Event-T2M: Event-level Conditioning for Complex Text-to-Motion Synthesis (ICLR 2026)

The official PyTorch implementation of the paper "Event-T2M: Event-level Conditioning for Complex Text-to-Motion Synthesis".

Setting an Environment

Details

1. Create Conda Environment

conda create -n event-t2m python==3.10.14
conda activate event-t2m

# install pytorch
pip install torch==2.2.2 torchvision==0.17.2 torchaudio==2.2.2 --index-url https://download.pytorch.org/whl/cu121


# install requirements
pip install -r requirements.txt

2. Download the Original Datasets

We conduct experiments on the HumanML3D and KIT-ML datasets. For both datasets, you can download them by following the instructions in here.

3. Prepare the HumanML3D-E Dataset

You can download the completed HumanML3D-E dataset from here.

If you want to prepare the dataset from scratch, follow the steps below:

Since an LLM (Gemini 2.5 flash) was used for HumanML3D-E data preprocessing, an API key is required. Please enter the issued API key on line 6 of src/tools/data_decompose.py.

GOOGLE_API_KEY = "" # your api key here

• For processing,

python src/tools/data_decompose.py

4. Preprocess the Datasets

python src/tools/data_preprocess_decomposed.py --dataset hml3d
python src/tools/data_preprocess_decomposed.py --dataset kit

This will add the following files to the directory:

./dataset/HumanML3D
├── ...
├── data_train.npy
├── data_val.npy
└── data_test.npy

Also, we have released test subsets based on the number of conditions for event-stratified evaluation.

./dataset/HumanML3D
├── ...
├── data_test_condition2.npy
├── data_test_condition3.npy
└── data_test_condition4.npy

5. Download Dependencies and Pre-trained Models

Download and unzip dependencies from here.

Download and unzip pre-trained models from here.

./
├── checkpoints
|   ├── hml3d.ckpt
|   ├── kit.ckpt
├── deps
|   ├── glove
|   ├── t2m_guo
└── ...

Training

Details

• For HumanML3D

python src/train.py trainer.devices=\"0,1\" logger=wandb data=hml3d_event_final \
    data.batch_size=128 data.repeat_dataset=5 trainer.max_epochs=600 \
    callbacks/model_checkpoint=t2m +model/lr_scheduler=cosine model.guidance_scale=4\
    model.noise_scheduler.prediction_type=sample trainer.precision=bf16-mixed

• For KIT-ML

python src/train.py trainer.devices=\"2,3\" logger=wandb data=kit_event_final \
    data.batch_size=128 data.repeat_dataset=5 trainer.max_epochs=1000 \
    callbacks/model_checkpoint=t2m +model/lr_scheduler=cosine model.guidance_scale=4\
    model.noise_scheduler.prediction_type=sample trainer.precision=bf16-mixed

Evaluation

Details

Set model.metrics.enable_mm_metric to True to evaluate Multimodality. Setting model.metrics.enable_mm_metric to False an speed up the evaluation.

python src/eval.py trainer.devices=\"0,\" data=hml3d_event_final data.test_batch_size=128 \
    model=event_final  \
    model.guidance_scale=4 model.noise_scheduler.prediction_type=sample\
    model.denoiser.stage_dim=\"256\*4\" \
    ckpt_path=\"checkpoints/hml3d.ckpt\" model.metrics.enable_mm_metric=true