HQ-50K: Baidu Netdisk Archive Mirror

English | 中文

English

This repository provides a Hugging Face mirror of an HQ-50K archive downloaded through the original project's Baidu Netdisk distribution channel, referenced in data_preprocess/README.md.

The currently available archive is TRAIN_BICUBIC.7z, approximately 338 GB. This repository hosts a data archive rather than only a list of image URLs, providing an alternative way to obtain the file through Hugging Face.

This is a third-party download mirror, not an official release repository maintained by the original authors. The current file listing does not establish that all HQ-50K training, test, or derived data have been mirrored. Please refer to the original project for authoritative dataset descriptions and processing instructions.

1. About HQ-50K

HQ-50K: A Large-scale, High-quality Dataset for Image Restoration is a large-scale, high-quality dataset for image restoration research. According to the original project, it includes:

• 50,000 training images with rich texture details and semantic diversity.
• 1,250 test images spanning different semantic categories and frequency ranges for fine-grained evaluation.

These counts describe the original HQ-50K dataset, not a verified inventory of the archive hosted here. This dataset card does not independently verify the archive's internal directories, image count, HR/LR pairing, or scale factors. Please inspect the extracted contents and consult the original documentation.

2. Sources and Links

Resource	Link
Original project and code	littleYaang/HQ-50K
Original download and preprocessing instructions	data_preprocess/README.md
Hugging Face dataset linked by the original project	YangQiee/HQ-50K
Paper	arXiv:2306.05390
This mirror	flashszn/HQ-50K

The original preprocessing documentation provides the following Baidu Netdisk links:

Split	Original download link	Access code
HQ-TRAIN	Baidu Netdisk	hq50
HQ-TEST	Baidu Netdisk	hq50

These are upstream source links; listing both does not mean this repository contains every file available through both links. Refer to the original project for link updates.

3. Repository Contents

As of 2026-11-10, the repository contains:

HQ-50K/
├── .gitattributes
├── README.md
└── TRAIN_BICUBIC.7z

File	Size	Description
TRAIN_BICUBIC.7z	Approximately 338 GB	A .7z archive obtained through the original project's Baidu Netdisk download channel

The size is the approximate value displayed by Hugging Face, not the extracted size. No separate HQ-TEST archive is currently listed. This repository should not be treated as an independently verified complete mirror of both training and test data.

4. Download

Option A: Hugging Face CLI

Install or update the client:

python -m pip install -U huggingface_hub

Download the archive:

hf download flashszn/HQ-50K TRAIN_BICUBIC.7z \
  --repo-type dataset \
  --local-dir ./HQ-50K

Option B: Python

from huggingface_hub import hf_hub_download

archive_path = hf_hub_download(
    repo_id="flashszn/HQ-50K",
    repo_type="dataset",
    filename="TRAIN_BICUBIC.7z",
    local_dir="./HQ-50K",
)
print(archive_path)

Option C: Browser

Select TRAIN_BICUBIC.7z under Files and versions, or use the download link.

Download notes:

• Ensure adequate bandwidth, free space, and filesystem support for very large individual files.
• Allow additional space for extracted contents and any download cache. The extracted size has not been verified here.
• A Git LFS/Xet pointer is not the archive itself. The Hugging Face download tools above are recommended for obtaining the actual file contents.
• This repository distributes a .7z archive, not a ready-to-preview sample table. Download and extract it first, then configure your data loader for the actual directory structure. Do not assume that datasets.load_dataset() can directly load it as training samples.

5. Integrity Check and Extraction

SHA-256

As of 2026-11-10, the Hugging Face file details page reports:

e359038b5eb8ab88d564074cfb3f9f2067aeaef5cb2499e7304507a5c658ab58

On Linux, compute the downloaded file's hash and compare it with the value above:

sha256sum ./HQ-50K/TRAIN_BICUBIC.7z

This value comes from Hugging Face metadata; it was not independently recomputed for this README. It can be used to check a download against the current mirrored file, but does not establish byte-for-byte identity with the Baidu Netdisk source. If the archive is updated, consult the file details for the corresponding revision.

List, Test, and Extract

After installing 7-Zip with .7z support:

# List the archive contents
7z l ./HQ-50K/TRAIN_BICUBIC.7z

# Test archive integrity; this may take a long time for a large file
7z t ./HQ-50K/TRAIN_BICUBIC.7z

# Extract while preserving the internal directory structure
7z x ./HQ-50K/TRAIN_BICUBIC.7z -o./HQ-50K/extracted

Some installations use 7zz instead of 7z; adjust the command accordingly. Inspect the file listing before configuring training or evaluation paths. This README does not assume specific extracted directory names, scale factors, or pairing rules.

