Segment Anything (SAM).

Select an object while keeping the original frame

To change the background behind a person or apply an effect only to one object, you first need its mask. Meta's Segment Anything handles this part of the work: it finds the objects you specify and marks the pixels belonging to them. You can then use the mask in your editor or application.

The current documentation describes SAM 3.1. It works with images and video, accepting a short text name for an object as well as box and point prompts. Text helps find all matching objects; a visual prompt lets you select a particular instance. Older versions in the family may have different capabilities.

What the model returns

The model returns masks and bounding boxes. A mask defines the object's area; a box marks its rectangular boundary. In the programmatic response, these are segmentation data that an application must process and display. SAM does not produce a finished edit or create a new image from a description.

In video, an identifier helps link the masks of one object across frames while the object stays in frame. Do not rely on the identifier remaining the same across full occlusion, the object's entry or exit, or very long clips. The “Mask” output in the tool comparison means pixel annotations, including those for individual frames, rather than a promise of a finished video file.

An example with two red mugs

For an illustrative exercise, use a photo of your own showing two red mugs and a blue bottle. The short prompt red mug specifies the type of object. The intended goal is to obtain a separate mask for each mug without selecting the bottle. This is an example of setting a task, not a test we performed.

Check the outer silhouette and the spaces inside the handles. If the wrong object is selected, refine it with a box or point prompt. Against a complex background, it helps to enlarge the selection and compare it with the original.

A demo, the API or your own environment

The official Segment Anything link leads to Meta AI Demos. Check the demonstrations currently available, sign-in requirements and conditions for uploading your own material on the platform itself.

For integration, Meta offers SAM 3.1 through Meta Model API. Its documentation explains how to submit an image or video clip with a prompt and read the result. One text request specifies one type of object; use separate requests for several types.

For self-hosting, Meta publishes code, examples and links to the weights. Downloading the reviewed checkpoints requires requesting access and authenticating on Hugging Face. You will need a compatible GPU environment; check the requirements for the version you choose.

What costs to expect

The reviewed weights are distributed under the SAM License, which permits use without license fees subject to its conditions. This is a separate access option: hardware, deployment and maintenance require resources. Read the usage and distribution rules for the weights you choose.

The official cloud API is billed under its current terms. The model page lists separate usage rates for images and video frames. We do not list fixed prices here: check billing, account availability and service limits before uploading work material.

How to check the result

For an image, inspect the edges, fine details and nearby similar objects. For video, also check the object entering the frame, occlusions and fast movement. If the mask loses the intended area, refine the result before editing or measurement.

The API documentation treats zero matches as a valid result: the stream can finish without finding any objects. An empty response alone does not establish a connection failure. First check that the object is distinguishable in the input and clearly named in the prompt.

Suitable tasks

• Prepare object masks for further editing
• Track a selected object across video frames
• Annotate images for your own application