arxiv: 2603.29773 · v2 · submitted 2026-03-31 · 💻 cs.CV

Recognition: 1 theorem link

· Lean Theorem

Beyond Ground-Truth: Leveraging Image Quality Priors for Real-World Image Restoration

Chengyu Fang, Chunming He, Fengyang Xiao, Guanyi Qin, Lei Xu, Peng Hu, Rihan Zhang, Sina Farsiu, XingE Guo, Yuqi Shen

Pith reviewed 2026-05-14 00:02 UTC · model grok-4.3

classification 💻 cs.CV

keywords real-world image restorationno-reference image quality assessmentimage quality priordual-branch codebookquality-conditioned transformerdiscrete optimizationperceptual quality

0 comments

The pith

Image quality priors from no-reference models can steer restoration past ground-truth averages.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Real-world image restoration models usually train against ground-truth images, yet those references often carry uneven perceptual quality that leads networks to converge on mediocre averages instead of the best possible outputs. The paper claims an Image Quality Prior pulled from pre-trained no-reference image quality assessment models can be injected as an explicit conditioning signal to push results toward higher perceptual fidelity. This prior is fused inside a quality-conditioned Transformer with a dual-branch codebook that splits generic structure from quality-sensitive details, followed by discrete optimization to limit artifacts. A reader should care because the method promises visibly cleaner restorations on uncontrolled photos without needing new ground-truth data or retraining entire pipelines.

Core claim

The IQPIR framework extracts an Image Quality Prior from pre-trained NR-IQA models and feeds it as a conditioning signal into a quality-conditioned Transformer, while a dual-branch codebook disentangles common structural features from high-quality attributes and a discrete representation optimization step prevents over-optimization artifacts, allowing the restored image to exceed the average perceptual quality of the ground-truth training set.

What carries the argument

Image Quality Prior (IQP) drawn from NR-IQA scores, used as a conditioning input to a Transformer inside a dual-branch codebook architecture with discrete optimization.

If this is right

Existing restoration networks gain a plug-and-play quality boost by adding the conditioning branch without any structural redesign.
Restored images can surpass the average perceptual fidelity present in the original training ground-truth.
The dual-branch codebook keeps structural information intact while separately optimizing quality-sensitive attributes.
Discrete optimization in the codebook avoids the artifact patterns typical of continuous latent-space training.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same quality-prior conditioning could be tested on video restoration where frame-to-frame ground-truth is especially noisy.
Combining the prior with semantic or scene-type signals might help in cases where quality and content are tightly coupled.
Running the method on multiple independent NR-IQA models as an ensemble could reveal how sensitive the gains are to the choice of prior source.

Load-bearing premise

Pre-trained no-reference image quality assessment models supply reliable signals that reliably steer restoration toward maximal perceptual quality without adding new artifacts.

What would settle it

On a held-out set of real degraded images, if the IQP-guided outputs receive lower perceptual quality scores than a plain ground-truth-supervised baseline under the same architecture, the central claim is false.

Figures

Figures reproduced from arXiv: 2603.29773 by Chengyu Fang, Chunming He, Fengyang Xiao, Guanyi Qin, Lei Xu, Peng Hu, Rihan Zhang, Sina Farsiu, XingE Guo, Yuqi Shen.

**Figure 1.** Figure 1: Score distributions given by different Image Quality Assessment (IQA) models across various ground truth datasets, where [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗

**Figure 3.** Figure 3: Results on blind face restoration. DifFace+ is the Dif [PITH_FULL_IMAGE:figures/full_fig_p002_3.png] view at source ↗

**Figure 2.** Figure 2: Structural comparison of different training paradigms, [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

**Figure 4.** Figure 4: Overall framework of IQPIR. (a) In the codebook learning stage, a dual-codebook architecture is proposed. The HQ+ codebook is learned to quantize Zh only when the quality score of xh is higher than the threshold Sthr. (b) In the codebook lookup stage, we input the quality score S as a condition into Transformer T, which predicts two code sequences at the same time. The two codebooks are leveraged to look u… view at source ↗

**Figure 5.** Figure 5: Visualizations on low light, underwater and backlit image restoration. [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗

**Figure 6.** Figure 6: Visualization of breakdown ablation, where (a), (b), (c), and [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗

