ContactGen: Contact-Guided Interactive 3D Human Generation for Partners (AAAI 2024)

Install

Docker

docker pull dongjunku/humaninter:cu11

Project Structure

Please organize your project in the following structure:

ContactGen
├── body_models
|   ├── smplx
|   |   ├── SMPLX_NEUTRAL.npz
|   |   ├── SMPLX_NEUTRAL.pkl
├── datasets
|   ├── chi3d
|   |   ├── train
|   |   |   ├── s02
|   |   |   |   ├── camera_parameters
|   |   |   |   ├── gpp
|   |   |   |   ├── joints3d_25
|   |   |   |   ├── smplx
|   |   |   |   ├── videos
|   |   |   |   ├── interaction_contact_signature.json
|   |   |   ├── s03
|   |   |   ├── s04
|   ├── chi3d_whoisactor.pkl
|   ├── contact_regions.json
|   ├── r_sym_pair.pkl
├── ci3d.py
├── loss.py
├── model.py
├── optimizer.py
├── params.py
├── sample.py
├── test_diffusion.py
├── test_guidenet.py
├── train_diffusion.py
├── train_guidenet.py
├── utils.py
├── visualize.py

You can get CHI3D dataset here

You can get SMPL-X here

You can get contact_regions.json here

Pretrained Model

The pretrained models can be downloaded here. After downloading, place checkpoint_diffusion_ci3d and checkpoint_guidenet_ci3d in the ContactGen/.

Training

Diffusion module should be trained first

python train_diffusion.py

Next you can train guidenet

python train_guidenet.py

Sampling

python sample.py

It will generate samples in the output_diffusion_epoch1000_ci3d

Visualizing

python visualize.py output_diffusion_epoch1000_ci3d/???_human_pred.pkl

You can visualize overall diffusion steps using visualize.py

dongjunKu / ContactGen