You can deploy this model and start earning money today!

ZeyuLing/
Motius-GEM-SMPL

Name: Motius-GEM-SMPL
Availability: InStock

motius

human-motion

monocular-motion-capture

Views

No views yet

GEM-SMPL

GEM-SMPL is the SMPL video-motion-estimation release of GEM: A Generalist Model for Human Motion, originally released as GENMO.

Paper: GEM: A Generalist Model for Human Motion
Official source: NVlabs/GENMO
Motius checkpoint: ZeyuLing/Motius-GEM-SMPL
Pinned source revision: 16bebf402d8893184249ee206d957b8248cd8310
Checkpoint SHA-256: 1d15cbe2864d6de61a75e83fdbfe83bec3c7b183eee3d3dcdbd9107e4456454a

Tasks: Monocular Motion Capture

Supported Tasks

Task	Public API	Input	Output
Monocular Motion Capture	`infer_monocular_motion_capture`	RGB video	`MonocularCaptureResult`

Checkpoint

The Motius Hugging Face artifact contains the exact official GEM-SMPL, HMR2, ViTPose, and YOLO checkpoint bytes and a manifest with every SHA-256. The runtime source is shipped inside the motius wheel at its pinned revision; inference never imports another repository checkout.

SMPL and SMPL-X files are license-gated and are not redistributed. Download them into checkpoints/body_models/ using checkpoints/body_models/README.md.

Motion Representation

GEM-SMPL predicts camera-space and gravity-aligned global body parameters. Motius preserves the native 21-joint body pose, root orientation, translation, and ten shape coefficients. Geometry materialization exposes:

SMPL-24 named joints;
6,890-vertex SMPL meshes;
camera and world root trajectories;
per-frame camera intrinsics;
the source video clock without temporal resampling.

The internal SMPL-X body layer is converted with the same fixed sparse SMPL-X-to-SMPL map used by the pinned implementation.

Usage

Create an isolated environment without cloning GENMO:

bash

python3.10 -m venv outputs/envs/gem-smpl
outputs/envs/gem-smpl/bin/pip install -e ".[gem-smpl]"

Run the standard task API:

python

from pathlib import Path

from motius.motion.representation.monocular_capture import (
    save_monocular_capture_result,
)
from motius.pipelines.gem_smpl import GemSmplPipeline

pipeline = GemSmplPipeline.from_pretrained(
    "ZeyuLing/Motius-GEM-SMPL",
    bundle_kwargs={
        "python_executable": "outputs/envs/gem-smpl/bin/python",
        "body_models_root": "checkpoints/body_models",
    },
)
result = pipeline.infer_monocular_motion_capture(
    "input.mp4",
    output_root="outputs/gem_smpl/run_001",
    materialize_geometry=True,
    render=True,
)
save_monocular_capture_result(
    result,
    Path("outputs/gem_smpl/run_001/result.npz"),
)

render=True requests the upstream in-camera/world previews. It can be left off for evaluation and batch inference.

Demo

This 768px, 30 FPS preview renders the world-space SMPL vertices returned by the public Motius pipeline.

Evaluation Results

Protocol: 3dpw_test_camera_v1, one inference item per official 3DPW person track using the released target-crop protocol.

Coverage	MPJPE ↓	PA-MPJPE ↓	Acceleration ↓
100.00%	64.46 mm	46.45 mm	5.713 m/s²

Stage Parity

The migration gate replays the same video, checkpoint bytes, licensed body models, precomputed visual tensors, and random seed through the pinned official source and the Motius package.

Boundary	Fields	Requirement	Result
Tracking	1	exact	pass
Keypoints	1	exact	pass
Visual features	2	exact	pass
Complete model input	15	exact	pass
Network and decoded output	9	exact	pass
SMPL geometry	4	exact	pass
Public result	10	exact	pass
Total	42	`rtol=0`, `atol=0`	pass

bash

python tools/verify_monocular_pipeline_parity.py \
  --reference outputs/parity/gem_smpl/reference_trace.npz \
  --candidate outputs/parity/gem_smpl/motius_trace.npz

License

The vendored source retains the NVIDIA OneWay Noncommercial license. Public weights retain the NVIDIA Open Model License. SMPL and SMPL-X have separate terms. See the GEM-SMPL attributions and the packaged license files before use.

Direct Loading

python

from motius import Pipeline

pipeline = Pipeline.from_pretrained("ZeyuLing/Motius-GEM-SMPL")