AlphaMocap model pack
The models AlphaMocap runs to turn one video of a person into body motion. AlphaMocap's installer downloads this pack; the files are ONNX, run by ONNX Runtime on an NVIDIA GPU, and are of no use to anything but the AlphaMocap solver without its code.
These are not alpha-studio's models. They are other people's work, some
exported or changed here to run without PyTorch, each under its own licence
-- read them in LICENSES/ before using any file on its own.
| File | What | From | Licence |
|---|---|---|---|
yolox.onnx |
YOLOX-X person detector | OpenMMLab, the MMPose RTMPose SDK, as shipped | Apache-2.0 |
yolox_raw.onnx |
the same detector cut before its non-maximum suppression (boxes and scores), for live capture | yolox.onnx, cut by AlphaMocap's cut_yolox.py |
Apache-2.0 |
vitpose.onnx + .data |
GEM-X 77-keypoint pose model | NVIDIA GEM-X, weights re-saved into one file | NVIDIA Open Model License |
gem.onnx + .data |
GEM-X motion regressor | NVIDIA GEM-X, its contact head restored | NVIDIA Open Model License |
sam3d.onnx + .data |
SAM 3D Body, to its pose token and its 70 keypoints (3D and 2D), with the MHR body model inside | Meta SAM 3D Body and MHR, exported whole here | SAM License; MHR Apache-2.0 |
sam3d_hand.onnx + .data |
SAM 3D Body's hand decoder, a hand's crop to its keypoints | the same, exported here | SAM License; MHR Apache-2.0 |
videomocap.bin |
decoder statistics, skeleton, rest-skeleton model | derived from GEM-X and NVIDIA SOMA | Apache-2.0 |
face_detector.onnx, face_landmarks.onnx, face_blendshapes.onnx, face_mesh.bin |
a face detector, face landmarks, the 52 ARKit expression weights from them, and the canonical face mesh | Google MediaPipe Face Landmarker, converted from TFLite | Apache-2.0 |
vitpose and gem: Licensed by NVIDIA Corporation under the NVIDIA Open Model License.
The NVIDIA Open Model License is in LICENSES/NVIDIA-Open-Model-License.txt.
sam3d: SAM 3D Body is licensed under the SAM License,
LICENSES/SAM-LICENSE.txt, which among other things does not allow its use
for military purposes and makes its use subject to trade controls.
manifest.json lists every file with its size and SHA-256.
alpha-studio is not affiliated with, or endorsed by, NVIDIA, Meta, Google or OpenMMLab.