Pith. sign in

REVIEW 5 major objections 6 minor 57 references

Leveraging Vision-Language Models to Select Trustworthy Super-Resolution Samples Generated by Diffusion Models

T0 review · 5 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The authors claim that vision-language models can pick the single trustworthy output from a diffusion model's many super-resolution candidates, and that a new hybrid Trustworthiness Score confirms the choice without ground truth.

desk verdict Useful VLM-driven sample selection for diffusion SR, but the TWS validation is partly circular and the metric requires ground truth despite claiming otherwise. read the letter →

arxiv 2506.20832 v1 pith:CBYVL3Q3 submitted 2025-06-25 cs.CV cs.AI

Cansu Korkmaz , Ahmet Murat Tekalp , Zafer Dogan This is my paper
classification cs.CVcs.AI
keywords super-resolutiondiffusionmodelsvision-languagesampleselectiontrustworthinessscoreimagequalityassessmentsemanticfidelityhumanevaluation
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 a fully automated pipeline for choosing one trustworthy high-resolution image from the many super-resolution candidates a diffusion model generates. It asks vision-language models to rank candidates by semantic correctness, artifact presence, and visual appeal, then averages the top-ranked few into a single output. To verify those selections, it introduces the Trustworthiness Score (TWS), a weighted combination of CLIP semantic similarity, edge-structure similarity, and a wavelet-based artifact penalty. The authors claim that TWS tracks human preference and that VLM-selected samples score higher on TWS than random diffusion samples or standard regressive and GAN-based super-resolution outputs. If true, this would replace costly human evaluation with a scalable automated judge for information-critical super-resolution.

What carries the argument

The mechanism has two parts. First, the selection pipeline: vision-language models (BLIP-2, GPT-4o, and variants) receive structured prompts asking what digit or letter is shown, whether artifacts are visible, and how visually appealing each candidate is; the top-ranked candidates are then averaged into one output. Second, the verification metric, the Trustworthiness Score, defined as $$T(I_{HR}, I_{SR}) = \lambda_{CLIP} S_{CLIP} + \lambda_{edge} S_{edge} - \lambda_{wavelet} S_{wavelet},$$ where $S_{CLIP}$ is the cosine similarity of CLIP embeddings of the high-resolution reference and the super-resolved image, $S_{edge}$ is SSIM computed on Sobel edge maps, and $S_{wavelet}$ is the normalized $\ell^1$ energy of high-frequency sub-bands from a two-level Daubechies-19 wavelet decomposition. The weights are $\lambda_{CLIP}=0.2$, $\lambda_{edge}=0.3$, and $\lambda_{wavelet}=0.5$, fixed after a targeted search on one representative image. TWS is the load-bearing validator: the paper uses it to argue that VLM-selected ensembles are more trustworthy than random diffusion samples and than conventional SR baselines.

What would settle it

On a held-out set of ambiguous digit and natural-image inputs, compute Spearman's rank correlation between TWS and human top-sample choices with the paper's weights fixed; if TWS does not rank the human-preferred samples higher than PSNR or LPIPS do, the claim that TWS verifies VLM-selected trustworthiness collapses.

Watch

Extended reading notes

Core claim

The paper's central claim is that trustworthiness in diffusion-model super-resolution can be automated: a vision-language model, prompted with questions about the digit or letter shown, visible artifacts, and visual appeal, ranks the candidate set, and pixel-averaging the top-ranked candidates yields one output that preserves critical semantic information while suppressing hallucinations. The authors state that these VLM-guided selections consistently achieve high values of their Trustworthiness Score, matching or exceeding the human-guided baseline, and that TWS itself agrees with human preference on ambiguous and natural images where conventional metrics such as PSNR, LPIPS, and DISTS do not. The proposed pipeline, named LDM-VLM, is positioned as a new benchmark for trustworthy generative super-resolution.

Load-bearing premise

The framework's validation rests on TWS being a faithful proxy for human judgment, but the weights were tuned on a single image and no direct TWS-to-human correlation is reported, so if TWS disagrees with human preference the central claim loses its evidence.

