SAM (Segment Anything v1)

Meta's first Segment Anything Model converts a click, a bounding box, or a rough mask into a precise object outline, and it can also propose masks for every object in an image. A heavy ViT encoder computes an embedding once per image, after which a small prompt encoder and mask decoder respond to each new prompt quickly; that split lets the decoder alone run inside a web page after ONNX conversion, and the repository ships both the conversion script and a sample React app that does it. Three checkpoints are offered, ViT-B, ViT-L, and ViT-H, trained on SA-1B, a set of about 11 million images with 1.1 billion masks. Installation is a pip install from GitHub on top of PyTorch and torchvision, a CUDA GPU is strongly advised for the encoder, and Hugging Face Transformers also includes a SamModel port. Code and checkpoints are Apache 2.0, while the SA-1B data carries a separate research-only license. The repository has been frozen since 2023 and its README now sends visitors to SAM 2, which adds video, yet the original remains a common building block in labeling tools and pipelines that only need masks for still images.