MuRI-Restore is a multi-source cultural heritage image restoration dataset covering museum objects, traditional New Year paintings, silk and brocade patterns, digital museum resources, overseas collections, and other cultural images. Each sample provides one high-resolution target together with a consistent set of restoration inputs for super-resolution, denoising, JPEG artifact removal, and inpainting.

The dataset is intended for supervised restoration, multi-task restoration, degradation-aware training, cross-collection evaluation, and cultural heritage image analysis. All image columns use the Hugging Face Image feature and are decoded as PIL images by datasets.

Dataset Summary

Property	Value
Dataset name	MuRI-Restore
Repository	`nomodeset/muri-hf-parquet`
Domain	Cultural heritage image restoration
Samples	5,168
Source collections	29
Default configuration	`all`
Collection configurations	29 named configurations
Split	`train`
Columns per sample	42
Metadata columns	15
Image columns	27
High-resolution targets	1 per sample
Degraded restoration inputs	22 per sample
Inpainting masks	4 per sample
Parquet size	253.1 GiB
License	CC BY-NC-SA 4.0
Access	Manually approved gated access

Tasks

Each degraded input is paired with the same hr_image target. The dataset contains all task variants listed below.

Task	Variants and parameters
Natural super-resolution	Scale factors `x2`, `x3`, `x4`, and `x8`; natural degradation includes blur, downsampling, noise, and JPEG artifacts
Bicubic super-resolution	Scale factors `x2`, `x3`, `x4`, and `x8`; uses bicubic downsampling
Natural denoising	Noise levels `sigma=15`, `sigma=25`, and `sigma=50`; natural degradation includes shot noise, read noise, channel gain variation, blur, and JPEG artifacts
Gaussian denoising	Noise levels `sigma=15`, `sigma=25`, and `sigma=50`; additive independent Gaussian noise
JPEG artifact removal	JPEG quality levels `Q=10`, `Q=20`, `Q=30`, and `Q=40`
Inpainting	Random brush mask, structure-line mask, pattern-defect mask, and manual/semi-automatic damage mask

JPEG variants are identified by quality factor Q, not noise standard deviation sigma.

Image Pairing

Element	Description
Target	`hr_image`, a center-cropped RGB image whose width and height are multiples of 24
Super-resolution input	A lower-resolution image paired with `hr_image`; dimensions are reduced by the named scale factor
Denoising input	A noisy image with the same dimensions as `hr_image`
JPEG input	A JPEG-compressed image with the same dimensions as `hr_image`
Inpainting input	An image with selected regions replaced by a light gray fill
Inpainting mask	A single-channel mask in which white pixels identify regions to restore

Task Columns

Super-Resolution

Column	Contents
`sr_bicubic_x2_lr`	Bicubic low-resolution input at `x2`
`sr_bicubic_x3_lr`	Bicubic low-resolution input at `x3`
`sr_bicubic_x4_lr`	Bicubic low-resolution input at `x4`
`sr_bicubic_x8_lr`	Bicubic low-resolution input at `x8`
`sr_natural_x2_lr`	Natural-degradation low-resolution input at `x2`
`sr_natural_x3_lr`	Natural-degradation low-resolution input at `x3`
`sr_natural_x4_lr`	Natural-degradation low-resolution input at `x4`
`sr_natural_x8_lr`	Natural-degradation low-resolution input at `x8`

Denoising

Column	Contents
`denoise_gaussian_sigma15_noisy`	Gaussian-noise input at `sigma=15`
`denoise_gaussian_sigma25_noisy`	Gaussian-noise input at `sigma=25`
`denoise_gaussian_sigma50_noisy`	Gaussian-noise input at `sigma=50`
`denoise_natural_sigma15_noisy`	Natural-noise input at nominal `sigma=15`
`denoise_natural_sigma25_noisy`	Natural-noise input at nominal `sigma=25`
`denoise_natural_sigma50_noisy`	Natural-noise input at nominal `sigma=50`

