MDM: Human Motion Diffusion Model

The official PyTorch implementation of the paper "Human Motion Diffusion Model".

Please visit our webpage for more details.

MDM is now 40X faster 🤩🤩🤩 (~0.4 sec/sample)

How come?!?

(1) We released the 50 diffusion steps model (instead of 1000 steps) which runs 20X faster with comparable results.

(2) Calling CLIP just once and caching the result runs 2X faster for all models. Please pull.

MDM results on HumanML3D to cite in your paper (The original model used in the MDM paper)

Performance improvement is due to an evaluation bug fix. BLUE marks fixed entries compared to the paper.

• You can use this .tex file.
• The fixed KIT results are available here.

Bibtex

🔴🔴🔴NOTE: MDM and MotionDiffuse are NOT the same paper! For some reason, Google Scholar merged the two papers. The right way to cite MDM is:

@inproceedings{
tevet2023human,
title={Human Motion Diffusion Model},
author={Guy Tevet and Sigal Raab and Brian Gordon and Yoni Shafir and Daniel Cohen-or and Amit Haim Bermano},
booktitle={The Eleventh International Conference on Learning Representations },
year={2023},
url={https://openreview.net/forum?id=SJ1kSyO2jwu}
}

News

📢 15/Apr/24 - Released a 50 diffusion steps model (instead of 1000 steps) which runs 20X faster 🤩🤩🤩 with comparable results.

📢 12/Apr/24 - MDM inference is now 2X faster 🤩🤩🤩 This was made possible by calling CLIP just once and caching the result, and is backward compatible with older models.

📢 25/Jan/24 - Fixed bug in evalutation code (#182) - Please use the fixed results when citing MDM.

📢 1/Jun/23 - Fixed generation issue (#104) - Please pull to improve generation results.

📢 23/Nov/22 - Fixed evaluation issue (#42) - Please pull and run bash prepare/download_t2m_evaluators.sh from the top of the repo to adapt.

📢 4/Nov/22 - Added sampling, training and evaluation of unconstrained tasks. Note slight env changes adapting to the new code. If you already have an installed environment, run bash prepare/download_unconstrained_assets.sh; conda install -y -c anaconda scikit-learn to adapt.

📢 3/Nov/22 - Added in-between and upper-body editing.

📢 31/Oct/22 - Added sampling, training and evaluation of action-to-motion tasks.

📢 9/Oct/22 - Added training and evaluation scripts. Note slight env changes adapting to the new code. If you already have an installed environment, run bash prepare/download_glove.sh; pip install clearml to adapt.

📢 6/Oct/22 - First release - sampling and rendering using pre-trained models.

Checkout MDM Follow-ups (partial list)

🐉 SinMDM - Learns single motion motifs - even for non-humanoid characters.

👯 PriorMDM - Uses MDM as a generative prior, enabling new generation tasks with few examples or even no data at all.

💃 MAS - Generating intricate 3D motions (including non-humanoid) using 2D diffusion models trained on in-the-wild videos.

🐒 MoMo - Monkey See, Monkey Do: Harnessing Self-attention in Motion Diffusion for Zero-shot Motion Transfer

Getting started

This code was tested on Ubuntu 18.04.5 LTS and requires:

• Python 3.7
• conda3 or miniconda3
• CUDA capable GPU (one is enough)

1. Setup environment

Install ffmpeg (if not already installed):

sudo apt update
sudo apt install ffmpeg

For windows use this instead.

