Center a person

createLocalFaceDetector from @mcut/media samples a video on device and returns the largest face at each sample time, or null when there is none. The YuNet face model ships inside the package, so detection needs no download. planCenterPerson from @mcut/editor turns those samples into a list of commands. The first is one smoothed setReframe, and Reframe covers the track it writes. Apply the list in one transaction so it undoes as one step.

import { planCenterPerson } from "@mcut/editor"
import { createLocalFaceDetector } from "@mcut/media"

const detector = createLocalFaceDetector()
const samples = await detector.detect(asset.src)

const plan = planCenterPerson(engine.project, { elementId: "e-talk" }, samples, {
  aspect: 9 / 16,
  smoothing: 0.5,
})

engine.transact(() => {
  for (const command of plan) engine.dispatch(command)
})

On a video, the plan crops to aspect and the crop follows the face. When the crop aspect is within 1% of the project aspect, the plan also scales the clip to cover the frame and centers it. At another aspect the clip keeps its size and position, since it is likely picture in picture. Set fill to true to fill at any aspect, or to false to keep the size. On a multicam, pass the camera source key as source in the target. The layout slot keeps its rect and aspect, and the camera framing inside it follows the face. smoothing runs from 0 for a tight follow to 1 for a steady frame. The planner throws when no sample has a face, so a failed run leaves the project untouched.

In Studio, right-click a video or multicam clip and choose Center person. Over the live bridge, agents call center_person, which follows the camera source of a multicam by default and takes the same fill override. Either way the edit is one undo step.