Coordinating Multiple Conditions for Trajectory-Controlled Human Motion Generation (TMM 2026)

This is the official PyTorch implementation for the paper "Coordinating Multiple Conditions for Trajectory-Controlled Human Motion Generation".

---

🛠 Preparation

1. Configure Conda Environment

We recommend using Conda to set up the environment. Run the following commands:

conda env create -f environment.yml
conda activate cmc

pip install git+https://github.com/openai/CLIP.git

2. Download Pre-trained Weights

Download the pre-trained weights for the HumanML3D and KIT datasets from our Google Drive link.

Extract the downloaded files into the output/ directory. The file structure should look exactly like this:

./output/
├── cmc_humanml3d/
│   ├── cmc_humanml3d_s1.pth
│   └── cmc_humanml3d_s2.pth
└── cmc_kit/
    ├── cmc_kit_s1.pth
    └── cmc_kit_s2.pth

3. Download the evaluator

Follow MoMask Github. The structure should look like this:

./checkpoints/
├── t2m/
│   ├── text_mot_match
└── kit/
    ├── text_mot_match

4. Soft link to your dataset path

ln -s /path/to/dataset/HumanML3D/ dataset/HumanML3D/
ln -s /path/to/dataset/KIT-ML/ dataset/KIT-ML/

---

🚀 Training

To train your own model from scratch, execute the following commands.

Training CMC HumanML3D Stage 1:

python train.py \
    --exp_name cmc_humanml3d_s1 \
    --batch_size 128 --gpu 2 --overwrite --save_iter 10000 \
    --total_iter 600000 --lr 1e-4 --lr-scheduler 300000 \
    --modeltype s1 --multi_joint_control

Training CMC HumanML3D Stage 2:

python train.py \
    --exp_name cmc_humanml3d_s2 \
    --batch_size 128 --gpu 2 --overwrite --save_iter 10000 \
    --total_iter 600000 --lr 1e-4 --lr-scheduler 300000 \
    --modeltype s2 --multi_joint_control

---

📊 Evaluation

We provide several evaluation scripts for different configurations. For multi-joint control, you can specify joints using --control_joint (e.g., --control_joint 0 20).

1. HumanML3D (Stage 1 + Stage 2) | DDPM 1000+1000 steps:

python eval_cmc.py \
    --resume_root output/cmc_humanml3d/cmc_humanml3d_s1.pth \
    --resume_trans output/cmc_humanml3d/cmc_humanml3d_s2.pth \
    --dataset_name t2m \
    --control_joint 0 --density 100 --gpu 3

2. HumanML3D (Stage 1 + Stage 2) | DDIM 100+100 steps:

python eval_cmc.py \
    --resume_root output/cmc_humanml3d/cmc_humanml3d_s1.pth \
    --resume_trans output/cmc_humanml3d/cmc_humanml3d_s2.pth \
    --dataset_name t2m \
    --control_joint 0 --density 100 --gpu 3 \
    --S1_diffusion_step 100 --S2_diffusion_step 100

3. HumanML3D (Stage 2 Only) | Text-to-Motion:

python eval_cmc.py \
    --resume_trans output/cmc_humanml3d/cmc_humanml3d_s2.pth \
    --dataset_name t2m \
    --gpu 3 \
    --only_t2m_s2 1

4. HumanML3D Sample:

python sample.py

This script will visualize motion in a web page and save html file.