This is the official code for the paper "PINO: Person-Interaction Noise Optimization for Long-Duration and Customizable Motion Generation of Arbitrary-Sized Groups".

This code was tested on Red Hat Enterprise Linux 9.4 and requires:

• Python 3.8

• uv (for environment management)

• NVIDIA H100 (or a similar CUDA-capable GPU)

If you don't have uv installed, you can install it with:

curl -LsSf https://astral.sh/uv/install.sh | sh

uv sync
. .venv/bin/activate

Please download the InterHuman dataset from here. If you use this dataset, please make sure to cite the original paper.

Run the shell script:

.prepare/download_pretrain_model.sh
.prepare/download_evaluation_model.sh

python tools/infer_opt.py --prompt "Two people danced at the party."

You need motion data that includes at least two generated people.

The first person will be used as a hub, and additional people will be generated one by one in interaction with them.

python tools/infer_opt.py --prompt "Two people danced at the party." --motion_path results/Two_people_danced_at_the_party_2.pt

For detailed motion control, refer to here.

Our code is mainly based on InterGen. In addition, part of the code is adapted from progmogen.

@inproceedings{ota2025pino,
    title     = {PINO: Person-Interaction Noise Optimization for Long-Duration and Customizable Motion Generation of Arbitrary-Sized Groups},
    author    = {Ota, Sakuya and Yu, Qing and Fujiwara, Kent and Ikehata, Satoshi and Sato, Ikuro},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    year      = {2025},
}