metadata
license: other
tags:
- human-motion-generation
- human-human-interaction
- text-to-motion
- diffusion
library_name: pytorch
SocialStructureHHI — Executor
Pretrained executor weights for SocialStructureHHI: Social Structure Matters in 3D Human–Human Interaction Generation.
The solo-to-social motion executor is the full social-planning model — a HunyuanMotion-Lite backbone with variable-length partner conditioning plus a pinned self-history prefix — that generates two-person interactions phase-by-phase in a Ping-Pong manner.
Files
| file | description |
|---|---|
ckpts/best.pt |
solo-to-social motion executor, ~1.86 GB |
Usage
# In the SocialStructureHHI repo:
# https://github.com/EngineeringAI-LAB/SocialStructureHHI
from huggingface_hub import hf_hub_download
ckpt = hf_hub_download("EngineeringAI-LAB/SocialStructure", "ckpts/best.pt")
python inference/run_inference.py --ckpt $ckpt --demo_id 5 \
--llm_model qwen3.5:35b --llm_base_url http://localhost:11435/v1 \
--out_dir outputs/handshake --render
The HunyuanMotion backbone, Qwen3-8B and CLIP-L are additional external dependencies — see
the code repository for setup. Dataset normalization stats
(data/motion_norm_stats.npz) ship with the code repo and are required to load the model.
Links
- 📄 Paper: https://arxiv.org/abs/2606.24255
- 💻 Code: https://github.com/EngineeringAI-LAB/SocialStructureHHI
- 🌐 Project page: https://engineeringai-lab.github.io/SocialStructureHHI/
- 🤗 Dataset: https://huggingface.co/datasets/EngineeringAI-LAB/SocialStructure
Citation
@misc{wang2026socialstructure,
title = {Social Structure Matters in 3D Human--Human Interaction Generation},
author = {Zhongju Wang and Beier Wang and Yatao Bian and Pichao Wang and Zhi Wang
and Daoyi Dong and Hongdong Li and Huadong Mo and Zhenhong Sun},
year = {2026},
eprint = {2606.24255},
archivePrefix = {arXiv},
primaryClass = {cs.CV}
}