creative-graphic-design/GenPoster100K

---
language:
- en
license: cc-by-nc-4.0
pretty_name: GenPoster100K
tags:
- graphic-design
- layout-generation
- multimodal
- poster-design
annotations_creators:
- machine-generated
language_creators:
- found
size_categories:
- 100K<n<1M
source_datasets:
- original
task_categories:
- text-to-image
- image-to-text
task_ids:
- other:content-aware-layout-generation
---

# Dataset Card for GenPoster100K

[](https://github.com/creative-graphic-design/huggingface-datasets/actions/workflows/ci.yaml)
[](https://github.com/creative-graphic-design/huggingface-datasets/actions/workflows/push_to_hub.yaml)

## Table of Contents

- [Dataset Description](#dataset-description)
- [Dataset Summary](#dataset-summary)
- [Supported Tasks and Leaderboards](#supported-tasks-and-leaderboards)
- [Languages](#languages)
- [Dataset Structure](#dataset-structure)
- [Data Instances](#data-instances)
- [Data Fields](#data-fields)
- [Data Splits](#data-splits)
- [Dataset Creation](#dataset-creation)
- [Curation Rationale](#curation-rationale)
- [Source Data](#source-data)
- [Initial Data Collection and Normalization](#initial-data-collection-and-normalization)
- [Who are the source language producers?](#who-are-the-source-language-producers)
- [Annotations](#annotations)
- [Annotation process](#annotation-process)
- [Who are the annotators?](#who-are-the-annotators)
- [Personal and Sensitive Information](#personal-and-sensitive-information)
- [Considerations for Using the Data](#considerations-for-using-the-data)
- [Social Impact of Dataset](#social-impact-of-dataset)
- [Discussion of Biases](#discussion-of-biases)
- [Other Known Limitations](#other-known-limitations)
- [Additional Information](#additional-information)
- [Dataset Curators](#dataset-curators)
- [Licensing Information](#licensing-information)
- [Citation Information](#citation-information)
- [Contributions](#contributions)

## Dataset Description

- **Homepage:** https://huggingface.co/datasets/BruceW91/GenPoster-100K
- **Repository:** https://github.com/creative-graphic-design/huggingface-datasets/tree/main/datasets/GenPoster100K
- **Paper (Preprint):** https://arxiv.org/abs/2510.15749
- **Paper (Conference):** https://openaccess.thecvf.com/content/ICCV2025/html/Wang_SEGA_A_Stepwise_Evolution_Paradigm_for_Content-Aware_Layout_Generation_with_ICCV_2025_paper.html
- **Leaderboard:** No official public leaderboard URL is documented for GenPoster-100K.
- **Point of Contact:** Original dataset/project authors (paper) and maintainers of the source HF dataset (`BruceW91/GenPoster-100K`).

### Dataset Summary

GenPoster-100K is a large-scale dataset for content-aware graphic layout generation introduced in the SEGA paper.
The paper describes it as a high-quality poster dataset with layer-parseable source materials and rich metadata.

This repository provides a Hugging Face `datasets` loader implementation that reads the source release (`BruceW91/GenPoster-100K`) and exposes normalized examples with:

- poster background image (`background_image`)
- PSD reference path (`psd_path`)
- region boxes (`regions`)
- layer-level annotations (`layers`) including text, bbox, typography, color, and per-layer rendered image

Implementation note: this loader uses `0503_raw_offline.pkl` and `part_*.tar.gz`, yielding 102,703 rows in `train`.

### Supported Tasks and Leaderboards

- `other:content-aware-layout-generation`: Generate or refine poster element layouts conditioned on background imagery and textual element metadata.
- `text-to-image`: Can be used in poster design workflows where textual content and structured attributes guide generated visual composition.
- `image-to-text`: Can support structured extraction/understanding tasks over design layers and poster metadata.

No official leaderboard URL specific to GenPoster-100K is currently provided in the public source materials.

### Languages

- Frontmatter language is set to English (`en`) because the released examples predominantly contain English poster text.
- The dataset may include additional languages in real-world templates, but a full language distribution is not documented in the source materials.

