Reframing horizontal video into vertical, done properly

The arithmetic nobody likes

A 1920×1080 frame cropped to 9:16 is 608 pixels wide. A third of the picture, give or take. Everything left and right of that window — the second presenter, the whiteboard, the thing the subject gestures toward — is gone unless you do something about it, and "doing something about it" is the entire craft of vertical reframing.

Which is why the centre crop has such a bad name. It is the right answer for exactly one shot: one person, centred, nothing important at the edges. It is the wrong answer for everything else, and it is the only thing most auto-reframe tools actually do. The rest of this post is about the other cases.

When a fixed crop is fine

Fixed crop — one window, no movement — is correct whenever the subject doesn't move relative to the frame: a presenter at a desk, a fixed cooking camera, a head-and-shoulders interview shot. Set the window with eyes on the upper third of the 9:16 frame and lock it. Movement in a vertical clip reads as shakiness even when it is the subject moving normally, so the boring crop is the good crop whenever it is available.

Check the crop's edges before trusting it: hands rest low, and slide decks put their content right where a vertical window will cut. If the frame is telling its story at the edges, no fixed window works — keep reading.

The pan-and-scan problem

Two-person shots broke the fixed crop, so editors started keyframing: window over the host for her lines, slide to the guest for his. That is a reframe by hand, and it has the same failure modes as a 1970s pan-and-scan transfer — a slide every four seconds becomes a metronome; a slide timed wrong shows you the wrong face mid-reveal; and it doubles the time the edit takes because you are redrawing a window for every cut.

The insight the automated tools caught up to: most of the editor's keyframing is just keeping a face in frame. Face detection tells you where the face is with near-perfect accuracy, and a window that follows it — smoothed so it glides instead of snapping — does that part of the job in real time. That is what SpeakerLock does on every Video To Reel render: the crop follows the face in shot, and when the edit cuts from host to guest close-up the frame follows the cut. On a single locked two-shot it holds one person steady rather than hopping between speakers — the details are on the Interview to shorts page.

What automated reframing still gets wrong

An honest tool should tell you where it fails, because pretending otherwise is how you end up posting a clip of an ear. The known hard cases:

Hands doing the talking. A chef, a mechanic, a product demo — the action leaves the face and the face-tracking crop follows the wrong thing. Fix: for action-heavy material, crop on the hands and let the head go out the top.

Three-plus people in a wide shot. Nobody can follow a conversation between three people in a 608-pixel window without cutting two of them out of the story. Fix: reframe the shot — tight two-shots or separate cameras — or cut horizontal-with-pan, or don't make that moment a clip.

Burned-in text and slide decks. Any text already in the frame was composed for horizontal. The crop will amputate it mid-word. Fix: shoot decks for vertical separately, or keyframe those sections by hand.

The smooth pan nobody noticed. The inverse failure: tracking tuned too smoothly lags the talker by a beat, so the viewer sees a second of the wrong face at every turn. If you are tuning your own tracker, latency under half a second matters more than perfect centreing.

Shooting for the crop you have not made yet

Every reframing pain above shrinks if you shoot with the vertical window in mind. Four rules that cost nothing at recording time: keep subjects roughly centred and chest-up, so the 9:16 window finds them anywhere in the frame. Leave headroom — eyes near the top third of a 16:9 frame become eyes in the top third of the crop, which is exactly right. Never put meaning at the far edges. And if the moment is a two-hander, shoot it in two tighter frames rather than one wide — two people in one wide horizontal frame is the one setup no reframe can rescue cleanly.

If none of that describes the footage you already have — the hour of uncut conversation sitting in your recordings folder is probably the definition of that footage — the pipeline that does transcript-first moment selection, face-tracking crop and burned captions in one pass is described on the Video To Reel homepage, and it is the same crop logic this post walks through, run at scale.

Frequently asked questions

Every how-to on this blog describes work you can hand to Video To Reel in one upload: the transcript, the moment picks, the 9:16 crop on the face in shot and the burned-in captions, returned as ready-to-post clips.