Editorial extensions

If this is right

  • Diffusion-based super-resolution systems could be deployed without human-in-the-loop selection: a VLM prompt ensemble picks the top few candidates and pixel-averaging fuses them into a single output.
  • TWS provides a numeric score for SR trustworthiness that the paper claims reflects whether critical information, such as a digit, survived reconstruction, unlike PSNR, LPIPS, and DISTS.
  • Because selection is prompt-driven rather than trained, the framework can transfer to new vision-language models and new SR generators without retraining, at the cost of prompt quality.
  • For ambiguous inputs where regressive and GAN models output the wrong digit, the VLM-selected ensemble resolves the digit correctly in the reported examples.
  • The claims imply a new evaluation protocol: judge super-resolution by semantic and artifact-aware criteria rather than average pixel or feature distances.

Reading between the lines

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

  • The paper asserts that TWS correlates strongly with human preference but does not report a correlation coefficient; a direct head-to-head study between TWS ranking and human mean opinion scores would be the natural next test.
  • Because TWS decomposes into semantic, structural, and artifact terms, it could be repurposed as a reward signal for guiding diffusion sampling toward trustworthy outputs, not just for selecting from a fixed candidate set.
  • The framework's extension to medical or remote-sensing imagery is plausible but untested; VLM reasoning about digits and natural scenes may not transfer to specialized artifacts or anatomical structures.
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

5 major / 6 minor

Summary. The paper proposes an automated framework for selecting a single trustworthy super-resolution (SR) output from a set of diffusion-generated samples. The selection is performed by vision-language models (BLIP, BLIP-2, GPT-4o, GPT-4o-mini, Llama OCR, O1) prompted with structured queries about digit identity, artifacts, and visual appeal; the top-ranked samples are then ensembled. To validate the selections, the authors introduce a Trustworthiness Score (TWS) that combines CLIP-based semantic similarity, edge-based SSIM, and a wavelet-based artifact term, with weights fixed after a search on one image. Experiments cover MNIST digit identification, Urban100 '45' identification, DIV2K natural-image SR, prompt robustness, a small human evaluation, and a GPT-4o Mini versus MOS correlation on KADID-10k. The central claims are that VLM-selected samples achieve high TWS and that TWS correlates strongly with human preference.

Significance. If the central claims were established, the framework would be a useful, scalable alternative to human sample selection for diffusion-based SR, particularly for information-critical tasks such as digit or character recognition. The paper deserves credit for including human evaluations, a prompt-robustness analysis with a prompt pool, and statistical tests in Appendix B; these are positive elements that go beyond a purely algorithmic contribution. However, the significance is currently conditional: the validation of VLM selections rests on TWS, and TWS is (i) defined as a full-reference metric despite no-reference claims, (ii) fitted on a single representative image, and (iii) never directly correlated with human judgments in any reported experiment. The independent human-preference comparisons in Tables I-II and Table IV provide some support for VLM alignment, but they are small and lack error bars. Therefore the paper's main evidence for 'trustworthiness' is not yet convincing.

