Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Proxies for Distortion and Consistency with Applications for Real-World Image Restoration

T0 review · 4 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that image restoration algorithms can be ranked by distortion and consistency without ground-truth images, using a trained degradation estimator and proxy measures whose rankings match true MSE and LPIPS.

desk verdict A practical, honestly written no-reference evaluation suite for blind restoration whose real-world rankings are conditional on the assumed degradation family—worth serious refereeing, but with that caveat front and center. read the letter →

arxiv 2501.12102 v1 pith:ASA3NUOR submitted 2025-01-21 cs.CV cs.AIcs.LGeess.IV

Sean Man , Guy Ohayon , Ron Raphaeli , Michael Elad This is my paper · ORCID
classification cs.CVcs.AIcs.LGeess.IV MSC 68U1094A08
keywords real-worldimagerestorationblindfaceno-referencedistortionmeasuredegradationestimationconsistencyplug-and-playdiffusionProxMSELPIPS
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper is trying to establish that real-world blind image restoration, where no ground-truth images exist, can still be evaluated and guided by principled tools rather than heuristics. It proposes a trained estimator that predicts the chain of degradations a degraded image has gone through, and shows how that estimator enables an approximate likelihood, a blind consistency measure, and a plug-and-play diffusion restoration method. The central theoretical result is that the proposed no-reference proxy ProxMSE equals true MSE up to a constant independent of the estimator, and because LPIPS is a squared error in feature space, ProxLPIPS inherits the same property. If these claims hold, restoration algorithms can be compared on real-world datasets—such as LFW, WebPhoto, and WIDER—by their approximate distortion and consistency, tasks previously impossible without paired ground truth. The paper also introduces ELAD, a diffusion-based restorer guided by the estimated likelihood, which it reports improves distortion and consistency over DifFace and PGDiff.

What carries the argument

The central object is the identity of Proposition 1: for any estimator $\hat{X}$ and the MMSE estimator $X^* = E[X|Y]$, the cross term in $E[\|X - \hat{X}\|^2] = E[\|(X - X^*) - (\hat{X} - X^*)\|^2]$ vanishes by the law of total expectation, yielding $\text{ProxMSE}(\hat{X}) = \text{MSE}(X,\hat{X}) - d^*$. This identity is what lets a no-reference quantity rank estimators exactly like the full-reference MSE. Carrying the whole pipeline is the trained degradation estimator $a_\theta(y)$, which predicts the four degradation parameters (blur, scale factor, noise, JPEG quality) and feeds both the approximate likelihood and the ProxCMSE measure. For the proxy measures, the approximate MMSE estimator $\tilde{X}^*$ stands in for $X^*$; the paper bounds the induced error in Appendix C, and LPIPS is recast as a squared error in feature space so that ProxLPIPS obeys the same logic.

What would settle it

Take a collection of real-world degraded images for which true ground-truth versions are available (for example, images captured before and after an unknown degradation), run several candidate restoration algorithms, and compare the ordering given by true MSE and LPIPS against the ordering given by ProxMSE and ProxLPIPS. If the orderings disagree substantially, or if the proxy ranking flips when the set includes images with a degradation outside Eq. (1) such as haze, the paper's central claim that the proxies rank methods like the true measures would be refuted.

Watch

Extended reading notes

Core claim

The central claim is that with a trained degradation estimator $a_\theta(y)$ and an approximate MMSE estimator, one can compute no-reference proxies whose ranking of restoration algorithms equals the ranking by true distortion measures. Specifically, for any estimator $\hat{X}$ and the MMSE estimator $X^* = E[X|Y]$, the paper proves $\text{ProxMSE}(\hat{X}) = \text{MSE}(X,\hat{X}) - d^*$, where $d^* = E[\|X - X^*\|^2]$ is a constant independent of $\hat{X}$; this identity, credited to Freirich et al. [14], makes the ranking order of estimators by ProxMSE identical to their ranking by true MSE. Since LPIPS is shown to be a squared error in a VGG feature space, the same argument yields ProxLPIPS. The same degradation estimator provides an approximate log-likelihood $\ell(y,x) \approx -\|y - \mu_Y(x, a_\theta(y))\|^2$, giving ProxCMSE as a blind consistency measure and serving as guidance for ELAD, a plug-and-play restoration method that extends DifFace by taking diffusion steps along the estimated likelihood gradient. The paper validates the proxies on synthetic data with known ground truth, demonstrates their alignment with true MSE and LPIPS, and then uses them to rank real-world methods, reporting that ELAD improves ProxMSE, ProxLPIPS, and ProxCMSE over DifFace and PGDiff.

Load-bearing premise

The load-bearing premise is that the parametric degradation model in Eq. (1), with the chosen parameter ranges, actually covers the degradations present in real-world inputs; the paper itself notes that some WebPhoto images contain haze that this model does not account for.

