Segment Video
AI rotoscoping: track and segment named objects through footage (SAM 3) and output a matte video. Spends AI tokens, scaled by input duration.
Category: AI Menu path: AI > Segment Video
Ports
| Port | Type | Direction | Description |
|---|---|---|---|
video_in | imageRgba16f | input | Required — connect a Video Source (or another AI video node). The video file is uploaded for processing. |
prompt_in | string | input | Optional — a String node overrides the typed prompt. |
out | imageRgba16f | output | The segmented matte video. |
Parameters
| Param | Type | Default | Description |
|---|---|---|---|
prompt | string | "" | Required — what to segment, comma-separated: "person", "red car, dog". The model tracks each named object across frames, cuts, and occlusions. |
model | string | (first available) | Single curated model (SAM 3 video). |
resultPath | string | "" | Hidden — the saved matte clip under generated/. |
Standard rows expose Video Source-style playback params.
How It Works
This is text-prompted rotoscoping: name the objects, and the segmentation model finds and tracks them through the entire clip — no per-frame masks, no spline work. The job runs async (upload → queue → segment → download), with status streaming into the button. The result is a matte video you wire into masking workflows. Failures refund automatically.
Usage Examples
Track-matte isolation
Footage layer + Segment Video ("person") on a second layer → set the footage layer's track matte to the segment layer (Luminance). Effects, color, or replacement backgrounds now apply through the roto.
Selective stylization
Segment Video → Matte node alongside the original footage → run Posterize/Colorize only inside the matte — the subject stylizes while the world stays photographic.
Tips
- Comma-separate multiple objects to track them all in one pass (one cost).
- Be specific when scenes are busy: "person in red jacket" beats "person".
- The default detection threshold is tuned for precision; if the matte misses partial occlusions, ask — the threshold is tunable server-side per model.
Related Nodes
- VideoRemoveBackground — whole-foreground matting without prompting
- Matte — apply and combine mattes
- TrackPoints — point tracking on the same footage for attached graphics