Vision Segment
Segment people from an image with Apple Vision, output as a cutout on a transparent background.
What it does
Vision Segment finds people in an image and cuts them out onto a transparent background, using Apple's on-device Vision framework running on the Neural Engine. Any image wired in produces a live cutout, with no green screen and no manual masking.
Alongside the cutout it outputs the raw greyscale mask (white for person, black for background) for driving something else: a Composite node, a Distance Field, a particle emitter, anything needing the selection rather than the pixels.
When to use it
A person pulled out of a webcam feed or video without a green screen
A person-shaped mask driving another effect: blurring the background, keying in a new one, spawning particles off the silhouette
A live AR or VJ setup needing segmentation that updates frame to frame
| Parameter | Type | Default |
|---|---|---|
image | OneOf([FieldOf(Raster)]) | — |
quality | String | "balanced" |
invert | Boolean | false |
enabled | Boolean | true |
position | Vec3 | (0, 0, 0) |
strength | Number | 1 |
Gotchas
Quality trades resolution for speed. Fast segments at lower resolution for smoother real-time playback; Accurate runs at higher resolution and costs more per frame. Balanced is a reasonable default for most live setups.
Invert makes the mask, and the transparency, read background as the on region instead of the person, which keeps the background and cuts the person out.
Vision Segment runs on macOS only, on Apple's Neural Engine. On other platforms the node has nothing to compute from.
Worked example
Add a Vision Segment node and wire a Camera or Video node into Image.
Leave Quality on Balanced to start; drop to Fast if the live feed lags, or raise to Accurate for a cleaner still-frame edge.
Take Cutout straight into a Composite to place the person over a new background, or take Mask into another effect that needs the selection rather than the pixels.
See also
Vision Pose · Object Detect · Depth Estimate · Rasters (images)