Setup conda env:

conda env create -f environment.yml
conda activate mdm
python -m spacy download en_core_web_sm
pip install git+https://github.com/openai/CLIP.git

Download dependencies:

Text to Motion

bash prepare/download_smpl_files.sh
bash prepare/download_glove.sh
bash prepare/download_t2m_evaluators.sh

Action to Motion

bash prepare/download_smpl_files.sh
bash prepare/download_recognition_models.sh

Unconstrained

bash prepare/download_smpl_files.sh
bash prepare/download_recognition_models.sh
bash prepare/download_recognition_unconstrained_models.sh

2. Get data

Text to Motion

There are two paths to get the data:

(a) Go the easy way if you just want to generate text-to-motion (excluding editing which does require motion capture data)

(b) Get full data to train and evaluate the model.

a. The easy way (text only)

HumanML3D - Clone HumanML3D, then copy the data dir to our repository:

cd ..
git clone https://github.com/EricGuo5513/HumanML3D.git
unzip ./HumanML3D/HumanML3D/texts.zip -d ./HumanML3D/HumanML3D/
cp -r HumanML3D/HumanML3D motion-diffusion-model/dataset/HumanML3D
cd motion-diffusion-model

b. Full data (text + motion capture)

HumanML3D - Follow the instructions in HumanML3D, then copy the result dataset to our repository:

cp -r ../HumanML3D/HumanML3D ./dataset/HumanML3D

KIT - Download from HumanML3D (no processing needed this time) and the place result in ./dataset/KIT-ML

Action to Motion

UESTC, HumanAct12

bash prepare/download_a2m_datasets.sh

Unconstrained

HumanAct12

bash prepare/download_unconstrained_datasets.sh

3. Download the pretrained models

Download the model(s) you wish to use, then unzip and place them in ./save/.

Text to Motion

You need only the first one.

HumanML3D

humanml-encoder-512-50steps - Runs 20X faster with comparable performance!

humanml-encoder-512 (best model used in the paper)

humanml-decoder-512

humanml-decoder-with-emb-512

KIT

kit-encoder-512

Action to Motion

UESTC

uestc

uestc_no_fc

HumanAct12

humanact12

humanact12_no_fc

Unconstrained

HumanAct12

humanact12_unconstrained

Motion Synthesis

Text to Motion

Generate from test set prompts

python -m sample.generate --model_path ./save/humanml_trans_enc_512/model000200000.pt --num_samples 10 --num_repetitions 3

Generate from your text file

python -m sample.generate --model_path ./save/humanml_trans_enc_512/model000200000.pt --input_text ./assets/example_text_prompts.txt

Generate a single prompt

python -m sample.generate --model_path ./save/humanml_trans_enc_512/model000200000.pt --text_prompt "the person walked forward and is picking up his toolbox."

Action to Motion

Generate from test set actions

python -m sample.generate --model_path ./save/humanact12/model000350000.pt --num_samples 10 --num_repetitions 3

Generate from your actions file

python -m sample.generate --model_path ./save/humanact12/model000350000.pt --action_file ./assets/example_action_names_humanact12.txt

Generate a single action

python -m sample.generate --model_path ./save/humanact12/model000350000.pt --action_name "drink"

Unconstrained

python -m sample.generate --model_path ./save/unconstrained/model000450000.pt --num_samples 10 --num_repetitions 3

By abuse of notation, (num_samples * num_repetitions) samples are created, and are visually organized in a display of num_samples rows and num_repetitions columns.

You may also define:

• --device id.
• --seed to sample different prompts.
• --motion_length (text-to-motion only) in seconds (maximum is 9.8[sec]).

Running those will get you:

• results.npy file with text prompts and xyz positions of the generated animation
• sample##_rep##.mp4 - a stick figure animation for each generated motion.

It will look something like this:

You can stop here, or render the SMPL mesh using the following script.

Render SMPL mesh

To create SMPL mesh per frame run:

python -m visualize.render_mesh --input_path /path/to/mp4/stick/figure/file

This script outputs:

• sample##_rep##_smpl_params.npy - SMPL parameters (thetas, root translations, vertices and faces)
• sample##_rep##_obj - Mesh per frame in .obj format.

Notes:

• The .obj can be integrated into Blender/Maya/3DS-MAX and rendered using them.
• This script is running SMPLify and needs GPU as well (can be specified with the --device flag).
• Important - Do not change the original .mp4 path before running the script.

