Sapiens
MetaSapiens is Meta's family of high-resolution vision transformers for images of people, pretrained on a billion human images. The current generation, Sapiens2, spans 0.1B to 5B parameters and ships task checkpoints for whole-body pose with 308 keypoints, 29-part body segmentation, surface normals, pointmaps and human matting, along with fine-tuning code for each task.
Openness
3 medium confidence- weights
- open (0.1B to 5B pretrained backbones and pose, segmentation, normal, pointmap and matting checkpoints on the Hub, ungated)
- data
- closed (a curated set of one billion human images, not released)
- code
- partial (inference and per-task fine-tuning code; no pretraining code)
- license
- Sapiens2-License (code and weights; bars surveillance, biometric processing and re-identification, with no commercial or user cap)
- sapiens-v1
- caveat (the superseded first-generation Sapiens checkpoints stay on the Hub under CC-BY-NC-4.0)
Sapiens2's code and weights are under Meta's Sapiens2 License, which allows commercial use but forbids surveillance, biometric processing, re-identifying people and deepfakes. The billion-image pretraining set is not released, and the repository carries fine-tuning code for each task but no pretraining code.
- https://arxiv.org/abs/2604.21681 recorded 2026-11-12
Sapiens2 abstract: "we pretrain on a curated dataset of 1 billion high-quality human images"; no dataset release is described.
- https://huggingface.co/api/models?author=facebook&search=sapiens&limit=100&expand[]=downloads&expand[]=cardData&expand[]=gated recorded 2026-11-12
Every facebook/sapiens2-* repository gated false with license_name sapiens2-license; the first-generation facebook/sapiens-* repositories gated false and tagged cc-by-nc-4.0.
- https://raw.githubusercontent.com/facebookresearch/sapiens2/main/docs/MODEL_ZOO.md recorded 2026-11-12
Model zoo: pretrained, pose, segmentation, surface normal, pointmap and matting checkpoints, each linked to a facebook/sapiens2-* repository on the Hub.
- https://raw.githubusercontent.com/facebookresearch/sapiens2/main/LICENSE.md recorded 2026-11-12
Sapiens2 License: a "non-exclusive, worldwide, non-transferable and royalty-free limited license" to use, distribute and modify the Sapiens Materials, including "trained model weights"; use may not be "for purposes of surveillance", "for biometric processing" or "for attempting to identify or re-identify any individual".
- https://raw.githubusercontent.com/facebookresearch/sapiens2/main/README.md recorded 2026-11-12
README: "A family of high-resolution transformers pretrained on 1 billion human images"; Sapiens2-0.1B to 5B model card; inference and training docs for pose, segmentation, normal, pointmap and matting; "This project is licensed under the Sapiens2 License."
- https://ungh.cc/repos/facebookresearch/sapiens2/files/main recorded 2026-11-12
Repository tree of 331 paths: train.py and per-size train scripts for the pose and dense tasks under sapiens/pose and sapiens/dense; no pretraining script.
Adoption
3 medium confidenceHugging Face downloads over the trailing 30 days for the ten most-downloaded Sapiens2 checkpoints, led by the 1B pose model. The first-generation Sapiens checkpoints are not counted.
- https://huggingface.co/api/models?author=facebook&search=sapiens&limit=100&expand[]=downloads&expand[]=cardData&expand[]=gated recorded 2026-11-12
sapiens2-pose-1b 40,678, seg-0.4b 14,268, pose-5b 13,777, seg-1b 12,754, matting-1b 10,172, pose-0.4b 9,291, pointmap-5b 5,532, seg-5b 5,128, normal-1b 4,460 and pretrain-0.4b 3,582 downloads in the trailing 30 days (119,642 together); the 24 sapiens2 repositories sum to 135,120.
Capability
4 medium confidenceSapiens2 returns per-pixel geometry and a foreground matte for any photo of a person without training, as Depth Anything does for whole scenes. It covers human subjects only, and its pose and segmentation heads use fixed keypoint and body-part sets.
- https://raw.githubusercontent.com/facebookresearch/sapiens2/main/README.md recorded 2026-11-12
README vision tasks: Pose Estimation "308 whole-body keypoints", Body-Part Segmentation "29 body parts", Surface Normal Estimation "per-pixel normals", Pointmap Estimation "per-pixel 3D points", Human Matting "alpha matte + foreground"; all models trained at 1024x768.
Verified 2026-11-12