Pith. sign in

REVIEW 3 major objections 5 minor 36 references

Under signal-dependent camera noise, MSE-trained denoisers systematically under-reconstruct dark pixels—up to 6× worse relative to their per-band noise floor—and a drop-in loss, BBRD, corrects this by normalizing per-band error and dynamica

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 →

T0 review · deepseek-v4-flash

2026-08-02 04:42 UTC pith:OF2VXJVY

load-bearing objection Real phenomenon and a clean zero-parameter loss fix, but the headline +0.65 dB is measured against an undertrained MSE baseline and the causal proof has an unjustified independence step. the 3 major comments →

arxiv 2607.16320 v1 pith:OF2VXJVY submitted 2026-07-15 cs.CV

The Devil is in the Dark Pixels: Toward Brightness Bias-Robust Denoising

classification cs.CV
keywords image denoisingbrightness biassignal-dependent noiseGroup-DROper-band PSNRloss function designdark pixel recoveryPoisson-Gaussian noise
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.

The paper argues that standard MSE training is not brightness-neutral: under the Poisson–Gaussian noise of real cameras, dark pixels carry less noise but even less signal, and MSE-trained networks let bright regions dominate gradient updates, leaving dark pixels up to six times worse relative to their noise floor. The cause is two compounding factors: signal-dependent noise inflates bright-pixel residuals, and the network's Jacobian norm increases with brightness. The authors propose BBRD, a loss that partitions pixels into brightness bands, normalizes each band's error by its empirical noise variance, and uses softmax Group-DRO to upweight whichever band is currently worst. In their experiments across eight architectures and two real-world datasets, BBRD is the only method among 13 losses that improves every brightness band simultaneously, with up to +0.65 dB aggregate PSNR on SIDD and no inference cost. A sympathetic reader would care because it suggests the dark-region gap is a correctable loss-level artifact, not an architectural limit.

Core claim

The central claim is that brightness bias in learned denoisers is a loss-level artifact: under signal-dependent noise, the expected squared gradient magnitude for a pixel grows monotonically with brightness because both the residual and the network Jacobian norm increase with brightness, so bright pixels chronically dominate updates. BBRD corrects this by measuring per-band error on a noise-normalized scale and dynamically reweighting the worst band; the paper asserts that, in its experiments, this is the only tested objective that improves both dark and bright bands on SIDD and PolyU simultaneously, with zero additional parameters or inference cost.

What carries the argument

BBRD is a three-step drop-in MSE replacement. First, a 1-D Gaussian mixture model on smoothed clean-target intensities, with component count selected by BIC, partitions pixels into brightness bands using Gaussian soft assignments. Second, per-band MSE is divided by the empirical per-band noise variance (computed once from training pairs), giving a normalized ratio R_k that is about 1 with no denoising progress and approaches 0 for perfect reconstruction. Third, softmax Group-DRO with temperature η=5 reweights bands at every training step, minimizing Σ w_k R_k. The supporting analysis is Proposition 1, which attributes gradient imbalance to the product of residual variance and Jacobian norm.

Load-bearing premise

The diagnosis assumes that a pixel's residual and its Jacobian norm are independent, so their product can be split into two separable monotone factors; if they are actually correlated (both plausibly tied to activation magnitude), the two-factor story is not established by the backward-pass measurements.

What would settle it

Train a denoiser with BBRD and a control that uses fixed dark-band weights; if BBRD's dark-band gain is no larger than the control's, the dynamic DRO component is not necessary. Alternatively, find an architecture whose per-band Jacobian norm does not increase monotonically with brightness—if such a case still shows brightness bias, the proposed mechanism is undercut; also, if removing noise normalization while keeping DRO still improves all bands, the causal role of normalization is not confirmed.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • If BBRD is correct, any MSE-trained denoiser can gain dark-region performance by swapping only the training loss, with zero inference cost and no new parameters.
  • Difficulty-based reweighting methods (Focal, OHEM) and heteroscedastic NLL may actively amplify brightness bias; BBRD identifies noise normalization as the crucial corrective step.
  • The per-band PSNR evaluation protocol makes brightness disparity visible, so future denoiser comparisons can report dark and bright bands rather than a single aggregate.
  • Gains generalize across CNN, Transformer, and State-Space backbones without per-model tuning, supporting the claim that the bias is loss-intrinsic, not architecture-specific.
  • BBRD also improves performance under homoscedastic Gaussian noise, suggesting the mechanism addresses more than heteroscedasticity alone.