Editorial extensions

If this is right

  • Blind face restoration methods can now be compared on real-world datasets by distortion and consistency, not only by FID or subjective quality, using ProxMSE, ProxLPIPS, and ProxCMSE.
  • ELAD, which guides a diffusion prior with the estimated likelihood, attains lower ProxMSE, ProxLPIPS, and ProxCMSE than DifFace and PGDiff on LFW-Test, WebPhoto-Test, and WIDER-Test, with only a slight FID change.
  • On synthetic CelebA-Test datasets where ground truth is available, ProxMSE and ProxLPIPS rank a panel of end-to-end and plug-and-play methods in the same order as true MSE and LPIPS.
  • The degradation estimator reveals per-dataset degradation distributions, such as LFW having narrower blur and downsampling but stronger JPEG compression, enabling synthetic datasets that mimic real-world inputs more faithfully than uniform sampling.
  • The consistency guidance used by ELAD improves distortion as a side effect, suggesting that better consistency with the measurement and better distortion need not be in conflict in blind restoration.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The identity behind ProxMSE is generic: it holds for any distortion that is a squared error in some inner-product space, so analogous no-reference proxies could be built for other full-reference metrics that have a similar geometric structure.
  • A testable extension is to monitor the degradation estimator's predicted parameter distribution to detect distribution shift: if a real-world benchmark contains degradations outside the trained family, such as the haze the paper notes in WebPhoto, the proxy measures should be recalibrated before being trusted for ranking.
  • Because ProxMSE approximates true MSE up to a constant, it could be used directly as a training objective for restoration models on unpaired real-world data, effectively minimizing distortion without ground-truth pairs.
  • The synthetic datasets generated by sampling from estimated real-world degradation distributions could become a more realistic evaluation standard than uniform sampling, but their validity inherits the accuracy of the degradation estimator and the assumed degradation family.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes a suite of tools for blind real-world image restoration under the parametric degradation model of Eq. (1): a trained degradation estimator a_theta, an empirical likelihood approximation (ELA) that leads to a consistency measure ProxCMSE, no-reference distortion proxies ProxMSE and ProxLPIPS, and a plug-and-play diffusion restoration algorithm ELAD. Proposition 1, credited to Freirich et al., shows that ProxMSE equals MSE up to a constant when the true posterior mean is available, so ranking by ProxMSE matches ranking by MSE. Experiments on synthetic CelebA-Test datasets with uniform and dataset-mimicking degradation distributions show alignment between the proxies and true MSE/LPIPS/consistency, and the real-world datasets LFW-Test, WebPhoto-Test, and WIDER-Test are used to rank methods and evaluate ELAD. The paper also includes a synthetic-data generation procedure based on KDE estimates of real-world degradation parameters.

Significance. If the proposed tools are reliable outside their training distribution, this would be a meaningful contribution: it offers the first practical no-reference proxy framework for ranking blind image restoration methods by distortion and consistency, and a simple way to turn a diffusion prior into a blind plug-and-play restorer. The theoretical core is sound: Proposition 1 is correct and properly attributed to prior work, and the LPIPS-as-MSE-in-latent-space observation is clean and clearly proved. The synthetic validation is thorough in its coverage of multiple restoration methods and degradation distributions, and the paper explicitly acknowledges several of its own limitations, which is commendable. The main weakness is that the real-world claims currently rest on validation that uses the same degradation family used for training, so the paper's own admitted out-of-family cases (e.g., haze in WebPhoto) are not tested; the significance of the real-world results is therefore conditional on additional validation.

major comments (4)
  1. [§6 and Tables 1–2] The central real-world claims are not validated for degradations outside the family of Eq. (1). Section 6 concedes that "some images in WebPhoto seem to contain haze, which is not accounted in Eq. (1)"; since the degradation estimator, the KDE-based mimic datasets of Section 3.2, the ProxCMSE/ProxMSE/ProxLPIPS regressors, and the ELAD likelihood are all trained and validated on synthetic degradations from Eq. (1) (Figures 4 and 7), the real-world rankings in Tables 1 and 2 have no independent ground-truth reference. I would like to see either an evaluation on paired real-world low/high-quality images with true MSE, LPIPS, and consistency values, or a controlled out-of-family synthetic benchmark (e.g., haze, color shifts, spatially varying blur) demonstrating that proxy rankings and ELAD's guidance remain reliable under model misspecification.
  2. [§4.3 and Appendix C] Proposition 1 guarantees ranking equivalence only for the true posterior mean X*, whereas Section 4.3 uses the DifFace regressor as an approximate MMSE estimator. Appendix C's bound in Eq. (21) bounds the absolute error of ProxMSE but does not imply preservation of the ordering of estimators, and the paper states that the tightness of this bound is unclear in practice. The synthetic figures show visual alignment, but they do not quantify ranking fidelity with a rank-correlation measure, and they evaluate the approximation under its own training distribution. Please add explicit rank-correlation statistics (e.g., Kendall or Spearman between proxy and true values) and test the approximate MMSE regressor under a distribution different from its training distribution.
  3. [§3.3, §3.4, and §6] ProxCMSE and ELAD rely on a single point estimate a_theta(y) in Eq. (4), yet Section 6 notes that many different degradations may correspond to the same observed measurement. For real-world inputs with an erroneous or ambiguous degradation estimate, the consistency measure can be miscalibrated and the ELAD guidance step can be misdirected. Figure 4 only tests the case where the ground-truth degradation is drawn from the same family and is reasonably well estimated. A sensitivity analysis that perturbs a_theta(y), or averages over an approximate posterior over A, and that compares ProxCMSE against CMSE under deliberately misspecified parameters, would quantify the error introduced by the point-estimate approximation.
  4. [Table 1 and §3.4] The quantitative advantage of ELAD over DifFace and PGDiff in Tables 1 and 3 is measured with ProxCMSE, ProxMSE, and ProxLPIPS, all of which are built from the same degradation model and trained components used by ELAD's guidance term. Since ELAD directly minimizes ||y - mu(x, a_theta(y))||^2 at each step, the ProxCMSE comparison is at least partially aligned with ELAD's own objective, so the reported consistency gain is less informative as an independent evaluation. An independent consistency check on data with ground-truth degradation parameters, or on real paired data, is needed to confirm that the gain is not merely an artifact of optimizing the proxy.
minor comments (5)
  1. [Supplementary Figures 8–11] The captions refer to "GFPGAN [6]", but GFPGAN is reference [61], not [6]; the citation should be corrected.
  2. [Figure 7] The caption states that the ProxMSE regression line has slope one "following Proposition 1"; since Proposition 1 is an identity, the slope-one line is not empirical evidence of alignment. Reporting residual statistics or rank correlations would be more informative.
  3. [Table 3] The "Uniform Dist. #1" columns have missing CMSE/ProxCMSE values with only a parenthetical explanation; a dedicated footnote would improve readability.
  4. [Figure 10] The dataset name is misspelled as "WebPhot-Test"; it should be "WebPhoto-Test".
  5. [§4.3] The phrase "ELAD is the state-of-the-art plug-and-play method for BFR tasks" is stronger than the evidence supports, as Table 1 compares only PGDiff, DifFace, and ELAD; "state-of-the-art among the compared plug-and-play methods" would be more precise.

