jiuntian/IEBench

IEBench - a Human-Object Interaction Editing Benchmark

IEBench is released in conjunction with our paper, InteractEdit: Zero-Shot Editing of Human-Object Interactions in Images.

IEBench is designed to evaluate the capability of image editing models to modify Human-Object Interactions (HOI). While mainstream image editing methods have shown remarkable results in localized object modification or style transfer, they often struggle with the complex spatial and semantic dependencies inherent in HOIs.

The core challenge in HOI editing lies in the delicate balance of identity preservation and relational transformation. Existing methods frequently fail to maintain the consistent identity of the human subject and the target object while simultaneously modifying their interaction (e.g., changing "sitting on a chair" to "standing on a chair"). Achieving high fidelity in both the "identity" and "interaction" domains remains an open research problem that IEBench aims to address.

Key Features

• Diverse Interaction Scenarios: The benchmark consists of 28 source images covering 25 distinct actions and 13 object categories, creating a rich variety of physical and semantic relationships.

• Extensive Evaluation Pairs: We provide 100 unique ⟨source image, target interaction⟩ pairs, where each object is intentionally paired with multiple target interactions to test the model's flexibility and boundary cases.

• Granular Asset Decomposition: Each instance is meticulously structured with decomposed assets (subject, object, and background crops) to support part-based or mask-guided editing workflows.

• Broad Action Vocabulary: Supports a wide range of human behaviors, including complex movements like dribble, groom, and stand on, as well as static interactions like smell and lie on.

Dataset Structure

The dataset is organized by source image instance IDs (e.g., from the HICO-DET dataset). Each folder contains the necessary masks and cropped assets required for precise HOI editing.

Dataset Structure The dataset is organized by source image instance IDs (e.g., from the HICO-DET dataset). Each folder contains the necessary masks and cropped assets required for precise HOI editing.

IEBench/
├── HICO_train2015_00028163/      # Source image instance ID
│   ├── gt/                       # Ground truth assets
│   │   ├── asset0.png            # Subject crop
│   │   ├── asset1.png            # Object crop
│   │   └── bg.png                # Background crop
│   ├── jump/                     # Target action folder
│   ├── ride/                     # Target action folder
│   ├── info.json                 # Instance-specific info
│   ├── mask0.jpg                 # Subject mask
│   ├── mask1.jpg                 # Object mask
│   └── mask2.jpg                 # Background mask
├── HICO_train2015_00007695/
│   └── ...                       # Other instances
└── infos.json                    # Global metadata for all instances

Related Links

• InteractDiffusion (CVPR 2024): Interaction Control in Text-to-Image Diffusion Models
• InteractEdit: Zero-Shot Editing of Human-Object Interactions in Images (IEBench dataset)
• OneHOI (CVPR 2026): Unifying Human-Object Interaction Generation and Editing

If you find our dataset useful, feel free to ⭐ star this repo!

If you use our work in your research, please cite:

@misc{hoe2025interactedit,
    title={InteractEdit: Zero-Shot Editing of Human-Object Interactions in Images}, 
    author={Jiun Tian Hoe and Weipeng Hu and Wei Zhou and Chao Xie and Ziwei Wang and Chee Seng Chan and Xudong Jiang and Yap-Peng Tan},
    year={2025},
    eprint={2503.09130},
    archivePrefix={arXiv},
    primaryClass={cs.GR},
    url={https://arxiv.org/abs/2503.09130}, 
}
@inproceedings{hoe2026onehoi,
  title={OneHOI: Unifying Human-Object Interaction Generation and Editing},
  author={Hoe, Jiun Tian and Hu, Weipeng and Jiang, Xudong and Tan, Yap-Peng and Chan, Chee Seng},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2026}
}