SAM2 — Segmentation AI Model
SAM2 finds every distinct object in a picture without being told what to look for, coloring each one differently and listing them as metadata. Tiny, Small or Large trades accuracy against speed. 10-30 seconds.
What it does
SAM2 segments every distinct object in an image without being told what to look for. Upload a picture and each object comes back masked and coloured separately, with a list of them as metadata alongside.
That is the difference from SAM3 and SAM3.1, which start from a text prompt: this one starts from nothing and finds everything. Tiny, Small or Large trades accuracy against speed. Ten to thirty seconds.
When to use it
- Everything in a picture masked at once, with nothing to describe
- Any object category, since there is no vocabulary to train first
- Pixel-accurate boundaries rather than a rough bounding box
- Cluttered scenes, with many overlapping things in the frame
- Annotation work, faster than selecting pixels by hand
Details
- CategorySegmentation
- Year2024
- LicenseApache-2.0
Compliance & Provenance
- ProviderMeta (open) · Specialized
- EU AI Act RiskMinimal Risk
- Art. 50 TransparencyNot applicable
Inputs & Outputs
- ImageInputimage
Source image to segment
- ImageOutputimage
Segmentation mask overlay image
- TextOutputstring
Detected segment labels and metadata
Parameters
- versionstringdefault Largeoptions: Large, Small, Tiny
Variant (Tiny/Small/Large)
Model size. 'Large' (default) — most accurate, slowest. 'Small' — balanced. 'Tiny' — fastest, suitable for previews or edge.
Tags
Alternatives in Segmentation
- BiRefNet (Subject)
Cleanly extracts a single salient subject with high-resolution edges. ~1-3s.
- RF-DETR Seg Medium (Scene)
Real-time instance segmentation on 80 COCO classes (RF-DETR Seg Medium).
- SAM3 (Scene)
Language-prompted segmentation (e.g. "person", "red car"). 10-30s/class.
- SAM3.1 (Scene)
Language-prompted segmentation — SAM 3.1 Object Multiplex with improved accuracy over SAM3. 10-30s/class.