DEAL-300K Dataset

DEAL-300K: Diffusion-based Editing Area Localization with a 300K-Scale Dataset and Frequency-Prompted Baseline

Overview

DEAL-300K is a large-scale dataset for Diffusion-based Image Manipulation Localization (DIML) containing over 300,000 annotated images. The dataset is specifically designed to address the challenges of localizing regions edited by diffusion models, which often blend seamlessly with original content.

Key Features

• Scale: 330,979 training images, 3,989 validation images, and 5,500 test images
• Annotation: Pixel-level segmentation masks for edited regions
• Diversity: Covers a broad range of manipulations involving humans, animals, and objects
• Quality: Automated annotation pipeline using SAM-CD (Segment Anything Model with Change Detection)

Dataset Structure

The dataset is stored in Parquet format for efficient loading and compatibility with Hugging Face Datasets:

train-00000-of-00067.parquet  # Training set (67 files)
train-00001-of-00067.parquet
...
val-00000-of-00001.parquet    # Validation set (1 file)
test-00000-of-00002.parquet   # Test set (2 files)
test-00001-of-00002.parquet

Data Fields

Field	Type	Description
image	PIL.Image	RGB image (512×384 or similar resolution)
mask	PIL.Image	Grayscale binary mask (L mode)
label	ClassLabel	0=real, 1=fake
image_name	string	Original filename

Mask Details:

• For real images: all-black mask (pixel value 0)
• For fake images: segmentation mask of edited regions (white=255=edited, black=0=original)

Dataset Statistics

Split	Real Images	Fake Images	Total	Parquet Files
Train	115,814	215,165	330,979	67
Val	1,656	2,333	3,989	1
Test	1,901	3,599	5,500	2
Total	119,371	221,097	340,468	70

Usage

Loading from Hugging Face Hub (Recommended)

from datasets import load_dataset

# Load the full dataset
dataset = load_dataset("FlyHorseJ/DEAL-300K")

# Or load specific splits
train_dataset = dataset["train"]
val_dataset = dataset["val"]
test_dataset = dataset["test"]

# Access a sample
sample = train_dataset[0]
image = sample["image"]   # PIL Image (RGB)
mask = sample["mask"]     # PIL Image (grayscale mask)
label = sample["label"]   # 0 (real) or 1 (fake)
name = sample["image_name"]  # Original filename

PyTorch DataLoader Example

from torch.utils.data import DataLoader
from torchvision import transforms
import torch

def collate_fn(batch):
    transform = transforms.Compose([
        transforms.ToTensor(),
        transforms.Resize((512, 512))
    ])
    
    images = []
    masks = []
    labels = []
    
    for item in batch:
        images.append(transform(item["image"]))
        masks.append(transform(item["mask"]))
        labels.append(item["label"])
    
    return {
        "images": torch.stack(images),
        "masks": torch.stack(masks),
        "labels": torch.tensor(labels)
    }

# Load dataset from Hugging Face Hub
dataset = load_dataset("FlyHorseJ/DEAL-300K")

# Create DataLoader
train_loader = DataLoader(
    dataset["train"], 
    batch_size=32, 
    shuffle=True,
    collate_fn=collate_fn,
    num_workers=4
)

Visualization Example

import matplotlib.pyplot as plt
import numpy as np
from datasets import load_dataset

# Load from Hugging Face Hub
dataset = load_dataset("FlyHorseJ/DEAL-300K")

sample = dataset["train"][0]
fig, axes = plt.subplots(1, 3, figsize=(15, 5))

# Original image
axes[0].imshow(sample["image"])
axes[0].set_title(f"Image ({'Real' if sample['label'] == 0 else 'Fake'})")
axes[0].axis('off')

# Mask
axes[1].imshow(sample["mask"], cmap='gray')
axes[1].set_title("Mask")
axes[1].axis('off')

# Overlay
img_array = np.array(sample["image"])
mask_array = np.array(sample["mask"].convert('RGB'))
overlay = (img_array * 0.7 + mask_array * 0.3).astype(np.uint8)

axes[2].imshow(overlay)
axes[2].set_title("Overlay")
axes[2].axis('off')

plt.tight_layout()
plt.show()

Dataset Generation Pipeline

1. Instruction Generation: Fine-tuned Qwen-VL generates editing instructions for MS COCO images
2. Image Editing: InstructPix2Pix performs mask-free editing based on instructions
3. Annotation: SAM-CD (Segment Anything Model with Change Detection) generates pixel-level masks

Comparison with Existing Datasets

Dataset	Year	Source Images	Edited Images	Image Size	Scenario	Generative Model
CoCoGlide	2023	512	512	256×256	General	GLIDE (Mask-Required)
AutoSplice	2023	2,273	3,621	256×256–4232×4232	General	DALL-E2 (Mask-Required)
MagicBrush	2023	5,313	10,388	1024×1024	General	DALL-E2 (Mask-Required)
Repaint-P2/CelebA-HQ	2024	10,800	41,472	256×256	Face	Repaint (Mask-Required)
DEAL-300K	2025	119,371	221,097	128×512–512×576	General	InstructPix2Pix (Mask-Free)

Citation

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

@article{zhang2025deal300k,
  title={DEAL-300K: Diffusion-based Editing Area Localization with a 300K-Scale Dataset and Frequency-Prompted Baseline},
  author={Zhang, Rui and Wang, Hongxia and Liu, Hangqing and Zhou, Yang and Zeng, Qiang},
  journal={arXiv preprint arXiv:2511.23377},
  year={2025}
}

Contact

For questions or issues about the dataset, please:

• Open an issue on GitHub: https://github.com/ymhzyj/DEAL-300K

License

This dataset is released for academic research purposes. Please refer to the original MS COCO and MagicBrush licenses for usage restrictions.

Related Links

• Paper: https://arxiv.org/abs/2511.23377
• GitHub: https://github.com/ymhzyj/DEAL-300K
• HuggingFace Dataset: https://huggingface.co/datasets/FlyHorseJ/DEAL-300K
• MS COCO: https://cocodataset.org/
• MagicBrush: https://github.com/OSU-NLP-Group/MagicBrush