major comments (5)
  1. [III-C] The TWS definition is full-reference, not no-reference. Equation T(I_HR, I_SR) uses S_CLIP = cosine(CLIP(I_HR), CLIP(I_SR)) and S_edge = SSIM(E(I_HR), E(I_SR)), both requiring the ground-truth HR image. Yet the Introduction and Section III-C claim that TWS is usable 'in the absence of ground truth' and 'without requiring human-labeled ground truth.' Either provide and validate a genuine no-reference variant of TWS, or remove these claims.
  2. [IV-C] The abstract states that 'TWS correlates strongly with human preference,' but no experiment in Section IV computes a correlation between TWS and human judgments. The reported 0.71 correlation is between GPT-4o Mini scores and MOS on KADID-10k, which validates the VLM as an IQA proxy, not TWS as a trustworthiness oracle. Table V is an ablation that compares TWS values under different lambda configurations only; it never compares TWS with human preference or with digit-recognition correctness. Please report a direct TWS-human correlation (for example, Spearman correlation on the human selections in Tables I-II or on the natural-image preference data), or soften the claim accordingly.
  3. [III-C and IV-E] The TWS weights lambda_CLIP=0.2, lambda_edge=0.3, lambda_wavelet=0.5 were obtained by 'a targeted weight search on a single representative case—image 45' and then applied as the validation instrument for LDM-VLM on all subsequent benchmarks. No held-out weight fitting or cross-validation is reported. Because the metric certifying the method was tuned on a case where the method was expected to score well, the high TWS values for LDM-VLM are partly a self-fulfilling result. Please validate the weighting on a separate image or show that the conclusions are stable across a range of weights.
  4. [III-C] The wavelet artifact term Swavelet is defined as the sum of absolute high-frequency wavelet coefficients over two decomposition levels, and TWS subtracts it. This penalizes all high-frequency energy, including legitimate sharp edges and fine texture, not only artifacts. A sharp, correct SR image can therefore receive a lower TWS than a blurry one, which is opposite to the intended notion of trustworthiness. Please demonstrate on controlled examples that TWS separates genuine detail from artifacts, or modify the wavelet term to target artifact-specific statistics.
  5. [IV-B and Table IV] The human evaluations are reported as single percentages without confidence intervals or per-participant variability. For example, Tables I-II describe 65 participants but give only aggregate identification counts, and Table IV reports 'Human Agreement' percentages without sample sizes or error bars. Since the alignment of VLM selections with human preference is a central claim, please report confidence intervals, inter-participant agreement, or significance tests for these comparisons.
minor comments (6)
  1. [Table III] The table caption says '128×128 SR patches from the DIV2K validation set,' but Section IV-A states that the LDM generates 512×512 SR samples; please clarify the actual evaluation resolution.
  2. [IV-B1] The text 'While human-selected ensembles (VLM-HS)' should refer to LDM-HS, not VLM-HS.
  3. [III-C] The normalization of S_CLIP, S_edge, and Swavelet to the [0,1] range is stated but not specified; please provide the normalization formula and the exact ranges before normalization.
  4. [III-C] The sentence 'even in the absence of ground-truth references' repeats the no-reference claim after the metric has been defined with I_HR; align the wording with the actual definition.
  5. [Appendix B] The one-sample t-test for GPT-4o-mini reports t = -11.56, indicating performance significantly below chance on digit recognition; this is surprising and may indicate a scoring or labeling issue; please explain or verify the computation.
  6. [References] Reference [50] cites the web service Llama OCR without a stable peer-reviewed citation; please provide a precise model version, access date, or a more formal reference.

Circularity Check

2 steps flagged · score 5.0 of 10

TWS is a self-calibrated validation oracle: its weights are fitted on one flagship image and then used to certify both the method and the metric itself, while the claimed TWS-human correlation is never measured.

  1. fitted input called prediction [Section III.C, 'Trustworthiness Score for SR Evaluation', and Section IV.E, 'Ablation Study on TWS Weighting Strategy' (Table V)]
    "The relative importance of each component was determined through a targeted weight search on a single representative case—image 45 in Figure 1. The resulting weights, λCLIP = 0.2, λedge = 0.3, and λwavelet = 0.5, were then fixed and applied uniformly across all subsequent evaluations."

    TWS is simultaneously the quantity being validated and the validation instrument. The weights are fit on the flagship example where the authors' method already produces the desired reconstruction, then the same weighted sum is used to certify LDM-VLM on all later benchmarks (e.g., the claim that LDM-VLM achieves the highest average TWS) and to validate its own weighting in the ablation, where the success criterion is TWS itself. The weight search is therefore a fitted parameter renamed as an independent confirmation: the metric that 'proves' trustworthiness embeds the authors' preference from its calibration, and no held-out weight fitting or external validation of TWS against human opinion is reported.

  2. self definitional [Section III.C, TWS definition (Eq. 1)]
    "Given an input high-resolution image IHR and a super-resolved image ISR, the trustworthiness metric TWS(IHR, ISR) is formulated as follows ... This balanced formulation allows TWS to serve as both a verification tool for VLM selections and a general-purpose trustworthiness estimator, particularly in the absence of ground truth."

    By the paper's own definition, TWS is a full-reference function that takes the ground-truth high-resolution image IHR as an argument. The claim that TWS is usable 'particularly in the absence of ground truth' contradicts the defining equation: without IHR, the stated formula is undefined. The no-reference generality of TWS is thus asserted rather than derived, so any TWS-based 'confirmation' of trustworthiness in no-reference settings reduces to an undefined quantity by the paper's own construction.