Circularity Check

2 steps flagged · score 4.0 of 10

Real-world ELAD consistency is evaluated with the same fitted likelihood surrogate that guides ELAD; the central ProxMSE/LPIPS proxy claim retains independent synthetic grounding.

  1. fitted input called prediction [Section 3.4 (Algorithm 2, lines 5-7), Section 3.3 Eq. (6), Table 1]
    "ProxCMSE( ˆX) := E(ˆx,y)∼p ˆX,Y [∥y − µY (ˆx, aθ(y))∥2 2] ... g = ∇ˆxt ∥y − µ(ˆxt 0, ˆa)∥2 2 // compute score likelihood ... ELAD is better in terms of distortion (ProxMSE, ProxLPIPS) and consistency (ProxCMSE)."

    ELAD's guidance step in Algorithm 2 descends the gradient of the squared residual ∥y − µ(x̂t0, aθ(y))∥², using the same fitted degradation estimator aθ that defines ProxCMSE in Eq. (6). Table 1 then reports ELAD's real-world ProxCMSE advantage over DifFace and PGDiff as evidence of better consistency. The ranking is therefore a value of ELAD's own fitted objective, not an independent measurement; the synthetic validation in Figure 4 uses degradations drawn from the Eq. (1) family in which aθ was trained, so it cannot break this self-reference on real inputs. Section 6 concedes that some WebPhoto images contain haze 'not accounted in Eq. (1)', which would corrupt both the guidance and the measure together.

  2. fitted input called prediction [Section 3.2, Section 3.3 Figure 4, Section 6]
    "we estimate the degradations' parameters in the real-world datasets LFW-Test [22], WebPhoto-Test [61], and WIDER-Test [68, 75] and approximate their distribution using Kernel Density Estimation (KDE). Then, we synthesize degraded measurements from CelebA-Test [25, 33, 75], a dataset of clean images, by samplingA according to the predicted distribution corresponding to each real-world dataset."

    The mimic datasets used to validate ProxCMSE (and the other proxies) are synthesized by sampling A from a KDE fitted to aθ's own predictions on real datasets. Thus the validation distribution is generated by the very estimator whose correctness is the load-bearing premise; alignment of ProxCMSE with CMSE on these datasets only shows consistency of the estimator with itself on its own predicted degradation family. It does not certify the estimator on true real-world degradations (e.g., WebPhoto haze), so the real-world 'reliable proxy' claims rest on an untested equivalence rather than an external benchmark.

full rationale

The central ProxMSE/ProxLPIPS ranking claim has independent content: Proposition 1 is credited to the external result of Freirich et al. [14], and the synthetic experiments in Figure 7 compare ProxMSE and ProxLPIPS against true MSE and LPIPS across multiple restoration methods, so those rankings are not circular. The paper also explicitly notes that Ohayon et al. [46] did not assess the practical validity of ProxMSE, making the current validation a new contribution rather than a load-bearing self-citation. The circularity is confined to the real-world consistency evaluation of ELAD. ProxCMSE in Eq. (6) is built from the same fitted degradation estimator aθ and degradation-mean operator that ELAD uses as its likelihood guidance in Algorithm 2, so Table 1's claim that ELAD improves consistency on real-world datasets is partly a construction: the method optimizes the same surrogate that is used to measure it. The synthetic validation of ProxCMSE (Figure 4) uses datasets generated from aθ's own estimated degradation distributions (Section 3.2), so it does not independently establish the measure's validity on true out-of-family real-world degradations. This is reinforced by the paper's own Section 6 admission that WebPhoto contains haze not represented in Eq. (1), and that the Appendix C bound on the approximate MMSE regressor is of unclear tightness. These are acknowledged limitations rather than hidden reductions, and the main distortion-proxy claim remains externally anchored, so the overall circularity is partial rather than total.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The paper introduces trained models and assumptions rather than new physical entities. The central claims rest on the parametric degradation model, the accuracy of two learned estimators, and the orthogonality property of the MMSE. The free parameters listed are manual design choices in the ELAD algorithm and estimator training.

free parameters (4)
  • ELAD likelihood step size λ = 10^-2
    Manual hyperparameter in Algorithm 2 (Appendix E) controlling the strength of the consistency guidance step; results depend on it.
  • Diffusion start timestep T0 = 400 (of T=1000)
    Initial timestep for accelerated sampling, following DifFace; affects initialization noise and final quality.
  • Number of Monte Carlo samples for μ estimation = 16
    Used to empirically approximate the mean and diagonal covariance of p(Y|X,A) each denoising step; chosen without analysis.
  • Degradation estimator loss weights = 0.25 (LMain), 1 (LReg)
    Manual weights balancing parameter regression and mean-matching terms in Section 3.2.
assumptions (4)
  • domain assumption Real-world degradations follow the parametric model Y = JPEG_Q((K*x)↓S + N) with σ_K ∈ [0.1,15], S ∈ [1,32], σ_N ∈ [0,20/255], Q ∈ [30,100].
    Core model used for training the degradation estimator and for synthesizing validation sets (Section 2). The authors acknowledge haze in WebPhoto exceeds this family (Section 6).
  • domain assumption The degradation estimator a_θ generalizes from synthetic training on FFHQ to real-world images.
    All downstream tools (ProxCMSE, ELAD, proxy measures) rely on a_θ(y) being accurate on real-world images (Section 3.2).
  • domain assumption The DifFace regressor (and its LPIPS retrained variant) is an accurate approximation of the MMSE estimator X* (and latent Z*).
    ProxMSE and ProxLPIPS use these approximations (Section 4.2). Appendix C gives a bound involving the unknown error R = X~* - X*, which is not actionable.
  • domain assumption Restoration estimators are functions of Y only (plus independent noise), so Xhat is conditionally independent of X given Y.
    Needed for the orthogonality property used in Proposition 1 (Appendix B, Eqs. 13-16). Violated if an estimator uses extra information.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Proxies for Distortion and Consistency with Applications for Real-World Image Restoration." pith.science (2026). https://pith.science/paper/ASA3NUOR

