Pith. sign in

REVIEW 3 major objections 6 minor 90 references

Full-reference image quality metrics can be estimated from a single processed photo and its ISO metadata, without a clean reference image.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-07-31 23:43 UTC pith:XD3SETZU

load-bearing objection A solid, honest empirical study of proxy-reference metric estimation with one real gap: the blind-IQA comparison only holds in the training distribution, not on real pipelines. the 3 major comments →

arxiv 2607.23321 v1 pith:XD3SETZU submitted 2026-07-25 eess.IV cs.CV

A Reference-Free Framework for Evaluating Single-Frame ISP Pipelines

Yujin Cho , Sira Ferradans , Jean-Michel Morel , Gabriele Facciolo , Thomas Eboli This is my paper
classification eess.IV cs.CV MSC 68T4568U10
keywords reference-free IQAfull-reference metricsISP pipeline evaluationproxy referencePSNR/SSIM/LPIPS estimationLoRA fine-tuningblind image quality assessment
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper argues that full-reference image quality metrics—PSNR, SSIM, LPIPS—can be computed without a clean reference image by first learning to synthesize a proxy reference from a processed photo and its ISO value, then running the standard metric between photo and proxy. The authors train a restoration network on synthetic noisy-raw to sRGB pairs from a fixed ISP, and show that the learned estimator transfers to a real Adobe Lightroom pipeline after lightweight LoRA fine-tuning on a small amount of paired data. They report that this proxy-reference approach beats direct regression of metric scores and conventional blind IQA on ranking images from the same scene processed at different noise levels. The practical payoff is that camera pipelines could be evaluated from ordinary photos and metadata, without lab capture of aligned clean references.

Core claim

Full-reference metrics compare a processed image to a clean reference of the same scene. This paper's central claim is that the clean reference can be replaced by a learned proxy: a neural network takes the processed sRGB image plus its ISO setting and outputs a proxy reference image, after which PSNR, SSIM, and LPIPS are computed in their standard two-image form. The authors show that on a synthetic benchmark the proxy model reaches high correlation with true metrics, outperforming direct scalar regression, and that after LoRA fine-tuning on paired SIDD data it also tracks the true metrics on real Google Pixel photos processed with Adobe Lightroom, with near-0.95 PLCC for PSNR and SSIM. The

What carries the argument

The load-bearing component is the proxy-reference network, a SwinIR restoration network whose LayerNorm is replaced by ISO-conditioned Adaptive LayerNorm (AdaLN). It is trained with a reconstruction loss plus metric-consistency losses on PSNR and SSIM and a VGG perceptual loss, so that the predicted proxy image y_θ is not necessarily pixel-identical to the ground truth, but behaves like the ground truth under the metrics of interest. LoRA low-rank adaptation, applied to attention, MLP, and AdaLN projections with the backbone frozen, provides few-shot adaptation to new pipelines. The metric is then simply C(y, y_θ), preserving the original full-reference computation.

Load-bearing premise

To apply the framework to a real ISP, a small set of aligned noisy/reference pairs from that pipeline must be available for LoRA fine-tuning; for pipelines without such paired data, the zero-shot estimates of SSIM and LPIPS are much weaker than the fine-tuned results.

What would settle it

Take a processed image and its proxy reference, then adversarially modify the proxy in a way that keeps the reported PSNR/SSIM/LPIPS unchanged but visibly changes image content; or, equivalently, find two different ground-truth references that yield the same metrics against a fixed degraded image. If the proxy-based scores remain aligned with ground-truth metrics across such inputs, the method is robust; if not, the proxy is merely fitting the training pipeline's artifact distribution rather than recovering true reference statistics.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Camera-pipeline developers could run PSNR/SSIM/LPIPS-style evaluation on single processed images plus EXIF ISO, without collecting aligned reference photos.
  • A single base model can serve many pipelines: component swaps (denoiser, demosaicing, JPEG quality, PSF) are handled by hours-long LoRA fine-tuning rather than days of full retraining.
  • The approach reframes blind IQA as hidden-reference recovery: predicting a metric-consistent proxy is claimed to be easier and more reliable than regressing a scalar score directly.
  • Because training needs only synthetic pairs (clean raw + processed) and no human quality labels, the framework can scale to new sensors and pipelines as long as a noise profile is calibrated.
  • For fixed-content panels (same scene at different ISO), the proxy method orders images correctly where several existing blind IQA methods fail.