**Figure 8.** Figure 8: Generalization of our IQPIR. LQ Score=1 Score=3 Score=5 Score=7 Score=9 [PITH_FULL_IMAGE:figures/full_fig_p008_8.png] view at source ↗

read the original abstract

Real-world image restoration aims to restore high-quality (HQ) images from degraded low-quality (LQ) inputs captured under uncontrolled conditions. Existing methods typically depend on ground-truth (GT) supervision, assuming that GT provides perfect reference quality. However, GT can still contain images with inconsistent perceptual fidelity, causing models to converge to the average quality level of the training data rather than achieving the highest perceptual quality attainable. To address these problems, we propose a novel framework, termed IQPIR, that introduces an Image Quality Prior (IQP)-extracted from pre-trained No-Reference Image Quality Assessment (NR-IQA) models-to guide the restoration process toward perceptually optimal outputs explicitly. Our approach synergistically integrates IQP with a learned codebook prior through three key mechanisms: (1) a quality-conditioned Transformer, where NR-IQA-derived scores serve as conditioning signals to steer the predicted representation toward maximal perceptual quality. This design provides a plug-and-play enhancement compatible with existing restoration architectures without structural modification; and (2) a dual-branch codebook structure, which disentangles common and HQ-specific features, ensuring a comprehensive representation of both generic structural information and quality-sensitive attributes; and (3) a discrete representation-based quality optimization strategy, which mitigates over-optimization effects commonly observed in continuous latent spaces. Extensive experiments on real-world image restoration demonstrate that our method not only surpasses cutting-edge methods but also serves as a generalizable quality-guided enhancement strategy for existing methods. The code is available.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

IQPIR conditions restoration on NR-IQA priors plus dual codebooks to move past average GT quality, but the gains depend on whether those priors actually match human perception on real degradations.

read the letter

The core move here is using scores from pre-trained NR-IQA models as explicit conditioning to steer a restoration network toward higher perceptual quality instead of settling at the mean of the ground-truth data. They add a quality-conditioned transformer, split the codebook into common and HQ-specific branches, and switch to discrete optimization to reduce over-smoothing in the latent space. The plug-and-play framing is the practical part: the idea is to drop this on top of existing architectures without rewriting them. That motivation is sound and directly tackles a limitation people see when training on real-world pairs. The architecture description is straightforward and the three mechanisms are clearly separated. The soft spots sit in the assumptions. NR-IQA models are trained on their own distortion distributions, so it is not obvious they will give reliable steering signals for the camera-specific degradations in the target low-quality inputs. If the scores correlate poorly with human judgment, the conditioning could add noise or artifacts rather than improvement. The dual-branch and discrete steps are plausible on paper, but without seeing the ablations it is hard to know whether they actually isolate quality features or just trade one set of artifacts for another. The abstract claims clear superiority and generalizability, yet the provided text gives no numbers, datasets, or controls, so the central claim stays unverified for now. This is for people already working on real-world restoration pipelines who want to test quality-guided extensions. It deserves a serious referee because the problem is real and the proposed fix is concrete, even if the experiments will need close scrutiny on metrics and human studies.

Referee Report

3 major / 2 minor

Summary. The manuscript proposes IQPIR, a framework for real-world image restoration that extracts an Image Quality Prior (IQP) from pre-trained No-Reference Image Quality Assessment (NR-IQA) models to guide restoration beyond the average perceptual quality of ground-truth (GT) data. It integrates this prior via a quality-conditioned Transformer, a dual-branch codebook that disentangles generic and HQ-specific features, and a discrete representation-based optimization strategy to avoid over-optimization artifacts in continuous spaces. The approach is presented as a plug-and-play enhancement for existing architectures, with claims of outperforming state-of-the-art methods on real-world datasets.

Significance. If the empirical claims hold, the work could meaningfully shift restoration paradigms away from pure GT supervision toward explicit quality-prior guidance, offering a generalizable module that improves perceptual outcomes without architectural overhaul. The dual-branch codebook and discrete optimization ideas address known issues in latent-space restoration and could influence downstream tasks like enhancement and generation.

major comments (3)

[Abstract, §4] Abstract and §4 (Experiments): The central claim that the method 'surpasses cutting-edge methods' is unsupported by any reported quantitative metrics, PSNR/SSIM/LPIPS values, ablation tables, error bars, or dataset statistics in the abstract; without these, the superiority assertion cannot be evaluated and the plug-and-play enhancement claim remains unverified.
[§3.2] §3.2 (Quality-conditioned Transformer): The premise that NR-IQA scores from pre-trained models reliably steer outputs to maximal perceptual quality on real camera degradations is load-bearing but untested; no correlation analysis with human preference or alternative metrics (e.g., BRISQUE vs. MUSIQ) is shown to confirm the conditioning signal is faithful rather than mismatched to the target distribution.
[§3.3] §3.3 (Dual-branch codebook and discrete optimization): The claim that the discrete strategy 'mitigates over-optimization effects' lacks a concrete comparison (e.g., continuous vs. discrete latent ablation) or artifact analysis; without this, it is unclear whether the disentanglement actually preserves structural fidelity or merely trades one set of artifacts for quantization noise.

