Dataset Card for POSTA-PosterArt

Dataset Description

• Homepage: https://haoyuchen.com/POSTA
• Repository: https://github.com/creative-graphic-design/huggingface-datasets/tree/main/datasets/POSTAPosterArt
• Paper (arXiv): https://arxiv.org/abs/2503.14908

Dataset Summary

POSTA-PosterArt is the dataset introduced with POSTA, a framework for customized artistic poster generation. It contains two subsets:

• PosterArt-Design: poster backgrounds with professional layout and typography annotations extracted from PSD files.
• PosterArt-Text: poster title regions with artistic text captions, masks, and single-region mask images for text stylization and segmentation.

The dataset supports research on controllable poster generation, layout planning, typography prediction, artistic text stylization, and text segmentation in graphic design.

Languages

The captions and text descriptions are in English.

Dataset Structure

Configurations

The dataset provides two configurations:

• text (default): PosterArt-Text examples.
• design: PosterArt-Design examples.

Data Instances

text example:

{
  "id": "2785",
  "image": "<image path>",
  "caption": "artistic text style description",
  "mask": "<segmentation mask path>",
  "mask_img_single": "<single text-region mask image path>"
}

design example:

{
  "id": "Part1/0002",
  "background_image": "<background image path>",
  "poster_image": "<final poster image path>",
  "psd_filename": "0002.psd",
  "annotation": "{... raw PSD JSON annotation ...}",
  "text_layers": [
    {
      "path": "Title",
      "name": "Title",
      "text_content": "POSTA",
      "font_name": "PlayfairDisplay",
      "font_size": 45.5,
      "alignment": "center",
      "rotation": 0.0
    }
  ]
}

Data Fields

text fields:

• id (string): Shared file stem for the grouped example.
• image (Image): Source poster or background image.
• caption (string): Artistic text style caption.
• mask (Image): Pixel-level text segmentation mask.
• mask_img_single (Image): Single text-region mask image.

design fields:

• id (string): Shared file stem for a poster design.
• background_image (Image): Poster background without layout text.
• poster_image (Image): Final poster image with complete layout.
• psd_filename (string): Original PSD filename when available.
• annotation (string): Raw JSON annotation extracted from the PSD structure.
• text_layers (sequence): Flattened text layers with layer path, text content, bbox, font, color, alignment, and rotation attributes.

Data Splits

Both configurations expose a single train split.

design is distributed through Google Drive. The POSTA project page lists PosterArt-Design as 152.3GB, while the currently downloadable Part1.zip used by this loader is 18.2GB compressed and about 24.8GB uncompressed.

Dataset Creation

Source Data

PosterArt was created for the POSTA framework. PosterArt-Design contains artistic poster backgrounds with layout and typography information crafted by professional designers. PosterArt-Text contains artistic title text regions, captions generated with vision-language models, and manually segmented title masks.

Annotations

PosterArt-Design annotations include PSD-derived layer structure, text positions, font information, color values, alignment, and rotation. PosterArt-Text annotations include captions and pixel-level text segmentation masks.

Considerations for Using the Data

Known Limitations

The dataset is large. Loading the design configuration downloads multi-GB Google Drive files. CI and lightweight tests should use helper-level tests or opt-in full download tests.

The license was not clearly identified from the public project page or Drive files during this implementation, so the dataset card records it as unknown.

Additional Information

Citation Information

@inproceedings{Chen_2025_CVPR,
  title = {POSTA: A Go-to Framework for Customized Artistic Poster Generation},
  author = {Chen, Haoyu and Xu, Xiaojie and Li, Wenbo and Ren, Jingjing and Ye, Tian and Liu, Songhua and Chen, Ying-Cong and Zhu, Lei and Wang, Xinchao},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  month = {June},
  year = {2025},
  pages = {28694--28704},
  url = {https://openaccess.thecvf.com/content/CVPR2025/html/Chen_POSTA_A_Go-to_Framework_for_Customized_Artistic_Poster_Generation_CVPR_2025_paper.html}
}

Contributions

Thanks to the POSTA authors for creating and releasing the PosterArt dataset. This Hugging Face dataset implementation was created by the creative-graphic-design organization to make POSTA-PosterArt easier to load for research workflows.