Motius ProjFlow HumanML3D
Self-contained Motius artifact for ProjFlow, using the released ACMDM Raw Flow S PatchSize22 motion prior. It bundles the motion model, OpenAI CLIP ViT-B/32 tensors, and official joint statistics, so runtime inference does not clone or import the upstream repository.
from motius import Pipeline
pipe = Pipeline.from_pretrained(
"ZeyuLing/motius-projflow-humanml3d",
bundle_kwargs={"device": "cuda"},
)
completed = pipe.infer_temporal_motion_completion(
source_motion,
generation_mask=generation_mask,
captions=["a person walks forward"],
)
controlled = pipe.infer_kinematic_motion_control(
["a person follows the supplied path"],
source_motion,
control_mode="trajectory",
joint_indices=[0],
axes="xyz",
)
part_controlled = pipe.infer_part_level_motion_control(
["a person walks while following the wrist path"],
source_motion,
control_mode="dense",
joint_indices=[20],
axes="xyz",
)
The native motion representation is world-space HumanML3D SMPL-22 joint
positions with shape (T, 22, 3), metric units, Y up, and 20 fps. ProjFlow
supports at most 196 frames. model_index.json declares the trusted local
Motius pipeline and every required artifact file.
The complete Motius Model Card contains one rendered demo per task, the motion representation contract, and the published 4,012-case Temporal Motion Completion benchmark.
Provenance
- Paper: ProjFlow: Projection Sampling with Flow Matching for Zero-Shot Exact Spatial Motion Control
- Original code: Akihisa-Watanabe/ProjFlow
- Original checkpoint:
Akihisa-Watanabe/ProjFlow,checkpoints/t2m/ACMDM_Raw_Flow_S_PatchSize22/model/latest.tar - Upstream source revision audited:
9550501a439964a73063505b7a52e574ae11a43c - Original checkpoint SHA256:
8a452a0a9cd404ece3e1beb2c9a50dcf8842672ccc3fb5b6f7a59919a0d6ec8f
The upstream repository did not include a repository-level license file at the audited revision. Review the upstream terms before redistribution or commercial use. OpenAI CLIP retains its own license and attribution.