@misc{pith2026250112102,
  author       = {Pith},
  title        = {Pith review of: Proxies for Distortion and Consistency with Applications for Real-World Image Restoration},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ASA3NUOR}},
  note         = {Machine review of arXiv:2501.12102}
}
read the original abstract

Real-world image restoration deals with the recovery of images suffering from an unknown degradation. This task is typically addressed while being given only degraded images, without their corresponding ground-truth versions. In this hard setting, designing and evaluating restoration algorithms becomes highly challenging. This paper offers a suite of tools that can serve both the design and assessment of real-world image restoration algorithms. Our work starts by proposing a trained model that predicts the chain of degradations a given real-world measured input has gone through. We show how this estimator can be used to approximate the consistency -- the match between the measurements and any proposed recovered image. We also use this estimator as a guiding force for the design of a simple and highly-effective plug-and-play real-world image restoration algorithm, leveraging a pre-trained diffusion-based image prior. Furthermore, this work proposes no-reference proxy measures of MSE and LPIPS, which, without access to the ground-truth images, allow ranking of real-world image restoration algorithms according to their (approximate) MSE and LPIPS. The proposed suite provides a versatile, first of its kind framework for evaluating and comparing blind image restoration algorithms in real-world scenarios.

Figures

Figures reproduced from arXiv: 2501.12102 by the authors.

Figure 1
Figure 1. This work introduces several novel tools to help tackle [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Degradation estimator accuracy. We test our degra￾dation estimator on synthetic CelebA-Test datasets (Section 3.2). (a-d) Scatter plots and R 2 scores of the true vs. the predicted val￾ues for each type of operator in Eq. (1). (e) The mean and standard deviation of the PSNR between µY (x, a) and µY (x, aθ(y)). The estimator demonstrates high prediction accuracy, as reflected by the high PSNR and R 2 scores. proposed… view at source ↗
Figure 4
Figure 4. Proxy consistency measure. Each plot shows the CMSE versus ProxCMSE, evaluated on synthetic CelebA-Test datasets (Section 3.2). The strong alignment of the two suggests that ProxCMSE is a trustworthy approximation for the CMSE when the degradation process is unknown. is unavailable in our setting). As seen in [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figures from the paper (10 more)
Figure 5
Figure 5. Figure 5: Consistency of P&P real-world restoration methods. Left column: A synthetic degraded example from CelebA-Test. First row: The ground-truth image alongside the restorations of ELAD (our method), DifFace, and PGDiff. Second row: The mean of the likelihood defined using t…
Figure 6
Figure 6. Figure 6: Restoration examples of real-world images taken from WIDER-Test, WebPhoto-Test, and LFW-Test (top to bottom) for different [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: Proxy distortion measures. The plots compare the proxy measures with their true counterparts, for several state-of￾the-art methods evaluated on the synthetic CelebA-Test datasets (Section 3.2). A linear regression line is drawn for better clar￾ity (for ProxMSE the slop…
Figure 8
Figure 8. Figure 8: Restoration examples on CelebA-Test [61]. 7 [PITH_FULL_IMAGE:figures/full_fig_p019_8.png]
Figure 9
Figure 9. Figure 9: Restoration examples on LFW-Test [22]. 8 [PITH_FULL_IMAGE:figures/full_fig_p020_9.png]
Figure 10
Figure 10. Figure 10: Restoration examples on WebPhot-Test [61]. 9 [PITH_FULL_IMAGE:figures/full_fig_p021_10.png]
Figure 11
Figure 11. Figure 11: Restoration examples on WIDER-Test [68]. 10 [PITH_FULL_IMAGE:figures/full_fig_p022_11.png]
Figure 12
Figure 12. Figure 12: Degradation estimator architecture for BFR. [PITH_FULL_IMAGE:figures/full_fig_p023_12.png]
Figure 13
Figure 13. Figure 13: Images from real datasets paired with images from CelebA synthetically degraded by the predicted degradation. [PITH_FULL_IMAGE:figures/full_fig_p023_13.png]
Figure 14
Figure 14. Figure 14: More samples of real and synthetic images. The real images come from the LFW-Test, WebPhoto-Test, and WIDER-Test [PITH_FULL_IMAGE:figures/full_fig_p024_14.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Compressed Image Generation with Denoising Diffusion Codebook Models

    eess.IV 2025-02 conditional novelty 8.0 of 10

    Using fixed codebooks of noise vectors in diffusion sampling yields images that carry their own compressed bit-streams and enables a strong perceptual image codec.

Reference graph

Works this paper leans on

79 extracted references · 69 canonical work pages · cited by 1 Pith paper

  1. [1]

    Explorable Super Resolu- tion

    Yuval Bahat and Tomer Michaeli. Explorable Super Resolu- tion. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition , pages 2716–2725,

    work page

  2. [2]

    What’s in the Image? Ex- plorable Decoding of Compressed Images

    Yuval Bahat and Tomer Michaeli. What’s in the Image? Ex- plorable Decoding of Compressed Images. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2908–2917, 2021. 7

    work page 2021

  3. [3]

    The Perception-Distortion Tradeoff

    Yochai Blau and Tomer Michaeli. The Perception-Distortion Tradeoff. In Proceedings of the IEEE Conference on Com- puter Vision and Pattern Recognition , pages 6228–6237,

    work page

  4. [4]

    Tim Brooks, Ben Mildenhall, Tianfan Xue, Jiawen Chen, Dillon Sharlet, and Jonathan T. Barron. Unprocessing Im- ages for Learned Raw Denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11036–11045, 2019. 7

    work page 2019

  5. [5]

    Chain-of- Restoration: Multi-Task Image Restoration Models are Zero-Shot Step-by-Step Universal Image Restorers

    Jin Cao, Deyu Meng, and Xiangyong Cao. Chain-of- Restoration: Multi-Task Image Restoration Models are Zero-Shot Step-by-Step Universal Image Restorers. (arXiv:2410.08688), 2024. 7

    arXiv 2024

  6. [6]

