Meta releases Segment Anything Model (SAM)

Meta released the Segment Anything Model (SAM), a computer-vision system for isolating any object in an image, together with the SA-1B dataset it was trained on: more than 1.1 billion segmentation masks drawn from about 11 million licensed, privacy-screened photographs — Meta said roughly 400 times more masks than any prior segmentation dataset. SAM itself was released under the permissive Apache 2.0 licence, and the dataset was made available for research use.

Segmentation — outlining the precise pixels that make up a given object in a photo — had previously required models trained separately for each category of object a user wanted to identify, with substantial labelled data for each. SAM was built as a “promptable” foundation model instead: given a click, a box, or (in later extensions) a line of text pointing at part of an image, it would return a mask for the object at that location, generating several candidate masks when a prompt was ambiguous about which object was meant. Meta reported the model ran in roughly 50 milliseconds per prompt after an image had been processed once, fast enough for interactive use, and that it generalised to object types and image domains it had not been trained on — a zero-shot transfer capability the team presented as SAM’s central claim.

Meta positioned SAM explicitly as infrastructure rather than a product: a component other researchers and companies could drop into pipelines for tasks such as photo editing, scientific image analysis, and training data generation for other vision models, without having to build and label a segmentation dataset of their own. It came weeks after Meta’s LLaMA language model release and continued a run of open releases from Meta’s research arm that made the company, despite building closed frontier chat products elsewhere, one of the largest contributors to open computer-vision and language tooling that year.