personal use but thought that it would be great if i shared it with the community. Eventually i should release some updates even though i am very busy.

Enjoy!

Changelog

v1.7.x – v1.9.x

• v1.7.0 - 2026.10.25 - Dual axis, pan/zoom timeline, batch chapter detection, cheat sheet, Process button
• v1.7.1 - 2026.10.25 - Fix file picker .* filter for single video import, fix timeline axis switch
• v1.7.2 - 2026.10.25 - Axis 2 independent signal settings
• v1.7.3 - 2026.10.25 - Fix timeline axis selection reset on parameter change, restore Axis 2 header visibility on startup
• v1.8.0 - 2026.10.25 - Per-chapter axis mode, overview axis selector, individual parameter reset buttons, Ctrl+arrow precision nudge, Axis 2 signal defaults
• v1.9.0 - 2026.10.25 - Exclude chapter from Save
• v1.9.1 - 2026.10.25 - Bundled default layout: first run and reset now restore the correct docked layout

v2.0.0 – v2.0.x

• v2.0.0 - 2026.10.26 - Axis lock (resist tracking deviations from chapter axis-ref for consistency on multi-axis), remove y axis mode (use angle 90°), fix direction resolution applying to both PCA and forced-angle modes
• v2.0.1 - 2026.10.26 - Fix folder open path duplication (DPG bug on all platforms)
• v2.0.2 - 2026.10.26 - Debug layout copy, move create_context after ini copy, add _MEIPASS fallback
• v2.0.3 - 2026.10.26 - Fix default_layout.ini missing from CI build (–add-data not in workflow), fix reset layout overwritten by ImGui auto-save
• v2.0.4 - 2026.10.26 - Reset layout closes app immediately
• v2.0.5 - 2026.10.26 - Fix reset layout: delete layout.ini after destroy_context so ImGui cannot recreate it

[WhoDat21]

Looks very interesting. I’ve been checking out all the auto scripters lately to see how they are progressing. I’ll have to check it out and see how well it does.

[seen33]

OK after a lot of effort I got it working. It needs some QOL improvements for sure

• Open video doesnt work at all. It just says file not found and may crash the app. Open folder seems to work.
• There’s no way to use the keyboard arrow keys to set values on sliders by 1 over. So setting an exact value on like u/s for example is impossible. I cant hit 240 exactly.
• The Ui doesnt all fit together. Im at 2k resolution with zoom on windows and boxes are overlapping others.
• Theres no select all when selecting the videos in a folder. I had to checkmark 115 boxes, yikes.

[AtumaPingas (CoolUsernameGottaStayAnonymous)]

Thanks for the feedback.

• Open video - I had that issue from the beginning and i thought it was something with my PC, i need to fix it then thanks.

• keyboard arrow keys to set values on sliders - Good QOL one for precision, i need to implement that.

• Ui doesnt all fit together - Yeah it seems like the default layout just puts the windows floating, not actually making them move together, to fix this you can stick them together by dragging them onto the edges.

• Theres no select all when selecting the videos in a folder - I’m sure there is a button for that in the control window under the Batch section, it has to be moved from the control window over to the folder window for ease of use

[Pudknocker]

Any chance of multi-axis support? This looks really nice, and something that generates positions for different axes with guided support (as in, not purely automatic) would be neat.

[bmweyebrow]

whats the difference between proces abc recalculate? Seems like calculate also generates the script?
also any option to exclude a chapter?

[AtumaPingas (CoolUsernameGottaStayAnonymous)]

I read your commend and added a feature where it generates a second funscript for motion perpendicular to the primary axis of motion.

[AtumaPingas (CoolUsernameGottaStayAnonymous)]

Process calculates optical flow, detects chapters, angles for those chapters and generates the script

Recalculate only does the optical flow(the slow step) for the current chapter, it doesn’t generate the script. This is useful because you can zoom in on a region in specific chapters and by zooming in you need to recalculate the optical flow.

You can add and remove chapters but it will not influence the video, if you want a chapter to have no script i just implemented a checkbox to choose if it saves that specific chapter to the .funscript file or it ignores it when building the script

[bmweyebrow]

Thanks. Where is the Direction used for? It doesn’t seem to influence anything. “Draw where 100 points”??

[AtumaPingas (CoolUsernameGottaStayAnonymous)]

Every time it detects motion the script only knows the axis of motion(it doesn’t know where 0 or 100 is, it now has 2 choices, instead of pointing 100 to a random direction I implemented a chapter direction reference, it’s a way for the axis to choose one of the 2 possibilities and be consistent over the chapter.

Imagine you are doing a cowgirl scene, 0 is down 100 is up, that’s good, if you put the chapter 100 direction to down it will detect motion in the same axis but the 100 will always be pointing in the down direction, this will be inverted because in the funscript standard 0 means fully inserted.

This direction reference is chosen automatically per chapter, there is just a feature to manually set it for the chapter.

When the person being inserted in moves the arrow points to that person.
When the object that inserts moves the arrow points to the object, opposite from the person being inserted in.