DecHOI: Decoupled Generative Modeling for Human-Object Interaction Synthesis

🚀 Official PyTorch implementation of the CVPR 2026 paper Decoupled Generative Modeling for Human-Object Interaction Synthesis (CVPR 2026)

[ArXiv]

Hwanhee Jung, Seunggwan Lee, Jeongyoon Yoon, SeungHyeon Kim, Giljoo Nam, Qixing Huang, Sangpil Kim

Environment

The verified environment is Python 3.10, CUDA 11.7, and PyTorch 2.0.1.

Create and activate the Anaconda environment:

conda create -n dechoi_env python=3.10
conda activate dechoi_env

Install PyTorch:

pip install --index-url https://download.pytorch.org/whl/cu117 torch==2.0.1 torchvision==0.15.2 torchaudio==2.0.2

Install PyTorch3D:

conda install -c fvcore -c iopath -c conda-forge fvcore iopath
conda install -c bottler nvidiacub
pip install --no-index --no-cache-dir pytorch3d==0.7.4 -f https://dl.fbaipublicfiles.com/pytorch3d/packaging/wheels/py310_cu117_pyt201/download.html

Install human_body_prior:

git clone https://github.com/nghorbani/human_body_prior.git
pip install tqdm dotmap PyYAML omegaconf loguru
cd human_body_prior/
python setup.py develop

Install BPS and Chamfer distance:

pip install git+https://github.com/otaheri/chamfer_distance.git
pip install git+https://github.com/otaheri/bps_torch.git

Install remaining dependencies:

pip install -r requirements.txt

Prerequisites

Download SMPL-X and place the model under data/smpl_all_models/.

Download the processed data from the previous work CHOIS and place it under data/processed_data/. You can override this path by setting the DATA_ROOT_FOLDER environment variable.

To generate visualizations, install Blender and set the following environment variables as needed:

export DECHOI_BLENDER_PATH=/path/to/blender
export DECHOI_BLENDER_UTILS_ROOT=manip/vis
export DECHOI_BLENDER_SCENE_FOLDER=/path/to/blender_files

Pretrained Weights

You can download the pretrained DecHOI checkpoint from the link below:

• weight link

Common Commands

Evaluate the model:

sh scripts/test_dechoi_single_window.sh

You can override the defaults with environment variables:

DATA_ROOT_FOLDER=/path/to/data \
PRETRAINED_MODEL=/path/to/checkpoint.pt \
SAVE_RES_FOLDER=/path/to/output \
sh scripts/test_dechoi_single_window.sh

To toggle visualization output, set save_obj_only in the gen_vis_res_generic() method of trainer_dechoi.py.

Train:

sh scripts/train_dechoi.sh

Override training defaults:

DATA_ROOT_FOLDER=/path/to/data \
EXP_NAME=my_experiment \
sh scripts/train_dechoi.sh

Citation

If you find this project useful, please consider citing:

@InProceedings{Jung_2026_CVPR,
    author    = {Jung, Hwanhee and Lee, Seunggwan and Yoon, Jeongyoon and Kim, SeungHyeon and Nam, Giljoo and Huang, Qixing and Kim, Sangpil},
    title     = {Decoupled Generative Modeling for Human-Object Interaction Synthesis},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2026},
    pages     = {2253-2263}
}

Related Repos

We thank the fantastic works T2M, OMOMO, CHOIS, and HOIFHLI for their pioneer code release, which provide codebase for this benchmark.