README.md

---
language:
  - zh
license: cc-by-nc-4.0
pretty_name: CreativePSD
tags:
  - graphic-design
  - psd
  - creative-ai
  - tool-use
  - multimodal
  - modelscope
annotations_creators:
  - machine-generated
language_creators:
  - found
size_categories:
  - 1K<n<10K
source_datasets:
  - original
task_categories:
  - image-to-text
  - text-to-image
---

# Dataset Card for CreativePSD

## Dataset Description

- **Homepage:** https://modelscope.cn/datasets/song322/CreativePSD
- **Repository:** https://github.com/creative-graphic-design/huggingface-datasets/tree/main/datasets/CreativePSD
- **arXiv:** https://arxiv.org/abs/2603.25738
- **Paper (CVPR 2026):** https://openaccess.thecvf.com/content/CVPR2026/html/Shuai_PSDesigner_Automated_Graphic_Design_with_a_Human-Like_Creative_Workflow_CVPR_2026_paper.html

### Dataset Summary

CreativePSD is the PSD-derived graphic design dataset released with PSDesigner. Each example is a poster archive containing PSD tree text, structured layer metadata, tool-call trajectories, source image resources, and stepwise rendered images.

This loader keeps the contents of each `poster_*.zip` archive: all metadata text/JSON files, all `raw_resource` images, all `rendering_imgs` images, and a manifest of every member in the archive.

### Languages

The dataset contains Chinese text in poster designs and metadata fields.

## Dataset Structure

### Data Instances

Each row corresponds to one `poster_*.zip` archive:

```json
{
  "id": "poster_000175",
  "archive_filename": "poster_000175.zip",
  "psd_info": {
    "filename": "poster_000175.psd",
    "width": 800,
    "height": 800,
    "resolution": 72,
    "color_mode": "RGB"
  },
  "total_layers": 18,
  "origin_psd_tree": "...",
  "tool_trajectory_json": "[...]",
  "raw_resources": [{"filename": "raw_resource/1_asset.png", "image": "<image>"}],
  "rendering_images": [
    {
      "filename": "rendering_imgs/0_total.jpg",
      "image_path": "poster_000175.zip/rendering_imgs/0_total.jpg",
      "image_bytes": "<bytes>"
    }
  ],
  "final_rendering": "<image>"
}
```

### Data Fields

- `id`: Poster archive stem.
- `archive_filename`, `archive_path`, `archive_size_bytes`: Source archive metadata.
- `psd_info`: PSD filename, canvas size, resolution, color mode, and fill color.
- `total_layers`: Number of PSD layers from `metadata/layer_info.json`.
- `origin_psd_tree`, `deleted_psd_tree`, `grouped_psd_tree`: PSD tree text files.
- `group_child_ids_json`, `layer_info_json`, `rendering_id_json`, `tool_trajectory_json`, `render_second_values_json`: Original JSON files serialized as strings.
- `metadata_files`: All files under `metadata/` as text.
- `non_image_files`: Every non-image file in the archive as text and bytes.
- `raw_resources`: All source images under `raw_resource/`.
- `rendering_images`: All images under `rendering_imgs/`, including `0_total.jpg` and intermediate layer renders, stored as image bytes plus archive-relative paths.
- `final_rendering`: `rendering_imgs/0_total.jpg` when present.
- `all_files`: Manifest of every file member in the zip.

### Data Splits

The loader exposes a single `train` split. The PSDesigner paper reports 10,454 CreativePSD samples in Table 1, while the public ModelScope release currently exposes 7,978 root-level `poster_*.zip` archives through the file metadata API. This loader targets the public ModelScope archive release, not the unreleased remainder implied by the paper count.

When loading from an official ModelScope checkout, the loader validates the archive count and skips only the known unavailable source archives listed below. With the current source availability issue, non-streaming `load_dataset(...)` completes with 7,968 usable poster archives.

### Known Source Availability Issue

As of 2026-06-28, the paper-reported CreativePSD count, the public ModelScope release, and the loadable dataset rows differ:

- Paper Table 1: 10,454 CreativePSD samples.
- Public ModelScope file metadata: 7,978 root-level `poster_*.zip` archives.
- Current `load_dataset(...)` result: 7,968 rows after skipping the 10 known unavailable archives below.

Repeated downloads from the available ModelScope SDK, resolve/raw, Git LFS, and OSS access paths returned 10 of the public archives as 0-byte files in the local checkout. The loader skips these known unavailable archives when they are 0-byte or absent in an official ModelScope checkout. Any other invalid or missing poster archive still fails validation.

The currently unresolved 0-byte archives are:

- `poster_003266.zip`
- `poster_003281.zip`
- `poster_003287.zip`
- `poster_003288.zip`
- `poster_003294.zip`
- `poster_003296.zip`
- `poster_003299.zip`
- `poster_003302.zip`
- `poster_003305.zip`
- `poster_003306.zip`

## Usage

Download the dataset with ModelScope:

```python
from modelscope.msdatasets import MsDataset

MsDataset.load("song322/CreativePSD")
```

Then load from the directory containing `poster_*.zip` files:

```python
import datasets as ds

dataset = ds.load_dataset(
    "datasets/CreativePSD/CreativePSD.py",
    data_dir="/root/ghq/www.modelscope.cn/datasets/song322/CreativePSD",
    trust_remote_code=True,
)
```

## Considerations for Using the Data

The ModelScope dataset card states that the dataset is for non-commercial research use only. Users should review the original dataset terms before redistribution or model training.

## Additional Information

### Licensing Information

The source dataset card lists `CC-BY-NC-4.0`.

### Citation Information

```bibtex
@InProceedings{Shuai_2026_CVPR,
  title={PSDesigner: Automated Graphic Design with a Human-Like Creative Workflow},
  author={Shuai, Xincheng and Tang, Song and Huang, Yutong and Ding, Henghui and Tao, Dacheng},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  month={June},
  year={2026},
  pages={10165-10175},
}
```

### Contributions

Thanks to the PSDesigner and CreativePSD authors for creating and releasing this dataset.