mkjia/UHDBench

Dataset Card for Dataset Name

Ultra-high definition benchmark for zero-shot image reconstruction evaluation.

Dataset Description

• Total images: 2 293 at 2K resolution
• Source datasets: HRSOD, LIU4k, UAVid, UHDM, UHRSD
• Resolution filter: Only images ≥ 2560 × 1440 included
• Purpose: Zero-shot image reconstruction benchmarking.
• license: cc-by-sa-4.0

Dataset Sources

• Repository: https://github.com/MKJia/MGVQ
• Paper: https://arxiv.org/abs/2507.07997
• Dataset Repository: https://github.com/MKJia/UHDBench

Dataset Leaderboard

Call for Submissions! We're continuously expanding our public benchmark leaderboard and welcome contributions from the community.

Feel free to suggest other VQVAEs or VAEs. We're happy to assist with the evaluation. We also invite you to share your reconstruction results to be included in our leaderboard.

Method	Type	Ratio	rFID↓	PSNR↑
SD-VAE	Continuous	16	1.07	26.86
VQGAN	Discrete	16	5.95	22.91
LlamaGen	Discrete	16	5.59	23.90
OpenMagvit2	Discrete	16	4.18	23.91
VAR	Discrete	16	9.85	21.79
MGVQ-f16c32-g4	Discrete	16	1.59	28.27

Dataset Structure

UHDBench/
├── HRSOD_release/
├── LIU4k/
├── uavid_test/
├── UHDM/
├── HRSD_TE/
└── UHDBench.json # json file for image sources and paths

Dataset Creation

Source data producers

• HRSOD & UHRSD: By Xie et al. in “Pyramid Grafting Network for One-Stage High Resolution Saliency Detection”
• LIU4K: By Liu et al. in “A Comprehensive Benchmark for Single Image Compression Artifact Reduction”
• UAVid: By Lyuet al. in “UAVid: A semantic segmentation dataset for UAV imagery”
• UHDM: By Yu et al. in “Towards efficient and scale-robust ultra-high-definition image demoireing”

Source Data Licenses

• HRSOD & UHRSD: MIT
• LIU4K: CC0
• UAVid: CC-BY-SA-4.0
• UHDM: Apache 2.0

Citation

Please consider staring UHDBench&MGVQ and citing the following paper if you feel this dataset useful.

@article{jia2025mgvq,
  title={MGVQ: Could VQ-VAE Beat VAE? A Generalizable Tokenizer with Multi-group Quantization},
  author={Jia, Mingkai and Yin, Wei and Hu, Xiaotao and Guo, Jiaxin and Guo, Xiaoyang and Zhang, Qian and Long, Xiao-Xiao and Tan, Ping},
  journal={arXiv preprint arXiv:2507.07997},
  year={2025}
}