ductai199x/open-set-synth-img-attribution

---
architectures:
- stylegan
- stylegan2
- stylegan3
- stargan
- tam_trans
- stable_diffusion
- projected_gan
- progan
- cycleGAN
license:
- cc-by-nc-sa-4.0
pretty_name: OSSIA
size_categories:
- 1M < images < 10M
task_categories:
- image-classification
task_ids:
- multi-class-image-classification
category_size:
videos: 4000
frames: 120000
---

# Open-set Synthetic Image Attribution Dataset

## Dataset Description

- **Paper:** [Open Set Synthetic Image Source Attribution](https://proceedings.bmvc2023.org/659/)
- **Total amount of data used:** approx. 500GB

This dataset is used to evaluate the ability of different algorithm to attribute the source generator or generator's architecture of synthetic images. The dataset consists of synthetic images generated by seven different image generators: ProGAN, Projected-GAN, StyleGAN, StyleGAN2, StyleGAN3, Taming Transformer, and Stable Diffusion. We generated all images ourselves using the generators and datasets provided by the authors of the respective methods. We specifically avoid resizing the images to preserve the original synthetic image forensic traces.

## Usage Example

The Open-set Synthetic Image Attribution (OSSIA) Dataset can be downloaded and used as follows:

```py
import datasets

ossia = datasets.load_dataset("ductai199x/open-set-synth-img-attribution", "arch") # or "gen"

# see structure
print(ossia)

# access the first example
print(ossia["train"][0])
```

_NOTE: The dataset is already shuffled during packaging. There is no need to shuffle._

## Dataset Structure

### Data Instances

Some frame examples from this dataset:

### Data Fields
The data fields are the same among all splits.

- **image** (image): The synthetic image.
- **label** (int): The label of the image, which corresponds to the generator that created the image. To see the mapping between the label and the generator, please refer to the classes.py file contained in this repository.

### Data Splits
There are a total of 1,147,422 images in the dataset, divided into 803,195 images for training, 114,742 images for validation, and 229,485 images for testing. More details can be found in the paper.

### Licensing Information

All datasets are licensed under the [Creative Commons Attribution, Non-Commercial, Share-alike license (CC BY-NC-SA)](https://creativecommons.org/licenses/by-nc-sa/4.0/).

### Citation Information

```
@inproceedings{Fang_2023_BMVC,
author = {Shengbang Fang and Tai D Nguyen and Matthew c Stamm},
title = {Open Set Synthetic Image Source Attribution},
booktitle = {34th British Machine Vision Conference 2023, {BMVC} 2023, Aberdeen, UK, November 20-24, 2023},
publisher = {BMVA},
year = {2023},
url = {https://papers.bmvc2023.org/0659.pdf}
}
```

### Contact

For any questions, please contact Tai Nguyen at [@ductai199x](https://github.com/ductai199x) or by [email](mailto:[e-mail]).

原始信息汇总

Open-set Synthetic Image Attribution Dataset 概述

数据集描述

• 用途: 评估不同算法对合成图像来源生成器或生成器架构的识别能力。
• 生成器: 包含 ProGAN, Projected-GAN, StyleGAN, StyleGAN2, StyleGAN3, Taming Transformer, Stable Diffusion 七种生成器生成的合成图像。
• 数据量: 约500GB。
• 图像处理: 未对图像进行缩放，以保留原始的合成图像取证痕迹。

数据集结构

• 数据实例: 包含1,147,422张图像，分为803,195张训练图像，114,742张验证图像，229,485张测试图像。
• 数据字段:
  • image (图像): 合成图像。
  • label (整数): 图像标签，对应于生成该图像的生成器。标签与生成器之间的映射请参考本仓库中的classes.py文件。

许可证信息

• 许可证: 所有数据集均根据Creative Commons Attribution, Non-Commercial, Share-alike license (CC BY-NC-SA)授权。

引用信息

@inproceedings{Fang_2023_BMVC, author = {Shengbang Fang and Tai D Nguyen and Matthew c Stamm}, title = {Open Set Synthetic Image Source Attribution}, booktitle = {34th British Machine Vision Conference 2023, {BMVC} 2023, Aberdeen, UK, November 20-24, 2023}, publisher = {BMVA}, year = {2023}, url = {https://papers.bmvc2023.org/0659.pdf} }