Meta releases Segment Anything with separate model and data licences

Meta released Segment Anything on 5 April, offering a model that draws object boundaries from prompts such as clicks or boxes. The release pairs SAM with SA-1B, a dataset of 1.1 billion masks across 11 million images.

Meta headquarters, photographed in 2022. Photo: LPS.1, “Meta Platforms Headquarters Menlo Park California” (2022). CC0 1.0. Resized; cropped; responsive crops.

The two releases have different permissions. SAM uses Apache 2.0; Meta is making SA-1B available for research purposes. The model’s permissive licence doesn’t extend to the accompanying image collection.

The original paper’s dataset card explains what those masks contain: outlines without categories or text. The team also withheld the photographs’ captions under its agreement with the image provider. That makes the dataset useful for learning boundaries, while leaving anyone building a named-object recognition system with a separate labelling problem.

The distinction also runs through SAM’s design. A point on clothing can reasonably refer to the garment or its wearer, so the model can return several masks. Its job is to produce a valid segmentation, rather than determine a single correct meaning for every prompt.

The researchers describe text-to-mask experiments as exploratory and not fully robust. They also report missed fine structures and occasional disconnected fragments. Fast prompt responses depend on an image representation computed beforehand; the full process with a heavy image encoder isn’t real-time.

Meta’s data collection process moved from human-assisted annotation through mixed annotation to automatic mask generation. The scale comes from that final stage, rather than a billion individually hand-drawn outlines.