Where Pith is reading between the lines

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

  • If the proxy-reference trick generalizes, it suggests that full-reference metrics contain enough information in a single degraded image to invert the degradation; this may extend to other two-image comparison metrics beyond PSNR/SSIM/LPIPS, e.g., HDR-VDP or CameraVDP, provided suitable consistency losses exist.
  • The ISO conditioning via AdaLN implies noise level is the dominant latent variable; one could test whether the proxy network also implicitly learns the camera noise model and whether passing estimated noise parameters instead of ISO would improve calibration.
  • The method's reliance on paired fine-tuning data for new pipelines suggests a hybrid protocol: use the synthetic-trained base for ranking within a pipeline, and use few-shot LoRA only when cross-pipeline absolute metric values matter; zero-shot results are the real test of the 'reference-free' claim.
  • A testable extension: apply the proxy framework to video frames or burst-capture stacks where ISO varies per frame, to see whether per-frame proxy metrics can track temporal quality changes.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes a reference-free framework for estimating full-reference image quality metrics (PSNR, SSIM, LPIPS) from a single processed sRGB image and its ISO metadata. A restoration network with an AdaLN-conditioned SwinIR backbone is trained to predict a proxy reference image from the processed image; the target metrics are then computed between the processed image and this proxy. The model is trained on a synthetic ISP pipeline (SCUNet denoising, DemosaicNet, tone mapping, JPEG) and adapted to new pipelines via LoRA fine-tuning. Experiments on synthetic DIV2K data and real SIDD/Adobe Lightroom data report strong correlations with ground-truth metrics, beyond a direct regression baseline, and the authors claim higher agreement with full-reference rankings than conventional blind IQA methods.

Significance. If the claims hold, this would be a practically useful tool for camera-pipeline tuning: a single model could estimate several full-reference metrics without requiring aligned references at inference time, and LoRA adaptation would make it easy to switch pipelines. The paper has notable strengths: it reports confidence intervals throughout, uses a scene-disjoint split for the real-data fine-tuning, includes extensive ablations (conditioning, losses, LoRA layers/ranks), and evaluates on a realistic commercial pipeline (Adobe Lightroom). The core idea of learning a proxy reference instead of regressing scalar scores is intuitive and is supported by the experimental trend that restoration with metric-consistency losses outperforms direct regression. However, the headline comparison against blind IQA is currently demonstrated only on synthetic panels generated from the training pipeline, and the strong real-data results require paired noisy/reference data from the target pipeline. These two gaps are load-bearing for the abstract's central claims.

