DF2K_OST Dataset

High-quality 256x256 image dataset for training autoencoders.

Dataset Description

This dataset combines three high-quality image sources commonly used for image super-resolution and restoration tasks:

• DIV2K: 900 high-resolution images (800 train + 100 validation)
• Flickr2K: 2,650 high-resolution images
• OST (Outdoor Scene Training): 10,324 outdoor scene images

All images have been resized to 256x256 pixels using Lanczos resampling for optimal quality.

Dataset Structure

DF2K_OST/
├── train/          # ~90% of images ({len(dataset_dict['train'])} images)
└── validation/     # ~10% of images ({len(dataset_dict['validation'])} images)

Each sample contains:

• image: 256x256 RGB image
• source: Original dataset source (DIV2K, Flickr2K, or OST)
• filename: Original filename

Processing

All images were processed using:

• Target resolution: 256x256 pixels
• Resampling method: Lanczos (PIL.Image.LANCZOS)
• Color mode: RGB
• Train/validation split: 90/10 (stratified random)

Usage

from datasets import load_dataset

# Load full dataset
dataset = load_dataset("gperdrizet/DF2K_OST")

# Load only training split
train_data = load_dataset("gperdrizet/DF2K_OST", split="train")

# Access images
for sample in train_data:
    image = sample['image']  # PIL Image
    source = sample['source']  # Dataset source

Original Sources

DIV2K

• Citation: Agustsson, E., & Timofte, R. (2017). NTIRE 2017 Challenge on Single Image Super-Resolution: Dataset and Study.
• Paper: CVPR Workshops 2017
• Website: https://data.vision.ee.ethz.ch/cvl/DIV2K/

Flickr2K

• Citation: Agustsson, E. & Timofte, R. (2017). "NTIRE 2017 challenge on single image super-resolution: Dataset and study." In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW).
• Repository: https://github.com/limbee/NTIRE2017

OST (Outdoor Scene Training)

• Citation: Wang, X., Yu, K., Dong, C., & Loy, C. C. (2018). Recovering Realistic Texture in Image Super-resolution by Deep Spatial Feature Transform. arXiv:1804.02815
• Paper: https://arxiv.org/abs/1804.02815
• Repository: https://github.com/xinntao/SFTGAN

License

The compilation and processing are provided under Apache 2.0 license. Individual images retain their original licenses from source datasets.

Citation

If you use this dataset, please cite the original source datasets:

@inproceedings{{agustsson2017ntire,
  title={{NTIRE 2017 Challenge on Single Image Super-Resolution: Dataset and Study}},
  author={{Agustsson, Eirikur and Timofte, Radu}},
  booktitle={{CVPR Workshops}},
  year={{2017}}
}}

@misc{{wang2018recoveringrealistictextureimage,
  title={{Recovering Realistic Texture in Image Super-resolution by Deep Spatial Feature Transform}}, 
  author={{Xintao Wang and Ke Yu and Chao Dong and Chen Change Loy}},
  year={{2018}},
  eprint={{1804.02815}},
  archivePrefix={{arXiv}},
  primaryClass={{cs.CV}},
  url={{https://arxiv.org/abs/1804.02815}}
}}

Created By

Processed and compiled for the Autoencoders educational demo project by Dr. George Perdrizet. Repository: https://github.com/gperdrizet/autoencoders