    Towards Real-World Blind Face Restoration with Generative Diffusion Prior

    Xiaoxu Chen, Jingfan Tan, Tao Wang, Kaihao Zhang, Wenhan Luo, and Xiaochun Cao. Towards Real-World Blind Face Restoration with Generative Diffusion Prior. (arXiv:2312.15736), 2024. 7, 8, 9, 10

    arXiv 2024

  7. [7]

    Trainable Nonlinear Reac- tion Diffusion: A Flexible Framework for Fast and Effective Image Restoration

    Yunjin Chen and Thomas Pock. Trainable Nonlinear Reac- tion Diffusion: A Flexible Framework for Fast and Effective Image Restoration. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(6):1256–1272, 2017. 1

    work page 2017

  8. [8]

    Diffusion Pos- terior Sampling for General Noisy Inverse Problems

    Hyungjin Chung, Jeongsol Kim, Michael Thompson Mc- cann, Marc Louis Klasky, and Jong Chul Ye. Diffusion Pos- terior Sampling for General Noisy Inverse Problems. In The Eleventh International Conference on Learning Representa- tions, 2022. 2, 4, 5

    work page 2022

Show all 79 references
  1. [9]

    Parallel Diffusion Models of Operator and Image for Blind Inverse Problems

    Hyungjin Chung, Jeongsol Kim, Sehui Kim, and Jong Chul Ye. Parallel Diffusion Models of Operator and Image for Blind Inverse Problems. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 6059–6069, 2023. 8

    2023

  2. [10]

    Image Denoising by Sparse 3-D Transform-Domain Collaborative Filtering

    Kostadin Dabov, Alessandro Foi, Vladimir Katkovnik, and Karen Egiazarian. Image Denoising by Sparse 3-D Transform-Domain Collaborative Filtering. IEEE Transac- tions on Image Processing, 16(8):2080–2095, 2007. 1

    2007

  3. [11]

    ImageNet: A large-scale hierarchical im- age database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical im- age database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 3

    2009

  4. [12]

    Image Denoising Via Sparse and Redundant Representations Over Learned Dic- tionaries

    Michael Elad and Michal Aharon. Image Denoising Via Sparse and Redundant Representations Over Learned Dic- tionaries. IEEE Transactions on Image Processing, 15(12): 3736–3745, 2006. 1

    2006

  5. [13]

    Image Denoising: The Deep Learning Revolution and Beyond—A Survey Paper

    Michael Elad, Bahjat Kawar, and Gregory Vaksman. Image Denoising: The Deep Learning Revolution and Beyond—A Survey Paper. SIAM Journal on Imaging Sciences , 16(3): 1594–1654, 2023. 1

    2023

  6. [14]

    A Theory of the Distortion-Perception Tradeoff in Wasserstein Space

    Dror Freirich, Tomer Michaeli, and Ron Meir. A Theory of the Distortion-Perception Tradeoff in Wasserstein Space. In Advances in Neural Information Processing Systems , pages 25661–25672. Curran Associates, Inc., 2021. 6, 1

    2021

  7. [15]

    Weighted Nuclear Norm Minimization with Appli- cation to Image Denoising

    Shuhang Gu, Lei Zhang, Wangmeng Zuo, and Xiangchu Feng. Weighted Nuclear Norm Minimization with Appli- cation to Image Denoising. In Proceedings of the IEEE Con- ference on Computer Vision and Pattern Recognition, pages 2862–2869, 2014. 1

    2014

  8. [16]

    VQFR: Blind Face Restoration with Vector-Quantized Dictionary and Parallel Decoder

    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 Decoder. In Computer Vision – ECCV 2022, pages 126–143, Cham, 2022. Springer Nature Switzerland. 2, 3, 1, 4, 7, 8, 9, 10

    2022

  9. [17]

    Alarc´on

    Javier Gurrola-Ramos, Oscar Dalmau, and Teresa E. Alarc´on. A Residual Dense U-Net Neural Network for Im- age Denoising. IEEE Access, 9:31742–31754, 2021. 1

    2021

  10. [18]

    Scaling Laws and Compute-Optimal Training Beyond Fixed Training Du- rations

    Alexander H ¨agele, Elie Bakouch, Atli Kosson, Loubna Ben Allal, Leandro V on Werra, and Martin Jaggi. Scaling Laws and Compute-Optimal Training Beyond Fixed Training Du- rations. (arXiv:2405.18392), 2024. 3

    2024 arXiv

  11. [19]

    Under- water image restoration via contrastive learning and a real- world dataset

    Junlin Han, Mehrdad Shoeiby, Tim Malthus, Elizabeth Botha, Janet Anstee, Saeed Anwar, Ran Wei, Moham- mad Ali Armin, Hongdong Li, and Lars Petersson. Under- water image restoration via contrastive learning and a real- world dataset. Remote Sensing, 14(4297), 2022. 8

    2022

  12. [20]

    GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium

    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 Equilibrium. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 8, 1

    2017

  13. [21]

    Denoising Dif- fusion Probabilistic Models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Dif- fusion Probabilistic Models. In Advances in Neural Infor- mation Processing Systems, pages 6840–6851. Curran Asso- ciates, Inc., 2020. 5

    2020

  14. [22]

    Huang, Manu Ramesh, Tamara Berg, and Erik Learned-Miller

    Gary B. Huang, Manu Ramesh, Tamara Berg, and Erik Learned-Miller. Labeled faces in the wild: A database for studying face recognition in unconstrained environ- ments. Technical Report 07-49, University of Massachusetts, Amherst, 2007. 3, 8

    2007

  15. [23]

    Real-world super-resolution via kernel estimation and noise injection

    Xiaozhong Ji, Yun Cao, Ying Tai, Chengjie Wang, Jilin Li, and Feiyue Huang. Real-world super-resolution via kernel estimation and noise injection. In The IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR) Workshops, 2020. 2, 7, 8

    2020

  16. [24]

    DaLPSR: Leverage Degradation- Aligned Language Prompt for Real-World Image Super- Resolution

