SAM 3: Segment Anything with Concepts research paper summary (October 2026)

The SAM 3 paper from Meta presents a unified model that detects, segments and tracks all instances of a concept in images and videos. Users describe the concept with a short phrase or an example image, and the model returns masks and identities.

SAM 3 introduces promptable concept segmentation. A data engine annotated a dataset with 4 million unique concept labels, including hard negatives. The model pairs an image-level detector with a memory-based video tracker on a shared backbone and adds a presence head that separates recognition from localization. The authors report doubling the accuracy of existing systems on concept segmentation.

The paper was published in November 2025 and revised in March 2026. Code and checkpoints are in the facebookresearch/sam3 repository under Meta's SAM license, which had 11,818 stars and was updated on September 18, 2026.

• Finds every instance of a productGiven a phrase such as a product name or an example image, SAM 3 returns masks for all matching objects in a photo or video. Retail and logistics teams can count or locate items on shelves or pallets without training a detector for each product.
• Speeds up image labelingMasks from SAM 3 can serve as pre-labels that annotators only correct. Teams building their own vision datasets for quality inspection or document images reduce manual labeling effort.
• Tracks objects in videoThe video tracker keeps identities of objects across frames. Operations teams can follow parts on a conveyor or vehicles in a yard from recorded footage.

• Accepts open-vocabulary text prompts in addition to the clicks and boxes of earlier SAM versions.
• One model for images and video.
• Released under Meta's own SAM license, not a standard open source license.

The paper is freely available on arXiv, and the accompanying code runs on your own hardware in any country.

Meta's SAM license permits commercial use but adds conditions: no military use, no reverse engineering, compliance with trade controls and redistribution only under the same license. Checkpoints on Hugging Face require an access request. If you apply SAM 3 to video of people, for example in warehouses or stores, check data protection rules and employee representation requirements, since tracking can monitor behavior.

No reviews yet. Be the first to share how your team uses it.

Every review shows its verification level. We read each one before it goes live. How we check reviews