ZeyuLing/Motius-GEM-X

GEM-X

GEM-X is NVIDIA's monocular whole-body motion estimator built around the SOMA-X parametric body model.

• Paper lineage: GEM: A Generalist Model for Human Motion
• Official source: NVlabs/GEM-X
• Motius checkpoint: ZeyuLing/Motius-GEM-X
• Pinned source revision: 32992550dba114c62243fb55e361311972dce8f9
• Pinned SOMA-X revision: e0f8ff0ecfa3edbbb6058b1e0f08822ee2f84ee5
• Checkpoint SHA-256: 4c1f85ca8c1e11e6588aead49fbc024bf660708def670043e0b537c101ee298e

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 is complete: GEM-X, SAM-3D-Body, DINOv3, ViTPose, YOLOX, MHR, SOMA-X identity/corrective assets, and normalization statistics are stored under their expected paths with SHA-256 provenance. Pipeline.from_pretrained therefore needs no source checkout or second model download.

Motion Representation

GEM-X natively predicts SOMA-77:

• 77-joint axis-angle pose;
• 45 identity coefficients;
• 69 global/body-part scale parameters;
• camera and world root translations;
• 77 named joints and 4,505-vertex low-LOD SOMA meshes.

Motius keeps SOMA-X native. It does not manufacture SMPL vertices from a different topology. Cross-model 3DPW evaluation uses the audited common_hmr15_named_v1 joint subset; PVE is reported as unavailable.

Usage

Create an isolated environment without cloning GEM-X:

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

Run the standard task API:

from pathlib import Path

from motius.motion.representation.monocular_capture import (
    save_monocular_capture_result,
)
from motius.pipelines.gem_x import GemXPipeline

pipeline = GemXPipeline.from_pretrained(
    "ZeyuLing/Motius-GEM-X",
    bundle_kwargs={
        "python_executable": "outputs/envs/gem-x/bin/python",
    },
)
result = pipeline.infer_monocular_motion_capture(
    "input.mp4",
    output_root="outputs/gem_x/run_001",
    materialize_geometry=True,
    render=True,
)
save_monocular_capture_result(
    result,
    Path("outputs/gem_x/run_001/result.npz"),
)

render=True requests the upstream keypoint, in-camera, and world previews. Leave it off for evaluation and batch inference.

Demo

This 768px, 30 FPS preview renders the world-space SOMA-X mesh returned by the public Motius pipeline.

Evaluation Results

Protocol: 3dpw_test_camera_v1, all 24 test videos and 37 official person tracks, evaluated on common_hmr15_named_v1.

Coverage	MPJPE ↓	PA-MPJPE ↓	Acceleration ↓
100.00%	84.38 mm	53.20 mm	5.616 m/s²

The official demo emits an identity camera trajectory when no external visual odometry is supplied. These are camera-space metrics; world-space ranking is unavailable for that run.

Stage Parity

The strict gate compares all persisted boundaries. Deterministic fields are bitwise exact. The official contact IK/SOMA CUDA postprocess is non-deterministic across independent processes, so only its 15 explicitly named descendants use a hard 3e-6 absolute-error ceiling. No global tolerance is applied.

Boundary	Fields	Requirement	Result
Tracking	2	exact	pass
Keypoints and camera	2	exact	pass
Visual features	2	exact	pass
Complete model input	18	exact	pass
Raw network output and deterministic decoded fields	27	exact	pass
Contact-postprocessed model fields	7	atol ≤ 3e-6	max 6.71e-7
SOMA geometry	4	atol ≤ 3e-6	max 2.38e-6
Public result	9	exact except 4 descendants	max 9.54e-7
Total	71	field-scoped policy	pass

python tools/verify_monocular_pipeline_parity.py \
  --profile gem-x \
  --reference outputs/parity/gem_x/reference_trace.npz \
  --candidate outputs/parity/gem_x/motius_trace.npz

License

GEM-X source is Apache-2.0 and the public weights use the NVIDIA Open Model License. SAM-3D-Body, DINOv3, SOMA-X, MHR, and their assets retain their own terms. See the GEM-X attributions and the packaged third-party notices before use.

Direct Loading

from motius import Pipeline

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