Pith. sign in

REVIEW 3 major objections 7 minor 70 references

Incorporating Uncertainty-Guided and Top-k Codebook Matching for Real-World Blind Image Super-Resolution

T0 review · 3 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read UGTSR claims that uncertainty-guided top-k codebook matching lets codebook-based super-resolution beat GAN and diffusion methods on perceptual quality while staying far lighter.

desk verdict Plausible engineering contribution to codebook-based blind SR, but the 'best on all five datasets' claim contradicts its own Table I; needs correction and code release before acceptance. read the letter →

arxiv 2506.07809 v1 pith:QBI4AR3Q submitted 2025-06-09 cs.CV

Weilei Wen , Tianyi Zhang , Qianqian Zhao , Zhaohui Zheng , Chunle Guo , Xiuli Shao , Chongyi Li This is my paper
classification cs.CV
keywords real-worldimagesuper-resolutioncodebookprioruncertainty-guidedlearningtop-kfeaturematchingalign-attentionblindperceptualqualityvectorquantization
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

This paper proposes UGTSR, a two-stage framework for real-world blind image super-resolution built on a pre-trained codebook of high-quality texture features. It claims that three fixes remove the usual trade-offs of codebook-based SR: an uncertainty map focuses reconstruction effort on texture-rich regions, a top-k matching step fuses the three closest codebook entries instead of committing to a single nearest neighbor, and an attention module aligns low-resolution features with their high-resolution counterparts before decoding. On the paper's reported benchmarks, this combination achieves the best overall LPIPS across all five synthetic datasets for both 2x and 4x super-resolution and the best no-reference score on RealSet65, beating GAN- and diffusion-based rivals. The practical interest is that this perceptual quality is reached without diffusion models' long sampling times and without the painterly artifacts of earlier codebook-weighting methods.

What carries the argument

The machinery is a two-stage uncertainty-guided codebook pipeline. Stage I attaches a variance branch to the FeMaSR codebook reconstructor and trains with $L_{\mathrm{ESU}}=\frac{1}{N}\sum_i \exp(-s_i)\|x_i-f(y_i)\|_1+2s_i$, where $s_i=\ln\theta_i$ is the predicted log-variance, producing an uncertainty map over texture-rich regions. Stage II keeps the codebook and decoder fixed and trains the SR generator with $L_{\mathrm{UDL}}=\frac{1}{N}\sum_i \hat{s}_i\|x_i-f(y_i)\|_1$, with $\hat{s}_i$ the non-negative shifted uncertainty, so high-uncertainty pixels contribute more. Around this core sit the Top-k matching strategy (find the $k$ nearest codebook vectors, fuse them with lightweight attention, then re-quantize) and the Align-Attention module (low-resolution feature as query, high-resolution codebook feature as key and value). The paper's quantitative engine is the measured rise in matching accuracy from about 21% (top-1) to 42% (top-3) to 54% (top-5), which motivates the whole Top-k design.

What would settle it

Look at the top-k distances on a validation set: if high-uncertainty pixels are exactly those where even the fifth nearest codebook vector is far from the low-resolution feature, the map is encoding codebook coverage gaps, not recoverable texture; then reweighting those pixels should hurt rather than help, and a control using a fixed edge map would replicate the reported gains.

Watch

Extended reading notes

Core claim

The paper's central claim is that the bottleneck in codebook-based real-world super-resolution is not the codebook itself but how the model looks up and fuses it. UGTSR treats nearest-neighbour matching as an unreliable first step: measuring the FeMaSR baseline, the correct code appears among the top three candidates in about 42% of cases and among the top five in about 54%, versus about 21% for the single nearest neighbour. Fusing the top-k candidates before re-quantizing, weighting the reconstruction loss by a learned uncertainty map, and aligning low- and high-resolution features in the decoder jointly let a vector-quantized codebook model beat diffusion-based methods on LPIPS while staying lightweight. The paper reports the best overall performance on all five synthetic benchmarks for both scales and state-of-the-art NIQE on RealSet65.

Load-bearing premise

The load-bearing premise is that the uncertainty map learned in stage I marks the places where extra reconstruction effort will make textures more realistic, rather than simply the places where the stored set of texture examples has no good match.

Editorial extensions

If this is right

  • Codebook-based SR can be positioned as the perceptual-quality leader in blind SR: the reported LPIPS numbers beat diffusion baselines at both scales, so the usual assumption that diffusion is required for realistic texture is challenged.
  • For any codebook-based restoration model, replacing single nearest-neighbour lookup with top-k fusion and re-quantization is a cheap accuracy lever, since it raises the chance of hitting the correct code from about 21% toward 42% without a Transformer's quadratic cost.
  • The uncertainty map provides a free, annotation-free saliency for texture-rich regions, which the loss reweights; the same stage-I and stage-II recipe can be reused on top of any frozen codebook reconstructor.
  • The ablations tie every module to a measurable gain, with uncertainty learning producing the largest single LPIPS improvement; the full system depends on all three rather than any one trick.

