SAM3.1-Video — Video Segmentation AI Model

Name what to follow and SAM3.1 segments it in every frame, tracking up to sixteen objects at once and keeping their identities across the clip. Comma-separate the categories. Roughly one to three minutes per 200 frames.

Try in Studio

Details

  • CategoryVideo Segmentation
  • Year2026
  • LicenseCustom (SAM License)

Compliance & Provenance

  • ProviderMeta (open) · Specialized
  • EU AI Act RiskMinimal Risk
  • Art. 50 TransparencyNot applicable

Inputs & Outputs

  • VideoInput · video

    Source video to segment and track

  • TextInput · string

    Noun phrase(s) naming the foreground objects to track (e.g. 'person', 'the red car'). Comma- or semicolon-separated for multiple categories ('person, dog'); each category becomes a tracked text prompt. Categories that match nothing are named in the metadata output, and if none match the run fails with that reason.

  • VideoOutput · video

    Output video with each tracked object painted in a distinct color on a black background, across all frames. With binary_mask=true, all objects are merged into a single white-on-black mask instead.

  • TextOutput · string

    Per-object metadata. Each line: '<prompt>|<prompt>_<idx>|object_<global>,(R G B),frames=<n>,avg_area=<px>' — three pipe-separated alias labels (category / instance / global) followed by overlay color, frames the object was present in, and average mask area in pixels. Categories that matched nothing appear as 'no_match: <category>' lines.

Tags

  • video-segmentation
  • segment-anything
  • tracking
  • text-guided
  • multimodal
  • prompt-based

Alternatives in Video Segmentation

  • Subject Follow (9:16)

    Click one person and get a 9:16 vertical video that follows only them, audio intact.

Resources