Datasets

Comparison table

Feature	DIV2K	LIU4K v1	LIU4K v2	UHD-IQA
Resolution	2K	4K	4K–6K	4K (3840×2160)
Format	PNG (lossless)	PNG (lossless)	PNG (lossless)	JPEG (lossy)
Images (val)	100	80	400	~2000
Images (train)	800	800	1600	~4000
Size (val)	449 MB	1.3 GB	15 GB	~3.5 GB
Size (train)	3.5 GB	10 GB	60 GB	~7 GB
Storage	Direct HTTP	Google Drive	Google Drive	Direct HTTP
License	Unknown	CC BY-NC-ND 4.0	CC BY-NC-ND 4.0	CC0
Post-process	None	None	Extract multi-part zips	None

DIV2K (2K)

Source: ETH Zurich Computer Vision Lab
Resolution: 2K (variable aspect ratios)
Format: PNG (lossless)
License: Unknown — check DIV2K website
IDs: div2k-valid (100 images, 449 MB), div2k-train (800 images, 3.5 GB)

Best for development and testing due to manageable size and direct HTTP download.

LIU4K v1 (4K)

Source: Peking University — LIU4K Benchmark
Resolution: 4K
Format: PNG (lossless)
License: CC BY-NC-ND 4.0 (attribution required, non-commercial, no derivatives)
Storage: Google Drive (requires gdown)
IDs: liu4k-v1-valid (80 images, 1.3 GB), liu4k-v1-train (800 images, 10 GB)

Recommended for 4K research. Single-file zip archive, simpler than v2.

LIU4K v2 (4K–6K)

Source: Peking University — LIU4K v2
Resolution: 4K–6K (minimum 3K)
Format: PNG (lossless)
License: CC BY-NC-ND 4.0
Storage: Google Drive — multi-part zip archives by category
IDs: liu4k-v2-valid (400 images, 15 GB), liu4k-v2-train (1600 images, 60 GB)
Requires: 7z for multi-part zip extraction (pre-installed in dev container)

Larger and higher resolution than v1. Categories: Animal, Building, Capture, Mountain, Street.

UHD-IQA (4K JPEG)

Source: University of Konstanz — UHD-IQA Database
Resolution: UHD-1 (3840×2160)
Format: JPEG (lossy, sourced from Pixabay)
License: CC0 (Public Domain)
ID: uhd-iqa-full (6073 images, 10.7 GB)
Includes: Quality ratings (MOS) and metadata

Largest dataset by image count. However, JPEG source images contain pre-existing compression artifacts, making this dataset less suitable for unbiased format comparison. Best for large-scale testing and quality assessment research.

Recommendations by use case

Use case	Recommended dataset
Development and testing	`div2k-valid` — fast download, manageable size
4K format research	`liu4k-v1-valid` — lossless 4K, reasonable size
Large-scale 4K study	`liu4k-v2-train` — 1600 lossless images up to 6K
Scale testing	`uhd-iqa-full` — 6000+ images (note: JPEG source)
Unbiased format comparison	DIV2K or LIU4K (lossless PNG, no pre-existing artifacts)

Storage location

All datasets are stored under data/datasets/:

data/datasets/
├── DIV2K_valid/
├── DIV2K_train/
├── LIU4K_valid/
├── LIU4K_train/
└── UHD_IQA/