Reading between the lines

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

  • Editorial extension: the measured top-1 versus top-k accuracy gap suggests the codebook prior itself is under-used at lookup time; an orthogonal extension would be to enlarge or adapt the codebook per dataset and re-measure whether the 54% ceiling moves.
  • Editorial extension: a clean control experiment would replace the learned uncertainty map with a fixed high-frequency or edge map; if the LPIPS gains persist, the active ingredient is region weighting rather than uncertainty estimation, and stage I could be dropped.
  • Editorial extension: the RealSet65 NIQE gains imply the method suppresses real degradations better than the compared baselines; a natural next test is a human perceptual study on the returned photographs, since NIQE is a proxy rather than a viewer.
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

3 major / 7 minor

Summary. The paper proposes UGTSR, a two-stage codebook-based blind super-resolution framework built on FeMaSR. Stage I estimates a sparse uncertainty map using an ESU loss; Stage II uses this map in a UDL loss to weight texture-rich regions, introduces a top-k nearest-neighbor codebook matching strategy with a fusion module, and adds an Align-Attention module that treats LQ features as queries and HQ codebook features as keys/values. The method is evaluated on synthetic benchmarks (Set5, Set14, BSD100, Urban100, Manga109, DIV2K-Valid) and real datasets (Canon, Nikon, RealSet65) against GAN-, diffusion-, and codebook-based baselines, with ablations and an efficiency comparison. The paper claims best overall performance on all five synthetic datasets at both 2x and 4x, and state-of-the-art results on real datasets.

Significance. If the quantitative claims were supported, the contribution would be moderate: a lightweight codebook-based RISR method that improves over FeMaSR while adding modest components. The framework is plausible and builds on standard formulations (Eqs. (2), (3), (6)-(9)); the authors include ablation studies, report using official baseline implementations, and state that code will be released. However, the headline empirical claim is not supported by the paper's own Table I, and the feature-matching accuracy evidence is internally inconsistent. The significance as stated therefore cannot be accepted without revision.

major comments (3)
  1. [Section IV-D, Table I] The claim that 'our method achieved the best overall performance on all five datasets for both 2x and 4x super-resolution tasks' is contradicted by Table I. On LPIPS, which the paper calls the primary perceptual metric, FeMaSR is better on BSD100 at 2x (0.3581 vs 0.3587), SeeSR is better on Urban100 at 2x (0.2610 vs 0.2867), DiffBIR is better on Set5 at 4x (0.2175 vs 0.2238), AdaCode is better on Set14 at 4x (0.3597 vs 0.3619), and SeeSR is better on Urban100 at 4x (0.3018 vs 0.3398). PSNR and SSIM also favor baselines in several rows, for example RealESRGAN on BSD100 2x (PSNR 23.505 vs 23.320). Moreover, no error bars or statistical tests are provided for any table, so the small margins in the remaining rows do not support 'best' claims. The empirical claims need to be reworded or supported with repeated runs and significance analysis.
  2. [Section III-B and Section IV-E, Figure 7] The reported feature-matching accuracy is internally inconsistent. Section III-B states that nearest-neighbor matching in FeMaSR has about 21% accuracy, Top-3 42%, and Top-5 54%, which corresponds to gains of roughly 21 and 33 percentage points. Section IV-E, however, says that replacing Top-1 with Top-3 and Top-5 led to 'approximate accuracy improvements of 3% and 4%, respectively,' and Figure 7 shows accuracy values in a much lower range. The discrepancy is too large to be a rounding issue; the authors should define accuracy precisely, report the raw numbers, and reconcile the claims.
  3. [Section III-A, Eqs. (2)-(3)] The uncertainty mechanism assumes that the Stage I uncertainty map identifies regions where Stage II should invest additional modeling capacity. But L_ESU is trained so that high-uncertainty pixels are downweighted, and the text itself notes that Stage I images are less sharp in rich-texture regions; Stage II then upweights exactly those pixels via L_UDL. The paper provides no check that high uncertainty corresponds to recoverable texture rather than to regions the fixed FeMaSR codebook simply cannot represent, in which case upweighting could amplify codebook-mismatch artifacts. Table V shows an overall gain for the full framework, but that experiment changes several components at once. I request a direct validation, such as comparing uncertainty maps with edge/texture-density maps, or a control experiment that upweights low-uncertainty regions.
minor comments (7)
  1. [Section IV-A] The text reads 'we develop a trining set'; this should be 'training set.'
  2. [Section IV-D] The claim that our method outperforms AdaCode 'by margins of 0.0147 and 0.149' should read 0.0149 for the 4x LPIPS margin according to Table II.
  3. [Section IV-E] The ablation text states that 'SSIM improved by 0.076,' but Table V shows an improvement of 0.0076 from the baseline to the uncertainty-added model.
  4. [Section III-B, Eq. (6)] The notation 'arg min_{j1,j2,j3}' should specify that the three indices are distinct and sorted; otherwise the argmin is not uniquely defined.
  5. [Section IV-E, Figure 8] The x-axis of Figure 8 is labeled 'Feature Size' while the text discusses growth with codebook size; please clarify which quantity is varied and report the codebook size and feature dimension.
  6. [Section II-B] RTCNet and ITER are discussed in the related work but do not appear in the experimental comparisons; please clarify why they are omitted.
  7. [Throughout] The term 'VQV AE' should be written as 'VQ-VAE' or 'VQVAE' for consistency with the standard notation.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the method is benchmarked against external baselines, and the overstated SOTA claim is an empirical inconsistency, not a derivation that reduces to its inputs.