Where Pith is reading between the lines

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

  • The same gradient-imbalance mechanism likely affects other signal-dependent restoration tasks—low-light enhancement, deblurring, super-resolution—where dark regions are similarly under-weighted; a BBRD-style normalized DRO could transfer.
  • If Jacobian-norm monotonicity is truly architecture-agnostic, the bias may also appear in non-denoising losses (e.g., segmentation or depth estimation) on low-light images, making per-band normalization a general recipe.
  • BBRD's dark-band gain may partly follow by construction from upweighting the worst band; a sharper test would compare against a fixed dark-band weight tuned per dataset to isolate the dynamic DRO's contribution.
  • LumaSet-700, if adopted, could become a standard brightness-balanced benchmark, but its value depends on whether the community uses per-band metrics as a primary target rather than a supplement to aggregate PSNR.

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 / 5 minor

Summary. The paper identifies a brightness bias in MSE-trained denoisers under signal-dependent camera noise: dark pixels are reconstructed disproportionately worse relative to their per-band noise floor. It attributes this to two compounding factors — signal-dependent residuals and a brightness-correlated Jacobian norm — and proposes BBRD, a training-loss replacement that partitions pixels into brightness bands via a GMM, normalizes per-band error by empirical noise variance, and applies softmax Group-DRO to upweight the currently worst band. Experiments across 8 architectures and 2 datasets report that BBRD improves every brightness band simultaneously, with up to +0.45 dB on dark bands, +0.32 dB on bright bands, and +0.65 dB aggregate PSNR on SIDD, at zero inference cost.

Significance. If the reported gains survive a properly converged comparison, BBRD is a simple and broadly applicable correction to a real and underappreciated failure mode of MSE-based denoising. The paper's strengths include the unusually broad empirical scope (8 architectures, 13 losses, 2 datasets), the component ablation showing that both normalization and DRO are necessary, the proposed LumaSet-700 benchmark, and the promise of released code. The central claim is plausible and the method is easy to adopt, but the current evidence is weakened by the baseline-convergence issue and by an unproven factorization in the theoretical analysis.

major comments (3)
  1. [§4.1 (Table 1) and §4 (Training)] The headline gains are computed against an undertrained MSE baseline. NAFNet+MSE reaches 39.47 dB on SIDD in Table 1, whereas standard training of the same architecture reports ~40.30 dB; all models are trained for only 100 epochs, and Fig. 6a shows the MSE dark/bright gap still widening at epoch 100. The +0.65 dB aggregate and per-band gains thus measure BBRD against a non-converged MSE. Please add convergence-matched comparisons (e.g., train both until plateau, or report MSE at the same effective schedule) and show that the per-band and aggregate improvements persist. This is load-bearing for the drop-in-replacement and systematic-bias claims.
  2. [§3.2, Prop. 1 / Eq. (2)] The proof asserts E[r_i^2 ||J_i||^2] ∝ E[r_i^2] E[||J_i||^2] without stating or justifying independence of the residual magnitude and Jacobian norm; both are plausibly correlated with activation magnitude. Moreover, the reported per-band Jacobian norm is measured by a backward pass that accumulates gradients from all band pixels and then takes the squared Frobenius norm, so cross terms between pixels are included; this is not the per-pixel E[||J_i||^2] used in the proposition. Please either state the independence as an explicit assumption with a justified mechanism, or measure per-pixel Jacobian norms and verify the factorization empirically (e.g., compare E[r^2 J^2] with E[r^2]E[J^2] across bands). The monotone-Jacobian claim currently rests on an aggregate quantity.
  3. [§3.3 / Eq. (6) and §4.1] Because the DRO step explicitly upweights the worst normalized band, which in practice is the dark band, the PSNR-D improvement is to a significant degree a property of the objective, not an emergent benefit. The non-tautological claims are the simultaneous PSNR-B gain and aggregate gain. This should be acknowledged explicitly (the paper partly does in its ablation), and the bright-band/aggregate results should be emphasized as the decisive evidence. It would also be useful to report the per-band R_k and weights at convergence to show the mechanism.
minor comments (5)
  1. [Fig. 2 caption] Formatting issue: 'PSNR-D37.96, PSNR-B38.69' should read 'PSNR-D 37.96, PSNR-B 38.69'.
  2. [Abstract and §3.3] The phrase 'reconstructing dark pixels up to 6× worse relative to their per-band noise floor' uses the normalized ratio R_k before R_k is defined. Define it at first use or add a forward pointer.
  3. [§1 / Fig. 2] The text says the Jacobian norm 'increases monotonically with brightness,' but Figs. 2a and 2b appear to compare only the brightest and darkest bands. State explicitly whether monotonicity holds across all five bands or only B1<B5.
  4. [§4.1] The claim 'per-band gains reaching up to +0.89 dB' is not directly visible in Table 1; indicate which architecture/band yields this value, or correct the statement.
  5. [Eq. (5)] Ensure notation is consistent: earlier in the text y denotes clean target and x noisy input; in Eq. (5) the empirical variance uses (x_i - y_i)^2, which is correct under that convention but should be explicitly restated.