## Dataset Structure

### Data Instances

Each example includes image assets and structured layer metadata.

```json
{
"id": 0,
"background_image": "<image>",
"background_image_relpath": "big_poster/poster_metadata/3841272.psd_0_11775f75bf_bg.png",
"psd_path": "big_poster/meta_psd/3841272.psd",
"regions": [[1656, 481, 2545, 855]],
"layers": [
{
"layer_name": "&#wText&#wTitle",
"text": "Super price!",
"bbox": [1754, 573, 2423, 689],
"angle": 0,
"psd_size": [3508, 2480],
"stroke_width": 0.0,
"font": "Aftaserif",
"font_size": 113.29,
"tracking": 0.0,
"justification": 1,
"fill_color": [0.0, 0.0, 0.0, 1.0],
"layer_image": "<image>",
"layer_image_relpath": "big_poster/poster_metadata/3841272.psd_0_11775f75bf_3.png",
"label": "Calls to Action"
}
]
}
```

### Data Fields

- `id` (`int32`): Example identifier assigned by loader order.
- `background_image` (`Image`): Rendered background image for the poster.
- `background_image_relpath` (`string`): Relative path of the background image in source assets.
- `psd_path` (`string`): Relative PSD path recorded in annotations.
- `regions` (`Sequence[Sequence[int32]]`): Region boxes as `[x1, y1, x2, y2]`.
- `layers` (`Sequence[struct]`): Layer-level annotations.
- `layer_name` (`string`)
- `text` (`string`)
- `bbox` (`Sequence[int32]`, length=4)
- `angle` (`int32`)
- `psd_size` (`Sequence[int32]`, length=2)
- `stroke_width` (`float32`)
- `font` (`string`)
- `font_size` (`float32`)
- `tracking` (`float32`)
- `justification` (`int32`)
- `fill_color` (`Sequence[float32]`, length=4)
- `layer_image` (`Image`)
- `layer_image_relpath` (`string`)
- `label` (`string`)

### Data Splits

This implementation exposes a single `train` split from the upstream release.

| Split | Rows |
| --- | ---: |
| train | 102,703 |

Notes:

- The paper reports 105,456 posters in the broader GenPoster-100K corpus.
- The HF source release used by this loader provides `0503_raw_offline.pkl` with 102,703 records.
- The original source data is `train`-only, but our Hugging Face Hub publication splits it into `train/validation/test = 8:1:1` and pushes that split dataset.

## Dataset Creation

### Curation Rationale

According to the SEGA paper, GenPoster-100K was introduced to improve data quality and scale for content-aware layout generation.
The paper highlights limitations in earlier datasets (for example, artifacts from inpainted backgrounds and less structured metadata) and positions GenPoster-100K as a higher-fidelity, large-scale alternative with rich component-level information.

### Source Data

The source release includes poster metadata archives (`part_*.tar.gz`) and annotation pickle files.
A disclaimer in the source dataset indicates copyright belongs to the original owner (Freepik) and commercial use may require additional permission.

#### Initial Data Collection and Normalization

From public materials:

- The paper describes data built from layer-parseable source materials with hierarchical metadata.
- The source HF dataset release distributes image assets split across 79 archive parts (`part_0.tar.gz` ... `part_78.tar.gz`).

In this loader implementation:

- Annotation source is `0503_raw_offline.pkl`.
- URL query strings are normalized/removed and image paths are resolved against extracted archive contents.
- BBox/size/color fields are normalized to fixed lengths and numeric dtypes.

#### Who are the source language producers?

The textual content appears to originate from poster templates designed by content creators in the source design corpus.
No demographic metadata for these creators is provided in the public release.

### Annotations

The release provides machine-readable layer metadata per poster example, including geometry and typography attributes.
The dataset card metadata marks annotation creation as machine-generated.

#### Annotation process

Public paper/source materials indicate hierarchical metadata extracted from PSD-parseable design sources.
The implementation-level annotation fields include:

- text and layer names
- bounding boxes and angle
- typography attributes (font, size, tracking, justification, stroke)
- RGBA fill color
- region boxes
- references to rendered background/layer images

Detailed internal annotation tooling and QA workflow are not fully specified in the public documents.

#### Who are the annotators?

Annotations are primarily machine-generated from source design assets.
Named individual annotators are not documented.

### Personal and Sensitive Information

The dataset is composed of poster design assets and textual elements.
It is not released as a personal-data dataset, but real-world template text may include names, brands, or contact-like strings depending on source content.
Users should perform downstream filtering/redaction if their use case requires stricter privacy constraints.

## Considerations for Using the Data

### Social Impact of Dataset

Potential positive impact:

- Enables research on automated graphic design and multimodal layout understanding.
- Supports reproducible benchmarking for content-aware layout generation.

Potential risks:

- May be used to generate misleading or low-quality promotional content at scale.
- Could be misused for style imitation or copyright-sensitive commercial outputs.

### Discussion of Biases

Likely biases include:

- Domain/style bias toward stock-template aesthetics and marketing layouts.
- Language/domain imbalance (predominantly English and commercial poster styles).
- Visual-cultural bias inherited from source platforms and curation choices.

