Utility
Available now

SAM 3.1

by Meta

Meta's Segment Anything model.
Track any subject as a mask that follows it.

SAM 3.1 runs in Compose, Morphic's timeline editor, on the Canvas, and through Copilot.

Key features

Every clip here was made on Morphic. The source footage is MiniMax H3 Max Turbo, the mask is SAM 3.1, and the two are composited so the tracked subject keeps its colour while the rest of the frame is dimmed. The clip at the top of the page shows the raw output instead: the picture on the left, the mask SAM 3.1 returned on the right.

Technical specifications

Video + stills

Clips in Compose, images on the Canvas

Click/box/text

Up to 16 clicks, or one box

Both ways

Forward, backward or both

Mask video

White keeps, black hides

Up to 16

One per tracked subject

500 chars

Short noun phrases work best

Editable

Feather, expand, invert, combine

Editor, Canvas

And through Copilot

Use cases

Fix one thing, not the take

Track the subject, then grade or relight only inside the mask. The rest of the frame keeps the look it was rendered with.

Isolate one of two people

A click and a short description pick the person you mean when two share the frame. A rectangle cannot, and a drawn shape drifts.

Swap the background

Invert the tracked mask and the subject lifts off its plate, ready to sit over another clip, a gradient or a flat colour.

Hold a subject in a crowd

People cross in front of her and the mask picks her up again after. Long occlusions are where it slips, so check the whole clip.

Backlit and in silhouette

A dark shape against bright mist is the edge a keyer loses. The mask read the horse and its rider together, as one subject.

Overview

SAM 3.1 is the current version of Meta's Segment Anything model, and what it gives you is a mask: a black-and-white picture of exactly where a subject is in the frame. Point at the subject on one frame, or describe it in a few words, and the model finds it and keeps hold of it as it moves, returning a mask video that lines up with your footage frame for frame.

That is the step that used to be slow. Drawing a shape around a moving subject means redrawing it every few frames, and the shape never quite fits. With a tracked mask you can grade, blur, relight or replace just that subject, and leave the rest of the shot exactly as it was rendered.

Meta released SAM 3.1 on March 27, 2026 as a drop-in replacement for SAM 3, and Morphic added it on November 4, 2026.

How SAM 3.1 reads a subject

SAM 3 introduced what Meta calls promptable concept segmentation: instead of only taking points and boxes, it takes a short text phrase and finds every instance of that concept. Morphic gives you both. Clicks and a box say where to look, and a phrase says what to look for, which is how you separate one person from another in the same shot. Meta is explicit about the limit: short noun phrases work, long compositional ones like "the second to last book from the right" do not.

The 3.1 release is about doing more of this at once. Meta says object multiplexing tracks up to 16 objects in a single forward pass rather than one pass per object, raising throughput from 16 to 32 frames per second on a single H100 GPU.

SAM 3.1 on Morphic

In Compose, Morphic's timeline editor, select a clip and add a Track subject mask, or ask Copilot to track something for you. Mark the subject on the frame where it is clearest, pick Forward, Backward or Both, and the finished mask arrives on the clip as a matte you can feather, invert and combine with other masks. On the Canvas, the same model handles stills: Auto-segment splits a picture into layers, and Separate as layer pulls one object out.

For a walkthrough with every option, read the SAM 3.1 guide, or start from AI rotoscoping if the job is isolating a subject rather than the model itself. For what the 3.1 release changed, see SAM 3.1 vs SAM 3.

Simple pricing

Get started for free today, with the option to upgrade or cancel anytime.

Basic

$19/ month
billed as $228 per year

2400 monthly credits

1 user only

All models

Workflows

Standard

$24/ month
billed as $288 per year

3625 monthly credits

1 user only

All models

Workflows

Pro

$45/ month
billed as $540 per year

6350 shared monthly credits

1 user

+ up to 4 more at extra cost

All models

Workflows

Max

$170/ month
billed as $2040 per year

24650 shared monthly credits

1 user

+ up to 9 more at extra cost

All models

Workflows

Enterprise

For higher limits

Custom

pricing and billing terms

High-volume credits
Custom seat limits
All models
Workflows
Pricing Gradient

Free

For playing around

$0

forever free

Up to 20 credits
1 user only
Limited models
Workflows
Compare plan details

FAQs

What is SAM 3.1?
SAM 3.1 is the current version of Meta's Segment Anything model. It detects, segments and tracks objects in images and video: you point at a subject or describe it, and the model returns a mask, which is a black-and-white picture of where that subject is. In video it keeps that mask on the subject as it moves.
What does SAM 3.1 do on Morphic?
It powers subject tracking in [Compose](/resources/tools/ai-video-editor), Morphic's timeline editor, and object cutouts on the Canvas. In Compose, Track subject turns a marked subject into a mask video that follows the clip. The same model also moves a rectangle or ellipse mask along with the subject inside it, so a shape you drew once keeps up with the action.
How do I pick the subject?
Three ways, and you can combine the first two with the third. Click the subject to include it and Alt-click to exclude an area, up to sixteen clicks. Or drag one box around it. Then add a short description of what to look for, such as the man in the red jacket, to tell the model which thing you mean.
What changed between SAM 3 and SAM 3.1?
Meta describes SAM 3.1 as a drop-in replacement for SAM 3 whose addition is object multiplexing: tracking up to 16 objects in one forward pass instead of one pass per object. Meta says that raises throughput from 16 to 32 frames per second on a single H100 GPU. The way you prompt it did not change.
Can SAM 3.1 track more than one subject?
Yes. Each tracked subject is its own mask, and a clip holds up to sixteen masks. Track them one at a time, then combine them: add two masks to treat two people together, or subtract one from another to cut a hole in a shape.
Does SAM 3.1 work on still images?
Yes. On the Canvas, Auto-segment splits an image into separate layers, and Separate as layer lifts a single object out. Both run on SAM 3.1.
What kind of description works, and what does not?
Short noun phrases work: the cyclist, a red suitcase, the woman on the left. Meta says the model does not handle long compositional phrases such as the second to last book from the right on the top shelf, and that it is weaker on very fine-grained or unusual concepts. When words are not enough, click the subject instead.
What does a tracking run cost?
Each run costs credits, and running both directions costs two runs because the model is called twice. The pricing page has the current plans and credit amounts.
Does editing the clip break the mask?
No. The mask video is lined up against the clip's source frames rather than its position on the timeline, so trimming it, splitting it or changing its speed needs no new run.

More about SAM 3.1

Guides, prompts, and comparisons for SAM 3.1.

How to track a subject in a video with SAM 3.1

How to use SAM 3.1 on Morphic: mark a subject, track it through the clip, and shape the mask you get back. Plus shape-mask tracking and stills.

Other models

Explore the rest of the Morphic model catalog.

See all models →

Seed Audio 1.5

ByteDance

ByteDance's full-scene audio model. Dialogue, music and effects, as separate tracks.

FLUX Video Upscale

Black Forest Labs

Black Forest Labs' video upscaler. Up to 4K, with fine detail rebuilt and the sound kept.

Ideogram Vector

Ideogram

Ideogram's model for vector art. Logos, icons and illustrations as SVG files.

FLUX 3 Max

fal.ai

fal's post-trained FLUX 3. Tuned to follow the script, built for speed.