DISRQAD: Diffusion Image Super-Resolution Quality Assessment Dataset and Benchmark

Nikita Kukuzei; Artem Borisov; Evgeney Bogatyrev; Khaled Abud; Egor Chistov; Dmitriy Vatolin

cs.CVOct 6, 2026

DISRQAD: Diffusion Image Super-Resolution Quality Assessment Dataset and Benchmark

Authors: Nikita Kukuzei, Artem Borisov, Evgeney Bogatyrev, Khaled Abud, Egor Chistov, Dmitriy Vatolin

Organizations: Lomonosov Moscow State University Moscow, Russia · MSU Institute for AI, Lomonosov Moscow State University Moscow, Russia

Abstract

Diffusion-based image super-resolution (SR) can create visually plausible detail that is not supported by the low-resolution input. We introduce DISRQAD, a subjective-quality dataset and diagnostic benchmark for this setting. It contains mean opinion scores (MOS) for 14,000 SR outputs from ten diffusion and four non-diffusion methods, spanning four low-resolution degradation conditions and x2/x4 upscaling. We evaluate 51 standard full-reference and no-reference metric configurations and 11 adapted variants. Agreement with MOS is substantially weaker on diffusion outputs: the strongest standard no-reference baseline reaches 0.431 SRCC on diffusion SR versus 0.813 on non-diffusion SR. As a case study in benchmark use, a pruned and distilled Q-ReAlign-mini student reaches 0.496 SRCC on diffusion SR. DISRQAD measures perceived output quality, not faithfulness to the input; it enables analysis of metric behavior across generator families and input conditions. Our findings reveal a substantial gap in the assessment of diffusion-based SR and provide a basis for developing quality models sensitive to diffusion-specific artifacts.

Figures & tables

Figure 1: Comparison of diffusion-based super-resolution methods. The number in parentheses below each super-resolved image denotes its mean opinion score (MOS). (a) The super-resolved image exhibits significant artifacts. (b) The super-resolved image is perceptually comparable to the original image. (c) The super-resolved image is perceived as having higher visual quality than the original image.

Dataset	Sources	SR outputs	Non-diff. SR	Diffusion SR	Human input
CVIU-2017 ( Ma et al., 2017 )	30	1,620	9	0	Ratings
QADS ( Zhou et al., 2019 )	20	980	21	0	Pairs
SISRSet ( Shi et al., 2019 )	15	360	8	0	Pairs
RealSRQ ( Jiang et al., 2022 )	60	1,620	10	0	Pairs
SISAR ( Zhao et al., 2022 )	100	12,600 ∗	10 ∗	0	Ratings
SRIQA-Bench ( Chen et al., 2025b )	100	1,000	5	5	Pairs

Table 1: SR datasets with per-image scores. Ratings are absolute judgments; pairs are preferences. Sources are distinct LR inputs for SISAR and HR inputs otherwise.

Figure 2: DISRQAD construction: source selection, LR condition generation, SR outputs, and subjective subset selection.

Ratings per image	SROCC	95% confidence interval
		Lower bound	Upper bound
10	0.982547	0.969290	0.990110
11	0.978173	0.961658	0.987619
12	0.973870	0.954175	0.985165
13	0.983084	0.970229	0.990415
14	0.983596	0.971124	0.990706

Table 2: SROCC between MOS computed from different numbers of ratings and the 20-rating baseline, with 95% confidence intervals.

		Non-diffusion SR		Diffusion SR
Metric	Type	SRCC	PLCC	SRCC	PLCC
TOPIQ(FR) ( Chen et al., 2024 )	FR	$\phantom{-}\mathbf{0.811}$	$\phantom{-}\mathbf{0.789}$	$\phantom{-}\mathbf{0.397}$	$\phantom{-}\mathbf{0.417}$
AFINE(FR) ( Chen et al., 2025b )	FR	$\phantom{-}\underline{0.787}$	$\phantom{-}\underline{0.738}$	$\phantom{-}0.310$	$\phantom{-}\underline{0.410}$
VSI ( Zhang et al., 2014 )	FR	$\phantom{-}0.765$	$\phantom{-}0.663$	$\phantom{-}0.355$	$\phantom{-}0.366$
GMSD ( Xue et al., 2014 )	FR	$\phantom{-}0.765$	$\phantom{-}0.719$	$\phantom{-}0.317$	$\phantom{-}0.325$
LPIPS-VGG ( Zhang et al., 2018 )	FR	$\phantom{-}0.763$	$\phantom{-}0.719$	$\phantom{-}0.300$	$\phantom{-}0.329$