6. Preprocessing Reference

To download images from their URLs or prepare training data yourself, see the original preprocessing instructions:

• Use img2dataset to download images from URL lists. The original documentation notes that URLs may contain commas and describes the required delimiter adjustment.
• Use generate_bicubic.m to generate corresponding degraded low-resolution images.
• Use extract_subimages.py to crop high-resolution images; the documented settings are 768 × 768 crops with an overlap of 192.
• Use create_lmdb.py to build LMDB databases when needed, adjusting paths and configuration for your environment.

These are references to the upstream workflow, not a claim that every step has been applied to, or is still required for, this archive. Inspect the existing data first to avoid redundant processing or overwriting files.

7. License and Attribution

The original project README states that the dataset is released under CC BY 4.0. The license: cc-by-4.0 metadata in this dataset card follows that upstream statement; mirroring the archive does not relicense the original data.

When using, modifying, or redistributing the data, comply with the applicable license, retain attribution, source and license information, and indicate changes. The original code and its dependencies are subject to their respective licenses. This mirror does not grant additional third-party rights associated with the source images.

Credit for the dataset and paper belongs to the original authors. This repository only provides a download mirror.

8. Citation

If the dataset is useful for your research, please cite the original paper:

@misc{yang2023hq50k,
  title={HQ-50K: A Large-scale, High-quality Dataset for Image Restoration},
  author={Qinhong Yang and Dongdong Chen and Zhentao Tan and Qiankun Liu and Qi Chu and Jianmin Bao and Lu Yuan and Gang Hua and Nenghai Yu},
  year={2023},
  eprint={2306.05390},
  archivePrefix={arXiv},
  primaryClass={cs.CV}
}

---

中文

本仓库提供 HQ-50K 原项目百度网盘下载文件的 Hugging Face 镜像。文件来源对应原项目 data_preprocess/README.md 中提供的百度网盘下载入口。

当前提供的文件为 TRAIN_BICUBIC.7z，压缩包大小约 338 GB。 本仓库存放的是数据压缩包，而不是仅包含图片 URL 的列表，方便通过 Hugging Face 获取文件。

本仓库是第三方下载镜像，并非原作者维护的官方发布仓库。当前文件清单不代表已经镜像 HQ-50K 的全部训练、测试及其他衍生数据；数据集介绍和处理流程请以原项目为准。

1. 数据集介绍

HQ-50K: A Large-scale, High-quality Dataset for Image Restoration 是面向图像复原研究的大规模高质量图像数据集。根据原项目介绍，它包含：

• 50,000 张训练图像，具有丰富的纹理细节和语义多样性。
• 1,250 张测试图像，覆盖不同语义类别与频率范围，用于细粒度性能评估。

上述数量描述的是原始 HQ-50K 数据集，不是对本仓库压缩包内部图片数量的统计。本数据集卡未对压缩包内部目录、图片数量、HR/LR 配对方式或缩放倍率进行独立核验，请以实际解压内容及原项目说明为准。

2. 来源与相关链接

项目	链接
原始项目与代码	littleYaang/HQ-50K
原项目数据下载与预处理说明	data_preprocess/README.md
原项目提供的 Hugging Face 数据入口	YangQiee/HQ-50K
论文	arXiv:2306.05390
本镜像仓库	flashszn/HQ-50K

原项目预处理说明中提供的百度网盘入口如下：

数据划分	原始下载地址	提取码
HQ-TRAIN	百度网盘	hq50
HQ-TEST	百度网盘	hq50

以上为原项目公布的来源链接，不表示本仓库同时包含这两个入口中的所有文件；链接更新请以原项目为准。

3. 仓库文件

截至 2026-11-10，仓库内容如下：

HQ-50K/
├── .gitattributes
├── README.md
└── TRAIN_BICUBIC.7z

文件	大小	说明
TRAIN_BICUBIC.7z	约 338 GB	来自原项目百度网盘下载渠道的 .7z 压缩包

大小为 Hugging Face 页面显示的近似值，不是解压后大小。当前清单中没有单独的 HQ-TEST 压缩包；不要将本仓库视为已独立核验的完整训练集与测试集镜像。