full rationale

The VLM-selection component has genuine independent content: Tables I-II compare VLM choices against 65 human participants on MNIST and Urban100, and Table IV reports human-agreement percentages, so the core claim that VLMs pick samples preferred by humans is not circular. The circularity burden lies in the quantitative validation loop. TWS is introduced as the 'rigorous' instrument for assessing the validity of VLM-selected samples, yet its weights are tuned on a single representative case (image 45) and then reused to certify that LDM-VLM attains the highest TWS and to justify its own weighting in the ablation. No experiment computes a correlation between TWS and human preference: the 0.71 MOS correlation in Section IV.C is between GPT-4o Mini ratings and human MOS, not between TWS and human judgment. Hence the abstract's assertion that 'TWS correlates strongly with human preference' is unsupported, and the chain 'VLM selection is trustworthy because TWS is high; TWS is trustworthy because it correlates with humans' is not closed. Additionally, TWS is defined as a full-reference function of IHR yet advertised as usable 'in the absence of ground truth,' which is a self-contradictory definition that further weakens its role as an independent oracle. These issues are partially offset by the direct human/VLM agreement results, so the paper is not wholly circular; nevertheless, the metric-based 'confirmation' of the method reduces in part to the authors' own calibrated scoring function.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central validation rests on assumptions about human ground truth, sample diversity, and the validity of TWS as a trustworthiness oracle. The weights of TWS are fitted to one image, and no external calibration of TWS against human scores is provided.

free parameters (2)
  • TWS weights (λ_CLIP, λ_edge, λ_wavelet) = 0.2, 0.3, 0.5
    Chosen by targeted weight search on a single representative case (image 45 in Fig. 1) and then fixed across all datasets; the metric's judgments depend directly on these values.
  • Top-k ensembling count = 5
    The top-5 samples are ensembled; no sensitivity analysis over k is reported, and the human study likewise asked for top-2 or top-5 selections.
assumptions (3)
  • domain assumption Human preference is the correct ground truth for SR trustworthiness and can be measured reliably with 65 participants on digits and 15 natural images.
    Used to validate VLM selections in Section IV-B; the small, unpowered study is assumed representative.
  • domain assumption The pre-trained LDM produces a diverse sample set that contains trustworthy solutions.
    The framework's success depends on the candidate pool containing good samples; Section III-A asserts diversity but no coverage analysis is given.
  • ad hoc to paper TWS, as defined by the weighted combination of CLIP similarity, edge SSIM, and wavelet energy, measures trustworthiness.
    No external calibration to human ratings is provided; the weights are fitted to one image (Section III-C).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Leveraging Vision-Language Models to Select Trustworthy Super-Resolution Samples Generated by Diffusion Models." pith.science (2026). https://pith.science/paper/CBYVL3Q3

@misc{pith2026250620832,
  author       = {Pith},
  title        = {Pith review of: Leveraging Vision-Language Models to Select Trustworthy Super-Resolution Samples Generated by Diffusion Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CBYVL3Q3}},
  note         = {Machine review of arXiv:2506.20832}
}
read the original abstract

Super-resolution (SR) is an ill-posed inverse problem with many feasible solutions consistent with a given low-resolution image. On one hand, regressive SR models aim to balance fidelity and perceptual quality to yield a single solution, but this trade-off often introduces artifacts that create ambiguity in information-critical applications such as recognizing digits or letters. On the other hand, diffusion models generate a diverse set of SR images, but selecting the most trustworthy solution from this set remains a challenge. This paper introduces a robust, automated framework for identifying the most trustworthy SR sample from a diffusion-generated set by leveraging the semantic reasoning capabilities of vision-language models (VLMs). Specifically, VLMs such as BLIP-2, GPT-4o, and their variants are prompted with structured queries to assess semantic correctness, visual quality, and artifact presence. The top-ranked SR candidates are then ensembled to yield a single trustworthy output in a cost-effective manner. To rigorously assess the validity of VLM-selected samples, we propose a novel Trustworthiness Score (TWS) a hybrid metric that quantifies SR reliability based on three complementary components: semantic similarity via CLIP embeddings, structural integrity using SSIM on edge maps, and artifact sensitivity through multi-level wavelet decomposition. We empirically show that TWS correlates strongly with human preference in both ambiguous and natural images, and that VLM-guided selections consistently yield high TWS values. Compared to conventional metrics like PSNR, LPIPS, which fail to reflect information fidelity, our approach offers a principled, scalable, and generalizable solution for navigating the uncertainty of the diffusion SR space. By aligning outputs with human expectations and semantic correctness, this work sets a new benchmark for trustworthiness in generative SR.

Figures

Figures reproduced from arXiv: 2506.20832 by the authors.