minor comments (2)

[Abstract] Abstract: The acronym 'IQP' is introduced without an explicit expansion on first use, and the three key mechanisms are listed but not cross-referenced to specific subsections or equations.
[§2] §2 (Related Work): The discussion of prior NR-IQA-guided restoration methods omits recent works on perceptual quality metrics in diffusion-based restoration, weakening the novelty positioning.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback, which has helped us identify areas to strengthen the presentation and empirical support in our manuscript. We address each major comment point-by-point below, indicating revisions where appropriate.

read point-by-point responses

Referee: [Abstract, §4] Abstract and §4 (Experiments): The central claim that the method 'surpasses cutting-edge methods' is unsupported by any reported quantitative metrics, PSNR/SSIM/LPIPS values, ablation tables, error bars, or dataset statistics in the abstract; without these, the superiority assertion cannot be evaluated and the plug-and-play enhancement claim remains unverified.

Authors: We agree that the abstract would benefit from explicit quantitative support. In the revised version, we will add key metrics (e.g., average PSNR/SSIM/LPIPS gains on RealSR and DRealSR) directly into the abstract while retaining the high-level summary. Section 4 already contains the full tables, ablations, error bars, and dataset details; we will add explicit cross-references from the abstract to these results to verify both the superiority and plug-and-play claims. revision: yes
Referee: [§3.2] §3.2 (Quality-conditioned Transformer): The premise that NR-IQA scores from pre-trained models reliably steer outputs to maximal perceptual quality on real camera degradations is load-bearing but untested; no correlation analysis with human preference or alternative metrics (e.g., BRISQUE vs. MUSIQ) is shown to confirm the conditioning signal is faithful rather than mismatched to the target distribution.

Authors: This concern is well-taken. While the main experiments show perceptual gains from the MUSIQ-based conditioning, we did not previously report direct validation of the signal. In the revision we will add a correlation analysis (Pearson/Spearman) between the MUSIQ scores and human preference ratings on a subset of real-world degraded images, plus a brief comparison to BRISQUE to justify the choice of MUSIQ. These results will be placed in Section 3.2 or a new appendix. revision: yes
Referee: [§3.3] §3.3 (Dual-branch codebook and discrete optimization): The claim that the discrete strategy 'mitigates over-optimization effects' lacks a concrete comparison (e.g., continuous vs. discrete latent ablation) or artifact analysis; without this, it is unclear whether the disentanglement actually preserves structural fidelity or merely trades one set of artifacts for quantization noise.

Authors: We acknowledge the need for a direct head-to-head comparison. In the revised manuscript we will include an explicit ablation (continuous vs. discrete latent optimization) with both quantitative metrics and side-by-side visual artifact analysis. This will demonstrate that the dual-branch codebook plus discrete strategy reduces over-optimization artifacts while maintaining structural fidelity, rather than simply introducing quantization noise. revision: yes

Circularity Check

0 steps flagged

No circularity: external NR-IQA priors and architectural choices are independent

full rationale

The paper's core mechanisms rely on pre-trained external NR-IQA models to supply Image Quality Priors as conditioning signals, a dual-branch codebook for feature disentanglement, and discrete optimization to avoid continuous-space artifacts. None of these reduce to self-definition, fitted inputs renamed as predictions, or self-citation chains; the priors are drawn from independent pre-trained networks rather than derived or fitted within the paper's own data or equations. No equations are presented that would allow a prediction to equal its input by construction, and no uniqueness theorems or ansatzes are imported via self-citation. The framework is described as a plug-and-play addition to existing architectures, with performance claims resting on experimental comparisons rather than internal logical closure.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Review based on abstract only; no explicit free parameters, axioms, or invented entities are detailed in the provided text. The approach assumes NR-IQA models transfer reliably to restoration guidance.

axioms (1)

domain assumption Pre-trained NR-IQA models provide quality scores that can reliably steer restoration networks toward maximal perceptual quality.
Central to the quality-conditioned Transformer mechanism described in the abstract.

pith-pipeline@v0.9.0 · 5597 in / 1172 out tokens · 49435 ms · 2026-05-14T00:02:24.375642+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear
a quality-conditioned Transformer, where NR-IQA-derived scores serve as conditioning signals... dual-branch codebook structure... discrete representation-based quality optimization strategy

Reference graph

Works this paper leans on

68 extracted references · 68 canonical work pages

[1]

