What is Sapiens2?
Sapiens2 is Meta's second family of vision models built only for pictures of people, released on 24 April 2026 and accepted at ICLR 2026. The backbones were pretrained on a curated set of 1 billion human images, filtered from a pool of about 4 billion, and come in 0.1B, 0.4B, 0.8B, 1B and 5B sizes, plus a 1B variant trained at 4096 by 3072 pixels. They take images at 1024 by 768 pixels natively.
On top sit five task models. Pose estimation finds 308 keypoints per person, 274 of them on the face, plus hands and feet. Body-part segmentation labels 29 parts. Surface-normal and pointmap models estimate the shape and 3D position of every pixel, and a 1B matting model, added on 15 May 2026, cuts a person out with a soft alpha edge. Pose, segmentation, normals and pointmaps come in 0.4B, 0.8B, 1B and 5B sizes.
It replaces the original Sapiens from 2024, with large gains on every shared task, and borrows its training scheme from DINOv3. The weights download without a gate, but under the Sapiens2 License, Meta's own terms, which ban biometric processing, surveillance and deepfakes on top of the military and ITAR exclusions of Meta's other vision licences.
Is Sapiens2 worth it?
Sapiens2 is a clear step over its predecessor and over general backbones on anything involving people. In the paper, the 5B model reaches 82.3 mAP on 308-keypoint pose estimation against 78.3 for the original Sapiens-2B, and 82.5 mIoU on 29-class body-part segmentation against 58.2. Surface-normal error falls from 12.38 to 6.73 degrees on average, and even the 0.4B model, at 8.63, beats the strongest outside baseline, DAViD-L, at 10.73. All of these are self-reported on the authors' test sets.
It also beats the general model it learned from. With the backbone frozen and only a light head trained, Sapiens2-5B scores 74.7 mAP on pose against 68.2 for DINOv3-7B and 69.6 mIoU on segmentation against 67.6, despite being smaller. A team that only cares about people gets more from Sapiens2 than from DINOv3; a team that also needs cars, shelves or crops does not.
The pieces fit together like a toolkit rather than a single model. Pose is top-down, so it needs a person detector first; the repository points to Meta's DETR ResNet-101. Each task is a separate checkpoint, so a pipeline that wants pose, masks and normals runs three models. The matting model exists only at 1B, and the 4K backbone is released as a pretrained model with no task heads. The paper gives no speed figures, and the 5B model costs about 15.7 trillion operations per image.
The licence decides more than the benchmarks. Beyond the usual military and ITAR exclusions, the Sapiens2 License forbids surveillance, biometric processing, re-identifying people, deepfakes and inferring health or demographic data without consent, and lets Meta audit your use and order copies deleted. Pose tracking for a fitness app or virtual try-on probably sits inside those lines; anything that recognises or profiles individuals does not, and the boundary of "biometric processing" is a question for counsel.
Where it earns its place
A graphics engineer at a virtual try-on start-up
Get body shape and precise body-part masks from a single customer photo
Run the segmentation and pointmap checkpoints on each upload to get 29 part labels and per-pixel 3D positions to fit garments to. Meta reports 82.5 mIoU on segmentation for the 5B model; the 1B is the cheaper starting point.
A computer-vision lead at a fitness-coaching app
Score exercise form from video, including hand and foot position
Detect the person with DETR, then run Sapiens2 pose for 308 keypoints per frame. Have counsel confirm that form scoring falls outside the licence's ban on biometric processing before launch.
A VFX artist at a 12-person post-production studio
Pull clean mattes of actors without a green screen
Use the 1B matting model, which returns an alpha matte and foreground per frame, as a first pass before hand clean-up. The demo Space shows results before any local install; the licence's deepfake ban rules out face-swap work.
Before you commit
- Setup
- Clone the GitHub repo and pip install on Python 3.12+ and PyTorch 2.7+, download checkpoints from Hugging Face, and run the demo scripts; pose also needs the DETR ResNet-101 detector. The backbone alone loads from one standalone Python file.
- Integrations
- Hugging Face Spaces demos for pose, segmentation, normals, pointmaps and matting; training scripts for each task; built on OpenMMLab and Accelerate. No Transformers integration.
- Data and privacy
- Self-hosted, so images stay on your hardware. The Sapiens2 License bans surveillance, biometric processing, re-identification, deepfakes and inferring sensitive data without consent, requires compliance with GDPR, CCPA and biometric laws, and lets Meta audit use and demand deletion.
- Cost at scale
- The 1B task models hold about 3 GB of weights at 16-bit and the 5B about 10 GB, by our arithmetic, so one data-centre GPU serves either; the demo scripts default to two jobs per GPU across eight GPUs. Meta publishes no throughput figures.
- Limits
- Trained at 1024 by 768 pixels; the 4K backbone has no released task heads. Pose is top-down and fails where the person detector misses. Results are self-reported, and the paper states no limitations.
What’s good
- 82.3 mAP on 308-keypoint pose, +4 over Sapiens-2B (Meta)
- 82.5 mIoU on 29-part body segmentation, against 58.2 for Sapiens-2B (Meta)
- Five tasks, including 3D pointmaps and matting, from one backbone family
- No sign-up gate; a single-file backbone needs only torch and safetensors
What’s not
- Licence bans biometric processing and surveillance and allows Meta audits
- People only; no use for objects or scenes
- Pose needs a separate person detector; each task is its own checkpoint
- No published latency figures; 5B model is about 15.7 TFLOPs per image
Best for
Graphics, AR and fitness teams that need dense, high-resolution pose, body-part masks or 3D shape for people in photos and video frames.
Sapiens2 alternatives
Sapiens2: common questions
- Is Sapiens2 worth it?
- Sapiens2 is the most detailed open model for reading human bodies in a photo, under a licence strict enough that many of the obvious products need a lawyer first. We rate it 4.1 out of 5.
- How much does Sapiens2 cost?
- Free under the Sapiens2 License; no sign-up gate. Bans surveillance, biometric processing, re-identification, deepfakes, military and ITAR-related uses, and allows Meta to audit compliance. No hosted API.
- What are the downsides of Sapiens2?
- Licence bans biometric processing and surveillance and allows Meta audits; people only; no use for objects or scenes; pose needs a separate person detector; each task is its own checkpoint; no published latency figures; 5B model is about 15.7 TFLOPs per image.
- Who is Sapiens2 best for?
- Graphics, AR and fitness teams that need dense, high-resolution pose, body-part masks or 3D shape for people in photos and video frames.
- How much does Sapiens2 cost for a team or in production?
- The 1B task models hold about 3 GB of weights at 16-bit and the 5B about 10 GB, by our arithmetic, so one data-centre GPU serves either; the demo scripts default to two jobs per GPU across eight GPUs. Meta publishes no throughput figures.
- Is Sapiens2 safe for company data?
- Self-hosted, so images stay on your hardware. The Sapiens2 License bans surveillance, biometric processing, re-identification, deepfakes and inferring sensitive data without consent, requires compliance with GDPR, CCPA and biometric laws, and lets Meta audit use and demand deletion.
- What are the best Sapiens2 alternatives?
- The closest alternatives we have reviewed are SAM 3 (Segment Anything Model 3), DINOv3 and V-JEPA 2 and V-JEPA 2.1.
Answers checked Nov 9, 2026.
Make Sapiens2? Show this rating on your site