major comments (3)
  1. [§4.1, Table 3 (Comparison to Blind IQA Metrics)] The claim that the method 'achieves higher agreement with full-reference rankings than conventional blind IQA methods' is supported only on synthetic panels generated with the same pipeline F used for training the restoration model. Each panel consists of 20 noisy versions of the same clean image with only ISO/noise varying; because the model is conditioned on ISO and trained on F, this is a strongly favorable in-distribution comparison. No blind-IQA method is evaluated on the real SIDD/Adobe data of Table 5 and Figures 5–6, which is the stated practical setting. The paper should add blind-IQA correlations on the real test set or qualify the claim to 'on synthetic panels drawn from the training pipeline'; otherwise the abstract overstates the evidence.
  2. [§4.2, Table 5 and Conclusion] The strong real-camera results (PSNR PLCC 0.958, SSIM PLCC 0.957, LPIPS PLCC 0.923) are obtained only after LoRA fine-tuning on paired SIDD data with a scene-disjoint split. Zero-shot transfer from the synthetic pipeline to Adobe Lightroom yields SSIM PLCC 0.656 and LPIPS PLCC 0.719, only modestly better than the regression baseline for SSIM/LPIPS. The abstract's statement that 'the proxy-reference model can be learned from synthetic data and applied to real camera data' is therefore incomplete: high-accuracy real-pipeline evaluation is contingent on a small set of aligned pairs from the target pipeline. The authors should explicitly state this label requirement in the contribution list and temper the 'reference-free' framing, since a user evaluating a new ISP without paired data would not obtain the advertised accuracy.
  3. [§4.1, paragraph before Table 3] In the blind-IQA comparison, the LPIPS column for 'Ours' is not produced by a model trained to predict LPIPS; the text states 'To estimate LPIPS, we use our SSIM estimator as it shows high correlation with LPIPS.' This makes the claimed LPIPS advantage a cross-metric proxy rather than a direct validation of LPIPS estimation. The real-data LPIPS results (Table 5, Figure 5) should specify whether they come from a jointly trained model or from the SSIM-based proxy, and if the latter, the claim 'estimates PSNR, SSIM, and LPIPS' should be softened to 'estimates PSNR and SSIM, and provides a correlated surrogate for LPIPS.'
minor comments (6)
  1. [Introduction, first paragraph] Typo: 'withouot' should be 'without.'
  2. [Related Work, §2.3] Grammar: 'This idea was to adapted to IQA' should read 'This idea was adapted to IQA.'
  3. [Related Work, §2.3] The sentence 'Zheng et al. CKDN [ZYF∗21]. employ a coupled knowledge distillation framework' has an extra period after the citation; please fix.
  4. [Table 3] The rows for PAQ2PIQ and CLIPIQA contain identical correlation values (PLCC 0.147, SRCC 0.276, etc.). This is almost certainly a copy-paste error; please verify and correct.
  5. [Equation (7)] After Eq. (7), λ_PSNR and λ_SSIM are referenced but not explicitly defined in the text. Please define them (e.g., 'with λ_PSNR=0.1, λ_SSIM=0.5, λ_p=0.01') in the training section.
  6. [§4.2, Dataset description] The paper describes fine-tuning as 'few-shot' (e.g., 'with only a few shot images' in §4.2 and 'within a few hours' in §3.4). The actual number of SIDD scenes/patches used for LoRA adaptation is not stated. Please specify the data volume (e.g., number of scenes/patches) used in the scene-disjoint protocol.

Circularity Check

0 steps flagged

No significant circularity: the proxy-reference estimator is trained on held-out data and its predictions are not equivalent to its inputs by construction.

full rationale

Walking the derivation chain: Cθ(y)=C(y,φθ(y,ISO)) defines a learned estimator, not a derivation. The proxy φθ is trained with ℓdata and ℓC on synthetic pairs, and all reported correlations use held-out DIV2K test patches or held-out SIDD scenes with a scene-disjoint split (Table 5). The target metrics for those evaluations are not seen during training, so the predictions are not forced by construction. No equation reduces to its own input: the method never uses the identity C(y,y)=const, and it does not compute the target metric from itself. LoRA fine-tuning uses paired references for adaptation, but evaluation is on disjoint scenes, which is standard supervised adaptation rather than fitted-input-called-prediction. The only self-citation found is [EMF22] (Eboli, Morel, Facciolo) for PSF estimation in an auxiliary appendix experiment; it is not load-bearing for the central claim. The blind-IQA comparison (Table 3) is favorable to the method because the panels use the same synthetic pipeline F as training while blind IQA methods are not tuned for this task, but this is an evaluation-scope limitation, not circularity: the method's rankings are computed on held-out images and are not forced to match by construction. No claim is justified solely by a self-citation or by definition. Therefore no significant circularity is present.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The central claim rests on learned mappings fitted to synthetic data and on a calibrated noise model; the main free parameters are noise-profile constants, loss weights, and LoRA hyperparameters. No new physical entities are postulated.