Arniqa: Learning distortion mani- fold for image quality assessment

Lorenzo Agnolucci, Leonardo Galteri, Marco Bertini, and Alberto Del Bimbo. Arniqa: Learning distortion mani- fold for image quality assessment. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 189–198, 2024. 6

work page 2024
[2]

Progressive semantic- aware style transformation for blind face restoration

Chaofeng Chen, Xiaoming Li, Lingbo Yang, Xianhui Lin, Lei Zhang, and Kwan-Yee K Wong. Progressive semantic- aware style transformation for blind face restoration. InPro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11896–11905, 2021. 3

work page 2021
[3]

Topiq: A top-down approach from semantics to distortions for image quality assessment.IEEE Transactions on Image Processing,

Chaofeng Chen, Jiadi Mo, Jingwen Hou, Haoning Wu, Liang Liao, Wenxiu Sun, Qiong Yan, and Weisi Lin. Topiq: A top-down approach from semantics to distortions for image quality assessment.IEEE Transactions on Image Processing,

work page
[4]

Pugan: Physi- cal model-guided underwater image enhancement using gan with dual-discriminators.IEEE Transactions on Image Pro- cessing, 2023

Runmin Cong, Wenyu Yang, Wei Zhang, Chongyi Li, Chun- Le Guo, Qingming Huang, and Sam Kwong. Pugan: Physi- cal model-guided underwater image enhancement using gan with dual-discriminators.IEEE Transactions on Image Pro- cessing, 2023. 6

work page 2023
[5]

Pcgan: A noise robust conditional generative ad- versarial network for one shot learning.IEEE transactions on intelligent transportation systems, 23(12):25249–25258,

Lizhen Deng, Chunming He, Guoxia Xu, Hu Zhu, and Hao Wang. Pcgan: A noise robust conditional generative ad- versarial network for one shot learning.IEEE transactions on intelligent transportation systems, 23(12):25249–25258,

work page
[6]

Taming transformers for high-resolution image synthesis

Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. InPro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873–12883, 2021. 3

work page 2021
[7]

Real-world image dehazing with coherence-based label gen- erator and cooperative unfolding network.NeurIPS, pages arXiv–2406, 2024

Chengyu Fang, Chunming He, Fengyang Xiao, Yulun Zhang, Longxiang Tang, Yuelin Zhang, Kai Li, and Xiu Li. Real-world image dehazing with coherence-based label gen- erator and cooperative unfolding network.NeurIPS, pages arXiv–2406, 2024. 3

work page 2024
[8]

Rave: Residual vector embedding for clip-guided backlit im- age enhancement

Tatiana Gaintseva, Martin Benning, and Gregory Slabaugh. Rave: Residual vector embedding for clip-guided backlit im- age enhancement. InECCV, pages 412–428. Springer, 2024. 6

work page 2024
[9]

Scaling laws for reward model overoptimization

Leo Gao, John Schulman, and Jacob Hilton. Scaling laws for reward model overoptimization. InInternational Conference on Machine Learning, pages 10835–10866. PMLR, 2023. 4

work page 2023
[10]

Vqfr: Blind face restoration with vector-quantized dictionary and parallel de- coder

Yuchao Gu, Xintao Wang, Liangbin Xie, Chao Dong, Gen Li, Ying Shan, and Ming-Ming Cheng. Vqfr: Blind face restoration with vector-quantized dictionary and parallel de- coder. InEuropean Conference on Computer Vision, pages 126–143. Springer, 2022. 1, 3

work page 2022
[11]

Underwater ranker: Learn which is better and how to be better

Chunle Guo, Ruiqi Wu, Xin Jin, Linghao Han, Weidong Zhang, Zhi Chai, and Chongyi Li. Underwater ranker: Learn which is better and how to be better. InProceedings of the AAAI conference on artificial intelligence, pages 702–709,

work page
[12]

Mambair: A simple baseline for im- age restoration with state-space model

Hang Guo, Jinmin Li, Tao Dai, Zhihao Ouyang, Xudong Ren, and Shu-Tao Xia. Mambair: A simple baseline for im- age restoration with state-space model. InECCV, 2024. 6, 7

work page 2024
[13]

Hqg- net: Unpaired medical image enhancement with high-quality guidance.TNNLS, 2023

Chunming He, Kai Li, Guoxia Xu, Jiangpeng Yan, Longx- iang Tang, Yulun Zhang, Yaowei Wang, and Xiu Li. Hqg- net: Unpaired medical image enhancement with high-quality guidance.TNNLS, 2023. 3

work page 2023
[14]