No public bias audit report specific to this release is currently documented.

### Other Known Limitations

- The upstream source is train-only; this loader follows that source format, while our Hub publication additionally provides an `8:1:1` split.
- Paper-level total (105,456) and loader-level total (102,703) differ due to release artifacts used by this loader.
- `psd_path` is recorded as metadata path, but only a limited subset of raw PSD files is present in the public source release.
- License constraints limit direct commercial usage without additional permission from the copyright owner.

## Additional Information

### Dataset Curators

- Original dataset/paper authors: Wang et al. (SEGA, ICCV 2025).
- Source HF release contributor: [@BruceW91](https://huggingface.co/BruceW91).
- This repository contains a community-maintained Hugging Face dataset loader implementation.

### Licensing Information

- Dataset card license: `cc-by-nc-4.0`
- Source disclaimer indicates use is intended for academic purposes, and commercial use may require prior permission from the copyright owner (Freepik): https://www.freepik.com/

Always verify license compatibility with your intended use before redistribution or deployment.

### Citation Information

If your implementation is based on this dataset, please cite the original paper and the Hugging Face dataset implementation:

```bibtex
@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}
}
```

```bibtex
@misc{genposter100kdataset,
title = {GenPoster100K dataset},
author = {{Creative Graphic Design Lab} and Kitada, Shunsuke},
howpublished = {Hugging Face dataset},
year = {2025},
note = {URL: https://huggingface.co/datasets/creative-graphic-design/GenPoster100K},
}
```

### Contributions

Thanks to the original GenPoster-100K authors and [@BruceW91](https://huggingface.co/BruceW91) for releasing the source dataset.
This Hugging Face dataset implementation was created for the `creative-graphic-design/huggingface-datasets` monorepo.

---
语言：
- en
许可证：CC BY-NC 4.0（知识共享署名-非商业性使用4.0国际许可协议）
简称：GenPoster100K
标签：
- 平面设计（graphic design）
- 布局生成（layout generation）
- 多模态（multimodal）
- 海报设计（poster design）
注释创建者：
- 机器生成
语言创建者：
- 采集自现有资源
规模类别：
- 10万<样本数<100万
源数据集：
- 原始数据集
任务类别：
- 文本到图像（text-to-image）
- 图像到文本（image-to-text）
任务子类别：
- 其他：内容感知布局生成（content-aware layout generation）
---

# GenPoster100K 数据集卡片

[](https://github.com/creative-graphic-design/huggingface-datasets/actions/workflows/ci.yaml)
[](https://github.com/creative-graphic-design/huggingface-datasets/actions/workflows/push_to_hub.yaml)

## 目录

- [数据集概述](#数据集概述)
- [数据集摘要](#数据集摘要)
- [支持任务与排行榜](#支持任务与排行榜)
- [语言](#语言)
- [数据集结构](#数据集结构)
- [数据样本](#数据样本)
- [数据字段](#数据字段)
- [数据拆分](#数据拆分)
- [数据集构建](#数据集构建)
- [筛选依据](#筛选依据)
- [源数据](#源数据)
- [初始数据收集与标准化](#初始数据收集与标准化)
- [源语言生产者是谁？](#源语言生产者是谁？)
- [注释](#注释)
- [注释流程](#注释流程)
- [注释者是谁？](#注释者是谁？)
- [个人与敏感信息](#个人与敏感信息)
- [数据集使用注意事项](#数据集使用注意事项)
- [数据集的社会影响](#数据集的社会影响)
- [偏差讨论](#偏差讨论)
- [其他已知局限性](#其他已知局限性)
- [附加信息](#附加信息)
- [数据集维护者](#数据集维护者)
- [许可证信息](#许可证信息)
- [引用信息](#引用信息)
- [贡献声明](#贡献声明)

## 数据集概述

- **主页**：https://huggingface.co/datasets/BruceW91/GenPoster-100K
- **代码仓库**：https://github.com/creative-graphic-design/huggingface-datasets/tree/main/datasets/GenPoster100K
- **预印本论文**：https://arxiv.org/abs/2510.15749
- **会议论文**：https://openaccess.thecvf.com/content/ICCV2025/html/Wang_SEGA_A_Stepwise_Evolution_Paradigm_for_Content-Aware_Layout_Generation_with_ICCV_2025_paper.html
- **排行榜**：当前公开源材料中未提供针对 GenPoster-100K 的官方排行榜链接
- **联系方式**：原始数据集/项目作者（论文作者）与 Hugging Face 源数据集维护者（`BruceW91/GenPoster-100K`）

