WenhaoWang/OriPID

Summary

This is the dataset proposed in our paper Origin Identification for Text-Guided Image-to-Image Diffusion Models (ICML 2025).

Download

Training

You can download the images:

wget https://huggingface.co/datasets/WenhaoWang/OriPID/resolve/main/training/sd2_d_multi.tar.part_0{0..9}
cat sd2_d_multi.tar.part_* > sd2_d_multi.tar
tar -xvf sd2_d_multi.tar

Or you can directly download the features extracted by VAE in Stable Diffusion 2:

wget https://huggingface.co/datasets/WenhaoWang/OriPID/resolve/main/training/sd2_d_multi_feature.tar
tar -xvf sd2_d_multi_feature.tar

The features are extracted by:

# pip install torch torchvision torchaudio transformers diffusers accelerate
from diffusers import AutoPipelineForImage2Image
import torchvision
import torch
from PIL import Image
import requests

pipeline = AutoPipelineForImage2Image.from_pretrained("stabilityai/stable-diffusion-2", torch_dtype=torch.float32, variant="fp16", use_safetensors=True)
vae = pipeline.vae

mean, std = [0.485, 0.456, 0.406],[0.229, 0.224, 0.225]
transforms = torchvision.transforms.Compose([
  torchvision.transforms.Resize((256, 256)),
  torchvision.transforms.ToTensor(),
  torchvision.transforms.Normalize([0.5], [0.5]), 
])

url = "https://huggingface.co/datasets/WenhaoWang/AnyPattern/resolve/main/Irises.jpg"
image = Image.open(requests.get(url, stream=True).raw)
latents = vae.encode(transforms(image).unsqueeze(0)).latent_dist.sample() # torch.Size([1, 4, 32, 32])
features = latents.reshape(len(latents), -1) # torch.Size([1, 4096])

Query

wget https://huggingface.co/datasets/WenhaoWang/OriPID/resolve/main/query/colorful.tar
wget https://huggingface.co/datasets/WenhaoWang/OriPID/resolve/main/query/kk.tar
wget https://huggingface.co/datasets/WenhaoWang/OriPID/resolve/main/query/kolor.tar
wget https://huggingface.co/datasets/WenhaoWang/OriPID/resolve/main/query/opendalle.tar
wget https://huggingface.co/datasets/WenhaoWang/OriPID/resolve/main/query/sd2.tar
wget https://huggingface.co/datasets/WenhaoWang/OriPID/resolve/main/query/sd3.tar
wget https://huggingface.co/datasets/WenhaoWang/OriPID/resolve/main/query/sdxl.tar

tar -xvf colorful.tar
tar -xvf kk.tar
tar -xvf kolor.tar
tar -xvf opendalle.tar
tar -xvf sd2.tar
tar -xvf sd3.tar
tar -xvf sdxl.tar

Reference

wget https://huggingface.co/datasets/WenhaoWang/AnyPattern/resolve/main/reference/references_{0..19}.zip
for z in references_*.zip; do unzip $z; done
mv images/references reference_images

License

The dataset is licensed under the CC BY-NC 4.0 license. For commercial uses, please email [email protected].

Citation

@article{wang2025origin,
  title={Origin Identification for Text-Guided Image-to-Image Diffusion Models},
  author={Wang, Wenhao and Sun, Yifan and Yang, Zongxin and Tan, Zhentao and Hu, Zhengdong and Yang, Yi},
  journal={Forty-second International Conference on Machine Learning},
  year={2025},
  url={https://openreview.net/forum?id=46n3izUNiv}
}

Contact

If you have any questions, feel free to contact Wenhao Wang ([email protected]).