visualprior/DSR_QADB

DSR Benchmark — Public Release

DSR Benchmark is a dataset for studying the quality of images produced by classic and diffusion-based Super-Resolution (SR) methods. The public release contains all selected SR images, along with the references and subjective quality scores for 60% of the SR images in each category. The data associated with the remaining 40% are reserved for a future challenge.

Dataset summary

Component	Count
HR images in the source collection	10,036
LR images: 2 scales × 4 variants	80,288
All generated SR candidates: 10,036 × 8 × 14	1,124,032
Selected and subjectively evaluated Diffusion SR images	10,000
Selected and subjectively evaluated Classic SR images	4,000
Total selected SR images (all images are published)	14,000
SR images with public MOS values and references	60% of each SR category
SR images reserved for the challenge	40% of each SR category

The split is performed at the source HR-image level. All SR images sharing the same five-digit identifier xxxxx belong to the same partition; variants of a single HR image are never divided between the public and hidden partitions. The 60% target is calculated independently for Diffusion SR and Classic SR.

HR image sources

The 10,036 HR images were selected from the following datasets:

• KonIQ-10k;
• Microsoft COCO — images from the train, validation, and test splits were used;
• AVA: A Large-Scale Database for Aesthetic Visual Analysis;
• FLIVE, introduced in From Patches to Pictures (PaQ-2-PiQ).

Each HR image was assigned a five-digit identifier. HR filenames follow this pattern:

xxxxx_vg_dsr.jpg

Example:

00014_vg_dsr.jpg

LR generation

Each HR image was downscaled using bicubic interpolation at two scale factors:

• x2;
• x4.

Four degradation variants were prepared for each scale factor:

Label	LR variant
jpeg	JPEG compression
blur	blurring
noise	Gaussian noise
resized	bicubic downscaling without an additional distortion

Therefore, each HR image has eight corresponding LR images:

$$2\text{ scales}\times 4\text{ variants}=8\text{ LR images}\mathrm{.}2\backslash \; \backslash text\{scales\}\; \backslash times\; 4\backslash \; \backslash text\{variants\}\; =\; 8\backslash \; \backslash text\{LR\; images\}.$$

LR filenames follow this pattern:

xxxxx_vg_dsr_{distortion}_{x2/x4}.jpg

Examples:

00014_vg_dsr_blur_x2.jpg
00014_vg_dsr_noise_x4.jpg
00014_vg_dsr_resized_x4.jpg

Super-Resolution methods

Ten diffusion-based and four classic SR methods were applied to each LR image.

Diffusion SR

Method	Dataset label
ResShift	ResShift
SinSR	SinSR
SeeSR	SeeSR
PASD	PASD
PASD-SDXL	SPASD
DiT4SR	DiT4SR
SUPIR	SUPIR
CODSR	CODSR
InvSR	InvSR
PassionSR	PassionSR

Note that PASD-SDXL results are labeled SPASD in the filenames.

Classic SR

Method	Dataset label
HAT	HAT
Real-ESRGAN	Real-ESRGAN
SwinIR	SwinIR
RealSR	RealSR

SR filenames follow this pattern:

{method}_{x2/x4}_xxxxx_vg_dsr_{distortion}_{x2/x4}.{extension}

Examples:

CODSR_x2_00014_vg_dsr_blur_x2.jpg
SPASD_x4_00014_vg_dsr_noise_x4.png
SwinIR_x4_02310_vg_dsr_resized_x4.png

The filename components have the following meanings:

Component	Meaning
method	the SR method used
first x2 or x4	the SR upscaling factor
xxxxx	identifier shared by the corresponding HR and LR content
distortion	the LR degradation variant
final x2 or x4	the downscaling factor of the corresponding LR image
extension	the actual file format, such as jpg or png

Subjective evaluation

The following images were selected from the 1,124,032 generated candidates:

• 10,000 Diffusion SR images;
• 4,000 Classic SR images.

The images were evaluated using a no-reference MOS methodology: observers rated the visual quality of each SR image without seeing its corresponding LR or HR image. The subjective experiment used a scale from 1 to 5, where a higher score indicates higher perceived quality. The MOS for an SR image is obtained by aggregating its individual ratings.

Public annotations cover 60% of the Diffusion SR images and 60% of the Classic SR images. For the remaining 40%, the SR images themselves are available, but their corresponding MOS values, individual votes, LR images, and HR images are not published.

Note on score representation. The CSV files contain numeric values in the form in which they were exported and used to calculate MOS. The data collection procedure used a 1–5 scale. When using these files, take into account the score encoding employed by the particular dataset version.

Public repository structure

To stay within repository limits, image files are sharded by the first three digits of their five-digit source identifier. For example, all files associated with source ID 00014 are stored in shard 000. The shard name is denoted by {prefix} below and is always equal to xxxxx[:3].