full rationale

The paper's derivation chain is not circular. The three proposed components are (i) uncertainty-guided texture weighting via L_ESU and L_UDL (Eqs. 2 and 3), (ii) a Top-k codebook matching strategy motivated by measured retrieval statistics, and (iii) an Align-Attention module described in Eq. 9. None of these claims is defined in terms of the paper's own success metric. The uncertainty losses are explicitly credited to prior external work [51], and their usefulness is tested in Table V through ablations against the same FeMaSR baseline; the Top-k strategy is evaluated by reconstruction metrics and matching-accuracy measurements in Figure 7 and Table V; the Align-Attention module is isolated in the same ablation table. The core comparison is against external methods (AdaCode, FeMaSR, RealESRGAN, DiffBIR, SeeSR, SinSR, etc.) using official checkpoints on standard benchmarks. The only notable defect is that Section IV-D's claim that 'our method achieved the best overall performance on all five datasets for both 2x and 4x super-resolution tasks' is contradicted by Table I: for example, LPIPS on 2x BSD100 favors FeMaSR, 2x Urban100 favors SeeSR, and 4x Set5, 4x Set14, and 4x Urban100 favor DiffBIR, SinSR, AdaCode, and SeeSR. That is an evidentiary or correctness problem, not circularity. Self-citations to the authors' prior work appear only in background or related-work contexts and are not load-bearing premises; there is no fitted parameter renamed as a prediction, no uniqueness theorem imported from the authors, and no ansatz smuggled in solely by self-citation. The paper is therefore self-contained against external benchmarks, and the honest circularity finding is zero.

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

The framework inherits the frozen FeMaSR codebook and decoder and the uncertainty-loss formulation from prior work; these are assumed to be valid rather than re-derived. The manually chosen loss weights and the validation-selected k=3 are free parameters. No new physical entity is introduced; the uncertainty map and attention modules are internal network components. The central claim therefore stands on the FeMaSR prior, the transferability of uncertainty learning to codebook SR, and the loss-weight choices.

free parameters (4)
  • Top-k neighborhood size (k) = 3
    Authors compare k=1, 3, and 5 and choose k=3 in the final model ('Considering efficiency and performance, we ultimately adopted the Top-3 feature matching strategy'); the choice is validated on the same DIV2K-Valid split used for the headline comparison.
  • Gram matrix loss weight alpha = 1
    Set in Section III-D: 'the weights for each loss are set as: α=1, β=0.25, λadv=0.1'; no sensitivity analysis is provided.
  • Codebook L2 loss weight beta = 0.25
    Set in Section III-D with the other loss weights; no ablation over beta is reported.
  • Adversarial loss weight lambda_adv = 0.1
    Set in Section III-D with the other loss weights; no ablation over lambda_adv is reported.
assumptions (4)
  • domain assumption The frozen FeMaSR codebook and decoder provide a sufficient high-quality prior for real-world degraded images.
    UGTSR builds on FeMaSR and keeps the codebook Z and decoder fixed in both stages (Sections III-C and IV-C); if the prior cannot represent real-world textures, the added modules cannot recover them.
  • domain assumption Uncertainty values learned with L_ESU indicate where the SR model should concentrate its texture-generation capacity.
    This transfer from [51] is assumed in Section III-A, where the stage-I uncertainty map is reused as a weighting term in L_UDL; the paper does not re-derive or independently validate the link between ESU uncertainty and beneficial upweighting in the codebook setting.
  • domain assumption There is a well-defined ground-truth codebook code for each LR feature, so top-k accuracy (21%, 42%, 54%) is measurable.
    The motivation in Section III-B relies on counting how often the 'correct' code appears among the top-k neighbors; this presumes an oracle assignment of degraded features to HR codebook entries, which is nontrivial during blind degradation.
  • domain assumption Fusing the top-k nearest codes before re-quantization preserves or improves the information content of the matched representation.
    Equation (7) fuses Z_k1, Z_k2, Z_k3 and Equation (8) re-quantizes; the paper assumes the fused feature remains close to a valid codebook index and does not increase quantization error. This is empirically checked in Figure 7 but not analytically.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Incorporating Uncertainty-Guided and Top-k Codebook Matching for Real-World Blind Image Super-Resolution." pith.science (2026). https://pith.science/paper/QBI4AR3Q

