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

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

Image Sources (Train Split)

The dataset includes images from the following generators and sources:

Image Formats

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:

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

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

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

    

      

        text
        
          Copy
        
      
      
    # 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

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:

@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.

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
• 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
• GitHub Repository: github.com/H1merka/TIGAS

Changelog

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