### 数据集摘要

GenPoster-100K 是 SEGA 论文中提出的面向内容感知平面布局生成的大规模数据集。该论文将其描述为一个高质量海报数据集，具备可解析图层的源素材与丰富的元数据。

本仓库提供了一个 Hugging Face `datasets` 库的加载器实现，可读取源发布版本（`BruceW91/GenPoster-100K`）并输出经过标准化处理的样本，包含以下内容：
- 海报背景图像（`background_image`）
- PSD（Adobe Photoshop 文档格式）参考路径（`psd_path`）
- 区域边框（`regions`）
- 图层级注释（`layers`），包含文本、边界框、排版、色彩与单图层渲染图像

实现说明：该加载器使用 `0503_raw_offline.pkl` 与 `part_*.tar.gz` 文件，在训练拆分中生成102,703条样本。

### 支持任务与排行榜

- `其他：内容感知布局生成`：基于背景图像与文本元素元数据生成或优化海报元素布局
- `文本到图像（text-to-image）`：可应用于海报设计流程，通过文本内容与结构化属性指导生成视觉构图
- `图像到文本（image-to-text）`：可支持针对设计图层与海报元数据的结构化提取与理解任务

当前公开源材料中未提供针对 GenPoster-100K 的官方排行榜链接。

### 语言

- 元数据语言设置为英语（`en`），因为发布的样本中绝大多数海报文本为英文
- 数据集的真实模板中可能包含其他语言，但源材料中未记录完整的语言分布情况

## 数据集结构

### 数据样本

每个样本包含图像资源与结构化图层元数据。

json
{
"id": 0,
"background_image": "<图像>",
"background_image_relpath": "big_poster/poster_metadata/3841272.psd_0_11775f75bf_bg.png",
"psd_path": "big_poster/meta_psd/3841272.psd",
"regions": [[1656, 481, 2545, 855]],
"layers": [
{
"layer_name": "&#wText&#wTitle",
"text": "Super price!",
"bbox": [1754, 573, 2423, 689],
"angle": 0,
"psd_size": [3508, 2480],
"stroke_width": 0.0,
"font": "Aftaserif",
"font_size": 113.29,
"tracking": 0.0,
"justification": 1,
"fill_color": [0.0, 0.0, 0.0, 1.0],
"layer_image": "<图像>",
"layer_image_relpath": "big_poster/poster_metadata/3841272.psd_0_11775f75bf_3.png",
"label": "Calls to Action"
}
]
}

### 数据字段

- `id`（`int32`类型）：根据加载顺序分配的样本标识符
- `background_image`（`Image`类型）：海报的渲染背景图像
- `background_image_relpath`（字符串类型）：背景图像在源资源中的相对路径
- `psd_path`（字符串类型）：注释中记录的PSD文件相对路径
- `regions`（`int32`类型序列的序列）：区域边框，格式为`[x1, y1, x2, y2]`
- `layers`（结构体序列）：图层级注释
- `layer_name`（字符串类型）：图层名称
- `text`（字符串类型）：文本内容
- `bbox`（长度为4的`int32`类型序列）：边界框
- `angle`（`int32`类型）：旋转角度
- `psd_size`（长度为2的`int32`类型序列）：PSD画布尺寸
- `stroke_width`（`float32`类型）：笔触宽度
- `font`（字符串类型）：字体名称
- `font_size`（`float32`类型）：字体大小
- `tracking`（`float32`类型）：字间距
- `justification`（`int32`类型）：文本对齐方式
- `fill_color`（长度为4的`float32`类型序列）：RGBA填充色彩
- `layer_image`（`Image`类型）：图层渲染图像
- `layer_image_relpath`（字符串类型）：图层图像相对路径
- `label`（字符串类型）：图层标签

### 数据拆分

本实现仅从上游发布版本中提供一个`train`（训练）拆分。

| 拆分名称 | 样本数量 |
| --- | ---: |
| 训练集 | 102,703 |

备注：
- 论文中报道的完整 GenPoster-100K 语料库包含105,456张海报
- 本加载器使用的 Hugging Face 源发布版本中，`0503_raw_offline.pkl` 文件包含102,703条记录
- 原始源数据仅包含训练集，但我们在 Hugging Face Hub 上发布的版本将其划分为`训练集/验证集/测试集 = 8:1:1`并推送该拆分后的数据集

## 数据集构建

### 筛选依据

根据 SEGA 论文，GenPoster-100K 的提出旨在提升内容感知布局生成任务的数据质量与规模。该论文指出了早期数据集的局限性（例如，背景修复产生的伪影以及结构化程度不足的元数据），并将 GenPoster-100K 定位为具备丰富组件级信息的高保真大规模替代数据集。