4. 下载方式

方式一：Hugging Face CLI

安装或更新客户端：

python -m pip install -U huggingface_hub

下载压缩包：

hf download flashszn/HQ-50K TRAIN_BICUBIC.7z \
  --repo-type dataset \
  --local-dir ./HQ-50K

方式二：Python

from huggingface_hub import hf_hub_download

archive_path = hf_hub_download(
    repo_id="flashszn/HQ-50K",
    repo_type="dataset",
    filename="TRAIN_BICUBIC.7z",
    local_dir="./HQ-50K",
)
print(archive_path)

方式三：网页下载

在 Files and versions 中选择 TRAIN_BICUBIC.7z，或使用文件下载链接。

下载提示：

• 请确认网络、可用空间和文件系统支持超大单文件。
• 除压缩包本身外，还需要为解压结果及可能的下载缓存预留空间；解压后占用尚未核验。
• 不要将 Git 中的 LFS/Xet 指针文件当作实际压缩包。推荐使用上述 Hugging Face 下载工具获取文件内容。
• 本仓库目前以 .7z 压缩包提供数据，不是可直接预览的标准样本表；应先下载并解压，再根据实际目录配置数据加载逻辑。不要假定 datasets.load_dataset() 能直接加载为训练样本。

5. 完整性检查与解压

SHA-256 校验

截至 2026-11-10，Hugging Face 文件详情页 公布的 SHA-256 为：

e359038b5eb8ab88d564074cfb3f9f2067aeaef5cb2499e7304507a5c658ab58

Linux 下可计算下载文件的哈希并与上方值比较：

sha256sum ./HQ-50K/TRAIN_BICUBIC.7z

该值来自 Hugging Face 文件元数据，并非本 README 对完整文件重新计算所得。它可用于核对下载文件与当前镜像版本的一致性，不等同于已完成与百度网盘源文件的逐字节一致性验证。如果仓库文件更新，请以对应版本的文件详情为准。

查看、测试与解压

安装支持 .7z 格式的 7-Zip 后，可执行：

# 查看压缩包内部文件清单
7z l ./HQ-50K/TRAIN_BICUBIC.7z

# 测试压缩包完整性；大文件测试可能耗时较长
7z t ./HQ-50K/TRAIN_BICUBIC.7z

# 解压并保留内部目录结构
7z x ./HQ-50K/TRAIN_BICUBIC.7z -o./HQ-50K/extracted

部分系统的命令名称为 7zz，请按实际安装情况替换。请先查看文件清单，再根据实际目录配置训练或评估程序；本文不预设解压后的目录名称、缩放倍率或配对规则。

6. 数据预处理参考

如需按照原项目自行下载图片或生成训练数据，请参考其数据预处理说明：

• 使用 img2dataset 从图片 URL 列表下载数据；原项目特别说明 URL 可能含逗号，需按其说明调整读取分隔符。
• 使用 generate_bicubic.m 生成对应的低分辨率退化图像。
• 使用 extract_subimages.py 裁剪高分辨率图像；原文给出的裁剪设置为 768 × 768、重叠 192。
• 按需要使用 create_lmdb.py 构建 LMDB 数据库，并修改相应路径与配置。

上述内容是原项目的处理流程参考，不代表本压缩包已执行或仍需执行其中的每一步。使用本镜像时，请先检查已有数据，避免重复处理或覆盖文件。

7. 许可证与署名

原项目 README 声明，数据集采用 CC BY 4.0 许可证。本数据集卡的 license: cc-by-4.0 元数据依据这一上游声明填写；镜像分发不构成对原始数据的重新授权。

使用、修改或再分发数据时，请遵守适用许可，保留原作者署名、来源和许可信息，并注明所作修改。原项目代码及其依赖的许可证应分别遵循各自说明；本仓库不额外授予原始图像涉及的其他第三方权利。

数据集与论文贡献归原作者所有，本仓库仅提供便于下载的镜像。

8. 引用

如果该数据集对你的研究有帮助，请引用原论文：

@misc{yang2023hq50k,
  title={HQ-50K: A Large-scale, High-quality Dataset for Image Restoration},
  author={Qinhong Yang and Dongdong Chen and Zhentao Tan and Qiankun Liu and Qi Chu and Jianmin Bao and Lu Yuan and Gang Hua and Nenghai Yu},
  year={2023},
  eprint={2306.05390},
  archivePrefix={arXiv},
  primaryClass={cs.CV}
}