free parameters (5)
  • Noise curve parameters (α, β) = Calibrated from SIDD Google Pixel noise profile, scaled over ISO 100-12800, jittered during training
    The heteroscedastic Gaussian noise model in Eq. (1) determines the synthetic noisy inputs and therefore the learned proxy reference and metric targets.
  • Metric-consistency weights λ_PSNR, λ_SSIM, λ_p = 0.1, 0.5, 0.01
    Chosen by validation; they balance reconstruction, metric-matching, and perceptual losses in Eq. (7).
  • LoRA rank, scaling factor, dropout = 8, 16, 0.05
    Grid-searched on synthetic validation data; used for all pipeline adaptations.
  • Patch sharpness threshold = 0.6 for training and main evaluation (0.2/0.4/0.6 robustness in Table 6)
    Discards flat or out-of-focus patches, shaping the training and evaluation distributions.
  • JPEG quality factor Q = 90 in baseline pipeline; Q=50 and Q=10 for generalization tests
    Fixed choice in the training pipeline; generalization to other compression levels is tested in Table 9.
axioms (5)
  • domain assumption Heteroscedastic Gaussian approximation of Poisson-Gaussian noise, with ISO as a proxy for noise severity
    Eq. (1) and Section 3: all synthetic training pairs and the ISO-conditioning mechanism depend on this noise model being accurate for the target camera.
  • domain assumption Unprocessing sRGB images via Brooks et al. [BMX19] yields a valid noise-free raw x0
    Section 4.1: the training targets y_GT = F(x0) are generated from these inverted raw images.
  • domain assumption A proxy reference need not be pixel-identical to the ground truth; it only needs to match under the target metrics
    Section 3.3: this is the foundation of the method and justifies using metric-consistency losses instead of exact reconstruction.
  • domain assumption A synthetic pipeline with SCUNet, DemosaicNet, smooth-step gamma, and JPEG Q=90 transfers to a real pipeline such as Adobe Lightroom after LoRA fine-tuning
    Sections 3.1 and 4.2: zero-shot transfer is weak for SSIM and LPIPS, so the practical claim relies on this adaptation step.
  • domain assumption LPIPS can be approximated with a network trained for PSNR/SSIM consistency
    Table 3 uses the SSIM estimator to produce LPIPS rankings, which assumes the PSNR/SSIM model captures the LPIPS ordering.

pith-pipeline@v1.3.0-alltime-deepseek · 27099 in / 12419 out tokens · 119888 ms · 2026-07-31T23:43:50.589750+00:00 · methodology

0 comments
read the original abstract

Evaluating camera image signal processing (ISP) pipelines requires measuring low-level artifacts introduced by operations such as denoising, demosaicing, tone mapping, and compression. Blind image quality assessment (IQA) techniques can grade visual quality without a reference, but they typically focus on semantic and high-level visual cues or human perceptual scores rather than the low-level image-processing artifacts introduced by camera pipelines. In contrast, full-reference metrics such as PSNR and SSIM measure pixel-level differences and structural similarity, while LPIPS measures perceptual similarity in deep feature space. However, these metrics require perfectly aligned image pairs, which are difficult to collect in practical settings. We propose a reference-free learning framework that estimates full-reference image quality metrics from a processed sRGB image and its ISO metadata. Our method predicts a proxy sRGB reference, which is then compared with the processed image to compute PSNR, SSIM, and LPIPS in their standard full-reference form. Our experiments show that the proxy-reference model can be learned from synthetic data and applied to real camera data. We further show that lightweight LoRA fine-tuning enables efficient adaptation when ISP components or pipeline configurations are changed. The proposed method outperforms direct metric regression in estimating metric values and achieves higher agreement with full-reference rankings than conventional blind IQA methods. These results demonstrate the feasibility of reference-free estimation of full-reference metrics for practical camera-pipeline evaluation.

Figures

Figures reproduced from arXiv: 2607.23321 by Gabriele Facciolo, Jean-Michel Morel, Sira Ferradans, Thomas Eboli, Yujin Cho.

