Every tool that turns a wide recording into a vertical clip has to decide, many times a second, which slice of the picture to keep. When it decides well nobody notices. When it decides badly you get a clip of somebody's hand while a voice talks off screen. Here is how to find out which one you are buying, before you have cut fifty clips with it.
They are not equally common, and only the first is obvious in a still frame.
Something bright moves - a hand, a phone held up to the lens, a product being turned over - and the frame goes with it. The face leaves the top of the picture and the clip becomes a close-up of an object while the person explains it from somewhere off screen. This is the single most common failure and the easiest to provoke.
A face detector's answer moves a little every frame even when the person does not. A crop that follows it directly inherits that jitter and slides gently across the shot. You see it as a slow, unmotivated pan that nobody asked for, and it is most visible on a locked-off camera where the background gives you a reference.
The edit cuts to a different angle, or a different person. Whatever the crop was tracking no longer exists, but it keeps heading where it was going and arrives at nothing - so the first second of every new shot is framed on empty room, and only then does it correct.
Slides, gameplay, a screen recording, a wide shot of a room. There is no subject, and the right answer is to hold a steady frame. A tool that insists on finding one will wander around the picture looking for a face that is not there.
Four short clips, one question each. Run it against any tool, including this one.
Watch all four with the sound off. That is the whole trick. With audio running your brain fills in where the speaker is and forgives a frame that has lost them; in silence you see exactly what a scrolling viewer sees, which is a picture that either makes sense on its own or does not.
Test four is the one worth dwelling on, because sometimes the right move is not to crop at all.
A crowded beach with surf behind it. A stage with a band on it. A street. The picture is full of movement and none of it is the subject, because there is no subject - and a crop window placed anywhere in it is placed arbitrarily. A tool that insists on choosing will chase whichever thing moved last, and the result looks less like a camera and more like a fault.
The honest answer there is to stop cropping. Keep the whole frame and fill the sides, and the viewer sees the scene the way it was shot instead of a tight window onto a random corner of it. It costs screen height, which is why it should not be the default - but on the footage where it applies it is the only answer that is not a guess. Two people is the same decision in miniature: if both fit inside a vertical window, keep both; if they are at opposite ends of a table they do not fit, and picking one is how a conversation becomes a monologue.
A crop that works is doing three unglamorous things. It decides per shot rather than per frame, so a cut resets it and jitter does not move it. It prefers a person over motion, so a hand crossing the lens does not take the camera with it. And when it cannot find anybody it stops looking and holds a sensible frame, because a still picture of the right room beats a moving picture of the wrong corner.
SnapDub decides it this way: both people if both fit, the wide frame if they do not, and the wide frame again when the picture is busy and nobody in it is the subject. If it gets that wrong on your footage we would rather hear about it than not - tell us what it did.
Two pages next door to this one: vertical crop, on what a wide frame loses on the way to vertical, and video hooks, on where a clip should start.
Because movement is the easiest thing to detect and a face holding still is not. A hand crossing the lens is large, bright and fast; a face that does not move is none of those, so a crop that scores motion goes with the hand every time. The fix is not a better detector but a different priority: decide who the subject is once per shot, and let movement inside that shot change nothing.
It is a choice. Drift comes from following the detector's answer frame by frame, and that answer moves a little even when the person does not. Smoothing it costs nothing, and the difference shows on the first locked-off shot you try.
It depends on whether they fit. If both sit inside a vertical window, keeping both is right, and cutting to one turns a conversation into a monologue. If they are at opposite ends of a table they do not fit, and something has to give: either one of the two people or the whole frame with the sides filled in.
One video, and four moments out of it. You are not measuring accuracy, you are looking for the four failures above: each one shows in a second or two once you know what to watch for, and with the sound off not one of them gets past you.