JPEG Artifact Removal

Column	Contents
`jpeg_q10_compressed`	JPEG-compressed input at quality `Q=10`
`jpeg_q20_compressed`	JPEG-compressed input at quality `Q=20`
`jpeg_q30_compressed`	JPEG-compressed input at quality `Q=30`
`jpeg_q40_compressed`	JPEG-compressed input at quality `Q=40`

Inpainting

Mask family	Masked input column	Mask column
Random brush	`inpainting_random_brush_mask_masked`	`inpainting_random_brush_mask_mask`
Structure line	`inpainting_structure_line_mask_masked`	`inpainting_structure_line_mask_mask`
Pattern defect	`inpainting_pattern_defect_mask_masked`	`inpainting_pattern_defect_mask_mask`
Manual/semi-automatic damage	`inpainting_manual_semiauto_mask_masked`	`inpainting_manual_semiauto_mask_mask`

Data Fields

Identification and Metadata

Field	Type and description
`sample_id`	`string`; stable 16-character identifier
`subset`	`string`; collection configuration associated with the sample
`image_name`	`string`; human-readable image or object name
`source_path`	`string`; legacy provenance locator retained for traceability; it is not a portable URL
`relative_path`	`string`; collection-relative provenance locator
`sample_dir`	`string`; legacy sample grouping locator
`image_file_name`	`string`; original image file name
`image_extension`	`string`; original image extension
`source_width`	`int32`; original image width in pixels
`source_height`	`int32`; original image height in pixels
`hr_width`	`int32`; high-resolution target width in pixels
`hr_height`	`int32`; high-resolution target height in pixels
`hr_format`	`string`; encoding format of the high-resolution target
`info_json`	`string`; JSON-serialized cultural-object metadata, source references, labels, links, and other available annotations
`task_spec_json`	`string`; JSON-serialized target mapping, crop rule, image dimensions, task parameters, and inpainting mask semantics

Image Features

Field group	Type and description
`hr_image`	Hugging Face `Image`; RGB high-resolution target
`sr_bicubic_*`	Hugging Face `Image`; four bicubic super-resolution inputs
`sr_natural_*`	Hugging Face `Image`; four natural-degradation super-resolution inputs
`denoise_gaussian_*`	Hugging Face `Image`; three Gaussian denoising inputs
`denoise_natural_*`	Hugging Face `Image`; three natural denoising inputs
`jpeg_*`	Hugging Face `Image`; four JPEG artifact removal inputs
`inpainting_*_masked`	Hugging Face `Image`; four masked RGB inputs
`inpainting_*_mask`	Hugging Face `Image`; four single-channel restoration masks

Image Statistics

Property	Value
Original JPEG images	5,091
Original PNG images	77
HR target encoding	PNG for all 5,168 samples
Original width range	143 to 3,000 pixels
Original height range	133 to 3,087 pixels
HR width range	120 to 3,000 pixels
HR height range	120 to 3,072 pixels
Rows with `info_json` metadata	5,168

Configurations

The default all configuration combines every collection into one train split. A named configuration can be selected to load only one collection. Configuration names are preserved in their original language to maintain source identity.

Configuration	Samples
`all`	5,168
`上海博物馆_202003`	46
`中国丝绸博物馆_202003`	309
`中国南京`	63
`中国印刷博物馆_20200324`	104
`中国国家博物馆_202003`	327
`中国数字博物馆`	52
`中国美术馆_202003`	283
`中国记忆_202002`	255
`亚洲艺术品售卖网站_202003`	34
`南大_云锦_202001`	55
`博奥_云锦_202001`	100
`平山郁夫丝绸之路美术馆_三彩凤首瓶.jpg`	1
`微信图片`	15
`早稲田大学図書館_202004`	61
`正倉院_202004`	94
`清华大学艺术博物馆_202003`	104
`起承_凤翔年画_202003`	172
`起承_杨柳青年年画1_202003`	300
`起承_杨柳青年画10_202003`	111
`起承_杨柳青年画2_202003`	300
`起承_杨柳青年画3_202003`	300
`起承_杨柳青年画4_202003`	300
`起承_杨柳青年画5_202003`	300
`起承_杨柳青年画6_202003`	300
`起承_杨柳青年画7_202003`	300
`起承_杨柳青年画8_202003`	300
`起承_杨柳青年画9_202003`	300
`起承_桃花坞年画_202003`	21
`首都博物馆_202003`	261

