The short version

You record one continuous take. Unhanded sorts it into two kinds of moments: holds, where the scene is still and nobody is touching it, and intrusions, where hands are in the shot. It keeps one good frame from each hold and throws away the intrusions. Then it lays the keepers on a 24 fps timeline.

Nothing is painted out. We don't erase hands from frames. We remove the frames that have them, which is how you'd edit this by hand if you had an afternoon to spare.

One take, simplified. Striped blocks are hands in the shot. Teal blocks are holds. The small mustard blocks are the single frames kept from each hold.

Why hand detection alone isn't enough

The obvious approach is to run a hand detector on every frame and delete the hits. It breaks in the situations you actually shoot in. A detector misses hands that are partly out of frame, motion-blurred or gloved. It never sees a forearm or a sleeve. It has no idea that your shadow is crossing the set.

So Unhanded only calls a frame clean when two separate signals agree.

Signal 1: is there a hand?

Each frame goes through Apple's on-device hand pose detection, which looks for up to two hands and reports how confident it is. A frame with a confident hand is dirty.

Signal 2: is anything moving?

Each frame is shrunk down and compared with the one before it. Big changes in brightness mean something moved: an arm, a shadow, a blurry hand, or a model still settling after you nudged it. How much change counts as movement is calibrated from the clip's own noise level, so a grainy low-light take isn't treated as constant motion.

From signals to holds

Pad the dirty frames. Hands show up on screen a few frames before the detector is sure of them, so Unhanded widens every dirty stretch by about 150 milliseconds on each side.
Require a real pause. A run of clean frames only counts as a hold if it lasts at least about a quarter of a second. Quick flickers of clean in the middle of a hand movement are ignored.
Merge look-alikes. If a hand reached in and changed nothing, two holds look the same. They are merged so the pose isn't counted twice.
Pick the sharpest frame. Within each hold, Unhanded looks at the middle stretch and keeps the frame with the most fine detail. That avoids frames soft from the last bit of motion or the first bit of the next.
Lay it out. The keepers go on a 24 fps timeline. On twos, each pose is held for two frames, so you get 12 poses a second. Ones are available too.

Why capture is locked. A hand entering the frame can nudge a phone's automatic exposure, which both makes the footage flicker and confuses the motion signal. Unhanded locks focus, exposure and white balance when you start recording. More on that in the stop motion flicker guide.

Review before you export

After analysis you see a strip of the poses it found. If it kept one you don't want, or dropped one you do, switch it back.

What it can't do

It can't remove a hand that never leaves. If your hand is in every frame, there is nothing clean to keep. Each pose needs a short moment with your hands out of the shot.
It can't fix the light. If the lamp flickers or a cloud covers the window, those changes are in your footage. Locking the camera stops the camera reacting; it doesn't change the room.
It isn't proven on every kind of footage yet. We are shooting takes in different lighting, with different hands and with gloves, and measuring how many hand frames slip through and how many good poses get dropped. We'll launch when that holds up on footage we didn't shoot ourselves.

Where the processing happens

On your phone. Detection, motion analysis and export all run on the device, so your footage isn't uploaded anywhere and you don't need an account. See the privacy page.

Next: stop motion product video ideas, or how Unhanded compares with Stop Motion Studio.

How to remove hands from video, plainly