Figure 1
Figure 1. Ambiguity in SR: Results of the state-of-the-art models [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Vision-language models (VLMs) enable automatic selection of reliable SR samples out of a pool of samples generated [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. The block diagram of our proposed trustworthy SR framework, consisting of diffusion SR sample space generation, [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Illustration of diversity of SR samples generated by [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Position of samples generated by LDM [11], the pro [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]
Figure 6
Figure 6. Figure 6: Resolving ambiguity in SR: Identification of the digit [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]
Figure 7
Figure 7. Figure 7: Visual comparison of the proposed LDM-VLM method vs. the state-of-the-art regressive, GAN-based, flow-based, and [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]
Figure 8
Figure 8. Figure 8: Visual comparison of the proposed ensembled LDM-VLM method with the state-of-the-art for [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]
Figure 9
Figure 9. Figure 9: Visual comparison of the proposed ensembled LDM-VLM method with the SoTA for [PITH_FULL_IMAGE:figures/full_fig_p010_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

57 extracted references · 53 canonical work pages

  1. [1]

    Deep Learning for Image/Video Restoration and Super-resolution

    A. M. Tekalp. “Deep Learning for Image/Video Restoration and Super-resolution”. In: Foundations and Trends in Comp. Graph. & Vis. 13.1 (2022), pp. 1–110

    work page 2022

  2. [2]

    Image Super-Resolution Using Deep Convolutional Networks

    C. Dong et al. “Image Super-Resolution Using Deep Convolutional Networks”. In: IEEE Trans. on Pattern Analysis and Mach. Intell. 38 (2016), pp. 295–307

    work page 2016

  3. [3]

    Photo-Realistic Single Image Super- Resolution Using a Generative Adversarial Network

    C. Ledig et al. “Photo-Realistic Single Image Super- Resolution Using a Generative Adversarial Network”. In: IEEE/CVF Conf. on Comp. Vision and Patt. Recog. (CVPR) (2017), pp. 105–114

    work page 2017

  4. [4]

    Enhanced Deep Residual Networks for Single Image Super-Resolution

    B. Lim et al. “Enhanced Deep Residual Networks for Single Image Super-Resolution”. In: IEEE/CVF CVPR Workshops. 2017

    work page 2017

  5. [5]

    Image Super-Resolution Using Very Deep Residual Channel Attention Networks

    Y . Zhang et al. “Image Super-Resolution Using Very Deep Residual Channel Attention Networks”. In: IEEE/CVF ECCV. 2018

    work page 2018

  6. [6]

    ESRGAN: Enhanced Super-Resolution Generative Adversarial Networks

    X. Wang et al. “ESRGAN: Enhanced Super-Resolution Generative Adversarial Networks”. In: European Conf. on Comp. Vision (ECCV) Workshops . 2018. IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY , VOL. XX, NO. X, XXXX 2025 13

    work page 2018

  7. [7]

    ESRGAN+: Fur- ther improving enhanced super-resolution generative ad- versarial network

    N. Rakotonirina and A. Rasoanaivo. “ESRGAN+: Fur- ther improving enhanced super-resolution generative ad- versarial network”. In: ICASSP. 2020, pp. 3637–3641

    work page 2020

  8. [8]

    Perception-Oriented Single Image Super-Resolution Using Optimal Objec- tive Estimation

    S. Park, Y . Moon, and N. Cho. “Perception-Oriented Single Image Super-Resolution Using Optimal Objec- tive Estimation”. In: IEEE/CVF Conf. on Comp. Vision and Patt. Recog. (CVPR) . 2023, pp. 1725–1735

    work page 2023

Show all 57 references
  1. [9]

    Pixel-Aware Stable Diffusion for Real- istic Image Super-Resolution and Personalized Styliza- tion

    Tao Yang et al. “Pixel-Aware Stable Diffusion for Real- istic Image Super-Resolution and Personalized Styliza- tion”. In: Euro. Conf. on Comp. Vision (ECCV) . 2024

    2024

  2. [10]

    Structure-Preserving Super Resolution with Gradient Guidance

    C. Ma et al. “Structure-Preserving Super Resolution with Gradient Guidance”. In: IEEE/CVF Conf. Comp. Vis. and Patt. Recog. (CVPR) . 2020

    2020

  3. [11]

    High-resolution image synthesis with latent diffusion models

    R. Rombach et al. “High-resolution image synthesis with latent diffusion models”. In: IEEE Conf. Comp. Vis. Patt. Recog. (CVPR). 2022, pp. 10684–10695

    2022

  4. [12]

