Auto reframe for vertical video

On every checked frame the tool picks a subject: by default the largest person in shot, and only when there is no person, the largest object. A new person has to be clearly bigger, by about a third, before the crop switches to them, so two people of similar size don’t make it flick back and forth. Click someone in the preview and it locks onto them instead, tracking them forwards and backwards from the moment you clicked.

The window aims at the upper part of a person, roughly the face in a head-and-shoulders shot and the chest in a full-length one, because that is where viewers look. It never leaves the frame, so near an edge the person sits off-centre rather than showing black bars.

Moves are smoothed in both directions over the whole clip, so the camera starts turning a fraction before the person moves rather than lagging behind them, and never overshoots. At a hard cut the smoothing restarts: the window jumps to the new shot instead of panning slowly across the edit.

The export keeps the crop at the source’s own pixels and never enlarges it, because enlarging adds file size without adding detail. A 1080p landscape video holds a 9:16 slice 608 pixels wide, so that is what you get; Instagram scales it up to fill the screen. For a sharp 1080 × 1920 Reel, film in 4K.

Tip: Leave room around yourself when you film for both shapes: stand near the middle and keep gestures within shoulder width, and the 9:16 crop never has to cut off a hand.

• Clips can be up to 2 minutes. The export plays the clip once at normal speed, so a 90-second video takes 90 seconds to record; keep the tab in front, because browsers stop drawing video in background tabs.
• With WebGPU, tracking usually takes less time than the clip lasts. Without WebGPU it runs on the processor and takes several times longer, up to about a second per checked frame on a busy machine.
• It follows bodies, not voices. In a two-person conversation it stays on the larger person unless you click the other one; it can’t tell who is speaking.
• Chrome, Edge and Safari save an MP4. Firefox can only record WebM.
• Your video is never uploaded. The detector downloads once from Hugging Face and the runtime from a public CDN, and your browser keeps both.