Degradation- resistant unfolding network for heterogeneous image fusion

Chunming He, Kai Li, and Yulun Zhang. Degradation- resistant unfolding network for heterogeneous image fusion. InICCV, pages 611–621, 2023. 3

work page 2023
[15]

Reti-diff: Illumination degradation image restoration with retinex-based latent diffusion model.ICLR, 2025

Chunming He, Chengyu Fang, Yulun Zhang, Kai Li, Longx- iang Tang, Chenyu You, Fengyang Xiao, Zhenhua Guo, and Xiu Li. Reti-diff: Illumination degradation image restoration with retinex-based latent diffusion model.ICLR, 2025. 3, 6, 7, 8

work page 2025
[16]

Seg- ment concealed object with incomplete supervision.IEEE Trans

Chunming He, Kai Li, Yachao Zhang, Ziyun Yang, Longxi- ang Tang, Yulun Zhang, Linghe Kong, and Sina Farsiu. Seg- ment concealed object with incomplete supervision.IEEE Trans. Pattern Anal. Mach. Intell., 2025. 3

work page 2025
[17]

Reversible unfold- ing network for concealed visual perception with generative refinement.arXiv preprint arXiv:2508.15027, 2025

Chunming He, Fengyang Xiao, Rihan Zhang, Chengyu Fang, Deng-Ping Fan, and Sina Farsiu. Reversible unfold- ing network for concealed visual perception with generative refinement.arXiv preprint arXiv:2508.15027, 2025. 3

work page arXiv 2025
[18]

Unfoldir: Rethinking deep unfolding network in illu- mination degradation image restoration.arXiv preprint arXiv:2505.06683, 2025

Chunming He, Rihan Zhang, Fengyang Xiao, Chengyu Fang, Longxiang Tang, Yulun Zhang, and Sina Farsiu. Unfoldir: Rethinking deep unfolding network in illu- mination degradation image restoration.arXiv preprint arXiv:2505.06683, 2025. 3

work page arXiv 2025
[19]

Run: Reversible unfolding network for concealed object segmentation.ICML, 2025

Chunming He, Rihan Zhang, Fengyang Xiao, Chenyu Fang, Longxiang Tang, Yulun Zhang, Linghe Kong, Deng-Ping Fan, Kai Li, and Sina Farsiu. Run: Reversible unfolding network for concealed object segmentation.ICML, 2025. 3

work page 2025
[20]

Gcfsr: a generative and controllable face super resolution method without facial and gan priors

Jingwen He, Wu Shi, Kai Chen, Lean Fu, and Chao Dong. Gcfsr: a generative and controllable face super resolution method without facial and gan priors. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1889–1898, 2022. 3

work page 2022
[21]

Gans trained by a two time-scale update rule converge to a local nash equilib- rium.Advances in neural information processing systems, 30, 2017

Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilib- rium.Advances in neural information processing systems, 30, 2017. 6

work page 2017
[22]

Iqpfr: An image quality prior for blind face restoration and beyond.arXiv preprint arXiv:2503.09294, 2025

Peng Hu, Chunming He, Lei Xu, Jingduo Tian, Sina Farsiu, Yulun Zhang, Pei Liu, and Xiu Li. Iqpfr: An image quality prior for blind face restoration and beyond.arXiv preprint arXiv:2503.09294, 2025. 3

work page arXiv 2025
[23]

Face restoration via plug-and-play 3d facial priors.IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(12):8910–8926, 2021

Xiaobin Hu, Wenqi Ren, Jiaolong Yang, Xiaochun Cao, David Wipf, Bjoern Menze, Xin Tong, and Hongbin Zha. Face restoration via plug-and-play 3d facial priors.IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(12):8910–8926, 2021. 3

work page 2021
[24]

Labeled faces in the wild: A database forstudying face recognition in unconstrained environments

Gary B Huang, Marwan Mattar, Tamara Berg, and Eric Learned-Miller. Labeled faces in the wild: A database forstudying face recognition in unconstrained environments. InWorkshop on faces in’Real-Life’Images: detection, align- ment, and recognition, 2008. 6

work page 2008
[25]

Global structure-aware diffusion pro- cess for low-light image enhancement

HOU Jinhui, Zhiyu Zhu, Junhui Hou, LIU Hui, Huanqiang Zeng, and Hui Yuan. Global structure-aware diffusion pro- cess for low-light image enhancement. InNeurIPS, 2023. 6

work page 2023
[26]

A style-based generator architecture for generative adversarial networks

Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. InProceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 4401–4410, 2019. 3

work page 2019
[27]

Musiq: Multi-scale image quality transformer

Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 5148–5157, 2021. 5, 6

work page 2021
[28]

An underwater image enhancement benchmark dataset and beyond.IEEE Trans- actions on Image Processing, 29:4376–4389, 2019