Table 3: Overall signed correlations with MOS for 51 standard full-reference (FR) and no-reference (NR) IQA configurations on non-diffusion and diffusion SR. Rows are ordered by non-diffusion SRCC within each type. Best and second-best results within each type and column are shown in bold and underlined, respectively. Values are rounded to three decimal places.

Diffusion SR
		SRCC with MOS
Factor	Subset	PSNR (FR)	TOPIQ-FR	Q-Align (NR)	Q-ReAlign-mini (distilled, NR)
Overall	All outputs	$0.257$	$0.397$	$0.415$	$0.496$
SR method	ResShift	$\mathbf{0.412}$	$\mathbf{0.587}$	$0.571$	$0.613$
	SinSR	$0.342$	$0.506$	$0.428$	$0.515$
	SeeSR	$0.301$	$0.367$	$0.301$	$0.444$

Table 4: Within-subset Spearman correlation (SRCC) with MOS across SR generators, input degradations, and upscaling factors. PSNR and TOPIQ-FR use a reference image (FR); Q-Align and distilled Q-ReAlign-mini do not (NR). Each correlation is computed separately on its indicated subset. In the SR method rows, bold and underlined values mark the highest and lowest SRCC within each SR family and metric, respectively; they do not indicate statistical significance.

		Non-diffusion SR		Diffusion SR
Metric	Type	SRCC	PLCC	SRCC	PLCC
LPIPS-VGG(FT) ( Zhang et al., 2018 )	FR	$\phantom{-}\mathbf{0.778}$	$\phantom{-}\mathbf{0.727}$	$\phantom{-}0.267$	$\phantom{-}0.306$
LPIPS-AlexNet(FT) ( Zhang et al., 2018 )	FR	$\phantom{-}\underline{0.752}$	$\phantom{-}\underline{0.691}$	$\phantom{-}0.294$	$\phantom{-}0.331$
AHIQ(FT) ( Lao et al., 2022 )	FR	$\phantom{-}0.720$	$\phantom{-}0.687$	$\phantom{-}\underline{0.334}$	$\phantom{-}\underline{0.388}$
TOPIQ(FR)(FT) ( Chen et al., 2024 )	FR	$\phantom{-}0.694$	$\phantom{-}0.680$	$\phantom{-}\mathbf{0.436}$	$\phantom{-}\mathbf{0.474}$
PieAPP(FT) ( Prashnani et al., 2018 )	FR	$\phantom{-}0.539$	$\phantom{-}0.443$	$\phantom{-}0.306$	$\phantom{-}0.315$

Table 5: Overall signed correlations with MOS for 10 fine-tuned baselines and one distilled student on non-diffusion and diffusion SR. Rows are ordered by non-diffusion SRCC within each type. Best and second-best results within each type and column are shown in bold and underlined, respectively. Values are rounded to three decimal places.

Figure 3: The process of Q-Align pruning and fine-tuning.

Model	Params	VRAM	FPS	SRIQA-Bench	Diffusion	Non-diff.
Q-ReAlign-mini	0.8B	2.3 GB	3.76	0.479/0.504	0.341/0.393	0.683/0.597
Q-ReAlign-lite	4.0B	10.7 GB	2.69	0.475/0.519	0.405/0.458	0.772/0.719
Q-ReAlign-pro	9.0B	19.2 GB	2.65	0.428/0.488	0.431/0.469	0.760/0.716
Q-Align	8.2B	16.0 GB	5.39	0.624/0.677	0.415/0.487	0.813/0.758
Ours	0.34B	1.3 GB	6.37	0.755/0.794	0.496/0.542	0.729/0.747

Table 6: Accuracy and efficiency of the proposed student and its vision–language baselines. Accuracy is reported as SRCC/PLCC. SRIQA-Bench results use the 110-image validation split. Throughput and peak VRAM are measured on

