IMIG-100K: A Large-Scale Synthetic Dataset for Multi-Instance Image Generation with Detailed Annotation

Ruihang Xu, Dewei Zhou, Fan Ma^†, Yi Yang
ReLER Lab, CCAI, Zhejiang University

📄 Dataset Overview

The IMIG-100K dataset is a large-scale synthetic dataset designed for multi-instance image generation tasks. It contains more than 100,000 high-quality image samples, each annotated with masks and layout information. The dataset is organized into several sub-datasets with different difficulty levels.

🧩 Sub-Datasets and Structure

The IMIG-100K dataset is divided into the following sub-datasets:

• IMIG-Basic: This sub-dataset contains images with fewer instances derived from the composite images. The reference images closely resemble their corresponding instances in the composite image, making it suitable for basic multi-instance generation tasks.

• IMIG-Complex: This sub-dataset features images with more instances and increased complexity. The reference images in this set transform more significantly from their corresponding instances in the composite images, providing a greater challenge for multi-instance generation models.

• IMIG-Composite: This sub-dataset includes composite images created by combining multiple reference images. It also provides aligned face images to facilitate face-related generation tasks. The instance will be more flexibly composed from the reference images in this set.

🛠️ Quick Start

Unpack the shards using the provided script scripts/unpack_shards.py (pip dependencies: webdataset, tqdm). It will take some time to unpack all the shards. The unpacked dataset should be organized as follows:

IMIG-100K/
├── imig-basic/
│   ├── data/
│   │   ├── 000000.jpg
│   │   ├── 000001.jpg
│   │   └── ...
│   ├── instance_data/
│   │   ├── 000000_0.jpg
│   │   ├── 000000_1.jpg
│   │   └── ...
│   ├── masked_instance_data/
│   │   ├── 000000_0_mask.png
│   │   ├── 000000_1_mask.png
│   │   └── ...
│   ├── kontext_data/ # similar to instance_data
│   ├── masked_kontext_data/ # similar to masked_instance_data
│   ├── prompts.json
│   ├── prompts_with_valid_bboxes.json
│   └── filtered_prompts.json
├── imig-complex/ # similar to imig-basic
├── imig-composite/
│   ├── reference_images/
│   │   ├── 000000_0.jpg
│   │   ├── 000000_1.jpg
│   │   └── ...
│   ├── reference_masks/
│   │   ├── 000000_0.jpg
│   │   ├── 000000_0_mask.png
│   │   └── ...
│   ├── instance_masks/ # similar to reference_masks
│   ├── composite_images/
│   │   ├── 000000.jpg
│   │   ├── 000001.jpg
│   │   └── ...
│   ├── aligned_faces/
│   │   ├── 000000_0_aligned_face.png
│   │   └── ...
│   ├── prompts.json
│   ├── prompts_with_bboxes.json
│   └── filtered_prompts.json
└── imig-multicomposite/ # similar to imig-composite

You can customize your dataloader based on the corresponding code in our main repository. If you need to re-organize or re-filter the dataset, you can refer to out dataset construction repository.

📭 Citation

If you find IMIG-100K helpful to your research, please consider citing our paper:

@article{xu2025contextgencontextuallayoutanchoring,
      title={ContextGen: Contextual Layout Anchoring for Identity-Consistent Multi-Instance Generation},
      author={Ruihang Xu and Dewei Zhou and Fan Ma and Yi Yang},
      year={2025},
      eprint={2510.11000},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2510.11000},
}