Circularity Check

0 steps flagged

No significant circularity: BBRD's dark-band objective is close to its loss by design, but the central simultaneous all-band and aggregate gains are empirical and not force-fitted.

full rationale

The derivation chain is not circular. Proposition 1 is a stated sufficient-condition argument (residuals approximate noise; measured Jacobian norms increase with brightness), and Eq. (2) is an empirical decomposition rather than a hidden identity. The BBRD loss (Eqs. 5-6) does upweight the band with the largest normalized error R_k, so an improvement in that band is partly the training objective; however, the paper's headline claim is the simultaneous improvement of every brightness band and the aggregate/bright-band gains, which are not forced by the loss definition and are supported by controlled comparisons across 13 losses and 8 architectures. The dark-band gain alone would be close to tautological, but the paper's central claim is the Pareto-frontier result, which carries independent empirical content. The single self-citation (ref. [5], 'The devil is in the crack orientation') is used only in a broad application-listing sentence and is not load-bearing. The independence assumption in Prop. 1 (E[r_i^2]·E[||J_i||^2]) is a potential correctness risk because the per-band Jacobian measurement sums gradients over all band pixels and does not exclude cross terms, but that is an empirical/statistical issue rather than a reduction of the output to the input. No fitted parameter is renamed as a prediction, and no uniqueness or ansatz is imported from the authors' prior work.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 0 invented entities

The paper's empirical payload rests on the standard gradient decomposition plus two domain assumptions (A1 residuals≈noise; A2 brightness-monotone Jacobian) and an unstated independence step in Eq. (2). BBRD's design introduces hand-chosen constants (η, σg, ε, BIC-selected K) and a data-derived statistic σ̂²_k; none of these is a target of evaluation, so they count as design choices rather than fitted predictions, but the η sensitivity claim is internally contradicted. The metrics not forced by construction are the bright-band +0.32 dB, aggregate +0.65 dB, and homoscedastic-noise generalization, which give BBRD its independent empirical grounding. No invented physical or formal entities are postulated; LumaSet-700 is a new benchmark whose release is deferred.

free parameters (5)
  • DRO temperature η = 5
    Eq. (6). Chosen for 'balance between worst-band focus and overall performance'; §3.3 claims stability over [0.5,10] but Tab. 4's η=1 row shows PSNR-D 36.80 vs 37.96 at η=5, a >1 dB swing, so the value is consequential and the stability claim is contradicted or the table is malformed.
  • Gaussian soft-assignment width σg = 0.05
    Eq. (3). Hand-set; no σg ablation appears in the main text.
  • Evaluation band thresholds for PSNR-D/PSNR-B = SIDD [0,0.2)/[0.8,1.0]; PolyU [0,0.45)/[0.7,1.0]
    §3.4. Headline gains are threshold-dependent: the paper reports dark gains from +0.45 dB at [0,0.3) to +0.58 dB at [0,0.15).
  • GMM component count K (BIC-selected) = 5 (SIDD), 3 (PolyU)
    §3.3. Data-driven via BIC; K=4 loses 0.24 dB and K=8 loses 0.30 dB PSNR-D, so the choice affects results but is selected by a criterion rather than test-set tuning.
  • Variance clip ε = 1e-6
    Eq. (5). Never triggered on evaluated data; safety floor only.
axioms (6)
  • standard math MSE gradient at pixel i is (2/N)(fθ(xi)-yi)·∂fθ(xi)/∂θ
    Eq. (1), §3.2. Standard chain-rule decomposition; uncontroversial.
  • domain assumption A1: in early-to-mid training residuals approximate the noise (ri ≈ ni)
    §3.2, Prop. 1(A1). Plausible for denoisers before convergence, but not shown from training curves.
  • domain assumption A2: Jacobian norm E[∥J_i∥²] increases monotonically with brightness
    §3.2 (A2), supported only by Figs. 2a/2b measurements; the mechanism offered (monotone nonlinearities amplify gradients) is heuristic and not proven for attention or state-space blocks.
  • ad hoc to paper Independence factorization E[r²J²] = E[r²]·E[J²]
    Eq. (2) in Prop. 1's proof. No independence or uncorrelatedness argument is given; residuals and Jacobians are plausibly correlated through activation magnitude.
  • domain assumption Signal-dependent (Poisson-Gaussian) noise with per-band variance estimable from clean-noisy training pairs
    §3.3, Eq. (5). σ̂²_k is precomputed from (xi-yi)² and assumed stationary within each dataset (stability reported: CV<15%, ISO subgroups <8%).
  • domain assumption Softmax DRO inherits Group DRO's convergence guarantees
    §2/§3.3 citing Sagawa et al. [24]; the cited guarantees apply to the worst-group minimax objective, not to the softmax-reweighted sum in Eq. (6).

