README

This repository contains official code for our research paper "Synthesizing Long-Term Human Motions with Diffusion Models via Coherent Sampling" (https://dl.acm.org/doi/10.1145/3581783.3611887).

Environment setup

To set up the environment, we provide an requirements.txt file that you can use with pip:

pip install -r requirements.txt

Data preparation

Due to licensing restrictions, we are unable to provide pre-processed data directly. However, you can refer to TEACH for specific data processing methods.

Eventually, you should have a folder data with such a structure:

data
|-- babel
|   `-- babel_v2.1
|       `...
|   `-- babel-smplh-30fps-male 
|       `...
|
|-- smpl_models
|   `-- smplh
|       `--SMPLH_MALE.pkl

Besides, you should download the folder deps in TEACH to this project.

Pre-trained weights

We provide the pretrained models here: pretrained models. The 'pretrained.zip' file contains the pretrained model and training configurations used to report metrics in our paper, while 'MotionCLIP.zip' contains the model used for evaluation. You can put the pretrained model and training configuration file under ./save/pcmdm and put the MotipnClip.ckpt under ./motionclip_save.

Running the code

You can use the following three commands to obtain the results for the last three rows of the experimental results table in our paper:

# No special sampling
python eval.py --model_path ./save/pcmdm/model000600000.pt --guidance_param 2 --inpainting_frames 0

# Past inpainting sampling
python eval.py --model_path ./save/pcmdm/model000600000.pt --guidance_param 2 --inpainting_frames 2

# Compositional transition sampling
python eval.py --model_path ./save/pcmdm/model000600000.pt --guidance_param 2 --composition True --inter_frames 2

Besides, if you want to train a model from scratch, you can use this comman:

python train_diffusion.py --save_dir ./save/pcmdm --dataset babel --hist_frames 5 

Acknowledgments

Our code is based on TEACH and MDM. Thanks for their greate work!