🎨 OmniConsistency: Stylized Image Pair Dataset (22 Styles)

OmniConsistency is a large-scale multi-style image translation dataset featuring 22 distinct artistic styles. Each style includes aligned image pairs:

• src: the original image (e.g., photo or line sketch)
• tar: the stylized image
• prompt: a descriptive text representing the intended artistic style

This dataset is suitable for tasks such as:

• Style transfer
• Image-to-image generation
• Conditional generation with prompts
• Consistency learning

🗂️ Dataset Structure

Each style is a separate split (e.g., "Ghibli", "Pixel") in the DatasetDict format. Each sample in a split has:

{
  "src": "Ghibli/src/001.png",
  "tar": "Ghibli/tar/001.png",
  "prompt": "Ghibli Style, dreamy soft lighting, painterly landscape."
}

🔎 Note: Only image paths and prompts are provided. To access full image data, you must clone/download this repository.

🎨 Included Styles (22)

• 3D_Chibi
• American_Cartoon
• Chinese_Ink
• Clay_Toy
• Fabric
• Ghibli
• Irasutoya
• Jojo
• LEGO
• Line
• Macaron
• Oil_Painting
• Origami
• Paper_Cutting
• Picasso
• Pixel
• Poly
• Pop_Art
• Rick_Morty
• Snoopy
• Van_Gogh
• Vector

🧪 How to Use

Load a single style:

from datasets import load_dataset

ds = load_dataset("showlab/OmniConsistency", split="Ghibli")
print(ds[0])

Iterate through styles:

styles = ["3D_Chibi", "Pixel", "Ghibli", "Van_Gogh"]
for style in styles:
    ds = load_dataset("showlab/OmniConsistency", split=style)
    print(style, len(ds))

📷 Image Access

To work with the actual image files:

git lfs install
git clone https://huggingface.co/datasets/showlab/OmniConsistency

Make sure you have Git LFS installed to retrieve image content.

✨ Citation

@inproceedings{Song2025OmniConsistencyLS,
  title={OmniConsistency: Learning Style-Agnostic Consistency from Paired Stylization Data},
  author={Yiren Song and Cheng Liu and Mike Zheng Shou},
  year={2025},
  url={https://api.semanticscholar.org/CorpusID:278905729}
}