    Activating More Pixels in Image Super- Resolution Transformer

    X. Chen et al. “Activating More Pixels in Image Super- Resolution Transformer”. In: IEEE/CVF Conf. Comp. Vision and Patt. Recog. (CVPR) . 2023

    2023

  5. [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”. In: IEEE/CVF Conf. on Comp. Vision and Patt. Recog. (CVPR) . 2022, pp. 5657–5666

    2022

  6. [14]

    Trustworthy SR: Resolving ambiguity in image super-resolution via diffusion models and human feedback

    C. Korkmaz, A. M. Tekalp, and Z. Do ˘gan. “Trustworthy SR: Resolving ambiguity in image super-resolution via diffusion models and human feedback”. In: IEEE Int. Conf. on Image Processing (ICIP) . 2024

    2024

  7. [15]

    Single image super- resolution from transformed self-exemplars

    J. Huang, A. Singh, and N. Ahuja. “Single image super- resolution from transformed self-exemplars”. In: IEEE Conf. on Comp. Vision and Patt. Recog. (CVPR) . 2015

    2015

  8. [16]

    SRFlow-DA: Super- Resolution Using Normalizing Flow with Deep Convo- lutional Block

    Y . Jo, S. Yang, and S. Joo Kim. “SRFlow-DA: Super- Resolution Using Normalizing Flow with Deep Convo- lutional Block”. In: IEEE/CVF Conf. on Comp. Vision and Patt. Recog. (CVPR) Workshops . 2021

    2021

  9. [17]

    Image super-resolution via iterative refinement

    C. Saharia et al. “Image super-resolution via iterative refinement”. In: IEEE Trans. on Pattern Analysis and Machine Intelligence 45.4 (2022), pp. 4713–4726

    2022

  10. [18]

    Perception-Distortion Trade-Off in the SR Space Spanned by Flow Models

    C. Korkmaz et al. “Perception-Distortion Trade-Off in the SR Space Spanned by Flow Models”. In: IEEE Int. Conf. on Image Proc. (ICIP) . 2022, pp. 2396–2400

    2022

  11. [19]

    Image super-resolution via latent dif- fusion: A sampling-space mixture of experts and frequency-augmented decoder approach

    F. Luo et al. “Image super-resolution via latent dif- fusion: A sampling-space mixture of experts and frequency-augmented decoder approach”. In: arXiv preprint arXiv:2310.12004 (2023)

    2023 arXiv

  12. [20]

    Improving diffusion models for inverse problems using manifold constraints

    H. Chung et al. “Improving diffusion models for inverse problems using manifold constraints”. In: Adv. in Neural Info. Proc. Syst. (NeurIPS) 35 (2022), pp. 25683–25696

    2022

  13. [21]

    Learning continuous image representation with local implicit image func- tion

    Y . Chen, S. Liu, and X. Wang. “Learning continuous image representation with local implicit image func- tion”. In: IEEE/CVF Conf. on Computer Vision and Patt. Recog. (CVPR). 2021, pp. 8628–8638

    2021

  14. [22]

    One-Step Effective Diffusion Network for Real-World Image Super-Resolution

    Rongyuan Wu et al. “One-Step Effective Diffusion Network for Real-World Image Super-Resolution”. In: arXiv preprint arXiv:2406.08177 (2024)

    2024 arXiv

  15. [23]

    SeeSR: Towards semantics-aware real-world image super-resolution

    Rongyuan Wu et al. “SeeSR: Towards semantics-aware real-world image super-resolution”. In: IEEE/CVF conf. on comp. vis. and patt. recog. 2024, pp. 25456–25467

    2024

  16. [24]

    Exploiting diffusion prior for real- world image super-resolution

    Jianyi Wang et al. “Exploiting diffusion prior for real- world image super-resolution”. In: International Jour- nal of Computer Vision (2024), pp. 1–21

    2024

  17. [25]

    The Unreasonable Effectiveness of Deep Features as a Perceptual Metric

    R. Zhang et al. “The Unreasonable Effectiveness of Deep Features as a Perceptual Metric”. In: IEEE/CVF Conf. Comp. Vision and Patt. Recog. (CVPR) . 2018, pp. 586–595

    2018

  18. [26]

    Image Quality Assessment: Unifying Structure and Texture Similarity

    K. Ding et al. “Image Quality Assessment: Unifying Structure and Texture Similarity”. In: IEEE Trans. Patt Anal. Mach. Intel. 44 (2020), pp. 2567–2581

