For more information, please refer to our work FRoM-W1: Towards General Humanoid Whole-Body Control with Language Instructions.

Git Branches

• main: Original T2M
• humanml3d-smpl-dist: distributed training of T2M
• humanml3d-smplx-dist: distributed training of T2M-SMPL-X

We have devoted significant effort to training this version of the model very carefully. So if you find this version useful, please cite our work along with the original T2M paper:

@article{DBLP:journals/corr/abs-2601-12799,
  author       = {Peng Li and
                  Zihan Zhuang and
                  Yangfan Gao and
                  Yi Dong and
                  Sixian Li and
                  Changhao Jiang and
                  Shihan Dou and
                  Zhiheng Xi and
                  Enyu Zhou and
                  Jixuan Huang and
                  Hui Li and
                  Jingjing Gong and
                  Xingjun Ma and
                  Tao Gui and
                  Zuxuan Wu and
                  Qi Zhang and
                  Xuanjing Huang and
                  Yu{-}Gang Jiang and
                  Xipeng Qiu},
  title        = {FRoM-W1: Towards General Humanoid Whole-Body Control with Language
                  Instructions},
  journal      = {CoRR},
  volume       = {abs/2601.12799},
  year         = {2026},
  url          = {https://doi.org/10.48550/arXiv.2601.12799},
  doi          = {10.48550/ARXIV.2601.12799},
  eprinttype   = {arXiv},
  eprint       = {2601.12799},
  timestamp    = {Tue, 24 Mar 2026 08:45:06 +0100},
  biburl       = {https://dblp.org/rec/journals/corr/abs-2601-12799.bib},
  bibsource    = {dblp computer science bibliography, https://dblp.org}
}

Given a textual description for example, "the figure rises from a lying position and walks in a counterclockwise circle, and then lays back down the ground", our approach generates a diverse set of 3d human motions that are faithful to the provided text.

Anaconda is recommended to create this virtual environment.

conda create -f environment.yaml
conda activate text2motion_pub

If you cannot successfully create the environment, here is a list of required libraries:

Python = 3.7.9   # Other version may also work but are not tested.
PyTorch = 1.6.0 (conda install pytorch==1.6.0 torchvision==0.7.0 -c pytorch)  #Other version may also work but are not tested.
scipy
numpy
tensorflow       # For use of tensorboard only
spacy
tqdm
ffmpeg = 4.3.1   # Other version may also work but are not tested.
matplotlib = 3.3.1

After all, if you want to generate 3D motions from customized raw texts, you still need to install the language model for spacy.

python -m spacy download en_core_web_sm

If you just want to play our pre-trained models, you don't need to download datasets.

We are using two 3D human motion-language dataset: HumanML3D and KIT-ML. For both datasets, you could find the details as well as download link [here].
Please note you don't need to clone that git repository, since all related codes have already been included in current git project.

Download and unzip the dataset files -> Create a dataset folder -> Place related data files in dataset folder:

mkdir ./dataset/

Take HumanML3D for an example, the file directory should look like this:

./dataset/
./dataset/HumanML3D/
./dataset/HumanML3D/new_joint_vecs/
./dataset/HumanML3D/texts/
./dataset/HumanML3D/Mean.mpy
./dataset/HumanML3D/Std.npy
./dataset/HumanML3D/test.txt
./dataset/HumanML3D/train.txt
./dataset/HumanML3D/train_val.txt
./dataset/HumanML3D/val.txt  
./dataset/HumanML3D/all.txt 

Create a checkpoint folder to place pre-traine models:

mkdir ./checkpoints

./checkpoints/t2m/
./checkpoints/t2m/Comp_v6_KLD01/           # Text-to-motion generation model
./checkpoints/t2m/Decomp_SP001_SM001_H512/ # Motion autoencoder
./checkpoints/t2m/length_est_bigru/        # Text-to-length sampling model
./checkpoints/t2m/text_mot_match/          # Motion & Text feature extractors for evaluation

All intermediate meta files/animations/models will be saved to checkpoint directory under the folder specified by argument "--name".

python train_decomp_v3.py --name Decomp_SP001_SM001_H512 --gpu_id 0 --window_size 24 --dataset_name t2m

python train_decomp_v3.py --name Decomp_SP001_SM001_H512 --gpu_id 0 --window_size 24 --dataset_name kit

python train_length_est.py --name length_est_bigru --gpu_id 0 --dataset_name t2m

python train_length_est.py --name length_est_bigru --gpu_id 0 --dataset_name kit

python train_comp_v6.py --name Comp_v6_KLD01 --gpu_id 0 --lambda_kld 0.01 --dataset_name t2m

python train_comp_v6.py --name Comp_v6_KLD005 --gpu_id 0 --lambda_kld 0.005 --dataset_name kit

python train_tex_mot_match.py --name text_mot_match --gpu_id 1 --batch_size 8 --dataset_name t2m

python train_tex_mot_match.py --name text_mot_match --gpu_id 1 --batch_size 8 --dataset_name kit

python eval_comp_v6.py --name Comp_v6_KLD01 --est_length --repeat_time 3 --num_results 10 --ext default --gpu_id 1

where --est_length asks the model to use sampled motion lengths for generation, --repeat_time gives how many sampling rounds are carried out for each description. This script will results in 3x10 animations under directory ./eval_results/t2m/Comp_v6_KLD01/default/.

python gen_motion_script.py --name Comp_v6_KLD01 --text_file input.txt --repeat_time 3 --ext customized --gpu_id 1

This will generate 3 animated motions for each description given in text_file ./input.txt.

If you find problem with installing ffmpeg, you may not be able to animate 3d results in mp4. Try gif instead.

python final_evaluation.py 

This will evaluate the model performance on HumanML3D dataset by default. You could also run on KIT-ML dataset by uncommenting certain lines in ./final_evaluation.py. The statistical results will saved to ./t2m_evaluation.log.

Contact Chuan Guo at cguo2@ualberta.ca for any questions or comments.