@misc{pith2026250607809,
  author       = {Pith},
  title        = {Pith review of: Incorporating Uncertainty-Guided and Top-k Codebook Matching for Real-World Blind Image Super-Resolution},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QBI4AR3Q}},
  note         = {Machine review of arXiv:2506.07809}
}
read the original abstract

Recent advancements in codebook-based real image super-resolution (SR) have shown promising results in real-world applications. The core idea involves matching high-quality image features from a codebook based on low-resolution (LR) image features. However, existing methods face two major challenges: inaccurate feature matching with the codebook and poor texture detail reconstruction. To address these issues, we propose a novel Uncertainty-Guided and Top-k Codebook Matching SR (UGTSR) framework, which incorporates three key components: (1) an uncertainty learning mechanism that guides the model to focus on texture-rich regions, (2) a Top-k feature matching strategy that enhances feature matching accuracy by fusing multiple candidate features, and (3) an Align-Attention module that enhances the alignment of information between LR and HR features. Experimental results demonstrate significant improvements in texture realism and reconstruction fidelity compared to existing methods. We will release the code upon formal publication.

Figures

Figures reproduced from arXiv: 2506.07809 by the authors.

Figure 1
Figure 1. Qualitative comparison of three VQVAE-based methods on DIV2K [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. This paper presents UGTSR, a novel uncertainty-guided two-stage super-resolution framework. In the first stage, the model leverages a pre-trained [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Qualitative comparisons on DIV2K [32] dataset (2× SR). Our method demonstrates superior artifact removal and texture restoration compared to state-of-the-art approaches. Please zoom in for the best view. A. Real-World Super-Resolution In recent years, generative models such as Generative Ad￾versarial Networks (GAN [30]), diffusion models [18], and Vector Quantized Variational Autoencoders (VQVAE [24]) have attracted… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Qualitative comparisons on DIV2K [32] dataset (4× SR). Our method demonstrates superior artifact removal and texture restoration compared to state-of-the-art approaches. Please zoom in for the best view. quality and detail. VQGAN [25] introduced generative adver￾sarial…
Figure 5
Figure 5. Figure 5: Visual comparisons on RealSet65 [39] dataset under 4× SR. Our method achieves superior authenticity, effectively suppressing artifacts and reconstructing finer texture details compared to previous state-of-the-art approaches. Please zoom in for the best view. learning …
Figure 6
Figure 6. Figure 6: Ablation studies for 4× SR with different model variations on the DIV2K [32] dataset. Our method UGTSR with all the proposed modules outperforms other configurations. Top1 Top3 Top5 Top-k 18 20 22 24 26 28 30 Accuracy (%) FeMaSR FeMaSR UGTSR UGTSR Urban100 Manga109 [P…
Figure 7
Figure 7. Figure 7: Comparison of feature matching accuracy across different methods on [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 8
Figure 8. Figure 8: Computational complexity comparison between our proposed Top-k [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

70 extracted references · 64 canonical work pages

  1. [51]

    Uncertainty-driven loss for single image super-resolution,

    Q. Ning, W. Dong, X. Li, J. Wu, and G. Shi, “Uncertainty-driven loss for single image super-resolution,”Advances in Neural Information Processing Systems, vol. 34, pp. 16 398–16 409, 2021. 4, 5, 6, 10

    work page 2021

  2. [1]

    Enhanced deep residual networks for single image super-resolution,

    B. Lim, S. Son, H. Kim, S. Nah, and K. Mu Lee, “Enhanced deep residual networks for single image super-resolution,” inCVPRW, 2017, pp. 136–144. 1, 2

    work page 2017

  3. [2]

    Image super- resolution using very deep residual channel attention networks,

    Y . Zhang, K. Li, K. Li, L. Wang, B. Zhong, and Y . Fu, “Image super- resolution using very deep residual channel attention networks,” in ECCV, 2018, pp. 286–301. 1, 2

    work page 2018

  4. [3]

    Residual dense network for image super-resolution,

    Y . Zhang, Y . Tian, Y . Kong, B. Zhong, and Y . Fu, “Residual dense network for image super-resolution,” inCVPR, 2018, pp. 2472–2481. 1, 2

    work page 2018

  5. [4]

    Single image super-resolution via a holistic attention network,

    B. Niu, W. Wen, W. Ren, X. Zhang, L. Yang, S. Wang, K. Zhang, X. Cao, and H. Shen, “Single image super-resolution via a holistic attention network,” inECCV. Springer, 2020, pp. 191–207. 1

    work page 2020

  6. [5]

