Vision / Model

SAM 2 / SAM 2.1

Organisation: Meta AI
Family: Segment Anything · version 2 and 2.1
Released: 2024-07-29 (SAM 2 README: 'The previous SAM 2 checkpoints released on July 29, 2024'. SAM 2.1 checkpoints were released September 29, 2024.)
Checkpoints: sam2_hiera_tiny, sam2_hiera_small, sam2_hiera_base_plus, sam2_hiera_large, sam2.1_hiera_tiny, sam2.1_hiera_small, sam2.1_hiera_base_plus, sam2.1_hiera_large
Weights: Open · Apache-2.0
Links: Paper Code

Training data disclosure: Partial list

Level determined for 0% of benchmark cells.

TestTest data seen0
SeenSeen, split unclear0
TrainTraining data seen0
Rel.Related data seen0
NoNot seen0
?Unknown8

“Our released model is trained on SA-V manual + Internal and SA-1B.”

arxiv.org · paper · Appendix D.2.2 · retrieved 2026-10-15

“We also used internally available licensed video data to further augment our training set.”

arxiv.org · paper · Section 5.2 · retrieved 2026-10-15

Training corpora

SA-1B (Segment Anything 1-Billion masks) · 1 source (complete list) · pre-training · homepage
Evidence that SAM 2 / SAM 2.1 uses it

“We first pre-train SAM 2 on static images on the SA-1B dataset.”

arxiv.org · paper · Appendix D.2.1 · retrieved 2026-10-15
Contents of this corpus
- 11M licensed photographs from a photo provider · split: unknown
  The images come from a licensed photo collection, not a web crawl. The provider is not named.
  
  Evidence
  
  “We licensed a new set of 11M images from a provider that works directly with photographers.”
  
  arxiv.org · paper · Section 5 · retrieved 2026-10-15
SA-V (Segment Anything Video) · 1 source (complete list) · pre-training · homepage
Evidence that SAM 2 / SAM 2.1 uses it

“Our released model is trained on SA-V manual + Internal and SA-1B.”

arxiv.org · paper · Appendix D.2.2 · retrieved 2026-10-15
Contents of this corpus
- 50.9K crowdworker-filmed videos · split: unknown
  Evidence
  
  “We collected a new set of 50.9K videos captured by crowdworkers.”
  
  arxiv.org · paper · Section 5.2 · retrieved 2026-10-15
  
  “The videos in the dataset were collected via a contracted third-party vendor.”
  
  arxiv.org · paper · SA-V datasheet · retrieved 2026-10-15

Other declared training data

Internal (internally available licensed video data, 62.9K videos) (not yet matched to a dataset) · split: unknown
The content and provenance of this internal video set are not described.

Evidence

“We also used internally available licensed video data to further augment our training set.”

arxiv.org · paper · Section 5.2 · retrieved 2026-10-15

“Our released model is trained on SA-V manual + Internal and SA-1B.”

arxiv.org · paper · Appendix D.2.2 · retrieved 2026-10-15

Benchmark cells

Industrial anomaly detection

Unknown MVTec AD · training data not disclosed in enough detail

Unknown MVTec AD 2 · training data not disclosed in enough detail

Unknown MVTec LOCO AD · training data not disclosed in enough detail

Unknown VisA · training data not disclosed in enough detail

Unknown Real-IAD · training data not disclosed in enough detail

Unknown MPDD · training data not disclosed in enough detail

Unknown BTAD · training data not disclosed in enough detail

Unknown KolektorSDD2 · training data not disclosed in enough detail

Notes

The image encoder is initialised from MAE pre-trained Hiera, and the SAM 2 paper does not state that MAE's pre-training data. Open-source VOS data (DAVIS, MOSE, YouTubeVOS) appears only in a data-mix ablation, not in the released model. The paper v2 says its results are based on SAM 2.1 and does not give separate training data for 2.1, so one entry covers both.

Sources

SAM 2: Segment Anything in Images and Videos · paper · retrieved 2026-10-15
sam2 README · readme · retrieved 2026-10-15