ZeyuLing/Motius-MotionCanvas-0.46B

Motius MotionCanvas 0.46B

MotionCanvas is the Motius-native HYMotion M2M checkpoint for temporal motion completion, kinematic control, and text-guided motion editing. This epoch 2354 artifact is self-contained: it includes the motion transformer, learned M2M condition parameters, MotionCanvas-198 normalization statistics, SMPL-22 rest offsets, Qwen3, and CLIP.

Load

python -m pip install git+https://github.com/ZeyuLing/Motius.git
python -m pip install torchdiffeq einops mmcv-lite

import torch
from motius import Pipeline

pipe = Pipeline.from_pretrained(
    "ZeyuLing/Motius-MotionCanvas-0.46B",
    cache_dir="/path/to/large/huggingface-cache",
    bundle_kwargs={"device": "cuda", "text_dtype": "bf16"},
    num_steps=50,
)

source = torch.from_numpy(reference_motion_198)[None].cuda()
generation_mask = torch.zeros(1, source.shape[1], device="cuda")
generation_mask[:, 90:180] = 1

result = pipe.infer_temporal_motion_completion(
    source,
    generation_mask,
    captions="a person turns left and continues walking",
    seed=42,
)
motion_198 = result["motion_198"]

The complete artifact is about 20 GB. cache_dir must be a top-level loader argument; set HF_XET_CACHE too when the default Xet cache is on a small system disk.

generation_mask uses 1=generate and 0=preserve. It can have shape (B,T) or (B,T,198). The same artifact exposes:

• infer_temporal_motion_completion
• infer_motion_inbetweening
• infer_keyframe_motion_control
• infer_kinematic_motion_control
• infer_motion_editing
• infer_motion_repair
• infer_m2m

Demos

Temporal completion	Kinematic control

Motion editing	Trajectory editing

The demos/ directory also contains the original 1080p MP4 renders.

Representation

MotionCanvas-198 runs at 30 fps on a 360-frame canvas:

• 3 world-space root translation values
• 132 row-major local rotation-6D values for 22 SMPL joints
• 63 pelvis-relative FK joint-position values for 21 non-root joints

Mean.npy and Std.npy are the exact HY-Motion-201 statistics with the unused pelvis-RIC triplet removed. See normalization_provenance.md.

Evaluation

All rows use the 4,012-motion HumanML3D official test split. FID is computed in per-sample L2-normalized uTMR embedding space.

Setting	Text	R@1	R@2	R@3	FID	MM-Dist	Condition cm	Foot skating
First frame	yes	0.6914	0.8410	0.8988	0.0139	27.9444	0.0000	0.1380
Prefix 20%	yes	0.6931	0.8406	0.8993	0.0091	27.8049	0.0000	0.0840
Prefix 20%	no	0.3411	0.4731	0.5532	0.1043	40.0292	0.0000	0.0960
First + last frame	yes	0.6883	0.8380	0.8972	0.0135	28.0997	0.0000	0.1840
First + last 10%	yes	0.6941	0.8452	0.9030	0.0052	27.7324	0.0000	0.1100
First + last 10%	no	0.4283	0.5758	0.6562	0.0837	36.9094	0.0000	0.1250
Adaptive keyframes	yes	0.6875	0.8441	0.9031	0.0014	27.8142	0.0000	0.1184
Adaptive keyframes	no	0.6589	0.8218	0.8878	0.0031	28.6949	0.0000	0.1358

The complete protocol and physical metrics are maintained in the MotionCanvas Model Card and the Temporal Condition Leaderboard.

Provenance

MotionCanvas is implemented and trained in Motius. Its transformer and text-conditioning design build on HY-Motion 1.0. The M2M objective, arbitrary-condition data pipeline, clean-imputation solver, sparse rollout, and this checkpoint are Motius-native.