Loading the Dataset

This is a gated dataset. Request access on the repository page and authenticate with hf auth login before loading it.

Load a single collection:

from datasets import load_dataset

dataset = load_dataset(
    "nomodeset/muri-hf-parquet",
    "上海博物馆_202003",
    split="train",
)

row = dataset[0]
target = row["hr_image"]
natural_x4 = row["sr_natural_x4_lr"]
gaussian_noisy = row["denoise_gaussian_sigma25_noisy"]
jpeg_input = row["jpeg_q20_compressed"]
masked_input = row["inpainting_pattern_defect_mask_masked"]
mask = row["inpainting_pattern_defect_mask_mask"]

print(row["sample_id"])
print(target.size, natural_x4.size)

Stream the combined configuration without downloading the full dataset first:

from datasets import load_dataset

dataset = load_dataset(
    "nomodeset/muri-hf-parquet",
    "all",
    split="train",
    streaming=True,
)

row = next(iter(dataset))
print(row["subset"], row["hr_image"].size)

Parse the JSON metadata:

import json

object_metadata = json.loads(row["info_json"])
task_metadata = json.loads(row["task_spec_json"])

print(object_metadata.keys())
print(task_metadata["hr_size"])

Recommended Uses

Use	Notes
Single-task restoration	Train or evaluate one degradation family against `hr_image`
Multi-task restoration	Combine task columns while retaining a shared target
All-in-one restoration	Sample across super-resolution, denoising, JPEG, and inpainting inputs
Cross-collection evaluation	Train on selected configurations and evaluate on held-out collections
Metadata-aware restoration	Use parsed `info_json` fields as optional conditioning or analysis metadata
Cultural heritage analysis	Study image dimensions, visual domains, patterns, and collection diversity

Usage Notes

Topic	Details
Dataset split	The repository provides one `train` split
Collection sizes	Configuration sizes range from 1 to 327 samples
Source diversity	Resolution, acquisition conditions, object type, and metadata structure vary across collections
Inpainting masks	Four mask families define distinct restoration regions
Provenance fields	Locator fields preserve sample and collection provenance
Storage	The complete `all` configuration requires substantial disk space and bandwidth

License and Access

Dataset files are available to approved users under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license (CC BY-NC-SA 4.0). Access requires manual approval through the Hugging Face gated-dataset request workflow.

Requirement	Terms
Permitted purpose	Non-commercial use consistent with the approved access request
Attribution	Credit the dataset and its repository
Adaptations	Distribute permitted adaptations under the same or a compatible ShareAlike license
Redistribution	Do not mirror or publicly redistribute the raw dataset files without separate written authorization
Access control	Repository visibility does not itself grant permission to download or use dataset files
Revocation	Access may be revoked when the license, access terms, or approved purpose is violated

Users are responsible for reviewing the metadata and source information associated with individual samples and for ensuring that their use complies with applicable laws, institutional policies, and the stated license.

Citation

A formal paper citation is not yet provided. Until one is available, cite the Hugging Face dataset repository:

@dataset{muri_restore,
  title        = {MuRI-Restore},
  author       = {MuRI-Restore Dataset Authors},
  year         = {2026},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/nomodeset/muri-hf-parquet}
}

Downloads last month: 5

Total file size:

272 GB