Auto-reframe

Composition ▸ Auto-Reframe to …, or type the aspect into the command palette (9:16, 1:1, 4:5, 16:9, 4:3). Each entry greys itself out for a composition already at that aspect, which would have nothing to pan within.

What it produces#

A new composition at the target size, holding the original as a comp instance, scaled to cover and with a pan keyframed onto it. The source is not touched.

That is the whole point of the shape. A 16:9 cut retargeted to 9:16, 1:1 and 4:5 is four deliverables from one edit, and all four have to follow when the edit changes — as instances they do. Baking the crop into the original would have produced one file and destroyed the thing it came from.

The target size keeps the source's shorter edge, so retargeting never invents resolution: a 1920×1080 master becomes 1080×1920 vertical, not 2160×3840 upscaled from pixels that were never there. Both dimensions come out even, because every H.264/HEVC encoder wants them.

How it decides where to look#

The composition is rendered once at 160 px wide, 12 frames a second, through the same deterministic offline loop the exporter uses — the composition, not its footage, because a comp is titles, graphics and effects too and a reframe that ignores the lower third can crop it off.

Each sampled frame is scored on two cheap signals:

• Motion — what changed since the previous frame. The subject is the thing that moves; the background is the thing that does not. This carries almost every real shot.
• Detail — local gradient energy. It carries a locked-off shot: a talking head, a title card, a product on a table.

A centre prior is then added — not multiplied. That distinction is the whole trick, and a failing test is what found it: scaling the map by a centred bell does nothing to a lone blob's centroid, so a corner highlight in an otherwise empty frame won outright and the crop lurched into the corner. Adding a weak centred field gives the middle real mass to compete with, so an isolated speck is pulled back while a genuine subject still wins on its own energy.

Frames with nothing in them — a dip to black, a flat graphic, an empty sky — report low confidence and do not move the frame at all.

How it moves#

Finding the subject is the easy half. What separates a usable result from a seasick one is three rules:

1. A dead zone. The frame does not move until the subject has drifted meaningfully off centre. Without it the crop chases a centroid that wobbles by a pixel a frame, which reads as a handheld camera nobody asked for. Operators call this slop, and every shoulder rig has it deliberately.
2. Lag. Once it does move, it eases toward the target over about half a second rather than snapping.
3. Cuts are walls. At a shot change the frame jumps. Smoothing across a cut produces the worst artefact this feature can have — the crop visibly sliding to catch up during the first half-second of every new shot, which no editor would leave in.

Shot changes come from the same analysis pass that scores the frames (a luma histogram distance with an adaptive threshold — the same detector as Scene Edit Detection). Rendering twice to ask two questions about the same frames would double the only expensive part of this.

A subject merely moving is not a cut, and the detector is right about that: it compares luma distributions, which barely change when something slides across an otherwise unchanged frame.

The keyframes are yours#

The pan is written as ordinary position keyframes on an ordinary layer, thinned to the ones worth having: a 30-second shot that barely moves gets two, not 360. Cut boundaries are held on both sides so nothing ramps through them.

So the result is editable. If the automatic framing misses on one shot, drag its keyframes — you are not fighting a black box, and there is nothing to re-run.

Source map#