Datasets:
language:
- en
license: mit
size_categories:
- 10K<n<100K
task_categories:
- text-to-image
- image-feature-extraction
tags:
- diffusion models
- image copy detection
dataset_info:
features:
- name: Name
dtype: string
- name: Level
dtype: int64
- name: generated_images
dtype: image
- name: real_images
dtype: image
splits:
- name: Test
num_bytes: 2538590040
num_examples: 4000
- name: Train
num_bytes: 22265208436
num_examples: 36000
download_size: 24773596239
dataset_size: 24803798476
configs:
- config_name: default
data_files:
- split: Test
path: data/Test-*
- split: Train
path: data/Train-*
pretty_name: '='
Summary
This is the dataset proposed in our paper Image Copy Detection for Diffusion Models (NeurIPS 2024).
D-Rep consists of 40, 000 image-replica pairs, in which each replica is generated by a diffusion model. The 40, 000 image-replica pairs are manually labeled with 6 replication levels ranging from 0 (no replication) to 5 (total replication). We divide D-Rep into a training set with 90% (36, 000) pairs and a test set with the remaining 10% (4, 000) pairs.
Download
Automatical
Install the datasets library first, by:
pip install datasets
Then it can be downloaded automatically with
from datasets import load_dataset
dataset = load_dataset('WenhaoWang/D-Rep')
Manual
You can also download each file by wget
:
wget https://huggingface.co/datasets/WenhaoWang/D-Rep/resolve/main/training_pairs.tar
wget https://huggingface.co/datasets/WenhaoWang/D-Rep/resolve/main/test_pairs.tar
wget https://huggingface.co/datasets/WenhaoWang/D-Rep/resolve/main/labels.csv
Curators
D-Rep is created by Wenhao Wang, Dr. Yifan Sun, Zhentao Tan and Professor Yi Yang.
License
We release our D-Rep under the MIT license.
Citation
@article{wang2024icdiff,
title={Image Copy Detection for Diffusion Models},
author={Wang, Wenhao and Sun, Yifan and Tan, Zhentao and Yang, Yi},
booktitle={Thirty-eighth Conference on Neural Information Processing Systems},
year={2024},
url={https://openreview.net/forum?id=gvlOQC6oP1}
}
Contact
If you have any questions, feel free to contact Wenhao Wang ([email protected]).