ZeyuLing/Motius-TM2T-HumanML3D

<h1 align="center">TM2T Model Card</h1>

<p align="center"> <strong>Tokenized reciprocal motion-language translation, reproduced as a standalone Motius M2T pipeline.</strong> </p>

<p align="center"> <a href="https://arxiv.org/abs/2207.01696">Paper</a> | <a href="https://ericguo5513.github.io/TM2T/">Project Page</a> | <a href="https://github.com/EricGuo5513/TM2T">Original GitHub</a> | <a href="https://huggingface.co/ZeyuLing/Motius-TM2T-HumanML3D">Motius Checkpoint</a> </p>

TM2T is the ECCV 2022 reciprocal text-to-motion and motion-to-text method. The Motius release contains the HumanML3D VQ tokenizer, motion-to-text Transformer, vocabulary, statistics, and inference runtime. It does not import an original repository checkout.

Release Snapshot

Usage

import numpy as np
from motius.pipelines.tm2t import TM2TPipeline

pipe = TM2TPipeline.from_pretrained(
    "ZeyuLing/Motius-TM2T-HumanML3D",
    bundle_kwargs={"device": "cuda"},
)
motion = np.load("sample.npy")  # denormalized HumanML3D-263
caption = pipe.infer_m2t([motion], lengths=[len(motion)])[0]

M2T Evaluation

Full 4,400-sample evaluation follows the shared HumanML3D M2T protocol. Results are published only after the complete prediction set and metric artifact pass the population and sample-ID checks.

Motion Representation

TM2T normalizes HumanML3D-263 features with its released training statistics. The VQ encoder removes four contact dimensions, maps each clip to discrete motion tokens, and the reciprocal Transformer translates those tokens to text.

Motius Components

Citation

@inproceedings{guo2022tm2t,
  title={TM2T: Stochastic and Tokenized Modeling for the Reciprocal Generation of 3D Human Motions and Texts},
  author={Guo, Chuan and Zuo, Xinxin and Wang, Sen and Cheng, Li},
  booktitle={European Conference on Computer Vision},
  year={2022}
}

Direct Loading

from motius import Pipeline

pipeline = Pipeline.from_pretrained("ZeyuLing/Motius-TM2T-HumanML3D")