### 源数据

源发布版本包含海报元数据归档文件（`part_*.tar.gz`）与注释 pickle 文件。源数据集中的免责声明表明，版权归原所有者（Freepik）所有，商业使用可能需要额外授权。

#### 初始数据收集与标准化

基于公开资料：
- 论文中提到，该数据集基于具备层级化元数据且可解析图层的源素材构建
- Hugging Face 源数据集发布版本将图像资源拆分至79个归档文件中（`part_0.tar.gz` 至 `part_78.tar.gz`）

在本加载器实现中：
- 注释源文件为 `0503_raw_offline.pkl`
- URL 查询字符串会被标准化或移除，图像路径会根据解压后的归档内容进行解析
- 边界框、尺寸与色彩字段会被标准化为固定长度与数值数据类型

#### 源语言生产者是谁？

文本内容似乎源自源设计语料库中内容创作者设计的海报模板。公开发布版本中未提供这些创作者的人口统计元数据。

### 注释

该发布版本为每个海报样本提供了机器可读的图层元数据，包含几何与排版属性。数据集卡片的元数据标注注释创建方式为机器生成。

#### 注释流程

公开论文与源资料表明，层级化元数据从可解析PSD的设计源文件中提取。本实现中的注释字段包括：
- 文本与图层名称
- 边界框与旋转角度
- 排版属性（字体、字号、字间距、对齐方式、笔触）
- RGBA填充色彩
- 区域边框
- 渲染后的背景/图层图像的引用

公开文档中未完整说明内部注释工具与质量控制流程的细节。

#### 注释者是谁？

注释主要由源设计素材通过机器生成。文档中未提及具体的个体注释者。

### 个人与敏感信息

本数据集由海报设计素材与文本元素组成。本数据集并非以个人数据数据集的形式发布，但真实模板文本可能包含姓名、品牌或类似联系方式的字符串，具体取决于源素材内容。若用户的使用场景需要更严格的隐私保护，应在下游任务中对数据进行过滤或脱敏处理。

## 数据集使用注意事项

### 数据集的社会影响

潜在积极影响：
- 支持自动化平面设计与多模态布局理解的研究
- 为内容感知布局生成任务提供可复现的基准测试

潜在风险：
- 可能被用于大规模生成误导性或低质量的宣传内容
- 可能被滥用于风格模仿或涉及版权问题的商业产出

### 偏差讨论

可能存在的偏差包括：
- 领域与风格偏差：偏向于模板素材美学与营销类布局
- 语言与领域不平衡：以英文与商业海报风格为主
- 视觉文化偏差：继承自源平台与筛选规则的视觉文化偏向

当前未公开针对该发布版本的偏差审计报告。

### 其他已知局限性

- 上游源数据仅包含训练集；本加载器遵循该源格式，而我们在 Hugging Face Hub 上发布的版本额外提供了`8:1:1`的拆分
- 论文报道的总样本数（105,456）与本加载器的总样本数（102,703）存在差异，原因是本加载器使用的发布文件存在差异
- `psd_path` 字段记录为元数据路径，但公开源发布版本中仅包含有限数量的原始PSD文件
- 许可证限制：未经版权所有者额外授权，无法直接将数据集用于商业用途

## 附加信息

### 数据集维护者

- 原始数据集与论文作者：Wang 等人（SEGA，ICCV 2025）
- Hugging Face 源发布版本贡献者：[@BruceW91](https://huggingface.co/BruceW91)
- 本仓库包含由社区维护的 Hugging Face 数据集加载器实现

### 许可证信息

- 数据集卡片许可证：CC BY-NC 4.0（知识共享署名-非商业性使用4.0国际许可协议）
- 源数据集中的免责声明表明，本数据集仅用于学术用途，商业使用需事先获得版权所有者（Freepik）的授权：https://www.freepik.com/

在重新分发或部署前，请务必确认许可证与您的使用用途兼容。

### 引用信息

若您的实现基于本数据集，请引用原始论文与 Hugging Face 数据集实现：

bibtex
@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}
}

bibtex
@misc{genposter100kdataset,
title = {GenPoster100K dataset},
author = {{Creative Graphic Design Lab} and Kitada, Shunsuke},
howpublished = {Hugging Face dataset},
year = {2025},
note = {URL: https://huggingface.co/datasets/creative-graphic-design/GenPoster100K},
}

### 贡献声明

感谢原始 GenPoster-100K 数据集作者与 [@BruceW91](https://huggingface.co/BruceW91) 发布源数据集。本 Hugging Face 数据集实现是为 `creative-graphic-design/huggingface-datasets` 单仓库开发的。