Chongyi Li, Chunle Guo, Wenqi Ren, Runmin Cong, Junhui Hou, Sam Kwong, and Dacheng Tao. An underwater image enhancement benchmark dataset and beyond.IEEE Trans- actions on Image Processing, 29:4376–4389, 2019. 6

work page 2019
[29]

Interlcm: Low-quality images as intermediate states of latent consistency models for effective blind face restoration.ICLR, 2025

Senmao Li, Kai Wang, Joost van de Weijer, Fahad Shahbaz Khan, Chun-Le Guo, Shiqi Yang, Yaxing Wang, Jian Yang, and Ming-Ming Cheng. Interlcm: Low-quality images as intermediate states of latent consistency models for effective blind face restoration.ICLR, 2025. 3, 5, 7

work page 2025
[30]

Learning warped guidance for blind face restoration

Xiaoming Li, Ming Liu, Yuting Ye, Wangmeng Zuo, Liang Lin, and Ruigang Yang. Learning warped guidance for blind face restoration. InProceedings of the European conference on computer vision (ECCV), pages 272–289, 2018. 3

work page 2018
[31]

Blind face restoration via deep multi-scale component dictionaries

Xiaoming Li, Chaofeng Chen, Shangchen Zhou, Xianhui Lin, Wangmeng Zuo, and Lei Zhang. Blind face restoration via deep multi-scale component dictionaries. InEuropean conference on computer vision, pages 399–415. Springer,

work page
[32]

Enhanced blind face restoration with multi-exemplar images and adaptive spatial feature fusion

Xiaoming Li, Wenyu Li, Dongwei Ren, Hongzhi Zhang, Meng Wang, and Wangmeng Zuo. Enhanced blind face restoration with multi-exemplar images and adaptive spatial feature fusion. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2706– 2715, 2020. 1, 3

work page 2020
[33]

Iterative prompt learning for unsupervised backlit image enhancement

Zhexin Liang, Chongyi Li, Shangchen Zhou, Ruicheng Feng, and Chen Change Loy. Iterative prompt learning for unsupervised backlit image enhancement. InICCV, pages 8094–8103, 2023. 6, 7

work page 2023
[34]

Diff- bir: Toward blind image restoration with generative diffusion prior

Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Bo Dai, Fanghua Yu, Yu Qiao, Wanli Ouyang, and Chao Dong. Diff- bir: Toward blind image restoration with generative diffusion prior. InEuropean Conference on Computer Vision, pages 430–448. Springer, 2024. 3

work page 2024
[35]

Retinex-inspired unrolling with cooperative prior architecture search for low-light image enhancement

Risheng Liu, Long Ma, Jiaao Zhang, Xin Fan, and Zhongx- uan Luo. Retinex-inspired unrolling with cooperative prior architecture search for low-light image enhancement. In CVPR, pages 10561–10570, 2021. 8

work page 2021
[36]

Getting to know low- light images with the exclusively dark dataset.Computer Vision and Image Understanding, 178:30–42, 2019

Yuen Peng Loh and Chee Seng Chan. Getting to know low- light images with the exclusively dark dataset.Computer Vision and Image Understanding, 178:30–42, 2019. 8

work page 2019
[37]

Deep face super-resolution with iterative collabora- tion between attentive recovery and landmark estimation

Cheng Ma, Zhenyu Jiang, Yongming Rao, Jiwen Lu, and Jie Zhou. Deep face super-resolution with iterative collabora- tion between attentive recovery and landmark estimation. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 5569–5578, 2020. 1, 3

work page 2020
[38]

Toward fast, flexible, and robust low-light image enhancement

Long Ma, Tengyu Ma, Risheng Liu, Xin Fan, and Zhongx- uan Luo. Toward fast, flexible, and robust low-light image enhancement. InCVPR, pages 5637–5646, 2022. 8

work page 2022
[39]

Waveface: Authentic face restoration with efficient frequency recovery

Yunqi Miao, Jiankang Deng, and Jungong Han. Waveface: Authentic face restoration with efficient frequency recovery. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6583–6592, 2024. 5, 7

work page 2024
[40]

No-reference image quality assessment in the spatial domain.IEEE Transactions on image processing, 21(12): 4695–4708, 2012

Anish Mittal, Anush Krishna Moorthy, and Alan Conrad Bovik. No-reference image quality assessment in the spatial domain.IEEE Transactions on image processing, 21(12): 4695–4708, 2012. 5

work page 2012
[41]