    Adaptive blind super-resolution network for spatial-specific and spatial-agnostic degra- dations,

    W. Wen, C. Guo, W. Ren, H. Wang, and X. Shao, “Adaptive blind super-resolution network for spatial-specific and spatial-agnostic degra- dations,”IEEE TIP, vol. 33, pp. 4404–4418, 2024. 1

    work page 2024

  7. [6]

    Second-order attention network for single image super-resolution,

    T. Dai, J. Cai, Y . Zhang, S.-T. Xia, and L. Zhang, “Second-order attention network for single image super-resolution,” inCVPR, 2019, pp. 11 065–11 074. 1, 2

    work page 2019

  8. [7]

    Swinir: Image restoration using swin transformer,

    J. Liang, J. Cao, G. Sun, K. Zhang, L. Van Gool, and R. Timofte, “Swinir: Image restoration using swin transformer,” inICCVW, 2021. 1, 2

    work page 2021

Show all 70 references
  1. [8]

    Efficient long-range attention network for image super-resolution,

    X. Zhang, H. Zeng, S. Guo, and L. Zhang, “Efficient long-range attention network for image super-resolution,” inECCV. Springer, 2022, pp. 649–667. 1, 2

    2022

  2. [9]

    Dual aggregation transformer for image super-resolution,

    Z. Chen, Y . Zhang, J. Gu, L. Kong, X. Yang, and F. Yu, “Dual aggregation transformer for image super-resolution,” inICCV, 2023, pp. 12 312–12 321. 1, 2

    2023

  3. [10]

    Real-esrgan: Training real- world blind super-resolution with pure synthetic data,

    X. Wang, L. Xie, C. Dong, and Y . Shan, “Real-esrgan: Training real- world blind super-resolution with pure synthetic data,”ICCVW, 2021. 1, 3, 6, 8, 9

    2021

  4. [11]

    Desra: detect and delete the artifacts of gan-based real-world super-resolution models,

    L. Xie, X. Wang, X. Chen, G. Li, Y . Shan, J. Zhou, and C. Dong, “Desra: detect and delete the artifacts of gan-based real-world super-resolution models,”arXiv preprint arXiv:2307.02457, 2023. 1, 3

    2023 arXiv

  5. [12]

    Designing a practical degradation model for deep blind image super-resolution,

    K. Zhang, J. Liang, L. Van Gool, and R. Timofte, “Designing a practical degradation model for deep blind image super-resolution,”ICCV, 2021. 1, 3, 6, 8

    2021

  6. [13]

    Details or artifacts: A locally discriminative learning approach to realistic image super-resolution,

    J. Liang, H. Zeng, and L. Zhang, “Details or artifacts: A locally discriminative learning approach to realistic image super-resolution,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 5657–5666. 1, 3, 6, 8, 9

    2022

  7. [14]

    Perceptual losses for real-time style transfer and super-resolution,

    J. Johnson, A. Alahi, and L. Fei-Fei, “Perceptual losses for real-time style transfer and super-resolution,” inComputer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11- 14, 2016, Proceedings, Part II 14. Springer, 2016, pp. 694–711. 1, 3, 8

    2016

  8. [15]

    Generative adversarial nets,

    I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y . Bengio, “Generative adversarial nets,” Advances in neural information processing systems, vol. 27, 2014. 1, 8

    2014

  9. [16]

    Learning image-adaptive codebooks for class-agnostic image restoration,

    K. Liu, Y . Jiang, I. Choi, and J. Gu, “Learning image-adaptive codebooks for class-agnostic image restoration,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 5373–5383. 1, 4, 5, 6, 7, 8, 9

    2023

  10. [17]

    Real-world blind super-resolution via feature matching with implicit high-resolution priors,

    C. Chen, X. Shi, Y . Qin, X. Li, X. Han, T. Yang, and S. Guo, “Real-world blind super-resolution via feature matching with implicit high-resolution priors,” inACM MM, 2022. 1, 2, 4, 5, 6, 7, 8, 9

    2022

  11. [18]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” NeurIPS, vol. 33, pp. 6840–6851, 2020. 1, 3

    2020

  12. [19]

    High- resolution image synthesis with latent diffusion models,

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer, “High- resolution image synthesis with latent diffusion models,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 10 684–10 695. 1

    2022

  13. [20]

    Exploiting diffusion prior for real-world image super-resolution,

    J. Wang, Z. Yue, S. Zhou, K. C. Chan, and C. C. Loy, “Exploiting diffusion prior for real-world image super-resolution,”arXiv preprint arXiv:2305.07015, 2023. 1, 3

    2023 arXiv

  14. [21]

