Sudarshan2002/GenDS

[CVPR-2025] GenDeg: Diffusion-based Degradation Synthesis for Generalizable All-In-One Image Restoration

Dataset Card for GenDS dataset

The GenDS dataset is a large dataset to boost the generalization of image restoration models. It is a combination of existing image restoration datasets and diffusion-generated degraded samples from GenDeg.

---

Usage

The dataset is fairly large at ~360GB. We recommend having at least 800GB of free space. To download the dataset, git-lfs is required.

Download Instructions

# Install git lfs
git lfs install

# Clone the dataset repository
git clone https://huggingface.co/datasets/Sudarshan2002/GenDS.git
cd GenDS

# Pull the parts
git lfs pull

Extract the Dataset:

# Combine and extract
cat GenDS_part_* > GenDS.tar.gz
tar -xzvf GenDS.tar.gz

After extraction, rename GenDSFull to GenDS.

Dataset Structure

The dataset includes:

• train_gends.json: Metadata for the training data
• val_gends.json: Metadata for the validation data

Each JSON file contains a list of dictionaries with the following fields:

{
  "image_path": "/relpath/to/image",
  "target_path": "/relpath/to/ground_truth",
  "dataset": "Source dataset name",
  "degradation": "Original degradation type",
  "category": "real | synthetic",
  "degradation_sub_type": "GenDeg-generated degradation type OR 'Original' (if from existing dataset)",
  "split": "train | val",
  "mu": "mu value used in GenDeg",
  "sigma": "sigma value used in GenDeg",
  "random_sampled": true | false,
  "sampled_dataset": "Dataset name if mu/sigma are not random"
}

Example Usage:

import json

# Load train metadata
with open("/path/to/train_gends.json") as f:
    train_data = json.load(f)

print(train_data[0])

Citation

If you use GenDS in your work, please cite:

@article{rajagopalan2024gendeg,
  title={GenDeg: Diffusion-Based Degradation Synthesis for Generalizable All-in-One Image Restoration},
  author={Rajagopalan, Sudarshan and Nair, Nithin Gopalakrishnan and Paranjape, Jay N and Patel, Vishal M},
  journal={arXiv preprint arXiv:2411.17687},
  year={2024}
}