pith-pipeline@v1.3.0-alltime-deepseek · 12632 in / 30882 out tokens · 265228 ms · 2026-08-02T04:42:03.860831+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of The Devil is in the Dark Pixels: Toward Brightness Bias-Robust Denoising." pith.science (2026). https://pith.science/paper/OF2VXJVY

@misc{pith2026260716320,
  author       = {Pith},
  title        = {Pith review of: The Devil is in the Dark Pixels: Toward Brightness Bias-Robust Denoising},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OF2VXJVY}},
  note         = {Machine review of arXiv:2607.16320}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

In this paper, we reveal an important yet overlooked problem in image denoising: under signal-dependent camera noise models, dark regions suffer from inherently low Signal-to-Noise Ratio (SNR), as signal intensity decays far faster than noise variance diminishes, making detail recovery in dark areas fundamentally challenging. Yet rather than compensating for this difficulty, MSE-trained denoisers exacerbate it -- reconstructing dark pixels up to 6x worse relative to their per-band noise floor. This bias stems from two compounding factors: signal-dependent noise inflates bright-pixel residuals, and the network's Jacobian norm increases monotonically with brightness. Together, these cause bright regions to chronically dominate gradient updates at the expense of dark ones. To this end, we propose Brightness Bias-Robust Denoising (BBRD), a drop-in replacement for MSE loss that partitions pixels into brightness bands, normalizes per-band error by empirical noise variance, and applies Group Distributionally Robust Optimization (Group-DRO) to dynamically upweight whichever band is currently worst, with zero additional parameters or inference cost. Across 8 architectures and 2 datasets in our experiments, BBRD is the only method among 13 tested alternatives that improves each brightness band simultaneously, achieving up to +0.45 dB on dark bands, +0.32 dB on bright bands, and +0.65 dB aggregate Peak Signal-to-Noise Ratio (PSNR) on SIDD, with the largest per-band gains in the darkest regions where detail recovery matters most. Code is available at https://github.com/xmed-lab/BBRD

Figures

Figures reproduced from arXiv: 2607.16320 by Sungjun Cho, Xiaomeng Li, ZhuangZhuang Chen.

Figure 1
Figure 1. Figure 1: (a) Brightness distribution across datasets. (b) Under identical NAFNet train￾ing on SIDD dataset, MSE under-optimizes dark regions (1.3 dB gap at convergence); BBRD reduces this to 0.8 dB with zero additional parameters. share a common training paradigm: per-pixel MSE or ℓ1 minimization, evalu￾ated by a single aggregate PSNR. We reveal an important yet overlooked failure of this paradigm: brightness bias.… view at source ↗
Figure 2
Figure 2. Figure 2: Jacobian norm of the brightest band consistently exceeds that of the darkest band under MSE across 8 architectures (CNNs, Transformers, SSMs) on (a) SIDD and (b) PolyU, confirming the gradient imbalance is architecture-agnostic. (c) BBRD (PSNR-D 37.96, PSNR-B 38.69) strictly dominates all baselines—including the strongest, ℓ1+FFT (37.70, 38.28) and MSE (37.51, 38.37)—pushing the Pareto frontier outward on … view at source ↗
Figure 3
Figure 3. Figure 3: BBRD training pipeline. A smoothed brightness map partitions pixels into K bands via GMM and Gaussian soft assignments (orange). Per-band MSE is normal￾ized by empirical noise variance σˆ 2 k, yielding Rk (teal). Softmax DRO (η=5) upweights the worst band, forming LBBRD = P k wkRk (dark teal). Band computation is training￾only; inference runs fθ with zero overhead. naturally cluster in brightness space, pl… view at source ↗
Figure 4
Figure 4. Figure 4: Brightness band construction. (1) Input patch. (2) Smoothed brightness map I¯ via AvgPool. (3) GMM fitted to training-set intensities; boundaries placed at inter-component valleys (K∗=5 for SIDD, K∗=3 for PolyU). (4) Gaussian soft weights ϕik ensure smooth pixel-to-band assignment. Interpretation: The normalized ratio Rk has a natural interpretation: Rk ≈ 1 means no denoising progress; Rk → 0 is perfect re… view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparison (SIDD and PolyU). [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Training dynamics and DRO stability. (a) [PITH_FULL_IMAGE:figures/full_fig_p014_6.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