    Aiwen Jiang, Zhi Wei, Long Peng, Feiqiang Liu, Wenbo Li, and Mingwen Wang. DaLPSR: Leverage Degradation- Aligned Language Prompt for Real-World Image Super- Resolution. (arXiv:2406.16477), 2024. 7

    2024 arXiv

  17. [25]

    Progressive Growing of GANs for Improved Quality, Stabil- ity, and Variation

    Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive Growing of GANs for Improved Quality, Stabil- ity, and Variation. In International Conference on Learning Representations, 2018. 3

    2018

  18. [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. In Proceedings of the IEEE/CVF Conference 10 on Computer Vision and Pattern Recognition , pages 4401– 4410, 2019. 3

    2019

  19. [27]

    Stochastic Image Denoising by Sampling From the Posterior Distribution

    Bahjat Kawar, Gregory Vaksman, and Michael Elad. Stochastic Image Denoising by Sampling From the Posterior Distribution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1866–1875, 2021. 1, 2, 3, 7

    2021

  20. [28]

    Denoising Diffusion Restoration Models

    Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising Diffusion Restoration Models. Advances in Neural Information Processing Systems, 35:23593–23606,

  21. [29]

    Fast Diffusion EM: A Diffusion Model for Blind Inverse Prob- lems With Application to Deconvolution

    Charles Laroche, Andr ´es Almansa, and Eva Coupet ´e. Fast Diffusion EM: A Diffusion Model for Blind Inverse Prob- lems With Application to Deconvolution. In Proceedings of the IEEE/CVF Winter Conference on Applications of Com- puter Vision, pages 5271–5281, 2024. 8

    2024

  22. [30]

    SwinIR: Image Restoration Using Swin Transformer

    Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. SwinIR: Image Restoration Using Swin Transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1833– 1844, 2021. 1

    2021

  23. [31]

    Diff- BIR: Towards Blind Image Restoration with Generative Dif- fusion Prior

    Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Bo Dai, Fanghua Yu, Wanli Ouyang, Yu Qiao, and Chao Dong. Diff- BIR: Towards Blind Image Restoration with Generative Dif- fusion Prior. (arXiv:2308.15070), 2024. 2, 3, 6, 7, 1, 4, 8, 9, 10

    2024 arXiv

  24. [32]

    A Comprehensive Study on Robustness of Image Classification Models: Benchmarking and Rethink- ing

    Chang Liu, Yinpeng Dong, Wenzhao Xiang, Xiao Yang, Hang Su, Jun Zhu, Yuefeng Chen, Yuan He, Hui Xue, and Shibao Zheng. A Comprehensive Study on Robustness of Image Classification Models: Benchmarking and Rethink- ing. International Journal of Computer Vision, 2024. 3

    2024

  25. [33]

    Deep learning face attributes in the wild

    Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaoou Tang. Deep learning face attributes in the wild. In Proceedings of International Conference on Computer Vision (ICCV), 2015. 3

    2015

  26. [34]

    A ConvNet for the 2020s

    Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feicht- enhofer, Trevor Darrell, and Saining Xie. A ConvNet for the 2020s. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition , pages 11976–11986,

  27. [35]

    Decoupled Weight De- cay Regularization

    Ilya Loshchilov and Frank Hutter. Decoupled Weight De- cay Regularization. In International Conference on Learning Representations, 2018. 3

    2018

  28. [36]

    SRFlow: Learning the Super-Resolution Space with Normalizing Flow

    Andreas Lugmayr, Martin Danelljan, Luc Van Gool, and Radu Timofte. SRFlow: Learning the Super-Resolution Space with Normalizing Flow. In Computer Vision – ECCV 2020, pages 715–732, Cham, 2020. Springer International Publishing. 3, 7

    2020

  29. [37]

    NTIRE 2021 Learning the Super-Resolution Space Chal- lenge

    Andreas Lugmayr, Martin Danelljan, and Radu Timofte. NTIRE 2021 Learning the Super-Resolution Space Chal- lenge. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, pages 596–612, 2021

    2021

  30. [38]

    NTIRE 2022 Challenge on Learning the Super-Resolution Space

    Andreas Lugmayr, Martin Danelljan, Radu Timofte, Kang- wook Kim, Younggeun Kim, Jae-young Lee, Zechao Li, Jinshan Pan, Dongseok Shim, Ki-Ung Song, Jinhui Tang, Cong Wang, and Zhihao Zhao. NTIRE 2022 Challenge on Learning the Super-Resolution Space. In Proceedings of the IEEE/C...

    2022

  31. [39]

    High-Perceptual Quality JPEG Decoding via Posterior Sam- pling

    Sean Man, Guy Ohayon, Theo Adrai, and Michael Elad. High-Perceptual Quality JPEG Decoding via Posterior Sam- pling. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition , pages 1272–1282,

  32. [40]

    A Variational Perspective on Solving Inverse Problems with Diffusion Models

    Morteza Mardani, Jiaming Song, Jan Kautz, and Arash Vah- dat. A Variational Perspective on Solving Inverse Problems with Diffusion Models. In The Twelfth International Confer- ence on Learning Representations, 2023. 2

    2023

  33. [41]

    completely blind

    Anish Mittal, Rajiv Soundararajan, and Alan C. Bovik. Mak- ing a “completely blind” image quality analyzer. IEEE Sig- nal Processing Letters, 20(3):209–212, 2013. 8

    2013

  34. [42]

    Metric Learning Based Interactive Modulation for Real-World Super-Resolution

    Chong Mou, Yanze Wu, Xintao Wang, Chao Dong, Jian Zhang, and Ying Shan. Metric Learning Based Interactive Modulation for Real-World Super-Resolution. In Computer Vision – ECCV 2022, pages 723–740, Cham, 2022. Springer Nature Switzerland. 2, 7

    2022

  35. [43]

    GibbsDDRM: A Partially Collapsed Gibbs Sampler for Solving Blind Inverse Problems with Denoising Diffusion Restoration

    Naoki Murata, Koichi Saito, Chieh-Hsin Lai, Yuhta Takida, Toshimitsu Uesaka, Yuki Mitsufuji, and Stefano Ermon. GibbsDDRM: A Partially Collapsed Gibbs Sampler for Solving Blind Inverse Problems with Denoising Diffusion Restoration. In Proceedings of the 40th International Con-...

