ZeyuLing/Motius-MotionGPT3-HumanML3D

MotionGPT3 Model Card

A continuous-latent bimodal motion-language model, packaged for HumanML3D motion captioning.

Paper | Project Page | Original GitHub | Motius Checkpoint

MotionGPT3 separates text and motion processing into modality-specific branches with shared attention. Unlike tokenized motion-language models, it represents motion in a continuous VAE latent space. The Motius artifact packages the final official multi-task checkpoint and all model/tokenizer configuration required by Pipeline.from_pretrained.

Release Snapshot

Item	Value
Released task	M2T
Input representation	HumanML3D-263, 20 fps
Motion latent	Continuous temporal VAE latent
Language model	GPT-2-family bimodal Transformer
Checkpoint provenance	Official final MotionGPT3 checkpoint
Checkpoint	ZeyuLing/Motius-MotionGPT3-HumanML3D
Pipeline	motius.pipelines.motiongpt3.MotionGPT3Pipeline

Usage

import numpy as np
from motius.pipelines.motiongpt3 import MotionGPT3Pipeline

pipe = MotionGPT3Pipeline.from_pretrained(
    "ZeyuLing/Motius-MotionGPT3-HumanML3D",
    bundle_kwargs={"device": "cuda"},
)
motion = np.load("sample.npy")  # denormalized HumanML3D-263
caption = pipe.infer_m2t([motion], lengths=[len(motion)])[0]

M2T Evaluation

Protocol	Samples	BLEU-4	ROUGE-L	CIDEr	BERT F1	R@1	R@2	R@3	Matching
HumanML3D M2T	4,400	-	-	-	-	-	-	-	-

Motius Components

Component	Path
Pipeline	motius/pipelines/motiongpt3/pipeline.py
Bundle	motius/models/motiongpt3/bundle.py
Bimodal GPT runtime	motius/models/motiongpt3/mot_example_gpt2_sepattn.py
Generation runtime	motius/models/motiongpt3/mot_example_gpt2_sepattn_gen.py

Citation

@misc{zhu2025motiongpt3,
  title={MotionGPT3: Human Motion as a Second Modality},
  author={Zhu, Bingfan and Jiang, Biao and Wang, Sunyi and Tang, Shixiang and Chen, Tao and Luo, Linjie and Zheng, Youyi and Chen, Xin},
  year={2025},
  eprint={2506.24086},
  archivePrefix={arXiv}
}