Figure 1
Figure 1. Figure 1: Reference-free metric estimation for camera pipeline evaluation. A noisy RAW image is first processed by the ISP to produce an sRGB camera output. At inference time, the proposed method takes only this processed image and its ISO metadata as input, predicts a proxy reference, and estimates full-reference metrics without access to the ground-truth reference. We used a raw photograph taken with a Samsung Gal… view at source ↗
Figure 2
Figure 2. Figure 2: Explanatory diagram of our restoration method. Top: Main branch of our approach. A noisy raw image x is processed with the pipeline F to obtain the image y = F(x). The image y, together with its ISO value, is fed to the network φθ to predict the proxy reference yθ = φθ(y,ISO) and the score C(y, yθ). Bottom: Additional branch for the training and evaluation stages. Starting either from a high-quality raw im… view at source ↗
Figure 3
Figure 3. Figure 3: Visual examples of estimated references yθ. Two synthetic patches at ISO1600 and ISO3200 are shown: noisy inputs x, processed outputs y, and the corresponding proxy references predicted by the fine-tuned model φθ,ω. Blue boxes indicate ground-truth metric values, and green boxes the predicted ones. Although the estimated references differ from yGT in color and sharpness, the proxy references yield comparab… view at source ↗
Figure 4
Figure 4. Figure 4: Visual example of rankings with blind IQA methods. Each method computes a quality score for 4 noisy versions of a ref￾erence image that are degraded with increasing noise levels. Only our method and MANIQA predict the right order, highlighting the insensitivity of blind IQA methods to detect distortions caused by F. The reader is invited to zoom in [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Correlation scores on GP phone patches after Adobe fine-tuning (FT). Scatter plots compare ground-truth values against our predictions for PSNR, SSIM, and LPIPS on real pairs of 256 × 256 textured patches taken from SIDD [ALB18]. The Adobe fine￾tuned model attains high Pearson (PLCC) and Spearman (SRCC) correlations across all three metrics, indicating effective adaptation of the synthetic-pretrained model… view at source ↗
Figure 6
Figure 6. Figure 6: Aggregated quality metric results per image with the Adobe fine-tuned (FT) pipeline. We group the patch estimates per image, resulting in 25 boxes, and report medians (orange lines) for PSNR, SSIM, and LPIPS. These results complement the global correlation analysis in the main paper by highlighting per-image agreement and discrepancies [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Unrealistic denoising distortions in KADID dataset [ [PITH_FULL_IMAGE:figures/full_fig_p015_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Comparison of different LoRA configurations. Six variants are shown: (a) noisy raw x, (b) ground-truth reference yGT, (c) processed image y, (d) fine-tuning only attention q, k, v, (e) attention + MLP + AdaLN layers, and (f) attention + MLP + AdaLN layers with VGG perceptual loss [PITH_FULL_IMAGE:figures/full_fig_p016_8.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

90 extracted references · 9 linked inside Pith

  1. [1]

    Reference-free estimation of structural and perceptual metrics for single-frame isp pipelines , author=

  2. [2]

    ACM Transactions on Graphics (ToG) , volume=

    Deep joint demosaicking and denoising , author=. ACM Transactions on Graphics (ToG) , volume=. 2016 , publisher=

    2016

  3. [3]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    Unprocessing Images for Learned Raw Denoising , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

  4. [4]

    Hasinoff and Dillon Sharlet and Ryan Geiss and Andrew Adams and Jonathan T

    Samuel W. Hasinoff and Dillon Sharlet and Ryan Geiss and Andrew Adams and Jonathan T. Barron and Florian Kainz and Jiawen Chen and Marc Levoy , title =

  5. [5]

    Handheld multi-frame super-resolution , journal =

    Bartlomiej Wronski and Ignacio Garcia. Handheld multi-frame super-resolution , journal =

  6. [6]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Real-esrgan: Training real-world blind super-resolution with pure synthetic data , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  7. [7]

    completely blind

    Making a “completely blind” image quality analyzer , author=. 2012 , publisher=

    2012

  8. [8]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Rethinking noise synthesis and modeling in raw denoising , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  9. [9]

    Machine Intelligence Research , volume=

    Practical blind image denoising via Swin-Conv-UNet and data synthesis , author=. Machine Intelligence Research , volume=. 2023 , publisher=

    2023

  10. [10]

    Proceedings of the IEEE conference on computer vision and pattern recognition workshops , pages=

    Ntire 2017 challenge on single image super-resolution: Dataset and study , author=. Proceedings of the IEEE conference on computer vision and pattern recognition workshops , pages=