  36. [44]

    High Perceptual Quality Image De- noising With a Posterior Sampling CGAN

    Guy Ohayon, Theo Adrai, Gregory Vaksman, Michael Elad, and Peyman Milanfar. High Perceptual Quality Image De- noising With a Posterior Sampling CGAN. In Proceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 1805–1813, 2021. 1, 3, 7

    2021

  37. [45]

    Reasons for the Superiority of Stochastic Esti- mators over Deterministic Ones: Robustness, Consistency and Perceptual Quality

    Guy Ohayon, Theo Joseph Adrai, Michael Elad, and Tomer Michaeli. Reasons for the Superiority of Stochastic Esti- mators over Deterministic Ones: Robustness, Consistency and Perceptual Quality. In Proceedings of the 40th Inter- national Conference on Machine Learning , pages 26...

    2023

  38. [46]

    Posterior- Mean Rectified Flow: Towards Minimum MSE Photo- Realistic Image Restoration

    Guy Ohayon, Tomer Michaeli, and Michael Elad. Posterior- Mean Rectified Flow: Towards Minimum MSE Photo- Realistic Image Restoration. (arXiv:2410.00418), 2024. 6, 7, 1, 8, 9, 10

    2024 arXiv

  39. [47]

    Benchmarking Denoising Al- gorithms With Real Photographs

    Tobias Plotz and Stefan Roth. Benchmarking Denoising Al- gorithms With Real Photographs. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition , pages 1586–1595, 2017. 8

    2017

  40. [48]

    The Little Engine That Could: Regularization by Denoising (RED)

    Yaniv Romano, Michael Elad, and Peyman Milanfar. The Little Engine That Could: Regularization by Denoising (RED). SIAM Journal on Imaging Sciences , 10(4):1804– 1844, 2017. 2

    2017

  41. [49]

    Un- paired Real-World Super-Resolution With Pseudo Control- lable Restoration

    Andr ´es Romero, Luc Van Gool, and Radu Timofte. Un- paired Real-World Super-Resolution With Pseudo Control- lable Restoration. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 798–807, 2022. 7

    2022

  42. [50]

    Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, 11 Jonathan Ho, David J

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, 11 Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Pho- torealistic Text-to-Image Diffusion Models with Deep Lan-...

    2022

  43. [51]

    Mehdi S. M. Sajjadi, Olivier Bachem, Mario Lucic, Olivier Bousquet, and Sylvain Gelly. Assessing Generative Models via Precision and Recall. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2018. 8

    2018

  44. [52]

    Zero-Shot

    Assaf Shocher, Nadav Cohen, and Michal Irani. “Zero-Shot” Super-Resolution Using Deep Internal Learning. InProceed- ings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3118–3126, 2018. 7

    2018

  45. [53]

    Very Deep Con- volutional Networks for Large-Scale Image Recognition

    Karen Simonyan and Andrew Zisserman. Very Deep Con- volutional Networks for Large-Scale Image Recognition. (arXiv:1409.1556), 2015. 5

    2015 arXiv

  46. [54]

    Denois- ing Diffusion Implicit Models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denois- ing Diffusion Implicit Models. In International Conference on Learning Representations, 2020. 3, 5

    2020

  47. [55]

    Pseudoinverse-Guided Diffusion Models for Inverse Problems

    Jiaming Song, Arash Vahdat, Morteza Mardani, and Jan Kautz. Pseudoinverse-Guided Diffusion Models for Inverse Problems. In International Conference on Learning Repre- sentations, 2022. 5

    2022

  48. [56]

    NIMA: Neural image assessment

    Hossein Talebi and Peyman Milanfar. NIMA: Neural image assessment. IEEE Transactions on Image Processing, 27(8): 3998–4011, 2018. 8

    2018

  49. [57]

    Explore Image Deblurring via Blur Kernel Space

    Phong Tran, Anh Tran, Quynh Phung, and Minh Hoai. Explore Image Deblurring via Blur Kernel Space. (arXiv:2104.00317), 2021. 7

    2021 arXiv

  50. [58]

    Taming Gener- ative Diffusion for Universal Blind Image Restoration

    Siwei Tu, Weidong Yang, and Ben Fei. Taming Gener- ative Diffusion for Universal Blind Image Restoration. (arXiv:2408.11287), 2024. 8

    2024 arXiv

  51. [59]

    Venkatakrishnan, Charles A

    Singanallur V . Venkatakrishnan, Charles A. Bouman, and Brendt Wohlberg. Plug-and-Play priors for model based re- construction. In 2013 IEEE Global Conference on Signal and Information Processing, pages 945–948, 2013. 2

    2013

  52. [60]

    Jianyi Wang, Kelvin C. K. Chan, and Chen Change Loy. Ex- ploring CLIP for Assessing the Look and Feel of Images. Proceedings of the AAAI Conference on Artificial Intelli- gence, 37(2):2555–2563, 2023. 8

    2023

  53. [61]

    To- wards Real-World Blind Face Restoration With Generative Facial Prior

    Xintao Wang, Yu Li, Honglun Zhang, and Ying Shan. To- wards Real-World Blind Face Restoration With Generative Facial Prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 9168– 9178, 2021. 2, 3, 8, 1, 4, 7, 9

    2021

  54. [62]

    Real-ESRGAN: Training Real-World Blind Super- Resolution With Pure Synthetic Data

    Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-ESRGAN: Training Real-World Blind Super- Resolution With Pure Synthetic Data. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 1905–1914, 2021. 2, 8

    1905

  55. [63]

    Zero-Shot Image Restoration Using Denoising Diffusion Null-Space Model

    Yinhuai Wang, Jiwen Yu, and Jian Zhang. Zero-Shot Image Restoration Using Denoising Diffusion Null-Space Model. (arXiv:2212.00490), 2022. 2

