GenPoster-100K dataset

Contains 100,000 PSD-format posters with parseable layers.

Disclaimer

We declare that we do not hold the copyright to the data, and we are only permitted by the copyright owner to use it for academic purposes. If you wish to use it for commercial purposes, prior permission from the copyright owner is required. We shall not be liable for any disputes arising therefrom.

Usage Guide

Due to the requirement limitations of the Hugging Face platform for datasets uploading, we have split the contents of the poster_metadata folder (which contains all images) into 79 compressed files (from part_0.tar.gz to part_78.tar.gz). Therefore, you need to first untar all these compressed files into a self-created folder named poster_metadata.

📁 structure of dataset

/GenPoster-100K/  
├── poster_metadata/      # all images (need to untar)
│   ├── part_0.tar.gz       
│   ├── part_1.tar.gz       
│   ├── ...                 
│   └── part_78.tar.gz    
├── meta_psd              # some raw psd file (not used))
├── 0503_0.pkl
├── 0503_2.pkl
├── 0503_5.pkl
├── 0503_raw.pkl
├── 0503_raw_offline.pkl  # layer annotation file for images
└── README.md             # readme document

If you want to use this dataset, please cite our work:

@inproceedings{wang2025sega,
  title={SEGA: A Stepwise Evolution Paradigm for Content-Aware Layout Generation with Design Prior},
  author={Wang, Haoran and Zhao, Bo and Wang, Jinghui and Wang, Hanzhang and Yang, Huan and Ji, Wei and Liu, Hao and Xiao, Xinyan},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={19321--19330},
  year={2025}
}