    2017

  11. [11]

    European Conference on Computer Vision , pages=

    Multi-patch learning: looking more pixels in the training phase , author=. European Conference on Computer Vision , pages=. 2022 , organization=

    2022

  12. [12]

    2016 , publisher=

    Waterloo exploration database: New challenges for image quality assessment models , author=. 2016 , publisher=

    2016

  13. [13]

    2015 , publisher=

    Massive online crowdsourced study of subjective and objective picture quality , author=. 2015 , publisher=

    2015

  14. [14]

    2019 Eleventh International Conference on Quality of Multimedia Experience (QoMEX) , pages=

    KADID-10k: A large-scale artificially distorted IQA database , author=. 2019 Eleventh International Conference on Quality of Multimedia Experience (QoMEX) , pages=. 2019 , organization=

    2019

  15. [15]

    arXiv preprint arXiv:1803.08489 , year=

    KonIQ-10K: Towards an ecologically valid and large-scale IQA database , author=. arXiv preprint arXiv:1803.08489 , year=

    Pith/arXiv arXiv

  16. [16]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Perceptual quality assessment of smartphone photography , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  17. [17]

    Communications of the ACM , volume=

    Yfcc100m: The new data in multimedia research , author=. Communications of the ACM , volume=. 2016 , publisher=

    2016

  18. [18]

    Proceedings of the International COnverence on Advanced Concepts for Intelligent Vision Systems , pages=

    A new color image database TID2013: Innovations and results , author=. Proceedings of the International COnverence on Advanced Concepts for Intelligent Vision Systems , pages=

  19. [19]

    arXiv preprint arXiv:1711.05101 , year=

    Decoupled weight decay regularization , author=. arXiv preprint arXiv:1711.05101 , year=

    Pith/arXiv arXiv

  20. [20]

    arXiv preprint arXiv:1912.01703 , year=

    Pytorch: An imperative style, high-performance deep learning library , author=. arXiv preprint arXiv:1912.01703 , year=

    Pith/arXiv arXiv 1912

  21. [21]

    2006 , publisher=

    A statistical evaluation of recent full reference image quality assessment algorithms , author=. 2006 , publisher=

    2006

  22. [22]

    Most apparent distortion: Full-reference image quality assessment and the role of strategy , volume =

    Larson, Eric and Chandler, Damon , year =. Most apparent distortion: Full-reference image quality assessment and the role of strategy , volume =

  23. [23]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    From patches to pictures (PaQ-2-PiQ): Mapping the perceptual space of picture quality , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  24. [24]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    The unreasonable effectiveness of deep features as a perceptual metric , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  25. [25]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , year =

    Su, Shaolin and Yan, Qingsen and Zhu, Yu and Zhang, Cheng and Ge, Xin and Sun, Jinqiu and Zhang, Yanning , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , year =

  26. [26]

    2011 , publisher=

    Blind image quality assessment: From natural scene statistics to perceptual quality , author=. 2011 , publisher=

    2011

  27. [27]

    2012 , publisher=

    No-reference image quality assessment in the spatial domain , author=. 2012 , publisher=

    2012

  28. [28]

    2012 , publisher=

    Blind image quality assessment: A natural scene statistics approach in the DCT domain , author=. 2012 , publisher=

    2012

  29. [29]

    2012 IEEE conference on computer vision and pattern recognition , pages=

    Unsupervised feature learning framework for no-reference image quality assessment , author=. 2012 IEEE conference on computer vision and pattern recognition , pages=. 2012 , organization=

    2012

  30. [30]

    2017 , publisher=

    Deep neural networks for no-reference and full-reference image quality assessment , author=. 2017 , publisher=

    2017

  31. [31]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Blindly assess image quality in the wild guided by a self-adaptive hyper network , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  32. [32]

    Proceedings of the IEEE International Conference on Computer Vision , pages=

    Rankiqa: Learning from rankings for no-reference image quality assessment , author=. Proceedings of the IEEE International Conference on Computer Vision , pages=

  33. [33]