.
├── README.md
├── images/
│   ├── hr/
│   │   └── {prefix}/
│   │       └── xxxxx_vg_dsr.jpg
│   ├── lr/
│   │   ├── x2/
│   │   │   └── {prefix}/
│   │   │       └── xxxxx_vg_dsr_{distortion}_x2.jpg
│   │   └── x4/
│   │       └── {prefix}/
│   │           └── xxxxx_vg_dsr_{distortion}_x4.jpg
│   └── sr/
│       ├── diffusion/
│       │   └── {prefix}/
│       │       └── {method}_{scale}_xxxxx_vg_dsr_{distortion}_{scale}.{extension}
│       └── classic/
│           └── {prefix}/
│               └── {method}_{scale}_xxxxx_vg_dsr_{distortion}_{scale}.{extension}
└── metadata/
    ├── mos.csv
    ├── votes.csv
    ├── mos_csr.csv
    └── votes_csr.csv

The public repository intentionally does not include split_assignment.csv or sr_manifest.csv. Available SR, LR, and HR images can be matched using the five-digit xxxxx identifier in their filenames.

images/hr/{prefix}/

Contains HR references for the public partition only. HR images corresponding to the hidden 40% of SR images are not included.

images/lr/x2/{prefix}/ and images/lr/x4/{prefix}/

Contain the LR inputs for the public partition, organized by downscaling factor. LR images corresponding to the hidden 40% of SR images are not included.

images/sr/diffusion/{prefix}/

Contains all 10,000 selected Diffusion SR images, including images from both the public and challenge partitions.

images/sr/classic/{prefix}/

Contains all 4,000 selected Classic SR images, including images from both the public and challenge partitions.

Annotation files

metadata/mos_diffusion_SR.csv

Contains MOS values for public Diffusion SR images.

,img_name,mos
0,CODSR_x2_00014_vg_dsr_blur_x2.jpg,1.0

Column	Description
first unnamed column	row index added during export; it can be ignored
img_name	filename of the Diffusion SR image
mos	aggregated subjective quality score for the image

metadata/mos_classic_SR.csv

Uses the same structure as mos_diffusion_SR.csv, but contains MOS values for public Classic SR images.

metadata/votes_diffusion_SR.csv

Contains individual ratings for public Diffusion SR images.

,INPUT:image_a,OUTPUT:result,ASSIGNMENT:worker_id
0,CODSR_x2_00014_vg_dsr_blur_x2.jpg,1.0,db490266783ff3f7b7cbfe29cf9cbffa

Column	Description
first unnamed column	row index added during export; it can be ignored
INPUT:image_a	filename of the evaluated SR image
OUTPUT:result	individual numeric rating
ASSIGNMENT:worker_id	technical participant identifier stored by the annotation system

metadata/votes_classic_SR.csv

Uses the same structure as votes_diffusion_SR.csv, but contains individual ratings for public Classic SR images.

Matching SR, LR, and HR images

Files correspond to one another through the xxxxx identifier. For example, the SR image

CODSR_x2_00014_vg_dsr_blur_x2.jpg

has the following references when they are publicly available:

images/lr/x2/000/00014_vg_dsr_blur_x2.jpg
images/hr/000/00014_vg_dsr.jpg

If the corresponding LR and HR files are absent from the public repository, the SR image belongs to the challenge partition, and its MOS and votes are not published either.

Loading the annotations

from pathlib import Path

import pandas as pd

root = Path("/path/to/dataset")

diffusion_mos = pd.read_csv(
    root / "metadata/mos.csv",
    index_col=0,
)
classic_mos = pd.read_csv(
    root / "metadata/mos_csr.csv",
    index_col=0,
)

print(diffusion_mos.head())
print(classic_mos.head())

The path of an image can be reconstructed from its source identifier:

source_id = "00014"
prefix = source_id[:3]

hr_path = root / "images/hr" / prefix / f"{source_id}_vg_dsr.jpg"
lr_path = root / "images/lr/x2" / prefix / f"{source_id}_vg_dsr_blur_x2.jpg"
sr_path = (
    root
    / "images/sr/diffusion"
    / prefix
    / f"CODSR_x2_{source_id}_vg_dsr_blur_x2.jpg"
)

The source-content identifier can be extracted from a filename with a regular expression:

import re

SOURCE_ID_RE = re.compile(r"(?<!\d)(\d{5})_vg_dsr")

name = "CODSR_x2_00014_vg_dsr_blur_x2.jpg"
source_id = SOURCE_ID_RE.search(name).group(1)
print(source_id)  # 00014

Recommended use cases

• training and evaluating no-reference IQA models for SR;
• comparing the perceived quality of classic and diffusion-based SR methods;
• analyzing the effects of degradation type and scale factor;
• studying the generalization of quality metrics across SR method families;
• preparing for the challenge with hidden references and subjective scores.

Limitations

• MOS reflects the subjective perception of the participants and is not an objective measure of pixel-level fidelity or semantic accuracy.
• Diffusion-based methods may generate plausible details that were not present in the LR or HR images.
• The content distribution inherits the characteristics and potential biases of KonIQ-10k, COCO, AVA, and FLIVE.
• The public and challenge partitions are grouped by HR content; therefore, exactly 60% of the HR groups need not be public.
• Missing LR, HR, or MOS data in the public release are part of the challenge protocol and do not indicate dataset corruption.

Terms of use and attribution

The source images originate from KonIQ-10k, COCO, AVA, and FLIVE. The terms and restrictions of the respective source datasets continue to apply. This repository does not grant additional rights to the source images and does not replace the licensing and attribution requirements established by their authors.

When using this dataset, users should also consider the licenses and terms of the SR method implementations used to produce the derived images.