36 extracted references · 4 canonical work pages

  1. [1]

    In: CVPR (2018) 10

    Abdelhamed, A., Lin, S., Brown, M.S.: A high-quality denoising dataset for smart- phone cameras. In: CVPR (2018) 10

  2. [2]

    arXiv preprint arXiv:2309.15563 (2023) 10, 11

    Benjdira, B., Ali, A.M., Koubaa, A.: Guided frequency loss for image restoration. arXiv preprint arXiv:2309.15563 (2023) 10, 11

  3. [3]

    In: ICIP

    Charbonnier, P., Blanc-Féraud, L., Aubert, G., Barlaud, M.: Two deterministic half-quadratic regularization algorithms for computed imaging. In: ICIP. vol. 2, pp. 168–172 (1994) 5, 10, 11

  4. [4]

    In: ECCV (2022) 1, 2, 4, 10

    Chen, L., Chu, X., Zhang, X., Sun, J.: Simple baselines for image restoration. In: ECCV (2022) 1, 2, 4, 10

  5. [5]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)

    Chen, Z., Zhang, J., Lai, Z., Zhu, G., Liu, Z., Chen, J., Li, J.: The devil is in the crack orientation: A new perspective for crack detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 6653–6663 (Oct 2023) 1

  6. [6]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Dong, L., Fan, Q., Guo, Y., Wang, Z., Zhang, Q., Chen, J., Luo, Y., Zou, C.: TSD- SR: One-step diffusion with target score distillation for real-world image super- resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 23174–23184 (Jun 2025) 4

  7. [7]

    SIAM Journal on Imaging Sciences16(3), 1594–1654 (2023).https://doi.org/10.1137/23M15458591

    Elad, M., Kawar, B., Vaksman, G.: Image denoising: The deep learning revolution and beyond—A survey paper. SIAM Journal on Imaging Sciences16(3), 1594–1654 (2023).https://doi.org/10.1137/23M15458591

  8. [8]

    In: CVPRW (2024) 10, 11

    Fu, Y., et al.: NTIRE 2024 challenge on low light image enhancement: Methods and results. In: CVPRW (2024) 10, 11

  9. [9]

    In: CVPR (2025) 1, 4, 10

    Guo, H., Guo, Y., Zha, Y., Zhang, Y., Li, W., Dai, T., Xia, S.T., Li, Y.: Mam- baIRv2: Attentive state space restoration. In: CVPR (2025) 1, 4, 10

  10. [10]

    In: ECCV (2024) 1, 4, 10

    Guo, H., Li, J., Dai, T., Ouyang, Z., Ren, X., Xia, S.T.: MambaIR: A simple baseline for image restoration with state-space model. In: ECCV (2024) 1, 4, 10

  11. [11]

    In: NeurIPS

    Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS. vol. 33, pp. 6840–6851 (2020), arXiv:2006.11239 4

  12. [12]

    The Annals of Mathemat- ical Statistics35(1), 73–101 (1964) 5, 10, 11

    Huber, P.J.: Robust estimation of a location parameter. The Annals of Mathemat- ical Statistics35(1), 73–101 (1964) 5, 10, 11

  13. [13]

    Information Fusion (2025).https://doi.org/10.1016/j.inffus.2025.1008671

    Jiang, B., et al.: Eficient image denoising using deep learning: A brief survey. Information Fusion (2025).https://doi.org/10.1016/j.inffus.2025.1008671

  14. [14]

    In: ICCV (2021) 10, 11

    Jiang, L., Dai, B., Wu, W., Loy, C.C.: Focal frequency loss for image reconstruction and synthesis. In: ICCV (2021) 10, 11

  15. [15]

    In: ECCV (2016) 5

    Johnson, J., Alahi, A., Fei-Fei, L.: Perceptual losses for real-time style transfer and super-resolution. In: ECCV (2016) 5