    IEEE Transactions on Image Processing , volume =

    Uncertainty-aware blind image quality assessment in the laboratory and wild , author =. IEEE Transactions on Image Processing , volume =

  34. [34]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Musiq: Multi-scale image quality transformer , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  35. [35]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Swinir: Image restoration using swin transformer , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  36. [36]

    IET Image Processing , volume=

    Full-reference tone-mapped images quality assessment , author=. IET Image Processing , volume=. 2021 , publisher=

    2021

  37. [37]

    2014 , publisher=

    FSITM: A feature similarity index for tone-mapped images , author=. 2014 , publisher=

    2014

  38. [38]

    Human vision and electronic imaging , volume=

    Extending quality metrics to full luminance range images , author=. Human vision and electronic imaging , volume=. 2008 , organization=

    2008

  39. [39]

    2012 , publisher=

    Objective quality assessment of tone-mapped images , author=. 2012 , publisher=

    2012

  40. [40]

    Bovik and Hamid R

    Zhou Wang and Alan C. Bovik and Hamid R. Sheikh and Eero P. Simoncelli , title =

  41. [41]

    Signal, Image and Video Processing , volume=

    Sparsity-based no-reference image quality assessment for automatic denoising , author=. Signal, Image and Video Processing , volume=. 2018 , publisher=

    2018

  42. [42]

    Proceedings of the Computer Vision Conference , pages=

    No-reference image denoising quality assessment , author=. Proceedings of the Computer Vision Conference , pages=. 2020 , organization=

    2020

  43. [43]

    Rempel and Wolfgang Heidrich , title =

    Rafal Mantiuk and Kil Joong Kim and Allan G. Rempel and Wolfgang Heidrich , title =

  44. [44]

    Human Vision and Electronic Imaging X , volume=

    Predicting visible differences in high dynamic range images: model and its calibration , author=. Human Vision and Electronic Imaging X , volume=. 2005 , organization=

    2005

  45. [45]

    Proceedings of the SIGGRAPH Asia 2025 Conference Papers , pages=

    CameraVDP: Perceptual Display Assessment with Uncertainty Estimation via Camera and Visual Difference Prediction , author=. Proceedings of the SIGGRAPH Asia 2025 Conference Papers , pages=

    2025

  46. [46]

    Egiazarian , title =

    Alessandro Foi and Mejdi Trimeche and Vladimir Katkovnik and Karen O. Egiazarian , title =

  47. [47]

    Brown , title =

    Abdelrahman Abdelhamed and Stephen Lin and Michael S. Brown , title =. Proceedings of the

  48. [48]

    Alireza Golestaneh and Saba Dadsetan and Kris M

    S. Alireza Golestaneh and Saba Dadsetan and Kris M. Kitani , title =. Proceedings of the

  49. [49]

    2025 , booktitle =

    Augmenting Perceptual Super-Resolution via Image Quality Predictors , author =. 2025 , booktitle =

    2025

  50. [50]

    Proceedings of the

    Enhanced deep residual networks for single image super-resolution , author=. Proceedings of the

  51. [51]

    Proceedings of the

    Kaixuan Wei and Ying Fu and Jiaolong Yang and Hua Huang , title =. Proceedings of the

  52. [52]

    arXiv preprint arXiv:2410.08534 , year=

    Quality Prediction of AI Generated Images and Videos: Emerging Trends and Opportunities , author=. arXiv preprint arXiv:2410.08534 , year=

    Pith/arXiv arXiv

  53. [53]

    Hu and Yelong Shen and Phillip Wallis and Zeyuan Allen

    Edward J. Hu and Yelong Shen and Phillip Wallis and Zeyuan Allen. LoRA: Low-Rank Adaptation of Large Language Models , booktitle =

  54. [54]

    The annals of Statistics , pages=

    Estimation of the mean of a multivariate normal distribution , author=. The annals of Statistics , pages=. 1981 , publisher=

    1981

  55. [55]

    Proceedings of the

    Wang, Jianyi and Chan, Kelvin CK and Loy, Chen Change , title =. Proceedings of the

  56. [56]

    Blind Image Quality Assessment Using A Deep Bilinear Convolutional Neural Network , author=