A two- step framework for constructing blind image quality indices

Anush Krishna Moorthy and Alan Conrad Bovik. A two- step framework for constructing blind image quality indices. IEEE Signal processing letters, 17(5):513–516, 2010. 6

work page 2010
[42]

Opinion-unaware blind image qual- ity assessment using multi-scale deep feature statistics.IEEE Transactions on Multimedia, 2024

Zhangkai Ni, Yue Liu, Keyan Ding, Wenhan Yang, Hanli Wang, and Shiqi Wang. Opinion-unaware blind image qual- ity assessment using multi-scale deep feature statistics.IEEE Transactions on Multimedia, 2024. 6

work page 2024
[43]

Human-visual- system-inspired underwater image quality measures.IEEE Journal of Oceanic Engineering, 41(3):541–551, 2015

Karen Panetta, Chen Gao, and Sos Agaian. Human-visual- system-inspired underwater image quality measures.IEEE Journal of Oceanic Engineering, 41(3):541–551, 2015. 6

work page 2015
[44]

Data-efficient image quality assessment with attention-panel decoder

Guanyi Qin, Runze Hu, Yutao Liu, Xiawu Zheng, Haotian Liu, Xiu Li, and Yan Zhang. Data-efficient image quality assessment with attention-panel decoder. InAAAI, pages 2091–2100, 2023. 3

work page 2091
[45]

Going the extra mile in face image quality assess- ment: A novel database and model.IEEE Transactions on Multimedia, 2023

Shaolin Su, Hanhe Lin, Vlad Hosu, Oliver Wiedemann, Jin- qiu Sun, Yu Zhu, Hantao Liu, Yanning Zhang, and Dietmar Saupe. Going the extra mile in face image quality assess- ment: A novel database and model.IEEE Transactions on Multimedia, 2023. 5

work page 2023
[46]

Dual associated encoder for face restoration

Yu-Ju Tsai, Yu-Lun Liu, Lu Qi, Kelvin CK Chan, and Ming- Hsuan Yang. Dual associated encoder for face restoration. ICLR, 2024. 1, 3, 5, 7

work page 2024
[47]

Ex- ploring clip for assessing the look and feel of images

Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Ex- ploring clip for assessing the look and feel of images. InPro- ceedings of the AAAI Conference on Artificial Intelligence, pages 2555–2563, 2023. 5, 6

work page 2023
[48]

To- wards real-world blind face restoration with generative fa- cial prior

Xintao Wang, Yu Li, Honglun Zhang, and Ying Shan. To- wards real-world blind face restoration with generative fa- cial prior. InThe IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 3, 6

work page 2021
[49]

Restoreformer: High-quality blind face restoration from undegraded key-value pairs

Zhouxia Wang, Jiawei Zhang, Runjian Chen, Wenping Wang, and Ping Luo. Restoreformer: High-quality blind face restoration from undegraded key-value pairs. InProceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, pages 17512–17521, 2022. 1, 5

work page 2022
[50]

Dr2: Diffusion-based robust degradation remover for blind face restoration

Zhixin Wang, Ziying Zhang, Xiaoyun Zhang, Huangjie Zheng, Mingyuan Zhou, Ya Zhang, and Yanfeng Wang. Dr2: Diffusion-based robust degradation remover for blind face restoration. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1704– 1713, 2023. 1, 3, 5

work page 2023
[51]

Deep retinex decomposition for low-light enhancement

Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep retinex decomposition for low-light enhancement. arXiv preprint arXiv:1808.04560, 2018. 6

work page arXiv 2018
[52]

Q-align: Teaching lmms for visual scoring via discrete text-defined levels.ICLR, 2024

Haoning Wu, Zicheng Zhang, Weixia Zhang, Chaofeng Chen, Liang Liao, Chunyi Li, Yixuan Gao, Annan Wang, Erli Zhang, Wenxiu Sun, et al. Q-align: Teaching lmms for visual scoring via discrete text-defined levels.ICLR, 2024. 5, 6

work page 2024
[53]

Diffir: Efficient diffusion model for image restoration

Bin Xia, Yulun Zhang, Shiyin Wang, Yitong Wang, Xing- long Wu, Yapeng Tian, Wenming Yang, and Luc Van Gool. Diffir: Efficient diffusion model for image restoration. In ICCV, 2023. 6

work page 2023
[54]

Snr-aware low-light image enhancement

Xiaogang Xu, Ruixing Wang, Chi-Wing Fu, and Jiaya Jia. Snr-aware low-light image enhancement. InCVPR, pages 17714–17724, 2022. 8

work page 2022
[55]

You only need one color space: An efficient network for low-light image enhancement.CVPR, 2025

