Single-Subject Tracking + Automatic 9:16 Reframe
Click one person on the preview frame and get back a vertical video that follows only them, with the original audio intact. Runs on the same SAM 3.1 (Video) tracking model as the Video Segmentation node, wrapped as a one-click reframing tool — no masks to handle, no keyframes to set.
What it does
Subject Follow takes a horizontal (or any aspect) video and one click. You pick the person you want on the preview frame; the click is stored as an "x,y" pair of fractions of frame width and height. SAM 3.1 tracks that subject across the clip, and the node converts the track into a moving crop window that keeps the subject framed as the shot moves — then renders a 1080×1920 vertical video with the source audio carried through unchanged.
The difference from the Video Segmentation node is what comes out. Video Segmentation returns per-frame masks and IDs for you to composite with. Subject Follow returns a finished vertical video — the tracking is a means to an end, not the deliverable. Three controls shape the result: who (the picked point), how tight (a framing preset plus a punch-in multiplier), and how the crop is delivered (final 1080×1920, or the crop at its native size for an upscaling node downstream)
Problem it solves
- Horizontal footage → vertical, without an editor – Reframe interviews, webinars, lectures, and podcasts for Shorts / Reels / TikTok without manual keyframing or pan-and-scan work
- Keeps one person centered in a crowded frame – A single click isolates the subject; other people in shot don't pull the crop around
- Audio survives the reframe – The source audio track is preserved, so the clip is publishable as-is rather than needing a re-marry in an editor
- Consistent framing across a batch – The same preset (
full_body,upper_body,face) applied over many clips produces a uniform look, which manual cropping does not - Feeds an upscaling step cleanly –
nativeoutput hands the raw crop to a super-resolution node instead of baking in an upscale the node didn't optimize - Cheaper than full multi-object segmentation – One subject, one track; no mask assets to store or post-process
Input/Output
- Input: Video clip + one subject point
- Format: Standard video formats (MP4, MOV, etc.)
- Subject point: Click on the preview frame via Pick Person. Stored as
"x,y"fractions of frame width and height - Prompt type: Visual (point) only — this node does not take a text concept prompt. Use the Video Segmentation node for language-prompted selection
- Options:
- Pick Person (button) — opens the preview frame; click the person to follow. Stored as
"x,y"fractions of width and height - Framing (dropdown) — how tightly to frame the subject.
naturalkeeps the original size; the others punch in until the subject fills the frame that much natural(default) ·full_body·upper_body·face- Punch-in (slider,
1–4in steps of0.1, default1) — punch-in on the subject.1keeps the natural size; higher values crop tighter and enlarge more. At4the crop is a quarter of its natural width, so the subject fills the frame but the source pixels are stretched four times further - Tracker sampling rate (slider,
1–15in steps of0.5, default5) — how often the tracker looks. Higher follows fast movement more closely and costs more GPU time. Raise it for handheld or fast-moving subjects; lower it for a locked-off interview shot - Tracker input width (slider,
480–1920px in steps of60, default960, i.e. the midpoint) — width the tracker sees, in pixels. The crop is always taken from the full-resolution source, so this trades tracking cost against tracking precision, not output sharpness - Output size (dropdown, default
delivered) — deliver the finished 1080×1920 vertical video, or the crop at its own size for an upscaling node to enlarge delivered(default) ·native
- Output: A single vertical video following the picked subject
delivered: 1080×1920 vertical videonative: the crop at its own resolution, intended as input to an upscaling node- Audio: source audio track preserved
- No masks, IDs, or per-frame confidence scores are returned — use the Video Segmentation node if you need those
Performance
The figures below are for the underlying SAM 3.1 (Video) model. This node has not been separately benchmarked end to end; total runtime also includes crop computation and video encode.
Tracking inference time | ~1 – 3 minutes per 200 frames on typical hardware (SAM 3.1 Video) |
Objects tracked | 1 — the picked subject. Object Multiplex multi-object scaling does not apply to this node |
VOS benchmarks | SAM 3.1 improved on 6 of 7 standard Video Object Segmentation benchmarks vs. SAM 3 |
Tracking robustness | Persistent subject identity across frames, including under occlusion (inherited from SAM 3.1) |
Cost driver | Tracker sampling rate — raising it increases GPU time roughly in proportion to how often the tracker runs |
End-to-end node timing | Not yet measured — verify on your own footage before committing to a batch schedule |
Model Source
- Huggingface
- License: Custom (SAM License)
- Important: You are not the target of Trade Controls and your use of SAM Materials must comply with Trade Controls. You agree not to use, or permit others to use, SAM Materials for any activities subject to the International Traffic in Arms Regulations (ITAR) or end uses prohibited by Trade Controls, including those related to military or warfare purposes, nuclear industries or applications, espionage, or the development or use of guns or illegal weapon.
Compliance & Provenance
Provider | Meta (open) |
Provider type | Specialized |
License | |
EU AI Act risk class | Minimal Risk |
Art. 50 transparency | Not applicable |
Region availability | Available globally |
Training data summary |
For more on how we classify models and mark outputs, see our AI Policy.