Type a subject in plain language — "person," "red jacket," "the dog" — and a Segment layer tracks it across every frame of the clip in a single pass, instead of guessing frame by frame. It lives on the same layer stack as Color and Mask layers, so its matte grades, folds, refines, and exports like anything else.
A Segment layer holds one or more text-prompted subjects. Each is tracked together, in one temporal pass over the whole clip — not detected independently frame by frame — which is what keeps a matte from jumping to a different guess every time a subject moves or the framing changes. If a subject is briefly lost and reappears, Transduce checks how far the new detection is from where it was last seen before trusting it again, rather than silently locking onto whatever's closest.
Multiple subjects on one layer combine Add or Subtract, the same fold logic geometric shapes already use — union coverage, or punch a hole. The resulting matte is remapped onto your project's actual output frame the same way the graded picture itself is scaled and cropped, so a Segment layer's isolation lines up pixel-for-pixel with the shot, not just with the source file's native size.
Under the hood, subject detection and tracking is built on Meta's Segment Anything Model 3 (SAM 3).
| Aspect | Behaviour |
|---|---|
| Prompting | Plain-language subject names — no manual keyframing or box-drawing to start. |
| Tracking | One pass across the whole clip, not independent per-frame detection — subjects stay identified as they move. |
| Multiple subjects | Several prompts per layer, each combined Add or Subtract. |
| Re-acquisition guard | A subject that reappears far from its last known position is treated as untrusted rather than accepted outright. |
| Refinement | Despeckle, hole fill, feather, expand/contract, intensity — the same matte controls used everywhere else in Transduce. |
| Output-correct geometry | Remapped to match the project's actual resize/crop, not just the source file's native resolution. |
A fast, clean starting matte for selective grading — most useful exactly where colour qualifiers struggle: motion blur, inconsistent lighting, and the messier framing that comes with phone and AI-generated footage.
Segment layers use Meta's Segment Anything Model 3 (SAM 3) for subject detection and tracking. Transduce is not affiliated with, endorsed by, or a partner of Meta.
Text-prompted, whole-clip tracking, on the same layer stack as everything else you grade.