Motius-ViMoGen-1.3B-HumanML3D

<!-- This model card is synchronized from docs/model_zoo/vimogen.md by tools/sync_model_zoo_cards.py. -->

ViMoGen

Text-to-motion baseline integrated into the motius Model Zoo. The runtime is motius-native and does not import the upstream repository at inference time: the released ViMoGen transformer, scheduler, smoothing step, and DART276 motion representation bridge live under motius.models.motion.vimogen and motius.motion.representation.dart276.

---

Weights

Current motius artifact:

Load through the same from_pretrained surface as the other reproduced baselines:

from motius.pipelines.vimogen import ViMoGenPipeline

pipe = ViMoGenPipeline.from_pretrained(
    "ZeyuLing/Motius-ViMoGen-1.3B-HumanML3D",
    device="cuda",
)

motions_276 = pipe.infer_t2m(
    ["Full-body shot, stable camera. A person walks forward at an average pace."],
    [200],
    seed=0,
)

ViMoGenBundle.from_pretrained reads model_index.json. If the Wan2.1 base assets are not already available locally, the bundle resolves the public Wan-AI/Wan2.1-T2V-1.3B Hub repo declared by wan_repo_id.

---

Motion Representation

ViMoGen emits DART276, the global DART-style representation:

text -> UMT5-XXL embeddings -> WanVideoTM2M DiT -> denormalized DART276
     -> dart276_to_motion135(...) for SMPL mesh / MotionCLIP / physics
     -> motion135_to_motion272(...) for MotionStreamer-272 evaluator

The public conversion API is:

from motius.motion.representation.dart276 import dart276_to_motion135

motion_135 = dart276_to_motion135(motion_276, rotation_convention="row")

See docs/motion/representations.md for the DART276 channel layout and the root / coordinate-system convention.

---

Evaluation

The leaderboard row uses the HumanML3D official-test split (n=4042) and the shared corrected caption set. ViMoGen is sensitive to terse HumanML3D captions, so generation uses a ViMoGen-style prompt rewrite derived from the corrected caption. The rewrite adds presentation/context details such as camera, floor, and motion-capture clothing while preserving the original action content. The semantic evaluators are still computed against the same corrected HumanML3D caption protocol used by the other methods.

MotionStreamer-272 and MotionCLIP

Physical Diagnostics

---

Implementation Notes

• motius-native runtime: motius.models.motion.vimogen.network vendors the required ViMoGen transformer modules and scheduler.
• No ref_repo dependency: full-set HumanML3D inference uses scripts/eval/vimogen_t2m_humanml3d.py with ViMoGenBundle / ViMoGenPipeline.
• Prompt sensitivity: for leaderboard-quality generation, use the ViMoGen-style prompt rewrite workflow before inference. The plain corrected HumanML3D captions produce substantially weaker text following.
• Evaluator bridge: DART276 outputs are converted to repository motion_135, then to MotionStreamer-272 or MotionCLIP-135 for the shared cross-model leaderboard protocol.

Direct Loading

from motius import Pipeline

pipeline = Pipeline.from_pretrained("ZeyuLing/Motius-ViMoGen-1.3B-HumanML3D")