README.md

---
license: mit
task_categories:
  - image-classification
language:
  - en
tags:
  - ai-generated-image-detection
  - deepfake-detection
  - synthetic-image-detection
  - computer-vision
  - binary-classification
  - gan-detection
  - diffusion-model-detection
size_categories:
  - 100K<n<1M
pretty_name: TIGAS Dataset
dataset_info:
  features:
    - name: image_path
      dtype: string
    - name: label
      dtype: int64
  splits:
    - name: train
      num_examples: 128776
    - name: test
      num_examples: 14126
---

# TIGAS Dataset

<div align="center">

[![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](https://opensource.org/licenses/MIT)
[![Dataset Size](https://img.shields.io/badge/Images-142.9K-blue.svg)]()
[![Task](https://img.shields.io/badge/Task-Binary%20Classification-green.svg)]()

**A comprehensive dataset for training AI-generated image detection models**

[TIGAS Model](https://huggingface.co/H1merka/TIGAS) • [GitHub Repository](https://github.com/H1merka/TIGAS)

</div>

## Dataset Description

The TIGAS Dataset is a large-scale collection of real and AI-generated images designed for training and evaluating AI-generated image detection models. It contains **142,902 images** from diverse sources, including state-of-the-art generative models.

### Key Features

- **Binary classification task**: Real (label=0) vs AI-Generated/Fake (label=1)
- **Diverse generators**: 19 different image sources including GANs and diffusion models
- **Balanced split**: ~54% real, ~46% fake images
- **High-quality annotations**: CSV format with image paths and labels
- **Ready-to-use**: Compatible with PyTorch and standard ML pipelines

## Dataset Statistics

### Overall Distribution

| Split | Total Images | Real (label=0) | Fake (label=1) | Real % |
|-------|--------------|----------------|----------------|--------|
| **Train** | 128,776 | 69,772 | 59,004 | 54.2% |
| **Test** | 14,126 | 7,037 | 7,089 | 49.8% |
| **Total** | 142,902 | 76,809 | 66,093 | 53.7% |

### Image Sources (Train Split)

The dataset includes images from the following generators and sources:

| Source | Images | Type | Description |
|--------|--------|------|-------------|
| `art002_4` | 10,986 | Mixed | Artistic images subset 4 |
| `art002_1` | 10,801 | Mixed | Artistic images subset 1 |
| `VQDM` | 9,518 | Generated | Vector Quantized Diffusion Model |
| `sd14` | 9,517 | Generated | Stable Diffusion 1.4 |
| `Midjourney` | 9,516 | Generated | Midjourney AI |
| `Glide` | 9,513 | Generated | OpenAI GLIDE |
| `wuk` | 9,510 | Mixed | Mixed source images |
| `art002_3` | 8,295 | Mixed | Artistic images subset 3 |
| `gaugan` | 7,992 | Generated | NVIDIA GauGAN |
| `art002_2` | 6,911 | Mixed | Artistic images subset 2 |
| `sd15_1` | 6,353 | Generated | Stable Diffusion 1.5 subset 1 |
| `sd15_2` | 6,349 | Generated | Stable Diffusion 1.5 subset 2 |
| `art001` | 5,966 | Mixed | Artistic images |
| `ADM` | 4,756 | Mixed | Ablated Diffusion Model (ImageNet) |
| `biggan` | 3,200 | Generated | BigGAN |
| `stargan` | 3,198 | Generated | StarGAN (face manipulation) |
| `sd_xl` | 3,196 | Generated | Stable Diffusion XL |
| `face` | 1,600 | Mixed | Face images |
| `DALLE2` | — | Generated | DALL-E 2 (fake only in subset) |

### Image Formats

| Format | Count | Percentage |
|--------|-------|------------|
| PNG | 48,130 | 37.4% |
| JPG | 44,414 | 34.5% |
| JPEG | 34,632 | 26.9% |
| jpeg | 1,600 | 1.2% |

## Dataset Structure

```
TIGAS/
├── LICENSE                     # MIT License
├── README.md                   # This file
├── train/
│   ├── annotations01.csv       # Training annotations (128,776 entries)
│   └── images/
│       ├── ADM/
│       │   ├── 0_real/         # Real images from ImageNet
│       │   └── 1_fake/         # Generated by ADM
│       ├── art001/
│       │   ├── 0_real/
│       │   └── 1_fake/
│       ├── art002_1/ ... art002_4/
│       ├── biggan/
│       ├── DALLE2/
│       ├── face/
│       ├── gaugan/
│       ├── Glide/
│       ├── Midjourney/
│       ├── sd_xl/
│       ├── sd14/
│       ├── sd15_1/
│       ├── sd15_2/
│       ├── stargan/
│       ├── VQDM/
│       └── wuk/
└── test/
    └── annotations01.csv       # Test annotations (14,126 entries)
```

### Annotation Format

The CSV files contain two columns:

```csv
image_path,label
images\ADM\0_real\ILSVRC2012_val_00000005.JPEG,0
images\Midjourney\1_fake\image_001.png,1
```

- **image_path**: Relative path to the image file (Windows-style backslashes)
- **label**: Binary label where:
  - `0` = Real/Natural image
  - `1` = AI-Generated/Fake image

> **Note**: The `test` split uses the same `images/` directory as `train` but with different image subsets defined in its annotation file.

