SAM3 — Segmentation AI Model
SAM3 segments anything you can name — "person", "red car" — returning a color overlay with every match labelled per instance. Comma-separate categories to catch several at once; each adds 10-30 seconds.
What it does
SAM3 segments anything you can name. Describe what to find - "person", "red car", "person wearing glasses" - and it masks every match, labelled per instance; points, boxes and masks work as prompts too. It covers around four million concepts.
Comma-separate categories to catch several in one pass, at ten to thirty seconds each. SAM 3.1 is the drop-in successor with better detection accuracy at comparable cost, so start there unless you have a reason to pin this version.
When to use it
- Finding and masking everything that matches a described object
- Attributes: a person in white kept apart from a person in red
- Spatial words - the car on the left, not the one on the right
- Blurring faces, plates, screens or documents automatically
- Counting how many instances of something are in an image
Details
- CategorySegmentation
- Year2025
- LicenseCustom (SAM License)
Compliance & Provenance
- ProviderMeta (open) · Specialized
- EU AI Act RiskMinimal Risk
- Art. 50 TransparencyNot applicable
Inputs & Outputs
- ImageInputimage
Source image to segment
- TextInputstring
Noun phrase(s) naming the foreground objects to segment (e.g. 'person', 'the red car'). Comma- or semicolon-separated for multiple categories ('person, dog'); each category triggers its own inference call.
- ImageOutputimage
Colored overlay where each detected instance is painted in a distinct color. Pair with the Text output (which carries per-instance class+color metadata) into Masking-Image-Masker-By-Class for pixel-precise mask extraction.
- TextOutputstring
Per-instance metadata. Each line: '<prompt>|<prompt>_<idx>|object_<global>,(R G B),score,bbox[x0 y0 x1 y1]' — three pipe-separated alias labels (category / instance / legacy index) followed by the instance's overlay color, confidence, and bounding box.
Parameters
- binary_maskbooleandefault false
Merge to binary mask
When true, merges all detected instances into a single white-on-black binary mask suitable for direct use with Inpainting or Image-Multiply (no Masker-By-Class step needed). Default false returns the colored overlay so the user can still distinguish individual instances.
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).
- SAM2 (Scene)
Segment every object in an image (no labels needed). 10-30s.
- SAM3.1 (Scene)
Language-prompted segmentation — SAM 3.1 Object Multiplex with improved accuracy over SAM3. 10-30s/class.