512\times 512

inputs using an NVIDIA RTX A6000 in BF16.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Feature	Type	Description
SI (Spatial Information)	Numeric	Characterizes the amount of detail and texture in the scene.
Blockiness Wu and Yuen (1997)	Numeric	Evaluates how pronounced the artificial boundaries between adjacent image blocks are.
SAM Kirillov et al. (2023)	Numeric	Number of objects detected by the model in the image.
Q-Align	Numeric	Expresses not only the visual quality of the image but also its aesthetic component, as it was trained on both tasks.
ARNIQA Agnolucci et al. (2024a)	Embedding	Reflects information about image quality and visual distortions.
CLIP Radford et al. (2021)	Embedding	Expresses semantic information about the image.

Appendix

Table 7: Calculated image features and their descriptions.

Panel of Figure 4 — Figure 4: Feature distributions in medoid images.

Figure 5: Examples of HR–SR image pairs filtered by the highest-standard-deviation criterion during in-lab annotation. This filtering was performed when selecting examples for the assessors’ training and examination sets.

Panel of Figure 6 — Figure 6: LM-fitted relationships between MOS and pairwise Bradley–Terry scores for SR outputs associated with one reference image: (1) no-reference MOS and (2) full-reference MOS.

Panel of Figure 7 — Figure 7: LM-fitting error between MOS and pairwise Bradley–Terry scores as a function of the minimum number of judgments per test case. The upper panel uses full-reference MOS, and the lower panel uses no-reference MOS. The panels use separate color scales.

Explore similar work

TUDSR: Twice Upsampling-Diffusion for Higher Super-Resolution
Jun 8, 2026Zhiqiang Wu, Yitong Dong, Xian WeiReal-World Image Super-Resolution Upsampling
1East China Normal University · 2Zhejiang University
How Accurate are Video Quality Models for Diffusion-Based Video Super-Resolution?
May 25, 2026Benjamin Herb, Steve Göring, Alexander Raake +1Video Quality Upsampling
Audiovisual Technology Group, Technische Universität Ilmenau, Germany · Institute for Communications Engineering, RWTH Aachen University, Germany
SR-Ground: Image Quality Grounding for Super-Resolved Content
May 20, 2026Artem Borisov, Evgeney Bogatyrev, Khaled Abud +1Image Quality Assessment
Lomonosov Moscow State University Moscow, Russia · MSU AI Center, Lomonosov Moscow State University Moscow, Russia

Jun 8, 2026, cs.CVJ/K move · Enter open · S save

TUDSR: Twice Upsampling-Diffusion for Higher Super-Resolution

Zhiqiang Wu, Yitong Dong, Xian Wei

1East China Normal University · 2Zhejiang University

Real-World Image Super-Resolution Upsampling Generative Adversarial Network

Open paper AskPDF arXiv

Diffusion-based generative models have achieved remarkable success in real-world image super-resolution (SR). With tiled diffusion techniques, these models can produce high-resolution images that exceed their native-supported resolution. However, the quality of such high-resolution (e.g

2048^2

) outputs often remains extremely poor, primarily due to two factors we consider: the image upsampling ratio (e.g

\times8

) exceeding the model's native-supported upsampling ratio (e.g

\times4

), and the model's native-supported resolution. In practice, training a native high-resolution model requires larger architectures, which incur significant computational overhead and GPU memory costs, making it hard on limited-resource equipment. Thus, we present TUDSR, a Twice Upsampling-Diffusion framework for higher SR. The TUDSR framework mainly consists of two stages: the first involves training at

R

-resolution, and the second introduces a looped chunk-based training strategy at

NR

-resolution. Each stage adapts a one-step GAN architecture comprising a generator and a discriminator. Based on SD2.1-base, we develop TUDSR-S, which achieves state-of-the-art performance across multiple benchmarks. Extensive experiments further demonstrate that TUDSR-S generates high-quality images at the resolutions of

1024^2

and even

2048^2

, significantly outperforming existing approaches. Code is available at https://github.com/wuer5/TUDSR.

Abstract

Explore similar work

TUDSR: Twice Upsampling-Diffusion for Higher Super-Resolution

How Accurate are Video Quality Models for Diffusion-Based Video Super-Resolution?

SR-Ground: Image Quality Grounding for Super-Resolved Content