## Usage

### Loading with Python

```python
import pandas as pd
from pathlib import Path
from PIL import Image

# Load annotations
data_root = Path("TIGAS")
train_df = pd.read_csv(data_root / "train" / "annotations01.csv")
test_df = pd.read_csv(data_root / "test" / "annotations01.csv")

# Convert Windows paths to current OS format
train_df['image_path'] = train_df['image_path'].str.replace('\\', '/')

# Load an image
def load_image(row):
    img_path = data_root / "train" / row['image_path']
    image = Image.open(img_path).convert('RGB')
    label = row['label']
    return image, label

# Example
image, label = load_image(train_df.iloc[0])
print(f"Label: {'Real' if label == 0 else 'Fake'}")
```

### Loading with PyTorch

```python
import torch
from torch.utils.data import Dataset, DataLoader
from torchvision import transforms
import pandas as pd
from PIL import Image
from pathlib import Path

class TIGASDataset(Dataset):
    def __init__(self, root_dir, split='train', transform=None):
        self.root_dir = Path(root_dir)
        self.split = split
        self.transform = transform
        
        # Load annotations
        ann_path = self.root_dir / split / "annotations01.csv"
        self.annotations = pd.read_csv(ann_path)
        self.annotations['image_path'] = self.annotations['image_path'].str.replace('\\', '/')
        
        # Images are always in train/images/
        self.images_dir = self.root_dir / "train"
    
    def __len__(self):
        return len(self.annotations)
    
    def __getitem__(self, idx):
        row = self.annotations.iloc[idx]
        img_path = self.images_dir / row['image_path']
        image = Image.open(img_path).convert('RGB')
        label = row['label']
        
        if self.transform:
            image = self.transform(image)
        
        return image, label

# Example usage
transform = transforms.Compose([
    transforms.Resize((256, 256)),
    transforms.ToTensor(),
    transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5])
])

train_dataset = TIGASDataset("TIGAS", split='train', transform=transform)
test_dataset = TIGASDataset("TIGAS", split='test', transform=transform)

train_loader = DataLoader(train_dataset, batch_size=32, shuffle=True, num_workers=4)
test_loader = DataLoader(test_dataset, batch_size=32, shuffle=False, num_workers=4)
```

### Using with TIGAS Model

```python
from tigas import TIGAS

# Initialize model (auto-downloads from HuggingFace)
tigas = TIGAS(auto_download=True, device='cuda')

# Evaluate on dataset images
from pathlib import Path
import pandas as pd

data_root = Path("TIGAS")
test_df = pd.read_csv(data_root / "test" / "annotations01.csv")
test_df['image_path'] = test_df['image_path'].str.replace('\\', '/')

# Evaluate first 10 images
for i, row in test_df.head(10).iterrows():
    img_path = data_root / "train" / row['image_path']
    score = tigas(str(img_path))
    true_label = "Real" if row['label'] == 0 else "Fake"
    pred_label = "Real" if score > 0.5 else "Fake"
    print(f"{img_path.name}: Score={score:.4f}, True={true_label}, Pred={pred_label}")
```

## Generators Included

### Diffusion Models
- **Stable Diffusion 1.4, 1.5, XL** - Open-source text-to-image diffusion models
- **DALL-E 2** - OpenAI's text-to-image model
- **Midjourney** - Commercial text-to-image service
- **GLIDE** - OpenAI's guided language-to-image diffusion
- **ADM** - Ablated Diffusion Model (class-conditional on ImageNet)
- **VQDM** - Vector Quantized Diffusion Model

### GANs (Generative Adversarial Networks)
- **BigGAN** - Large-scale class-conditional GAN
- **GauGAN** - NVIDIA's semantic image synthesis
- **StarGAN** - Multi-domain face manipulation

## Citation

If you use this dataset in your research, please cite:

```bibtex
@dataset{tigas_dataset_2025,
  title={TIGAS Dataset: A Comprehensive Collection for AI-Generated Image Detection},
  author={Morgenshtern, Dmitrij},
  year={2025},
  publisher={HuggingFace},
  url={https://huggingface.co/datasets/H1merka/TIGAS-dataset}
}
```

## License

This dataset is released under the [MIT License](LICENSE).

**Important**: Individual images in this dataset may be derived from or generated using various models with their own licensing terms:
- ImageNet images (in `0_real` folders) are subject to [ImageNet terms of use](https://www.image-net.org/download.php)
- Generated images are outputs of the respective models (Stable Diffusion, Midjourney, etc.)

The annotations and dataset organization are MIT licensed.

## Related Resources

- **TIGAS Model**: [huggingface.co/H1merka/TIGAS](https://huggingface.co/H1merka/TIGAS)
- **GitHub Repository**: [github.com/H1merka/TIGAS](https://github.com/H1merka/TIGAS)

## Changelog

- **v1.0** (December 2025): Initial release with 142,902 images from 19 sources