Skip to content

Installation

Strands Sapiens has three install steps because Sapiens2 depends on a CUDA-enabled PyTorch that's platform-specific.

1. CUDA PyTorch

Install a CUDA-enabled PyTorch matching your hardware. Some examples:

pip install torch torchvision \
  --index-url https://download.pytorch.org/whl/cu124

# Use the NVIDIA-provided PyTorch wheels for JetPack.
# See: https://developer.nvidia.com/embedded/downloads
pip install --no-cache $TORCH_WHEEL_URL

pip install torch torchvision
# Note: Sapiens2 inference on CPU will be slow but functional.

2. Sapiens2 (upstream)

pip install git+https://github.com/facebookresearch/sapiens2.git

The sapiens package provides:

  • sapiens.backbones.standalone.sapiens2.Sapiens2 - raw backbone
  • sapiens.dense - seg / normal / albedo / pointmap
  • sapiens.pose - 308-keypoint top-down pose

3. Strands Sapiens

pip install git+ssh://git@github.com/cagataycali/strands-sapiens.git

git clone git@github.com:cagataycali/strands-sapiens.git
cd strands-sapiens
pip install -e .

pip install -e '.[dev]'
pytest -q

4. Verify

python -c "from strands_sapiens import sapiens_info; import json; print(json.dumps(sapiens_info(), indent=2))"

You should see:

{
  "status": "success",
  "content": [
    {"text": "sapiens info"},
    {"json": {
      "checkpoint_root": "/home/you/sapiens2_host",
      "checkpoint_root_exists": false,
      "available": {},
      "detector_present": false,
      "detector_type": "none",
      "cuda": { "available": true, "device_count": 1, "device_name": "NVIDIA ..." },
      "sapiens_package": true
    }}
  ]
}

If sapiens_package is false, go back to step 2. If cuda.available is false, go back to step 1.

The next thing to do is download your first checkpoint →