Segment Anything Model

The Segment Anything Model developed by Meta AI Research is a breakthrough image segmentation system capable of identifying and separating any object in an image without additional training, through an interactive interface where users provide prompts including click points, bounding boxes, or text descriptions, and the model instantly generates accurate segmentation masks delineating the target object with pixel-level precision. The promptable segmentation architecture accepts various prompt types as input including foreground and background point clicks that indicate which regions to include or exclude from the mask, bounding boxes that constrain the segmentation area, and text prompts that specify the object category to segment, enabling flexible workflows from fully automatic processing to precise manual refinement. The zero-shot generalization capability enables segmenting objects and categories never seen during training, from everyday items and animals to medical imagery, satellite photos, and industrial inspection scenes, without requiring task-specific fine-tuning or annotated examples. The automatic mode processes entire images by generating segmentation masks for all detectable objects, producing a comprehensive object inventory with quality scores enabling filtering by confidence. The model architecture uses a powerful image encoder that processes the image once, combined with a lightweight mask decoder that runs efficiently for each prompt, enabling real-time interactive segmentation after the initial encoding step. The training data comprises over one billion masks on eleven million images, making it one of the largest segmentation datasets ever created. The model weights are released openly for research and commercial use. The outputs integrate with downstream tasks including image editing, photo manipulation, and augmented reality. Designed for computer vision researchers, developers, designers, medical imaging professionals, and content creators.