Meta Sapiens2 Human-Centric Vision Model GitHub Repository

Meta's Sapiens2 is a family of high-resolution transformer-based vision models pretrained on 1 billion human images. It aims for strong performance across human-centric tasks such as pose estimation, body-part segmentation, surface normal, and pointmap, and was released as a GitHub repository alongside an ICLR 2026 paper.

The model cards are divided as follows.

• Sapiens2-0.1B: 0.114B params, 0.342T FLOPs
• Sapiens2-0.4B: 0.398B params, 1.260T FLOPs
• Sapiens2-0.8B: 0.818B params, 2.592T FLOPs
• Sapiens2-1B: 1.462B params, 4.715T FLOPs
• Sapiens2-1B (4K): 1.607B params
• Sapiens2-5B: 5.071B params, 15.722T FLOPs

All models were trained at 1024×768 (H×W) with patch size 16. The exception, Sapiens2-1B (4K), uses 4096×3072 resolution and use_tokenizer=True.

Running it is relatively simple. With only torch and safetensors, you can run the backbone forward pass, and sapiens/backbones/standalone/sapiens2.py and the v1 sapiens.py can be copied as-is into projects outside the repo.

• Installation requirements: Python 3.12+, PyTorch 2.7+
• Checkpoints: download from MODEL_ZOO.md and place in $SAPIENS_CHECKPOINT_ROOT (default ~/sapiens2_host)
• Pose inference: requires adding the detr-resnet-101-dc5 detector
• Inference/training documentation is separated for pose, segmentation, normal, and pointmap