Why automatic cropping cuts people's heads off

A 16:9 frame turned into 9:16 keeps about a third of its width. Something has to be thrown away, and which third you keep is a decision that gets made on every single frame.

The centre crop

The simplest approach keeps the middle and discards the sides. It works perfectly when the subject is centred and fails completely when they are not — and in most recorded conversations, nobody is centred. Two people sitting either side of a table both end up half out of frame, with the empty space between them filling the screen.

Face detection, and why it is not enough

The next approach finds faces and crops to one. Better, but it introduces a new failure: with two faces in shot, something has to choose. Choose wrong and the frame sits on the listener while the speaker talks off-screen. Choose by size or position and it will get it wrong whenever the camera setup is asymmetric.

The frame has to follow whoever is speaking, which means the crop decision depends on the audio, not only the picture.

The drift problem

Even with the right subject, a crop that follows a face continuously produces a picture that sways gently for the entire clip, because people move their heads while talking. It reads as a camera operator who has had too much coffee, and it is tiring to watch.

A better behaviour is to hold still and cut. Stay on one framing while a person is speaking, then cut cleanly to the other when the speaker changes. That is how an editor would do it, and it is invisible in a way that continuous tracking never is.

Good reframing looks like nothing. If you notice the crop moving, it is wrong.

How to check a tool in thirty seconds

• Find a clip where the two people interrupt each other. Watch whether the frame keeps up or lags behind.
• Watch a clip with the sound off and see whether you can tell who is speaking from the framing alone.
• Look at the edges. Heads cut at the crown, or chins at the bottom of frame, mean the crop is following a box rather than a person.

When to give up on vertical

Some shots cannot be cropped. A wide three-person panel, a whiteboard, a gameplay capture with information in the corners. Forcing those into 9:16 produces something worse than not posting. Square is often the compromise that keeps the clip usable.