motius-maskcontrol-humanml3d

Motius MaskControl HumanML3D

Self-contained Motius artifact for MaskControl, converted from the released HumanML3D all-anchor checkpoint. Runtime does not import the official exitudio/MaskControl repository.

from motius.pipelines.maskcontrol import MaskControlPipeline

pipe = MaskControlPipeline.from_pretrained(
    "ZeyuLing/motius-maskcontrol-humanml3d",
    bundle_kwargs={"device": "cuda"},
    device="cuda",
)
motion = pipe.infer_t2m(["a person walks forward"], [120], seed=42)[0]

The artifact contains the retrained base and control transformer, RVQ-VAE, residual transformer, length estimator, CLIP ViT-B/32, and HumanML3D normalization statistics. Supported Motius APIs are infer_t2m, infer_control, infer_temporal, infer_body_part, and infer_sequential.

The released checkpoint controls six anchors: pelvis, both feet, head, and both wrists. Body-part and sequential generation are zero-shot composition protocols and share this checkpoint. Sequential composition is limited to 392 frames at 20 fps.

See the full Motius Model Card for coordinate conventions, control masks, evaluation, and reproducible CLI commands.

License And Provenance

The upstream project states that its code is distributed under CC BY-NC-ND 4.0 and that dependencies retain their own licenses. The weights are third-party research artifacts; Motius does not relicense them. The audited upstream revision is 71586fdeb1146ffe6b744d87d573febb10274237.

Direct Loading

from motius import Pipeline

pipeline = Pipeline.from_pretrained("ZeyuLing/motius-maskcontrol-humanml3d")