byliutao/coco2014val_10k

COCO-2014 Val 10K (256×256)

A curated subset of the COCO 2014 Validation set containing 9,986 image-caption pairs, designed as a standard reference benchmark for FID (Fréchet Inception Distance) evaluation in text-to-image generation research.

Dataset Summary

Intended Use

This dataset serves as a reference image set for FID computation when evaluating text-to-image diffusion models. The captions in test.txt can be used as generation prompts, and the corresponding images serve as ground-truth references for computing FID scores.

Primary use case: evaluation benchmark for the NFT (Neural Flow Trajectories) distillation framework for diffusion models.

Dataset Structure

coco2014val_10k/
├── README.md              # This file
├── metadata.jsonl         # 9,986 records: {"file_name": "images/000000.png", "text": "..."}
├── test.txt               # 9,986 captions (one per line), usable as generation prompts
├── metadata.json          # Dataset statistics
└── images/                # 9,986 reference images (256×256 PNG)
    ├── 000000.png
    ├── 000001.png
    └── ...

Preprocessing

Images were preprocessed from the original COCO 2014 Val images using the following pipeline:

1. 1.Center-crop to square (using the shorter side)
2. 2.Resize to 256×256 using Lanczos interpolation
3. 3.Save as lossless PNG in RGB mode

Each image is paired with one randomly selected caption (out of 5 available per image in the original COCO dataset), sampled with random_seed=42 for reproducibility.

Loading the Dataset

from datasets import load_dataset

dataset = load_dataset("YOUR_USERNAME/coco2014val_10k", split="test")

# Access an example
example = dataset[0]
print(example["text"])       # caption string
example["image"].show()      # PIL Image (256x256)

License

• —Annotations (captions): Creative Commons Attribution 4.0 (CC BY 4.0)
• —Images: Images are sourced from the COCO 2014 dataset, which were originally collected from Flickr. Image copyrights belong to their respective Flickr owners. The COCO consortium does not own the copyright of the images. Use of the images must abide by the Flickr Terms of Use. The images in this dataset have been resized and cropped from the originals.

Citation

If you use this dataset, please cite the original COCO dataset and (optionally) the NFT paper:

@inproceedings{lin2014microsoft,
  title={Microsoft {COCO}: Common Objects in Context},
  author={Lin, Tsung-Yi and Maire, Michael and Belongie, Serge and Hays, James
          and Perona, Pietro and Ramanan, Deva and Doll{\'a}r, Piotr and Zitnick, C. Lawrence},
  booktitle={European Conference on Computer Vision (ECCV)},
  pages={740--755},
  year={2014},
  publisher={Springer}
}

Disclaimer

This dataset is a derived, preprocessed subset of the COCO 2014 Validation set. The images have been center-cropped and resized to 256×256 pixels — they are not the original full-resolution COCO images. This dataset is intended for research purposes only.