  57. [57]

    Proceedings of the

    MANIQA: Multi-dimension Attention Network for No-Reference Image Quality Assessment , author=. Proceedings of the

  58. [58]

    arXiv:2110.13266 , year=

    Image Quality Assessment using Contrastive Learning , author=. arXiv:2110.13266 , year=

    Pith/arXiv arXiv

  59. [59]

    Electronic Imaging , volume =

    Martin Renaudin and Anna-Cecilia Vlachomitrou and Gabriele Facciolo and Wolf Hauser and Clement Sommelet and Clement Viard and Frédéric Guichard , title =. Electronic Imaging , volume =

  60. [60]

    Proceedings of the

    Du Chen and Jie Liang and Xindong Zhang and Ming Liu and Hui Zeng and Lei Zhang , title =. Proceedings of the

  61. [61]

    Pacific Conference on Computer Graphics and Applications , pages=

    Exposure fusion , author=. Pacific Conference on Computer Graphics and Applications , pages=. 2007 , organization=

    2007

  62. [62]

    Communications of the ACM , volume=

    Local Laplacian filters: edge-aware image processing with a Laplacian pyramid , author=. Communications of the ACM , volume=. 2015 , publisher=

    2015

  63. [63]

    arXiv preprint arXiv:2001.08113 , year=

    DeepFL-IQA: Weak supervision for deep IQA feature learning , author=. arXiv preprint arXiv:2001.08113 , year=

    Pith/arXiv arXiv 2001

  64. [64]

    Arnaud and Derek Nowrouzezahrai and Jean

    Ethan Tseng and Felix Yu and Yuting Yang and Fahim Mannan and Karl St. Arnaud and Derek Nowrouzezahrai and Jean. Hyperparameter optimization in black-box image processing using differentiable proxies , journal =

  65. [65]

    2021 , publisher=

    Plug-and-play image restoration with deep denoiser prior , author=. 2021 , publisher=

    2021

  66. [66]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Learning to see in the dark , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  67. [67]

    Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , pages=

    ARNIQA: Learning Distortion Manifold for Image Quality Assessment , author=. Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , pages=

  68. [68]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Learning conditional knowledge distillation for degraded-reference image quality assessment , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  69. [69]

    Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , month =

    Lin, Kwan-Yee and Wang, Guanxiang , title =. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , month =

  70. [70]

    Applied Sciences , volume=

    No-reference image quality assessment based on a multitask image restoration network , author=. Applied Sciences , volume=. 2023 , publisher=

    2023

  71. [71]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    RAN4IQA: Restorative adversarial nets for no-reference image quality assessment , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

  72. [72]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Scalable diffusion models with transformers , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  73. [73]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Photo-realistic single image super-resolution using a generative adversarial network , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  74. [74]

    2017 IEEE International Conference on Multimedia and Expo (ICME) , pages=

    Color image demosaicking via deep residual learning , author=. 2017 IEEE International Conference on Multimedia and Expo (ICME) , pages=. 2017 , organization=

    2017

  75. [75]

    IEEE transactions on pattern analysis and machine intelligence , volume=

    Limits on super-resolution and how to break them , author=. IEEE transactions on pattern analysis and machine intelligence , volume=. 2002 , publisher=

    2002

  76. [76]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Structure-preserving super resolution with gradient guidance , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  77. [77]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=

    Exploiting deep generative prior for versatile image restoration and manipulation , author=. IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=. 2021 , publisher=

    2021

  78. [78]

    European Conference on Computer Vision , pages=

    Deep cyclic generative adversarial residual convolutional networks for real image super-resolution , author=. European Conference on Computer Vision , pages=. 2020 , organization=

    2020

  79. [79]

    Proceedings of the European conference on computer vision (ECCV) workshops , pages=

    Esrgan: Enhanced super-resolution generative adversarial networks , author=. Proceedings of the European conference on computer vision (ECCV) workshops , pages=

  80. [80]

    Advances in neural information processing systems , volume=

    Denoising diffusion restoration models , author=. Advances in neural information processing systems , volume=

Showing first 80 references.