Segment Anything Model

Meta's foundation model for zero-shot image segmentation. Point at anything in an image and get a mask, no training required. Works across domains (medical, satellite, photos) because it was trained on 1.1 billion masks. You get three model sizes (ViT-B is fast, ViT-H is accurate), flexible prompts (points, boxes, or previous masks), and automatic mask generation that finds every object. The image encoder runs once, then you can segment interactively in real time. Honest take: it's overkill if you already have labeled data for your specific task, but if you're building annotation tools or need segmentation to work on weird domains out of the box, nothing else comes close.