Image segmentation got a general-purpose base model in 2023. Trained on a billion masks, the system can isolate any object from a prompt such as a point or box. It brought the foundation-model paradigm to pixels.
Promptable masks
Interaction drives it. A click or box tells the model what to segment. It responds with a precise mask.
Vast training data
Scale underpins it. A billion masks across eleven million images taught broad generalization. Coverage is enormous.
Zero-shot transfer
It adapts unseen. The model segments objects it never explicitly learned. Flexibility is high.
A building block
It composes. Other systems use its masks as input for tracking or editing. It slots into pipelines.
Open release
Access spread. Weights and data were shared openly. A community formed quickly.
Video and beyond
It extended. Later work brought promptable segmentation to video. The paradigm keeps growing.
The bottom line
Segment Anything is a promptable foundation model for image segmentation trained on a billion masks. It transfers zero-shot and composes into larger systems. It brought foundation models to vision.