    Diffbir: Towards blind image restoration with generative diffusion prior,

    X. Lin, J. He, Z. Chen, Z. Lyu, B. Fei, B. Dai, W. Ouyang, Y . Qiao, and C. Dong, “Diffbir: Towards blind image restoration with generative diffusion prior,”arXiv preprint arXiv:2308.15070, 2023. 1, 3, 6, 9

    2023 arXiv

  15. [22]

    Pixel-aware stable diffusion for realistic image super-resolution and personalized stylization,

    T. Yang, P. Ren, X. Xie, and L. Zhang, “Pixel-aware stable diffusion for realistic image super-resolution and personalized stylization,”arXiv preprint arXiv:2308.14469, 2023. 1, 3

    2023 arXiv

  16. [23]

    Seesr: Towards semantics-aware real-world image super-resolution,

    R. Wu, T. Yang, L. Sun, Z. Zhang, S. Li, and L. Zhang, “Seesr: Towards semantics-aware real-world image super-resolution,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024, pp. 25 456–25 467. 1, 3, 6, 8, 9 12 IEEE TRANSACTIONS ON IMAGE ...

    2024

  17. [24]

    Neural discrete representation learning,

    A. v. d. Oord, O. Vinyals, and K. Kavukcuoglu, “Neural discrete representation learning,”NeurIPS, 2017. 1, 3

    2017

  18. [25]

    Taming transformers for high- resolution image synthesis,

    P. Esser, R. Rombach, and B. Ommer, “Taming transformers for high- resolution image synthesis,” inCVPR, 2021, pp. 12 873–12 883. 1, 4

    2021

  19. [26]

    Towards robust blind face restoration with codebook lookup transformer,

    S. Zhou, K. C. Chan, C. Li, and C. C. Loy, “Towards robust blind face restoration with codebook lookup transformer,” inNeurIPS, 2022. 1, 2, 4, 7, 11

    2022

  20. [27]

    Iterative token evaluation and refinement for real-world super-resolution,

    C. Chen, S. Zhou, L. Liao, H. Wu, W. Sun, Q. Yan, and W. Lin, “Iterative token evaluation and refinement for real-world super-resolution,” in Proceedings of the AAAI Conference on Artificial Intelligence, 2024. 1, 4

    2024

  21. [28]

    Lar-sr: A local autoregressive model for image super-resolution,

    B. Guo, X. Zhang, H. Wu, Y . Wang, Y . Zhang, and Y .-F. Wang, “Lar-sr: A local autoregressive model for image super-resolution,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 1909–1918. 1

    2022

  22. [29]

    Vqfr: Blind face restoration with vector-quantized dictionary and parallel decoder,

    Y . Gu, X. Wang, L. Xie, C. Dong, G. Li, Y . Shan, and M.-M. Cheng, “Vqfr: Blind face restoration with vector-quantized dictionary and parallel decoder,” inEuropean Conference on Computer Vision. Springer, 2022, pp. 126–143. 1, 2, 4, 7, 11

    2022

  23. [30]

    Generative adversarial networks,

    I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y . Bengio, “Generative adversarial networks,” Communications of the ACM, vol. 63, no. 11, pp. 139–144, 2020. 1, 3

    2020

  24. [31]

    Image super-resolution using deep convolutional networks,

    C. Dong, C. C. Loy, K. He, and X. Tang, “Image super-resolution using deep convolutional networks,”IEEE transactions on pattern analysis and machine intelligence, vol. 38, no. 2, pp. 295–307, 2015. 2

    2015

  25. [32]

    NTIRE 2017 challenge on single image super-resolution: Dataset and study,

    E. Agustsson and R. Timofte, “NTIRE 2017 challenge on single image super-resolution: Dataset and study,” inCVPRW, July 2017. 3, 4, 8, 9, 10

    2017

  26. [33]

    Denoising diffusion restoration models,

    B. Kawar, M. Elad, S. Ermon, and J. Song, “Denoising diffusion restoration models,”Advances in Neural Information Processing Sys- tems, vol. 35, pp. 23 593–23 606, 2022. 3

    2022

  27. [34]

    Image super-resolution via iterative refinement,

    C. Saharia, J. Ho, W. Chan, T. Salimans, D. J. Fleet, and M. Norouzi, “Image super-resolution via iterative refinement,”IEEE transactions on pattern analysis and machine intelligence, vol. 45, no. 4, pp. 4713–4726,

  28. [35]

    Generating diverse high- fidelity images with vq-vae-2,

    A. Razavi, A. van den Oord, and O. Vinyals, “Generating diverse high- fidelity images with vq-vae-2,” inNeurIPS, 2019, pp. 14 866–14 876. 3

    2019

  29. [36]

    Low-light image enhancement with multi-stage residue quantization and brightness-aware attention,

    Y . Liu, T. Huang, W. Dong, F. Wu, X. Li, and G. Shi, “Low-light image enhancement with multi-stage residue quantization and brightness-aware attention,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 12 140–12 149. 4

