Every stabiliser does three jobs: it measures how the camera moved, decides how it should have
moved, and redraws each frame to match. Here all three run on your own device, spread across
several processor cores and your graphics chip, and none of it is AI.
Track
Each frame is turned into a small greyscale copy on your graphics chip, and a few hundred well-spread corners are picked out of it: the edge of a window, a stone, a leaf. Every corner is followed into the next frame and then back again, and any that don’t return to where they started are thrown away. A robust fit then finds the one camera movement that most of the picture agrees on, so a car crossing the shot or a person walking past is out-voted by the background, as long as the background fills most of the frame. Cuts between shots are found along the way, and each shot is treated on its own.
Plan
Added up, those movements give the path the camera actually took. A solver then finds the calmest path that stays close enough to it for the moving edges never to need more zoom than Zoom limit allows: gentle curves for Smooth, still holds joined by straight glides for Cinematic, and a single fixed position per shot for Locked. The same step works out the zoom for every frame and how far each row of the picture must move to undo a rolling shutter. Planning is quick enough to run again every time you touch a control, so the preview changes as you drag a slider, with no second analysis.
Render and encode
Your graphics chip moves every frame onto the new path, using high-quality sampling to keep detail through the slight zoom, shifting the rows of the picture individually to straighten a rolling shutter, and bending small regions of it when Correction is set to Warp. For the export, the frames are drawn the same way and encoded by your browser’s own video encoder, in several chunks at once where the encoder runs faster that way, then joined into one file, with the sound copied across untouched whenever the format allows.
What it can’t do
- Motion blur stays. A frame blurred by a fast jolt is still blurred once the movement is gone.
- The edges have to come from somewhere. Steadying moves the frame, so it either zooms in a little or fills the gap from the frame’s own border; nothing recovers picture that was outside the shot.
- Near and far move differently. Walking past close objects causes parallax that no single correction can hold completely still.
- It needs something to follow. Clear sky, fog, open water and night scenes, or a crowd filling the frame, are harder to measure.
- It reads the picture, not the camera’s sensors. There’s no gyroscope or lens data, so very wide fisheye lenses can still wobble a little near the edges.
- HDR becomes standard video. HDR and 10-bit clips are saved as 8-bit standard dynamic range.
- It steadies the camera, not a subject. It won’t keep a chosen person centred in the frame.
- A phone has less room to work. Long 4K clips are slow on a phone and can run out of memory there; a computer copes better.