Vishal Pandey

Segment Anything 3 just silently dropped on ICLR 🤯 The first SAM let you click on an object to segment it. SAM 2 added video and memory. Now SAM 3 says: just describe what you want — “yellow school bus”, “striped cat”, “red apple” — and it will find and segment every instance of it across images or videos. It’s like segmentation finally learned language, but in a grounded, minimal way, by just saying concepts. Short noun phrases, image exemplars, or both. You say the concept, it understands what you mean, and draws the boundary everywhere it appears. Behind it is a massive data engine: 4 million unique concepts, 52 million verified masks, built through humans and LLMs checking each other’s work. It’s both old-school annotation and self-improving AI loops. The result: roughly 2× better than SAM 2, near-human recognition on open-vocab benchmarks, and real-time speed.