Awesome

MotionLCM: Real-time Controllable Motion Generation via Latent Consistency Model

Wenxun Dai^😎, Ling-Hao Chen^😎, Jingbo Wang^🥳, Jinpeng Liu^😎, Bo Dai^🥳, Yansong Tang^😎

^😎Tsinghua University, ^🥳Shanghai AI Laboratory (Correspondence: Jingbo Wang and Bo Dai).

🤩 Abstract

This work introduces MotionLCM, extending controllable motion generation to a real-time level. Existing methods for spatial-temporal control in text-conditioned motion generation suffer from significant runtime inefficiency. To address this issue, we first propose the motion latent consistency model (MotionLCM) for motion generation, building upon the latent diffusion model. By adopting one-step (or few-step) inference, we further improve the runtime efficiency of the motion latent diffusion model for motion generation. To ensure effective controllability, we incorporate a motion ControlNet within the latent space of MotionLCM and enable explicit control signals (e.g., initial poses) in the vanilla motion space to further provide supervision for the training process. By employing these techniques, our approach can generate human motions with text and control signals in real-time. Experimental results demonstrate the remarkable generation and controlling capabilities of MotionLCM while maintaining real-time runtime efficiency.

🤼‍♂ Arena

📢 News

• [2024/12/12] Release MotionLCM-V2 and Hugging Face Blog.
• [2024/08/15] Support the training of motion latent diffusion model (MLD).
• [2024/08/09] Support the training of motion VAE.
• [2024/07/02] MotionLCM is officially accepted by ECCV 2024.
• [2024/05/01] Upload paper and release code.

👨‍🏫 Quick Start

This section provides a quick start guide to set up the environment and run the demo. The following steps will guide you through the installation of the required dependencies, downloading the pretrained models, and preparing the datasets.

conda create python=3.10.12 --name motionlcm
conda activate motionlcm

Install the packages in requirements.txt.

pip install -r requirements.txt

We test our code on Python 3.10.12 and PyTorch 1.13.1.

If you have the sudo permission, install ffmpeg for visualizing stick figure (if not already installed):

sudo apt update
sudo apt install ffmpeg
ffmpeg -version  # check!

If you do not have the sudo permission to install it, please install it via conda:

conda install conda-forge::ffmpeg
ffmpeg -version  # check!

Run the following command to install git-lfs:

conda install conda-forge::git-lfs

Run the script to download dependencies materials:

bash prepare/download_glove.sh
bash prepare/download_t2m_evaluators.sh
bash prepare/prepare_t5.sh
bash prepare/download_smpl_models.sh

Run the script to download the pre-trained models:

bash prepare/download_pretrained_models.sh

The folders experiments_recons experiments_t2m and experiments_control store pre-trained models for motion reconstruction, text-to-motion and motion control respectively.

Visit the Google Driver to download the previous dependencies and models.

Please refer to HumanML3D for text-to-motion dataset setup. Copy the result dataset to our repository:

cp -r ../HumanML3D/HumanML3D ./datasets/humanml3d

After the whole setup pipeline, the folder structure will look like:

MotionLCM
├── configs
├── configs_v1
├── datasets
│   ├── humanml3d
│   │   ├── new_joint_vecs
│   │   ├── new_joints
│   │   ├── texts
│   │   ├── Mean.npy
│   │   ├── Std.npy
│   │   ├── ...
│   └── humanml_spatial_norm
│       ├── Mean_raw.npy
│       └── Std_raw.npy
├── deps
│   ├── glove
│   ├── sentence-t5-large
|   ├── smpl_models
│   └── t2m
├── experiments_control
│   ├── spatial
│   │   └── motionlcm_humanml
│   │       ├── motionlcm_humanml_s_all.ckpt
│   │       └── motionlcm_humanml_s_pelvis.ckpt
│   └── temproal
│   │   └── motionlcm_humanml
│   │       ├── motionlcm_humanml_t_v1.ckpt
│   │       └── motionlcm_humanml_t.ckpt
├── experiments_recons
│   └── vae_humanml
│       └── vae_humanml.ckpt
├── experiments_t2m
│   ├── mld_humanml
│   │   ├── mld_humanml_v1.ckpt
│   │   └── mld_humanml.ckpt
│   └── motionlcm_humanml
│       ├── motionlcm_humanml_v1.ckpt
│       └── motionlcm_humanml.ckpt
├── ...