    2020

  19. [27]

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

    Martin Heusel et al. “GANs trained by a two time-scale update rule converge to a local nash equilibrium”. In: Advances in Neural Info. Processing Systems 30 (2017)

    2017

  20. [28]

    Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large lan- guage models

    Junnan Li et al. “Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large lan- guage models”. In: International conference on machine learning. PMLR. 2023, pp. 19730–19742

    2023

  21. [29]

    Gpt-4 technical report

    Josh Achiam et al. “Gpt-4 technical report”. In: arXiv preprint arXiv:2303.08774 (2023)

    2023 arXiv

  22. [30]

    Residual dense network for image super-resolution

    Yulun Zhang et al. “Residual dense network for image super-resolution”. In: IEEE/CVF conf. on comp. vision and patt. recog. (CVPR) . 2018, pp. 2472–2481

    2018

  23. [31]

    SwinIR: Image restoration using swin transformer

    Jingyun Liang et al. “SwinIR: Image restoration using swin transformer”. In: IEEE/CVF Int. Conf. on Com- puter Vision (CVPR) . 2021, pp. 1833–1844

    2021

  24. [32]

    Generative Adversarial Nets

    I. Goodfellow et al. “Generative Adversarial Nets”. In: Adv. in Neural Info. Proc. Systems . V ol. 27. 2014

    2014

  25. [33]

    Training generative image super-resolution models by wavelet- domain losses enables better control of artifacts

    C. Korkmaz, A. M. Tekalp, and Z. Dogan. “Training generative image super-resolution models by wavelet- domain losses enables better control of artifacts”. In: IEEE/CVF Conf. Comp. Vis. and Patt. Recog. (CVPR) . 2024, pp. 5926–5936

    2024

  26. [34]

    Variational autoen- coder for reference based image super-resolution

    Z. Liu, W. Siu, and L. Wang. “Variational autoen- coder for reference based image super-resolution”. In: IEEE/CVF Conf. on Comp. Vision and Patt. Recog. (CVPR). 2021, pp. 516–525

    2021

  27. [35]

    FS-NCSR: Increasing diversity of the super-resolution space via frequency separation and noise-conditioned normalizing flow

    K. Song et al. “FS-NCSR: Increasing diversity of the super-resolution space via frequency separation and noise-conditioned normalizing flow”. In: IEEE/CVF Conf. Comp. Vis. and Patt. Recog. Workshops (CVPRW). 2022, pp. 967–976

    2022

  28. [36]

    Generative Pretraining From Pixels

    M. Chen et al. “Generative Pretraining From Pixels”. In: Int. Conf. on Mach. Learning . 2020, pp. 1691–1703

    2020

  29. [37]

    SinSR: diffusion-based image super- resolution in a single step

    Yufei Wang et al. “SinSR: diffusion-based image super- resolution in a single step”. In: IEEE/CVF Conf. on Comp. Vis. and Patt. Recog. 2024, pp. 25796–25805

    2024

  30. [38]

    SRDiff: Single image super-resolution with diffusion probabilistic models

    H. Li et al. “SRDiff: Single image super-resolution with diffusion probabilistic models”. In: Neurocomputing 479 (2022), pp. 47–59

    2022

  31. [39]

    Pseudoinverse-Guided Diffusion Mod- els for Inverse Problems

    S. Jiaming et al. “Pseudoinverse-Guided Diffusion Mod- els for Inverse Problems”. In: Int. Conf. on Learning Rep. (ICLR). 2023

    2023

  32. [40]

    BLIP: Bootstrapping language-image pre-training for unified vision-language understanding and generation

    Junnan Li et al. “BLIP: Bootstrapping language-image pre-training for unified vision-language understanding and generation”. In: Int. Conf. on Machine Learning (ICML). 2022, pp. 12888–12900. IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY , VOL. XX, NO. X, XXXX 2025 14

    2022

  33. [41]

    IQAGPT: Image quality assessment with vision-language and ChatGPT models

    Zhihao Chen et al. “IQAGPT: Image quality assessment with vision-language and ChatGPT models”. In: arXiv preprint arXiv:2312.15663 (2023)

    2023 arXiv

  34. [42]

    Vision-language models for vision tasks: A survey

    Jingyi Zhang et al. “Vision-language models for vision tasks: A survey”. In: IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)

