Stable Diffusion generated images - AIS-4SD dataset

AIS-4SD

AIS-4SD (AI Summit - 4 Stable Diffusion models) is a collection of 4.000 images, generated using a set of Stability AI text-to-image diffusion models

Context

This dataset was developed during the development of a collaborative project between PEReN and VIGINUM for the AI Summit held in Paris in February 2025. This open-source project aims at assessing generated images detectors performances and their robustness to different models and transformations. The code is free and open source, and contributions to connect additional detectors are also welcome.

Official repository: https://code.peren.gouv.fr/open-source/ai-action-summit/generated-image-detection.

Dataset summary

This dataset can be used to assess detection models performances, and in particular their robustness to successive updates of the generation model. 

Dataset description

1.000 generated images with four different versions of stability AI text-to-image diffusion model.

For each models, we generated:

500 portraits (👨) using SFHQ-T2I "random" prompts for faces (see Github repo, and dataset on Kaggle),

500 more general content images (🖼️) using captions of Google's Conceptual Captions dataset.

Model
Number of images

stabilityai/stable-diffusion-xl-base-1.0
500 👨 + 500 🖼️

stabilityai/stable-diffusion-2-1
500 👨 + 500 🖼️

stabilityai/stable-diffusion-3-medium-diffusers
500 👨 + 500 🖼️

stabilityai/stable-diffusion-3.5-large
500 👨 + 500 🖼️

Reproducibility

The scripts used to generated these images can be found on our open-source repository (see this specific file). After setting-up our project, you can run:

$ poetry run python scripts/generate_images.py

With minor updates to these scripts you can enrich this dataset with your specific needs.

Dataset structure

One zip file with the following structure, each directory containing the associated 500 images:

AIS-4SD/
├── generation_metadata.csv
├── StableDiffusion-2.1-faces-20250203-1448
├── StableDiffusion-2.1-other-20250203-1548
├── StableDiffusion-3.5-faces-20250203-1012
├── StableDiffusion-3.5-other-20250203-1603
├── StableDiffusion-3-faces-20250203-1545
├── StableDiffusion-3-other-20250203-1433
├── StableDiffusion-XL-faces-20250203-0924
└── StableDiffusion-XL-other-20250203-1727

The metadata for generated images (see generation_metadata.csv) are:

model: model used for generation,

prompt: prompt used for generation (ie Conceptual Captions caption or sfhqt2i prompt, with some minor prompt engineering),

guidance_scale: guidance scale of diffusion process,

num_inference_steps: number of inference steps of diffusion process,

generated_img_relative_path: relative path to image in zip structure.

Project status

Project is under ongoing development. A preliminary blog post can be found here: https://www.peren.gouv.fr/en/perenlab/2025-02-11_ai_summit/.

AIS-4SD（AI峰会——4款Stable Diffusion模型）是一个包含4000张图像的数据集，所有图像均通过Stability AI旗下多款文本到图像扩散模型生成。

### 背景
本数据集是为2025年2月在巴黎举办的AI峰会，由PEReN与VIGINUM合作开发的开源项目产物。本开源项目旨在评估生成式图像检测器的性能，以及其对不同生成模型与图像变换的鲁棒性。项目代码完全免费开源，欢迎社区贡献代码以接入更多检测器。官方仓库地址：https://code.peren.gouv.fr/open-source/ai-action-summit/generated-image-detection.

### 数据集摘要
本数据集可用于评估检测模型的性能，尤其适用于测试模型对生成模型迭代更新的鲁棒性。

### 数据集详情
本数据集包含4款Stable Diffusion文本到图像扩散模型各自生成的1000张图像。针对每款模型，我们生成了：
- 500张肖像图像（👨），采用SFHQ-T2I的“随机”人脸提示词（详见GitHub仓库及Kaggle上的对应数据集）；
- 500张通用内容图像（🖼️），采用谷歌Conceptual Captions数据集的标注文本作为提示词。

| 模型名称 | 图像数量 |
| --- | --- |
| stabilityai/stable-diffusion-xl-base-1.0 | 500张肖像 + 500张通用内容 |
| stabilityai/stable-diffusion-2-1 | 500张肖像 + 500张通用内容 |
| stabilityai/stable-diffusion-3-medium-diffusers | 500张肖像 + 500张通用内容 |
| stabilityai/stable-diffusion-3.5-large | 500张肖像 + 500张通用内容 |

### 可复现性
用于生成该数据集的脚本可在本开源仓库中获取（详见指定文件）。完成项目环境配置后，可通过以下命令运行脚本：
bash
poetry run python scripts/generate_images.py

对脚本进行少量修改后，即可根据自身需求扩充该数据集。

### 数据集结构
数据集打包为单个ZIP文件，内部目录结构如下，每个目录包含对应的500张图像：

AIS-4SD/
├── generation_metadata.csv
├── StableDiffusion-2.1-faces-20250203-1448
├── StableDiffusion-2.1-other-20250203-1548
├── StableDiffusion-3.5-faces-20250203-1012
├── StableDiffusion-3.5-other-20250203-1603
├── StableDiffusion-3-faces-20250203-1545
├── StableDiffusion-3-other-20250203-1433
├── StableDiffusion-XL-faces-20250203-0924
└── StableDiffusion-XL-other-20250203-1727

生成图像的元数据（存于generation_metadata.csv）包含以下字段：
- model：生成所用的模型名称
- prompt：生成所用的提示词（即Conceptual Captions标注文本或SFHQ-T2I提示词，经过少量提示工程优化）
- guidance_scale：扩散过程的引导系数
- num_inference_steps：扩散过程的推理步数
- generated_img_relative_path：图像在ZIP包内的相对路径

### 项目状态
本项目仍在持续开发中。相关初步博客文章可参阅：https://www.peren.gouv.fr/en/perenlab/2025-02-11_ai_summit/