🚨 Following is based on MotionLCM-V2 (Check configs_v1 for V1).

🎬 Demo

MotionLCM provides three main functionalities: motion reconstruction, text-to-motion and motion control. The following commands demonstrate how to use the pre-trained models to generate motions. The outputs will be stored in ${cfg.TEST_FOLDER} / ${cfg.NAME} / demo_${timestamp} (experiments_t2m_test/motionlcm_humanml/demo_2024-04-06T23-05-07).

If you haven't completed the data preparation in 👨‍🏫 Quick Start, make sure to use the following command to download a tiny humanml3d dataset.

bash prepare/prepare_tiny_humanml3d.sh

python demo.py --cfg configs/vae.yaml

python demo.py --cfg configs/mld_t2m.yaml --example assets/example.txt
python demo.py --cfg configs/motionlcm_t2m.yaml --example assets/example.txt

python demo.py --cfg configs/mld_t2m.yaml
python demo.py --cfg configs/motionlcm_t2m.yaml

The following command is for MotionLCM with motion ControlNet.

python demo.py --cfg configs/motionlcm_control_s.yaml

The following command is for MotionLCM with consistency latent tuning (CLT).

python demo.py --cfg configs/motionlcm_t2m_clt.yaml --optimize

After running the demo, the output folder will store the stick figure animation for each generated motion (e.g., assets/example.gif).

To record the necessary information about the generated motion, a pickle file with the following keys will be saved simultaneously (e.g., assets/example.pkl):

• joints (numpy.ndarray): The XYZ positions of the generated motion with the shape of (nframes, njoints, 3).
• text (str): The text prompt.
• length (int): The length of the generated motion.
• hint (numpy.ndarray): The trajectory for motion control (optional).

To create SMPL meshes for a specific pickle file, let's use assets/example.pkl as an example:

python fit.py --pkl assets/example.pkl

The SMPL meshes (numpy array) will be stored in assets/example_mesh.pkl with the shape (nframes, 6890, 3).

You can also fit all pickle files within a folder. The code will traverse all .pkl files in the directory and filter out files that have already been fitted.

python fit.py --dir assets/

Refer to TEMOS-Rendering motions for blender setup (only Installation section).

We support three rendering modes for SMPL mesh, namely sequence (default), video and frame.

YOUR_BLENDER_PATH/blender --background --python render.py -- --pkl assets/example_mesh.pkl --mode sequence --num 8

You will get a rendered image of num=8 keyframes as shown in assets/example_mesh.png. The darker the color, the later the time.

YOUR_BLENDER_PATH/blender --background --python render.py -- --pkl assets/example_mesh.pkl --mode video --fps 20

You will get a rendered video with fps=20 as shown in assets/example_mesh.mp4.

YOUR_BLENDER_PATH/blender --background --python render.py -- --pkl assets/example_mesh.pkl --mode frame --exact_frame 0.5

You will get a rendered image of the keyframe at exact_frame=0.5 (i.e., the middle frame) as shown in assets/example_mesh_0.5.png.

🚀 Train your own models

We provide the training guidance for motion reconstruction, text-to-motion and motion control tasks. The following steps will guide you through the training process.

The parameters required for model training and testing are recorded in the corresponding YAML file (e.g., configs/motionlcm_t2m.yaml). Below are some of the important parameters in the file:

• ${FOLDER}: The folder for the specific training task (i.e., experiments_recons, experiments_t2m and experiments_control).
• ${TEST_FOLDER}: The folder for the specific testing task (i.e., experiments_recons_test, experiments_t2m_test and experiments_control_test).
• ${NAME}: The name of the model (e.g., motionlcm_humanml). ${FOLDER}, ${NAME}, and the current timestamp constitute the training output folder (for example, experiments_t2m/motionlcm_humanml/2024-04-06T23-05-07). The same applies to ${TEST_FOLDER} for testing.
• ${TRAIN.PRETRAINED}: The path of the pre-trained model.
• ${TEST.CHECKPOINTS}: The path of the testing model.

