model_card.md

HOIDiNi Dataset: Human-Object Interaction through Diffusion Noise Optimization

Dataset Description

This repository contains the datasets and trained models for HOIDiNi, a text-driven diffusion framework for synthesizing realistic and plausible human-object interaction (HOI).

Key Features

• Text-driven HOI generation: Generate human-object interactions from natural language descriptions
• Two-phase optimization: Object-centric phase for contact planning, human-centric phase for full-body motion
• Diffusion Noise Optimization (DNO): Test-time optimization in the noise space of pretrained diffusion models
• High accuracy: Millimeter-level precision in contact generation

Repository Structure

├── models/                                    # Trained model weights
│   └── cphoi_05011024_c15p100_v0/            # CPHOI model experiment
│       ├── model000120000.pt                 # Main CPHOI model (748MB)
│       ├── eps.pt                            # Normalization parameters
│       ├── mean.pt                           # Mean parameters
│       └── args.json                         # Training arguments
├── datasets/                                  # Compressed datasets
│   ├── GRAB_RETARGETED_compressed/          # Compressed GRAB dataset (2.7GB)
│   └── MANO_SMPLX_vertex_ids.pkl            # Hand-object mapping data
└── README.MD                                  # Setup and usage instructions

Quick Start

1. Clone this dataset repository:

git clone https://huggingface.co/datasets/Roey/hoidini
cd hoidini

2. Set environment variables:

export GRAB_DATA_PATH=$(pwd)/datasets/GRAB_RETARGETED_compressed
export MANO_SMPLX_VERTEX_IDS_PATH=$(pwd)/datasets/MANO_SMPLX_vertex_ids.pkl

3. Use the trained models: The models are available in the models/cphoi_05011024_c15p100_v0/ directory. For code and usage instructions, please refer to the main HOIDiNi repository.

Model Details

• Architecture: Contact-Pairs Human-Object Interaction (CPHOI) diffusion model
• Training Data: GRAB dataset with contact pair annotations
• Input: Text descriptions + object geometry
• Output: Full-body human motion + object trajectories + contact pairs
• Compression: Dataset compressed from 67GB to 2.7GB (24.8x reduction)

Citation

@article{ron2025hoidini,
  title={HOIDiNi: Human-Object Interaction through Diffusion Noise Optimization},
  author={Ron, Roey and Tevet, Guy and Sawdayee, Haim and Bermano, Amit H},
  journal={arXiv preprint arXiv:2506.15625},
  year={2025}
}

License

This project is licensed under the MIT License.

Links

• Project Page: hoidini.github.io
• Paper: arXiv:2506.15625
• Video: YouTube