Asketla/AIGC_Image_Steganography_Dataset

---
license: apache-2.0
task_categories:
- image-classification
- unconditional-image-generation
tags:
- steganography
- aigc
- text-to-image
- deepfake-detection
- image-forensics
size_categories:
- 10K<n<100K
---

# AIGC Image Steganography Dataset

## 📖 Dataset Description

This dataset is specifically designed for research in Artificial Intelligence Generated Content (AIGC) image steganography, steganalysis, and image forensics.

To construct a highly diverse and standardized dataset, we selected **10 prominent domestic and international text-to-image (T2I) large models** and batch-generated the images via their official APIs.

During the generation process, we carefully defined **10 typical image styles**. For each style, every model was tasked with generating 500 images featuring diverse content. This systematic approach resulted in a comprehensive dataset of **50,000 images** in total (10 models × 10 styles × 500 images).

To ensure experimental uniformity and rigorously evaluate the performance of steganography algorithms, the resolution of all generated images was strictly fixed at **1024 × 1024**. This prevents resolution variations from introducing unwanted interference into steganographic experiments.

## 🤖 Source Models

The 50,000 images in this dataset were generated by the following 10 advanced T2I models:
1. **Baidu (百度)**
2. **LiblibAI (liblib)**
3. **Seedream 3.0 (火山大模型)**
4. **Seedream 4.0 (火山大模型)**
5. **Star3 Batch**
6. **Tencent Hunyuan (混元)**
7. **Kuaishou Kling (可灵)**
8. **Alibaba Qwen (千问)**
9. **iFLYTEK Spark (讯飞)**
10. **Zhipu AI (智谱)**

## 📁 Dataset Structure

The dataset is organized into 10 main directories based on the source T2I models. Each directory contains the generated images for that specific model.

```text
AIGC_Image_Steganography_Dataset/
├── img_baidu/ # 5,000 images
├── img_liblib/ # 5,000 images
├── img_seedream3.0/ # 5,000 images
├── img_seedream4.0/ # 5,000 images
├── img_star3_batch/ # 5,000 images
├── img_混元/ # 5,000 images
├── img_可灵/ # 5,000 images
├── img_千问/ # 5,000 images
├── img_讯飞/ # 5,000 images
└── img_智谱/ # 5,000 images
```

## 🚀 How to Use (Python)

You can easily download and load this dataset using the Hugging Face `datasets` library or the `huggingface_hub` tool.

### Method 1: Using the `datasets` library (Recommended for ML pipelines)
First, install the library in your terminal:
```bash
pip install datasets
```

Then, load the dataset in your Python script:
```python
from datasets import load_dataset

# Load the dataset
dataset = load_dataset("Asketla/AIGC_Image_Steganography_Dataset")

# Print the dataset information
print(dataset)
```

### Method 2: Downloading raw files via `huggingface_hub`
If you prefer to download the raw image folders directly to your local machine:
First, install the library in your terminal:
```bash
pip install huggingface_hub
```

Then, run the following Python script:
```python
from huggingface_hub import snapshot_download

# Download the entire dataset repository to a local directory
local_dir = snapshot_download(
repo_id="Asketla/AIGC_Image_Steganography_Dataset",
repo_type="dataset",
local_dir="./aigc_steganography_data", # Specify your desired local path
max_workers=4 # Adjust based on your network speed
)

print(f"Dataset successfully downloaded to: {local_dir}")
```

许可证: Apache-2.0
任务类别:
- 图像分类
- 无条件图像生成
标签:
- 隐写术
- AIGC
- 文本到图像
- 深度伪造检测
- 图像取证
规模类别:
- 10K<n<100K

# AIGC图像隐写数据集

## 📖 数据集描述
本数据集专为人工智能生成内容（Artificial Intelligence Generated Content, AIGC）图像隐写、隐写分析及图像取证研究设计。
为构建高多样性且标准化的数据集，我们选取了**10款国内外领先的文本到图像（Text-to-Image, T2I）大模型**，通过其官方API批量生成图像。
在生成过程中，我们精心定义了**10种典型图像风格**，针对每种风格，每个模型需生成500张内容多样的图像。通过该系统化流程，最终得到总计**50000张图像**的完整数据集（10个模型 × 10种风格 × 500张图像）。
为确保实验一致性并严格评估隐写算法性能，所有生成图像的分辨率均严格固定为**1024 × 1024**，以避免分辨率差异对隐写实验引入不必要的干扰。

## 🤖 源模型
本数据集的50000张图像由以下10款先进的T2I模型生成：
1. **百度（Baidu）**
2. **LiblibAI（liblib）**
3. **Seedream 3.0（火山大模型）**
4. **Seedream 4.0（火山大模型）**
5. **Star3 Batch**
6. **腾讯混元（Tencent Hunyuan）**
7. **快手可灵（Kuaishou Kling）**
8. **阿里千问（Alibaba Qwen）**
9. **讯飞星火（iFLYTEK Spark）**
10. **智谱AI（Zhipu AI）**

## 📁 数据集结构
本数据集基于源T2I模型划分为10个主目录，每个目录包含对应模型生成的图像。
text
AIGC_Image_Steganography_Dataset/
├── img_baidu/ # 5000张图像
├── img_liblib/ # 5000张图像
├── img_seedream3.0/ # 5000张图像
├── img_seedream4.0/ # 5000张图像
├── img_star3_batch/ # 5000张图像
├── img_混元/ # 5000张图像
├── img_可灵/ # 5000张图像
├── img_千问/ # 5000张图像
├── img_讯飞/ # 5000张图像
└── img_智谱/ # 5000张图像

## 🚀 使用方法（Python）
您可通过Hugging Face的`datasets`库或`huggingface_hub`工具轻松下载并加载本数据集。

### 方法1：使用`datasets`库（推荐用于机器学习流水线）
首先在终端安装该库：
bash
pip install datasets

随后在Python脚本中加载数据集：
python
from datasets import load_dataset

# 加载数据集
dataset = load_dataset("Asketla/AIGC_Image_Steganography_Dataset")

# 打印数据集信息
print(dataset)

### 方法2：通过`huggingface_hub`下载原始文件
若您希望直接将原始图像文件夹下载至本地设备：
首先在终端安装该库：
bash
pip install huggingface_hub

随后运行以下Python脚本：
python
from huggingface_hub import snapshot_download

# 将整个数据集仓库下载至本地目录
local_dir = snapshot_download(
repo_id="Asketla/AIGC_Image_Steganography_Dataset",
repo_type="dataset",
local_dir="./aigc_steganography_data", # 指定您期望的本地路径
max_workers=4 # 可根据网络速度调整
)

print(f"数据集已成功下载至：{local_dir}")