Please update the parameters in configs/vae.yaml and configs/mld_t2m.yaml. Then, run the following commands:

python -m train_vae --cfg configs/vae.yaml
python -m train_mld --cfg configs/mld_t2m.yaml

3.1. Ready to train MotionLCM

Please first check the parameters in configs/motionlcm_t2m.yaml. Then, run the following command:

python -m train_motionlcm --cfg configs/motionlcm_t2m.yaml

3.2. Ready to train motion ControlNet

Please update the parameters in configs/motionlcm_control_s.yaml. Then, run the following command:

python -m train_motion_control --cfg configs/motionlcm_control_s.yaml

This command by default uses the Pelvis joint for motion control training. If you want to utilize all the joints defined in OmniControl (i.e., Pelvis, Left foot, Right foot, Head, Left wrist, and Right wrist), you need to modify the TRAIN_JOINTS in DATASET.HUMANML3D.CONTROL_ARGS in the configs/motionlcm_control_s.yaml.

TRAIN_JOINTS: [0] -> [0, 10, 11, 15, 20, 21]

This is also the reason we provide two checkpoints for testing in experiments_control/spatial/motionlcm_humanml.

CHECKPOINTS: 'experiments_control/spatial/motionlcm_humanml/motionlcm_humanml_s_pelvis.ckpt'  # Trained on Pelvis
CHECKPOINTS: 'experiments_control/spatial/motionlcm_humanml/motionlcm_humanml_s_all.ckpt'  #  Trained on All

During validation, the default testing joint is Pelvis, and the testing density is 100.

TEST_JOINTS: [0]  # choice -> [0], [10], [11], [15], [20], [21] (ONLY when trained on all)
TEST_DENSITY: 100  # choice -> [100, 25, 5, 2, 1]

TEST_DENSITY refers to the density level of control points selected from the ground truth (GT) trajectory. Specifically, 100 and 25 correspond to percentage, while 5, 2, and 1 correspond to number. The logic of the code is as follows:

# MotionLCM/mld/data/humanml/dataset.py (Text2MotionDataset)
length = joints.shape[0]
density = self.testing_density
if density in [1, 2, 5]:
    choose_seq_num = density
else:
    choose_seq_num = int(length * density / 100)

4.1. Motion Reconstruction:

python -m test --cfg configs/vae.yaml

4.2. Text-to-Motion:

python -m test --cfg configs/mld_t2m.yaml
python -m test --cfg configs/motionlcm_t2m.yaml

If you want to change the number of inference steps, change the num_inference_steps in the following configs:

configs/modules/scheduler_ddim.yaml  # MLD
configs/modules/scheduler_lcm.yaml   # MotionLCM

4.3. Motion Control:

The following command is for MotionLCM with motion ControlNet.

python -m test --cfg configs/motionlcm_control_s.yaml

The following command is for MotionLCM with consistency latent tuning (CLT).

python -m test --cfg configs/motionlcm_t2m_clt.yaml --optimize

For CLT, we default to using num_inference_steps=1 and batch_size=1. Do not modify these two parameters.

For our default motion control test (i.e., simply run the commands above), it is based on the pelvis joint with density=100. If you want to obtain complete results, please adjust the testing joint and testing density according to the motion control training tutorial.

📊 Results

We provide the quantitative and qualitative results in the paper.

TBD

TBD

🌹 Acknowledgement

We would like to thank the authors of the following repositories for their excellent work: MLD, latent-consistency-model, OmniControl, TEMOS, HumanML3D, UniMoCap, joints2smpl.

📜 Citation

If you find this work useful, please consider citing our paper:

@article{motionlcm,
  title={MotionLCM: Real-time Controllable Motion Generation via Latent Consistency Model},
  author={Dai, Wenxun and Chen, Ling-Hao and Wang, Jingbo and Liu, Jinpeng and Dai, Bo and Tang, Yansong},
  journal={arXiv preprint arXiv:2404.19759},
  year={2024},
}

📚 License

This code is distributed under an MotionLCM LICENSE, which not allowed for commercial usage. Note that our code depends on other libraries and datasets which each have their own respective licenses that must also be followed.

🌟 Star History

If you have any question, please contact at Wenxun Dai and cc to Ling-Hao Chen and Jingbo Wang.