Notes for 3d makers:

• You have two ways to animate the sequence:
  1. Use the SMPL add-on and the theta parameters saved to sample##_rep##_smpl_params.npy (we always use beta=0 and the gender-neutral model).
  2. A more straightforward way is using the mesh data itself. All meshes have the same topology (SMPL), so you just need to keyframe vertex locations. Since the OBJs are not preserving vertices order, we also save this data to the sample##_rep##_smpl_params.npy file for your convenience.

Motion Editing

• This feature is available for text-to-motion datasets (HumanML3D and KIT).
• In order to use it, you need to acquire the full data (not just the texts).
• We support the two modes presented in the paper: in_between and upper_body.

Unconditioned editing

python -m sample.edit --model_path ./save/humanml_trans_enc_512/model000200000.pt --edit_mode in_between

You may also define:

• --num_samples (default is 10) / --num_repetitions (default is 3).
• --device id.
• --seed to sample different prompts.
• --edit_mode upper_body For upper body editing (lower body is fixed).

The output will look like this (blue frames are from the input motion; orange were generated by the model):

• As in Motion Synthesis, you may follow the Render SMPL mesh section to obtain meshes for your edited motions.

Text conditioned editing

Just add the text conditioning using --text_condition. For example:

python -m sample.edit --model_path ./save/humanml_trans_enc_512/model000200000.pt --edit_mode upper_body --text_condition "A person throws a ball"

The output will look like this (blue joints are from the input motion; orange were generated by the model):

Train your own MDM

Text to Motion

HumanML3D

python -m train.train_mdm --save_dir save/my_humanml_trans_enc_512 --dataset humanml

KIT

python -m train.train_mdm --save_dir save/my_kit_trans_enc_512 --dataset kit

Action to Motion

python -m train.train_mdm --save_dir save/my_name --dataset {humanact12,uestc} --cond_mask_prob 0 --lambda_rcxyz 1 --lambda_vel 1 --lambda_fc 1

Unconstrained

python -m train.train_mdm --save_dir save/my_name --dataset humanact12 --cond_mask_prob 0 --lambda_rcxyz 1 --lambda_vel 1 --lambda_fc 1  --unconstrained

• Use --diffusion_steps 50 to train the faster model with less diffusion steps.
• Use --device to define GPU id.
• Use --arch to choose one of the architectures reported in the paper {trans_enc, trans_dec, gru} (trans_enc is default).
• Add --train_platform_type {ClearmlPlatform, TensorboardPlatform} to track results with either ClearML or Tensorboard.
• Add --eval_during_training to run a short (90 minutes) evaluation for each saved checkpoint. This will slow down training but will give you better monitoring.

Evaluate

Text to Motion

• Takes about 20 hours (on a single GPU)
• The output of this script for the pre-trained models (as was reported in the paper) is provided in the checkpoints zip file.

HumanML3D

python -m eval.eval_humanml --model_path ./save/humanml_trans_enc_512/model000475000.pt

KIT

python -m eval.eval_humanml --model_path ./save/kit_trans_enc_512/model000400000.pt

Action to Motion

• Takes about 7 hours for UESTC and 2 hours for HumanAct12 (on a single GPU)
• The output of this script for the pre-trained models (as was reported in the paper) is provided in the checkpoints zip file.

python -m eval.eval_humanact12_uestc --model <path-to-model-ckpt> --eval_mode full

where path-to-model-ckpt can be a path to any of the pretrained action-to-motion models listed above, or to a checkpoint trained by the user.

Unconstrained

• Takes about 3 hours (on a single GPU)

python -m eval.eval_humanact12_uestc --model ./save/unconstrained/model000450000.pt --eval_mode full

Precision and recall are not computed to save computing time. If you wish to compute them, edit the file eval/a2m/gru_eval.py and change the string fast=True to fast=False.

Acknowledgments

This code is standing on the shoulders of giants. We want to thank the following contributors that our code is based on:

guided-diffusion, MotionCLIP, text-to-motion, actor, joints2smpl, MoDi.

License

This code is distributed under an MIT LICENSE.

Note that our code depends on other libraries, including CLIP, SMPL, SMPL-X, PyTorch3D, and uses datasets that each have their own respective licenses that must also be followed.