Qingsen Yan, Yixu Feng, Cheng Zhang, Pei Wang, Peng Wu, Wei Dong, Jinqiu Sun, and Yanning Zhang. You only need one color space: An efficient network for low-light image enhancement.CVPR, 2025. 6, 7, 8

work page 2025
[56]

An underwater color im- age quality evaluation metric.IEEE Transactions on Image Processing, 24(12):6062–6071, 2015

Miao Yang and Arcot Sowmya. An underwater color im- age quality evaluation metric.IEEE Transactions on Image Processing, 24(12):6062–6071, 2015. 6

work page 2015
[57]

Sparse gradient regularized deep retinex network for robust low-light image enhancement.IEEE Transactions on Image Processing, 30:2072–2086, 2021

Wenhan Yang, Wenjing Wang, Haofeng Huang, Shiqi Wang, and Jiaying Liu. Sparse gradient regularized deep retinex network for robust low-light image enhancement.IEEE Transactions on Image Processing, 30:2072–2086, 2021. 6

work page 2072
[58]

Scaling up to excellence: Practicing model scaling for photo- realistic image restoration in the wild

Fanghua Yu, Jinjin Gu, Zheyuan Li, Jinfan Hu, Xiangtao Kong, Xintao Wang, Jingwen He, Yu Qiao, and Chao Dong. Scaling up to excellence: Practicing model scaling for photo- realistic image restoration in the wild. InCVPR, pages 25669–25680, 2024. 3

work page 2024
[59]

Difface: Blind face restoration with diffused error contraction.IEEE Transac- tions on Pattern Analysis and Machine Intelligence, 2024

Zongsheng Yue and Chen Change Loy. Difface: Blind face restoration with diffused error contraction.IEEE Transac- tions on Pattern Analysis and Machine Intelligence, 2024. 2, 5, 6, 7

work page 2024
[60]

Dual adversarial network: Toward real-world noise removal and noise generation

Zongsheng Yue, Qian Zhao, Lei Zhang, and Deyu Meng. Dual adversarial network: Toward real-world noise removal and noise generation. InECCV, pages 41–58, 2020. 3

work page 2020
[61]

Learning enriched features for real image restoration and enhancement

Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Learning enriched features for real image restoration and enhancement. InECCV, pages 492–511. Springer, 2020. 8

work page 2020
[62]

Restormer: Efficient transformer for high-resolution image restoration

Syed Waqas Zamir, Aditya Arora, Salman Khan, Mu- nawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. InCVPR, pages 5728–5739, 2022. 8

work page 2022
[63]

The unreasonable effectiveness of deep features as a perceptual metric

Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shecht- man, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. InProceedings of the IEEE conference on computer vision and pattern recogni- tion, pages 586–595, 2018. 7

work page 2018
[64]

Towards authentic face restora- tion with iterative diffusion models and beyond.ICCV, 2023

Yang Zhao, Tingbo Hou, Yu-Chuan Su, Xuhui Jia Li, Matthias Grundmann, et al. Towards authentic face restora- tion with iterative diffusion models and beyond.ICCV, 2023. 1, 3

work page 2023
[65]

Empowering low- light image enhancer through customized learnable priors

Naishan Zheng, Man Zhou, Yanmeng Dong, Xiangyu Rui, Jie Huang, Chongyi Li, and Feng Zhao. Empowering low- light image enhancer through customized learnable priors. In ICCV, pages 12559–12569, 2023. 6

work page 2023
[66]

Underwater camera: Improv- ing visual perception via adaptive dark pixel prior and color correction.International Journal of Computer Vision, pages 1–19, 2023

Jingchun Zhou, Qian Liu, Qiuping Jiang, Wenqi Ren, Kin- Man Lam, and Weishi Zhang. Underwater camera: Improv- ing visual perception via adaptive dark pixel prior and color correction.International Journal of Computer Vision, pages 1–19, 2023. 6

work page 2023
[67]

Towards robust blind face restora- tion with codebook lookup transformer.Advances in Neural Information Processing Systems, 35:30599–30611, 2022

Shangchen Zhou, Kelvin Chan, Chongyi Li, and Chen Change Loy. Towards robust blind face restora- tion with codebook lookup transformer.Advances in Neural Information Processing Systems, 35:30599–30611, 2022. 1, 3, 5, 6, 8

work page 2022
[68]

Adapt or perish: Adaptive sparse transformer with attentive feature refinement for image restoration

Shihao Zhou, Duosheng Chen, Jinshan Pan, Jinglei Shi, and Jufeng Yang. Adapt or perish: Adaptive sparse transformer with attentive feature refinement for image restoration. In CVPR, pages 2952–2963, 2024. 6, 7

work page 2024