4KLSDB: A Large-Scale Dataset for 4K Image Restoration and Generation

DataCV @ CVPR 2026 · Accepted 🎉

Zihao Zhu¹, Kuan-Ru Huang¹, Zhaoming Xu¹, Renjie Li¹, Bo Wu¹, Ruizheng Bai¹, Mingyang Wu¹, Sayak Paul², Zhengzhong Tu^†,1

¹Texas A&M University    ²Hugging Face   

4KLSDB is the first openly released native-4K image dataset that scales to 129k+ training images and is designed for both image restoration and generation research. Every model in our paper — HiT-SR, SwinIR, MambaIR, OSEDiff, SeeSR, and Sana — gets a consistent and substantial boost when fine-tuned on 4KLSDB.

• 🖼️ Dataset: 129,484 train / 2,000 val / 1,984 test, all native-4K, with captions, on 🤗 Hugging Face.
• 🧱 Pre-trained checkpoints: every SR/T2I model released under 🤗 taco-group.
• 🚀 One-click inference: ready-to-run scripts for each model under scripts/.
• 🏋️ One-click training: reproducible YAML configs and shell scripts under models/.

• 2026-05 — Public release of dataset, code, and pretrained weights.
• 2026-05 — Paper released on arXiv.

• Highlights
• Dataset
• Pre-trained Models
• Quick Start
  • Environment
  • Download Everything (one script)
  • Classical SR Inference (HiT-SR / SwinIR / MambaIR)
  • Real-World SR Inference (OSEDiff / SeeSR)
  • 4K T2I Inference (Sana)
• Training
• Benchmark Results
• Repository Structure
• Citation
• Acknowledgements
• License

• Native 4K — every image meets a minimum dimension of 3840 px and a 3840 × 2160 pixel budget.
• Scale — 129,484 training images, 22× larger than DIV8K, 150× larger than DIV2K.
• Quality pipeline — Q-Align aesthetic scoring + Laplacian/Sobel texture filtering + two human annotators per image.
• Dual-purpose — works out-of-the-box for classical SR (×4 / ×8 / ×16), real-world SR, and 4K T2I generation.
• Reproducibility — all training configs, blind-degradation pipeline, and evaluation code are included.

The 4KLSDB dataset is hosted on Hugging Face:

https://huggingface.co/datasets/SingleBicycle/4KLSDB

from datasets import load_dataset

# Streaming (recommended — the train split is ~1.5 TB)
ds = load_dataset("SingleBicycle/4KLSDB", split="train", streaming=True)
for ex in ds:
    print(ex["image"], ex["caption"])
    break

Categories covered: nature, urban scenes, people, food, artwork, CGI, animals, architecture.

All 4KLSDB-fine-tuned checkpoints live alongside the dataset under SingleBicycle/4KLSDB/ckpts/.

One-shot download of every model:

bash scripts/download_all_ckpts.sh        # → release_ckpts/<model>/

New (May 2026): the dataset now ships an authoritative metadata.jsonl with Qwen2.5-VL-7B recaptions for all 129,484 training images. Use this for T2I fine-tuning instead of the older caption column in the parquet shards.

We strongly recommend using a separate conda environment per model family.

# Classical SR & Real-SR (HiT-SR / SwinIR / MambaIR / OSEDiff / SeeSR)
conda env create -f envs/4k_sr.yml
conda activate 4k_sr

# 4K T2I generation (Sana)
conda env create -f envs/Sana_training.yml
conda activate Sana

bash scripts/download_all_ckpts.sh                # → release_ckpts/
huggingface-cli download SingleBicycle/4KLSDB \
    --repo-type=dataset --local-dir ./data/4KLSDB

bash scripts/inference_classical_sr.sh \
    --model hit_sr          \   # or swinir / mambair
    --scale 4               \   # 4, 8, or 16
    --input  data/4KLSDB/test/LR_x4 \
    --output results/hit_sr_x4

bash scripts/inference_real_sr.sh \
    --model seesr           \   # or osediff
    --scale 4               \
    --input  data/4KLSDB/test/LR_real_x4 \
    --output results/seesr_x4

bash scripts/inference_sana_4k.sh \
    --prompt "A serene mountain lake at sunrise, 4K, photorealistic" \
    --resolution 4096 \
    --output results/sana_4k.png

Run bash scripts/<name>.sh --help for the full list of options on any script.