    2022 arXiv

  56. [64]

    Zhou Wang and Alan C. Bovik. Mean squared error: Love it or leave it? A new look at Signal Fidelity Measures. IEEE Signal Processing Magazine, 26(1):98–117, 2009. 1

    2009

  57. [65]

    Bovik, H.R

    Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: From error visibility to structural similarity. IEEE Transactions on Image Processing , 13(4): 600–612, 2004. 2

    2004

  58. [66]

    RestoreFormer++: Towards Real- World Blind Face Restoration From Undegraded Key-Value Pairs

    Zhouxia Wang, Jiawei Zhang, Tianshui Chen, Wenping Wang, and Ping Luo. RestoreFormer++: Towards Real- World Blind Face Restoration From Undegraded Key-Value Pairs. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(12):15462–15476, 2023. 3, 6, 7, 1, 4, 8, 9, 10

    2023

  59. [67]

    PGDiff: Guiding Diffusion Models for Versatile Face Restoration via Partial Guidance

    Peiqing Yang, Shangchen Zhou, Qingyi Tao, and Chen Change Loy. PGDiff: Guiding Diffusion Models for Versatile Face Restoration via Partial Guidance. Ad- vances in Neural Information Processing Systems , 36: 32194–32214, 2023. 4, 5, 7, 8, 9, 10

    2023

  60. [68]

    WIDER FACE: A Face Detection Benchmark

    Shuo Yang, Ping Luo, Chen-Change Loy, and Xiaoou Tang. WIDER FACE: A Face Detection Benchmark. In Proceed- ings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5525–5533, 2016. 3, 10

    2016

  61. [69]

    DifFace: Blind Face Restoration with Diffused Error Contraction

    Zongsheng Yue and Chen Change Loy. DifFace: Blind Face Restoration with Diffused Error Contraction. (arXiv:2212.06512), 2023. 2, 3, 4, 5, 7, 8, 9, 10

    2023 arXiv

  62. [70]

    Restormer: Efficient Transformer for High-Resolution Im- age Restoration

    Syed Waqas Zamir, Aditya Arora, Salman Khan, Mu- nawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient Transformer for High-Resolution Im- age Restoration. In Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition , pages 5728–...

    2022

  63. [71]

    Beyond a Gaussian Denoiser: Residual Learning of Deep CNN for Image Denoising

    Kai Zhang, Wangmeng Zuo, Yunjin Chen, Deyu Meng, and Lei Zhang. Beyond a Gaussian Denoiser: Residual Learning of Deep CNN for Image Denoising. IEEE Transactions on Image Processing, 26(7):3142–3155, 2017. 1

    2017

  64. [72]

    Designing a Practical Degradation Model for Deep Blind Image Super-Resolution

    Kai Zhang, Jingyun Liang, Luc Van Gool, and Radu Timo- fte. Designing a Practical Degradation Model for Deep Blind Image Super-Resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4791– 4800, 2021. 2

    2021

  65. [73]

    Efros, Eli Shecht- man, and Oliver Wang

    Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shecht- man, and Oliver Wang. The Unreasonable Effectiveness of Deep Features as a Perceptual Metric. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recogni- tion, pages 586–595, 2018. 2, 5, 3

    2018

  66. [74]

    Learning to restore hazy video: A new real-world dataset and a new method

    Xinyi Zhang, Hang Dong, Jinshan Pan, Chao Zhu, Ying Tai, Chengjie Wang, Jilin Li, Feiyue Huang, and Fei Wang. Learning to restore hazy video: A new real-world dataset and a new method. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, page...

    2021

  67. [75]

    Towards Robust Blind Face Restoration with Codebook Lookup Transformer

    Shangchen Zhou, Kelvin Chan, Chongyi Li, and Chen Change Loy. Towards Robust Blind Face Restoration with Codebook Lookup Transformer. Advances in Neural Information Processing Systems, 35:30599–30611, 2022. 2, 3, 6, 8, 1, 4, 5, 7, 9, 10

    2022

  68. [76]

    Denoising Dif- fusion Models for Plug-and-Play Image Restoration

    Yuanzhi Zhu, Kai Zhang, Jingyun Liang, Jiezhang Cao, Bi- han Wen, Radu Timofte, and Luc Van Gool. Denoising Dif- fusion Models for Plug-and-Play Image Restoration. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1219–1229, 2023. 2,...

    2023

  69. [77]

    (23) and ⟨ ˆX, X∗⟩ in Eq

    Since X ∗ = ˜X ∗ − R, it holds that ∥ ˆX − X ∗∥2 2 = ∥ ˆX∥2 2 + ∥X ∗∥2 2 − 2⟨ ˆX, X∗⟩ (22) = ∥ ˆX∥2 2 + ∥ ˜X ∗∥2 2 + ∥R∥2 2 − 2⟨ ˜X ∗, R⟩ −2⟨ ˆX, X∗⟩ (23) = ∥ ˆX∥2 2 + ∥ ˜X ∗∥2 2 + ∥R∥2 2 − 2⟨ ˜X ∗, R⟩ −2⟨ ˆX, ˜X ∗⟩ + 2⟨ ˆX, R⟩, (24) where we expanded ∥X ∗∥2 2 in Eq. (23) and ...

  70. [78]

    Uniform Dist. #1

    (31) This is equivalent to an MSE between flattened feature vectors. Denote by z = [ vec(z1), . . . ,vec(zL)]⊤, where zl = 1√HlWl wl ⊙ fl, then ∆LPIPS(x, ˆx) = ∥z − ˆz∥2 2 = ∆SE(z, ˆz). (32) E. Implementation details E.1. Degradation estimator Our degradation estimator consist...

  71. [79]

    Datasets analysis & synthesis F .1

    // compute effective noise 9 xt−1 = DDIMStep(ˆxt 0, ˆε, η) // perform DDIM step 10 end F. Datasets analysis & synthesis F .1. Real-world datasets analysis Prior work [75] considered the degradations in LFW simpler than those in WebPhoto and WIDER. However, they could not justi...

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.