Segment Anything, Explained: The Architecture Behind SAM 1, 2 & 3

Meta's Segment Anything Model went from a fixed-category system into a class-agnostic foundation for vision AI — and in this video we break down exactly how it works across SAM 1, 2, and 3. We unpack the architecture behind image encoding, prompt translation, and mask decoding, trace how the data engine bootstrapped millions of masks without manual labeling, and show how SAM 2 solved the video problem with a memory bank. Then we get into SAM 3 (released November 2025) and its leap to Promptable Concept Segmentation — where you stop clicking objects and start describing them with text. By the end, you'll understand why the shift from "segment these 80 classes" to "segment anything on command" is being called the GPT moment for computer vision. ⏱️ CHAPTERS 0:00 – Introduction 0:36 – SAM 1 1:03 – Segment Anything Architecture 3:19 – SA-1B Segmentation Dataset 4:20 – SAM 2 & SA-V Dataset 5:36 – SAM 3 6:04 – Recap 🔑 WHAT YOU'LL LEARN - How SAM's image encoder, prompt encoder, and mask decoder fit together - Why class-agnostic segmentation beats fixed label sets - How Meta's data engine scaled to billions of masks - How SAM 2's memory mechanism enables video tracking - What changed in SAM 3: text/exemplar prompts, the presence token, and the decoupled detector–tracker design 📚 RESOURCES SAM 3 paper: https://arxiv.org/abs/2511.16719 SAM 3 code & checkpoints: https://github.com/facebookresearch/sam3 Meta announcement: https://ai.meta.com/blog/segment-anything-model-3/ Segment Anything Playground: try it on your own images/videos If this helped, leave a comment with what you'd like explained next, and subscribe for more deep dives into foundation models. #SegmentAnything #SAM3 #ComputerVision #MetaAI #DeepLearning #MachineLearning #AIExplained #ImageSegmentation #ObjectDetection #FoundationModels