Motius-VerMo-HumanML3D

<h1 align="center">VerMo Model Card</h1> <p align="center"> <strong>A Motius-native autoregressive motion-language baseline aligned on HumanML3D captions.</strong> </p> <p align="center"> <a href="https://github.com/ZeyuLing/Motius/tree/main/motius/models/vermo">Motius Implementation</a> | <a href="https://huggingface.co/ZeyuLing/Motius-VerMo-HumanML3D">Motius Checkpoint</a> </p>

VerMo is a Motius-native research baseline rather than a reproduction of an external paper. The released M2T checkpoint uses a Llama-3.2-1B-Instruct language backbone, a 16K motion tokenizer, and an explicit SMPL-22 motion processor. No external paper or original repository is claimed for this row.

Release Snapshot

Usage

import numpy as np
from motius.pipelines.vermo import VermoPipeline

pipe = VermoPipeline.from_pretrained(
    "ZeyuLing/Motius-VerMo-HumanML3D",
    bundle_kwargs={"device": "cuda"},
    smpl_model_dir="checkpoints/body_models/smpl",
)
motion = np.load("sample.npy")  # denormalized HumanML3D-263
caption = pipe.infer_m2t([motion], lengths=[len(motion)])[0]

M2T Evaluation

Motion Representation

VerMo-138 stores absolute root translation (3), frame-to-frame root translation (3), and 22 local joint rotations in column-major 6D form (132). HumanML3D inputs are recovered to SMPL-22 joints, solved to motion135 with position IK, then repacked explicitly from row-major to column-major 6D.

Motius Components

Direct Loading

from motius import Pipeline

pipeline = Pipeline.from_pretrained("ZeyuLing/Motius-VerMo-HumanML3D")