    2023

  30. [37]

    Ridcp: Re- vitalizing real image dehazing via high-quality codebook priors,

    R.-Q. Wu, Z.-P. Duan, C.-L. Guo, Z. Chai, and C. Li, “Ridcp: Re- vitalizing real image dehazing via high-quality codebook priors,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 22 282–22 291. 4

    2023

  31. [38]

    Blind image super- resolution with rich texture-aware codebook,

    R. Qin, M. Sun, F. Zhang, X. Wen, and B. Wang, “Blind image super- resolution with rich texture-aware codebook,” inProceedings of the 31st ACM International Conference on Multimedia, 2023, pp. 676–687. 4

    2023

  32. [39]

    Resshift: Efficient diffusion model for image super-resolution by residual shifting,

    Z. Yue, J. Wang, and C. C. Loy, “Resshift: Efficient diffusion model for image super-resolution by residual shifting,”arXiv preprint arXiv:2307.12348, 2023. 5, 6, 8, 9

    2023 arXiv

  33. [40]

    Regression with input- dependent noise: A gaussian process treatment,

    P. Goldberg, C. Williams, and C. Bishop, “Regression with input- dependent noise: A gaussian process treatment,”Advances in neural information processing systems, vol. 10, 1997. 4

    1997

  34. [41]

    Bayesian approach to neural-network modeling with input uncertainty,

    W. Wright, “Bayesian approach to neural-network modeling with input uncertainty,”IEEE Transactions on Neural Networks, vol. 10, no. 6, pp. 1261–1270, 1999. 4

    1999

  35. [42]

    Regression with input-dependent noise: A bayesian treatment,

    C. Bishop and C. Quazaz, “Regression with input-dependent noise: A bayesian treatment,”Advances in neural information processing systems, vol. 9, 1996. 4

    1996

  36. [43]

    What uncertainties do we need in bayesian deep learning for computer vision?

    A. Kendall and Y . Gal, “What uncertainties do we need in bayesian deep learning for computer vision?”Advances in neural information processing systems, vol. 30, 2017. 4, 5

    2017

  37. [44]

    Segnet: A deep con- volutional encoder-decoder architecture for image segmentation,

    V . Badrinarayanan, A. Kendall, and R. Cipolla, “Segnet: A deep con- volutional encoder-decoder architecture for image segmentation,”IEEE transactions on pattern analysis and machine intelligence, vol. 39, no. 12, pp. 2481–2495, 2017. 4, 5

    2017

  38. [45]

    Data uncertainty learning in face recognition,

    J. Chang, Z. Lan, C. Cheng, and Y . Wei, “Data uncertainty learning in face recognition,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 5710–5719. 4, 5

    2020

  39. [46]

    Active learning combining uncertainty and diversity for multi-class image classification,

    Y . Gu, Z. Jin, and S. C. Chiu, “Active learning combining uncertainty and diversity for multi-class image classification,”IET Computer Vision, vol. 9, no. 3, pp. 400–407, 2015. 4, 6

    2015

  40. [47]

    Probabilistic face embeddings,

    Y . Shi and A. K. Jain, “Probabilistic face embeddings,” inProceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 6902–6911. 4

    2019

  41. [48]

    Gram: Gradient rescaling attention model for data uncertainty estimation in single image super resolution,

    C. Lee and K.-S. Chung, “Gram: Gradient rescaling attention model for data uncertainty estimation in single image super resolution,” in 2019 18th IEEE International Conference On Machine Learning And Applications (ICMLA). IEEE, 2019, pp. 8–13. 4

    2019

  42. [49]

    Spatial-wise and channel-wise feature uncertainty for occluded person re-identification,

    Y . Shi, W. Tian, H. Ling, Z. Li, and P. Li, “Spatial-wise and channel-wise feature uncertainty for occluded person re-identification,”Neurocomput- ing, vol. 486, pp. 237–249, 2022. 4

    2022

  43. [50]

    Learning degra- dation uncertainty for unsupervised real-world image super-resolution

    Q. Ning, J. Tang, F. Wu, W. Dong, X. Li, and G. Shi, “Learning degra- dation uncertainty for unsupervised real-world image super-resolution.” inIJCAI, 2022, pp. 1261–1267. 4, 5

    2022

  44. [52]

    Uncertainty-guided transformer reasoning for camouflaged object detection,

    F. Yang, Q. Zhai, X. Li, R. Huang, A. Luo, H. Cheng, and D.-P. Fan, “Uncertainty-guided transformer reasoning for camouflaged object detection,” inProceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 4146–4155. 4

    2021

  45. [53]

    Sinsr: diffusion-based image super- resolution in a single step,

    Y . Wang, W. Yang, X. Chen, Y . Wang, L. Guo, L.-P. Chau, Z. Liu, Y . Qiao, A. C. Kot, and B. Wen, “Sinsr: diffusion-based image super- resolution in a single step,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024, pp. 25 796–25 805. 6, 9