Each model lives as a submodule under models/<name> with its own training entry-point. The shell scripts below wrap the upstream configs and inject 4KLSDB-specific paths so a single command reproduces the paper.

# Classical SR
bash scripts/train_hit_sr.sh   --scale 4   --data data/4KLSDB
bash scripts/train_swinir.sh   --scale 8   --data data/4KLSDB
bash scripts/train_mambair.sh  --scale 16  --data data/4KLSDB

# Real-World SR (blind degradation pipeline)
bash scripts/train_osediff.sh  --scale 4   --data data/4KLSDB
bash scripts/train_seesr.sh    --scale 4   --data data/4KLSDB

# 4K T2I
bash scripts/train_sana_4k.sh  --resolution 4096  --data data/4KLSDB

Detailed per-model docs:

• models/sana/README.md — Sana 4K fine-tuning + Gemma-2 caption embedding pre-compute.
• dataset/README.md — curation pipeline (resolution / Q-Align / Laplacian / Sobel / manual review).

Double-blind user study win rate of Sana + 4KLSDB over Sana: 57.3% overall, 60.9% detail, 74.3% realism, 64.4% fewer artifacts, 52.3% alignment.

4KLSDB/
├── README.md                     # this file
├── docs/                         # GitHub Pages project page (index.html)
│   ├── index.html                # https://4klsdb.github.io/
│   └── assets/                   # teaser & figure JPGs used by the project page
├── envs/
│   ├── 4k_sr.yml                 # classical SR + real-SR (HiT-SR / SwinIR / MambaIR / OSEDiff / SeeSR)
│   ├── Sana_training.yml         # 4K T2I (Sana)
│   └── 4k_data_curation.yml      # dataset filtering / Q-Align scoring
├── dataset/
│   ├── README.md                 # dataset curation pipeline doc
│   ├── preprocessing/            # Q-Align / Laplacian / Sobel filters
│   └── validation/               # manual inspection Flask app
├── models/
│   ├── README.md
│   ├── sana/                     # 4K T2I submodule (NVlabs/Sana)
│   ├── hit_sr/                   # → submodule placeholder
│   ├── swinir/                   # → submodule placeholder
│   ├── mambair/                  # → submodule placeholder
│   ├── seesr/                    # → submodule placeholder
│   └── osediff/                  # → submodule placeholder
├── scripts/                      # one-click inference / training / download wrappers
│   ├── download_all_ckpts.sh
│   ├── inference_classical_sr.sh
│   ├── inference_real_sr.sh
│   ├── inference_sana_4k.sh
│   ├── train_hit_sr.sh
│   ├── train_swinir.sh
│   ├── train_mambair.sh
│   ├── train_osediff.sh
│   ├── train_seesr.sh
│   └── train_sana_4k.sh
└── LICENSE

If you find 4KLSDB useful for your research, please cite:

@misc{zhu20264klsdblargescaledataset4k,
      title={4KLSDB: A Large-Scale Dataset for 4K Image Restoration and Generation}, 
      author={Zihao Zhu and Kuan-Ru Huang and Zhaoming Xu and Renjie Li and Bo Wu and Ruizheng Bai and Mingyang Wu and Sayak Paul and Zhengzhong Tu},
      year={2026},
      eprint={2605.24762},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2605.24762}, 
}

Our work builds on a number of excellent open-source projects:

• HiT-SR, SwinIR, MambaIR
• OSEDiff, SeeSR
• Sana
• Q-Align
• Image sources: LAION-2B, Photo Concept Bucket, PD12M

The project page is adapted from the SparkVSR template.

The code in this repository is released under the MIT License. The 4KLSDB dataset is released for research purposes only; please refer to the dataset card for the full terms and source-dataset licenses.