ZeyuLing/Motius-HYMotion-V2M

HYMotion-V2M for Motius

This artifact packages the HYMotion-V2M generator, motion statistics, SAM-3D-Body/MHR image encoder assets, and YOLOX detector for the first-class Motius pipeline. It does not require an external source checkout.

• Paper: https://arxiv.org/abs/2512.23464
• Official source: https://github.com/Tencent-Hunyuan/HY-Motion-1.0
• Motius source: https://github.com/ZeyuLing/Motius
• Source checkpoint SHA-256: ed301af6b6fc6bc22dc69d8c7f48c1d6b2fff31d48a1717c8f5bc5ea92bc71df

Usage

from motius import Pipeline

pipeline = Pipeline.from_pretrained(
    "ZeyuLing/Motius-HYMotion-V2M",
    bundle_kwargs={"device": "cuda"},
)
result = pipeline.infer_monocular_motion_capture(
    "input.mp4",
    work_dir="outputs/hymotion_v2m/run_001",
)

The checkpoint embeds the FK and mesh buffers required for inference. It does not redistribute the original licensed SMPL-H parameter file. See ATTRIBUTIONS.md, License.txt, and the bundled third-party license files before use or redistribution.

Numerical parity

Motius compares named intermediate stages with exact element-wise equality (rtol=0, atol=0). The release check covers tracking, camera conditioning, SAM-3D tokens, model conditioning, every inference window, stitching, SMPL-H forward kinematics, ground alignment, and the public result contract.