Dataset Card for CGL-Dataset v2

Dataset Description

• Homepage: https://github.com/liuan0803/RADM
• Repository: https://github.com/creative-graphic-design/huggingface-datasets/tree/main/datasets/CGLDatasetV2
• Hugging Face Dataset: https://huggingface.co/datasets/creative-graphic-design/CGL-Dataset-v2
• Paper (arXiv): https://arxiv.org/abs/2306.09086
• Paper (CIKM 2023): https://dl.acm.org/doi/10.1145/3583780.3615028

Dataset Summary

CGL-Dataset v2 is an advertising-poster layout dataset released with Relation-Aware Diffusion Model for Controllable Poster Layout Generation. The paper argues that poster layouts should account for both visual-textual relationships and geometry relationships between elements. This version extends CGL-Dataset with richer element annotations, text annotations, and text features for controllable poster layout generation.

Supported Tasks and Leaderboards

The dataset supports poster layout generation, layout understanding, and relation-aware controllable generation. No public leaderboard is bundled with this Hugging Face dataset.

Languages

Poster text and annotations are primarily Chinese (zh).

Dataset Structure

Data Fields

The default config contains poster images and COCO-style instance annotations with optional text_annotations and text_features.

The ralf-style config exposes the same data in a layout-generation format with original posters, inpainted posters, saliency maps, and annotations.

Data Splits

Dataset Creation

The original release was created for relation-aware diffusion research on controllable poster layout generation. Posters are annotated with visual elements such as logos, text, underlays, embellishments, and highlighted text, and the dataset supports generation under user constraints.

Considerations for Using the Data

The dataset is focused on advertising posters and Chinese e-commerce-style visual content. Models trained on it may inherit the visual conventions and category distribution of the source data.

Additional Information

Licensing Information

The dataset license is not specified in the local loader metadata. Users should verify the upstream terms before redistribution or commercial use.

Citation Information

@inproceedings{li2023relation,
  title={Relation-Aware Diffusion Model for Controllable Poster Layout Generation},
  author={Li, Fengheng and Liu, An and Feng, Wei and Zhu, Honghe and Li, Yaoyu and Zhang, Zheng and Lv, Jingjing and Zhu, Xin and Shen, Junjie and Lin, Zhangang},
  booktitle={Proceedings of the 32nd ACM international conference on information & knowledge management},
  pages={1249--1258},
  year={2023}
}

Contributions

Thanks to liuan0803 for creating the original dataset.