    2024

  46. [54]

    Taming transformers for high- resolution image synthesis,

    P. Esser, R. Rombach, and B. Ommer, “Taming transformers for high- resolution image synthesis,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 12 873–12 883. 7

    2021

  47. [55]

    Restormer: Efficient transformer for high-resolution image restoration,

    S. W. Zamir, A. Arora, S. Khan, M. Hayat, F. S. Khan, and M.-H. Yang, “Restormer: Efficient transformer for high-resolution image restoration,” inCVPR, 2022. 7

    2022

  48. [56]

    Ntire 2017 challenge on single image super-resolution: Methods and results,

    R. Timofte, E. Agustsson, L. Van Gool, M.-H. Yang, and L. Zhang, “Ntire 2017 challenge on single image super-resolution: Methods and results,” inCVPRW, 2017, pp. 114–125. 8

    2017

  49. [57]

    Div8k: Diverse 8k resolution image dataset,

    S. Gu, A. Lugmayr, M. Danelljan, M. Fritsche, J. Lamour, and R. Tim- ofte, “Div8k: Diverse 8k resolution image dataset,” inICCVW. IEEE, 2019, pp. 3512–3516. 8

    2019

  50. [58]

    A style-based generator architecture for generative adversarial networks,

    T. Karras, S. Laine, and T. Aila, “A style-based generator architecture for generative adversarial networks,” inCVPR, 2019, pp. 4401–4410. 8

    2019

  51. [59]

    Low- complexity single-image super-resolution based on nonnegative neighbor embedding,

    M. Bevilacqua, A. Roumy, C. Guillemot, and M. L. Alberi-Morel, “Low- complexity single-image super-resolution based on nonnegative neighbor embedding,”British Machine Vision Conference, 2012. 8

    2012

  52. [60]

    On single image scale-up using sparse-representations,

    R. Zeyde, M. Elad, and M. Protter, “On single image scale-up using sparse-representations,” inInternational conference on curves and sur- faces, 2010, pp. 711–730. 8

    2010

  53. [61]

    A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics,

    D. Martin, C. Fowlkes, D. Tal, and J. Malik, “A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics,” inICCV, vol. 2, 2001, pp. 416–423. 8

    2001

  54. [62]

    Single image super-resolution from transformed self-exemplars,

    J.-B. Huang, A. Singh, and N. Ahuja, “Single image super-resolution from transformed self-exemplars,” inProceedings of the IEEE confer- ence on computer vision and pattern recognition, 2015, pp. 5197–5206. 8

    2015

  55. [63]

    Sketch-based manga retrieval using manga109 dataset,

    Y . Matsui, K. Ito, Y . Aramaki, A. Fujimoto, T. Ogawa, T. Yamasaki, and K. Aizawa, “Sketch-based manga retrieval using manga109 dataset,” Multimedia Tools and Applications, vol. 76, no. 20, pp. 21 811–21 838,

  56. [64]

    Towards real-world blind face restoration with generative facial prior,

    X. Wang, Y . Li, H. Zhang, and Y . Shan, “Towards real-world blind face restoration with generative facial prior,” inCVPR, 2021, pp. 9168–9178. 8

    2021

  57. [65]

    Image quality assessment: from error visibility to structural similarity,

    Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli, “Image quality assessment: from error visibility to structural similarity,”IEEE TIP, vol. 13, no. 4, pp. 600–612, 2004. 9

    2004

  58. [66]

    The unreasonable effectiveness of deep features as a perceptual metric,

    R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang, “The unreasonable effectiveness of deep features as a perceptual metric,” in CVPR, 2018. 9

    2018

  59. [67]

    Image quality assessment: Unifying structure and texture similarity,

    K. Ding, K. Ma, S. Wang, and E. P. Simoncelli, “Image quality assessment: Unifying structure and texture similarity,”IEEE TPAMI, vol. 44, no. 5, pp. 2567–2581, 2020. 9

    2020

  60. [68]

    Gans trained by a two time-scale update rule converge to a local nash equilibrium,

    M. Heusel, H. Ramsauer, T. Unterthiner, B. Nessler, and S. Hochreiter, “Gans trained by a two time-scale update rule converge to a local nash equilibrium,”Advances in neural information processing systems, vol. 30, 2017. 9

    2017

  61. [69]

    Making a “completely blind

    A. Mittal, R. Soundararajan, and A. C. Bovik, “Making a “completely blind” image quality analyzer,”IEEE Sign. Process. Letters, vol. 20, no. 3, pp. 209–212, 2012. 9

    2012

  62. [70]

    Adam: A method for stochastic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” arXiv preprint arXiv:1412.6980, 2014. 9

    2014 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.