    2024

  35. [43]

    Quality Assessment in the Era of Large Models: A Survey

    Zicheng Zhang et al. “Quality Assessment in the Era of Large Models: A Survey”. In: arXiv preprint arXiv:2409.00031 (2024)

    2024 arXiv

  36. [44]

    Assessing GPT-4 multimodal perfor- mance in radiological image analysis

    Dana Brin et al. “Assessing GPT-4 multimodal perfor- mance in radiological image analysis”. In: European Radiology (2024), pp. 1–7

    2024

  37. [45]

    ChatGPT in healthcare: a taxonomy and systematic review

    Jianning Li et al. “ChatGPT in healthcare: a taxonomy and systematic review”. In: Computer Methods and Programs in Biomedicine (2024), p. 108013

    2024

  38. [46]

    The 2018 PIRM challenge on perceptual image super-resolution

    Yochai Blau et al. “The 2018 PIRM challenge on perceptual image super-resolution”. In: European Conf. on Computer Vision (ECCV) Workshops . 2018

    2018

  39. [47]

    NTIRE 2021 Learning the Super- Resolution Space Challenge

    A. Lugmayr et al. “NTIRE 2021 Learning the Super- Resolution Space Challenge”. In: IEEE Conf. on Comp. Vision and Patt. Recog. Workshops (CVPRW) . 2021, pp. 596–612. DOI: 10.1109/CVPRW53098.2021.00072

    2021

  40. [48]

    The MNIST database of handwritten digit images for machine learning research

    Li Deng. “The MNIST database of handwritten digit images for machine learning research”. In: IEEE Signal Processing Magazine 29.6 (2012), pp. 141–142

    2012

  41. [49]

    Hierarchical conditional flow: A unified framework for image super-resolution and image rescal- ing

    J. Liang et al. “Hierarchical conditional flow: A unified framework for image super-resolution and image rescal- ing”. In: IEEE Int. Conf. on Computer Vision . 2021

    2021

  42. [50]

    https://llamaocr.com/

    Llama OCR. https://llamaocr.com/. 2024

    2024

  43. [51]

    On Single Image Scale-Up Using Sparse-Representations

    Roman Zeyde, Michael Elad, and Matan Protter. “On Single Image Scale-Up Using Sparse-Representations”. In: Curves and Surfaces. Ed. by Jean-Daniel Boissonnat et al. Berlin, Heidelberg: Springer, 2012, pp. 711–730

    2012

  44. [52]

    A database of human segmented natural images and its application to evaluating segmen- tation algorithms and measuring ecological statistics

    D. Martin et al. “A database of human segmented natural images and its application to evaluating segmen- tation algorithms and measuring ecological statistics”. In: IEEE Int. Conf. on Computer Vision. (ICCV). V ol. 2. 2001, 416–423 vol.2

    2001

  45. [53]

    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”. In: IEEE/CVF Conf. Comp. Vis. and Patt. Recog. (CVPR) Workshops. 2017

    2017

  46. [54]

    Implicit diffusion models for continuous super-resolution

    S. Gao et al. “Implicit diffusion models for continuous super-resolution”. In: IEEE/CVF Conf. Comp. Vis. and Patt. Recog. 2023, pp. 10021–10030

    2023

  47. [55]

    KADID- 10k: A Large-scale Artificially Distorted IQA Database

    Hanhe Lin, Vlad Hosu, and Dietmar Saupe. “KADID- 10k: A Large-scale Artificially Distorted IQA Database”. In: Int. Conf. on Quality of Multimedia Experience (QoMEX). IEEE. 2019, pp. 1–3

    2019

  48. [56]

    Beyond a Gaussian denoiser: Residual learning of deep CNN for image denoising

    Kai Zhang et al. “Beyond a Gaussian denoiser: Residual learning of deep CNN for image denoising”. In: IEEE Trans. on Image Proc. 26.7 (2017), pp. 3142–3155

    2017

  49. [57]

    A Threshold Selection Method from Gray-Level Histograms

    Nobuyuki Otsu. “A Threshold Selection Method from Gray-Level Histograms”. In: IEEE Trans. on Systems, Man, and Cybernetics 9.1 (1979), pp. 62–66. DOI: 10. 1109/TSMC.1979.4310076. Cansu Korkmaz (M’19) received her B.S. and M.S. degrees in Electrical and Electronics Engineering ...

    1979

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.