dwchoi/div2k800-ffhq1000

DIV2K-800 + FFHQ-1000 (with prompts)

A convenience mirror of the two ground-truth image sets used as benchmarks for diffusion-based inverse problems (super-resolution, Gaussian/motion deblurring, inpainting, …), bundled with the text prompts that accompany each set.

Everything here is the clean ground truth — no degraded / measurement images are included. Degradations are meant to be generated on the fly by the operator A of whichever task you are running.

Contents

Path	Files	Size	Resolution	Prompt
DIV2K_train_HR/	800 PNG, 0001.png … 0800.png	3.3 GB	variable, W 1116–2040 (2040 for 713/800), H 648–2040	per-image, see prompts/DIV2K_prompts.txt
FFHQ_1000/	1000 PNG, 00000.png … 00999.png	1.3 GB	1024 × 1024	one shared prompt, see prompts/FFHQ_prompt.txt
prompts/DIV2K_prompts.txt	800 lines	—	—	line i ↔ {i:04d}.png (line 1 ↔ 0001.png)
prompts/FFHQ_prompt.txt	1 line	—	—	a high quality photo

Each image directory also carries a metadata.jsonl ({"file_name": ..., "prompt": ...}) so that 🤗 datasets returns image/prompt pairs directly. The plain-text files under prompts/ are the originals and are the ones to use if your pipeline reads a prompt list by index.

DIV2K prompts

DAPE-style tag prompts, one line per image, comma-separated, e.g.

close-up, coral reef, sea, pink, purple, red, reef, starfish, clean, high-resolution, 8k
horse, horseback, person, lake, man, mountain, mountain landscape, reflection, ride, water, clean, high-resolution, 8k

All 800 lines end with the fixed positive suffix clean, high-resolution, 8k; content tags average ≈ 12 per image (4 – 23).

FFHQ prompt

FFHQ is a single-domain (aligned face) set, so every image uses the same prompt:

a high quality photo

It is stored as a one-line file rather than 1000 duplicated lines. FFHQ_1000/metadata.jsonl does repeat it per image, purely so the datasets loader can pair it up.

Usage

With 🤗 datasets (image + prompt pairs)

from datasets import load_dataset

div2k = load_dataset("dwchoi/div2k800-ffhq1000", "div2k", split="train")
ffhq  = load_dataset("dwchoi/div2k800-ffhq1000", "ffhq",  split="train")

print(div2k[0]["prompt"])   # 'close-up, coral reef, sea, ...'
div2k[0]["image"]           # PIL.Image, 2040x1404

As plain folders (indexed pipelines)

from huggingface_hub import snapshot_download

root = snapshot_download(repo_id="dwchoi/div2k800-ffhq1000", repo_type="dataset")
# root/DIV2K_train_HR/0001.png ... 0800.png
# root/FFHQ_1000/00000.png ... 00999.png
# root/prompts/DIV2K_prompts.txt   (800 lines, line i -> {i:04d}.png)

Or one subset only:

hf download dwchoi/div2k800-ffhq1000 --repo-type dataset \
    --include "FFHQ_1000/*" "prompts/*" --local-dir ./data

Reading the prompt list by index

prompts = open("prompts/DIV2K_prompts.txt").read().rstrip("\n").split("\n")
assert len(prompts) == 800
prompt_for_0042 = prompts[42 - 1]

Note. The file has no trailing blank line in some earlier copies, so wc -l may report 799. The file really does contain 800 records — always split on \n after rstrip, or use metadata.jsonl.

Provenance

• DIV2K_train_HR — the complete 800-image training split of the DIV2K super-resolution dataset, unmodified original HR PNGs.
• FFHQ_1000 — the first 1000 images (00000–00999) of Flickr-Faces-HQ at 1024 × 1024, unmodified.

No resizing, re-encoding, or cropping was applied to either set — pixels are bit-identical to the sources. (Benchmarks that need a fixed input size, e.g. Resize(768) + CenterCrop, apply that at load time, not here.)

License & attribution

This repository redistributes two third-party datasets; the original licenses govern, and this mirror adds no rights of its own.

Subset	Source	License
DIV2K	DIV2K, NTIRE 2017 (ETH Zürich)	Released for academic research use; see the DIV2K page for terms
FFHQ	Flickr-Faces-HQ (NVIDIA)	CC BY-NC-SA 4.0 (non-commercial). Individual photographs remain under their original Flickr licenses and copyright of their respective photographers

FFHQ contains photographs of real people. It is provided for non-commercial research only. Do not use it to identify, profile, or generate content about the individuals depicted. If you are a subject of an FFHQ photo and want it removed, follow the takedown procedure on the NVlabs/ffhq-dataset repository — removals there should be mirrored here.

Please cite the originals:

@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}
}

@inproceedings{karras2019style,
  title     = {A Style-Based Generator Architecture for Generative Adversarial Networks},
  author    = {Karras, Tero and Laine, Samuli and Aila, Timo},
  booktitle = {CVPR},
  year      = {2019}
}

The DIV2K tag prompts follow the DAPE / SeeSR prompt convention:

@inproceedings{wu2024seesr,
  title     = {SeeSR: Towards Semantics-Aware Real-World Image Super-Resolution},
  author    = {Wu, Rongyuan and Yang, Tao and Sun, Lingchen and Zhang, Zhengqiang and Li, Shuai and Zhang, Lei},
  booktitle = {CVPR},
  year      = {2024}
}