  16. [16]

    Cho et al

    Kendall, A., Gal, Y.: What uncertainties do we need in Bayesian deep learning for computer vision? In: NeurIPS (2017) 5, 10, 11 16 S. Cho et al

  17. [17]

    In: ECCV (2022) 5

    Khor, H.Q., Jiang, J., Ong, E.J.: Wavelet-based dual-branch network for image demoireing. In: ECCV (2022) 5

  18. [18]

    In: ICCVW (2025), arXiv:2510.06601 1

    Li, F., Li, J., Conde, M., et al.: AIM 2025 challenge on real-world RAW image denoising. In: ICCVW (2025), arXiv:2510.06601 1

  19. [19]

    In: ICCVW (2021) 1, 4, 5, 10

    Liang, J., Cao, J., Sun, G., Zhang, K., Gool, L.V., Timofte, R.: SwinIR: Image restoration using swin transformer. In: ICCVW (2021) 1, 4, 5, 10

  20. [20]

    In: ICCV (2017) 3, 5, 10, 11

    Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollár, P.: Focal loss for dense object detection. In: ICCV (2017) 3, 5, 10, 11

  21. [21]

    In: ECCV (2024), arXiv:2308.15070 4

    Lin, X., He, J., Chen, Z., Lyu, Z., Dai, B., Yu, F., Qiao, Y., Ouyang, W., Dong, C.: DiffBIR: Toward blind image restoration with generative diffusion prior. In: ECCV (2024), arXiv:2308.15070 4

  22. [22]

    Sensors25(8), 2615 (Apr 2025).https://doi.org/10.3390/s25082615, pMCID: PMC12031399 2

    Mao, J., Sun, L., Chen, J., Yu, S.: Overview of research on digital image denoising methods. Sensors25(8), 2615 (Apr 2025).https://doi.org/10.3390/s25082615, pMCID: PMC12031399 2

  23. [23]

    Circuits, Systems, and Signal Processing (2026).https://doi.org/10.1007/s00034-025-03477-z1

    Pathak, K., Bhandari, A.K.: Advances in deep learning and filtering models for medical image denoising: A review of current and future trends. Circuits, Systems, and Signal Processing (2026).https://doi.org/10.1007/s00034-025-03477-z1

  24. [24]

    In: ICLR (2020) 5

    Sagawa, S., Koh, P.W., Hashimoto, T.B., Liang, P.: Distributionally robust neu- ral networks for group shifts: On the importance of regularization for worst-case generalization. In: ICLR (2020) 5

  25. [25]

    In: CVPR (2016) 3, 5, 10, 11

    Shrivastava, A., Gupta, A., Girshick, R.: Training region-based object detectors with online hard example mining. In: CVPR (2016) 3, 5, 10, 11

  26. [26]

    In: CVPRW (2025), arXiv:2504.12276 1

    Sun, L., Guo, H., et al.: The tenth NTIRE 2025 image denoising challenge report. In: CVPRW (2025), arXiv:2504.12276 1

  27. [27]

    In: CVPR (2022) 5, 10, 11

    Wang, Z., Cun, X., Bao, J., Zhou, W., Liu, J., Li, H.: Uformer: A general U-shaped transformer for image restoration. In: CVPR (2022) 5, 10, 11

  28. [28]

    IEEE TIP13(4), 600–612 (2004) 5, 10, 11

    Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: From error visibility to structural similarity. IEEE TIP13(4), 600–612 (2004) 5, 10, 11

  29. [29]

    In: Asilomar Conference on Signals, Systems and Computers

    Wang, Z., Simoncelli, E.P., Bovik, A.C.: Multiscale structural similarity for image quality assessment. In: Asilomar Conference on Signals, Systems and Computers. pp. 1398–1402 (2003) 5, 10, 11

  30. [30]

    arXiv preprint arXiv:2406.08177 (2024).https:// doi.org/10.48550/arXiv.2406.081774

    Wu, R., Sun, L., Ma, Z., Zhang, L.: One-step effective diffusion network for real- world image super-resolution. arXiv preprint arXiv:2406.08177 (2024).https:// doi.org/10.48550/arXiv.2406.081774

  31. [31]

    arXiv preprint arXiv:1804.02603 (2018) 10

    Xu, J., Li, H., Liang, Z., Zhang, D., Zhang, L.: Real-world noisy image denoising: A new benchmark. arXiv preprint arXiv:1804.02603 (2018) 10

  32. [32]

    In: CVPR (2022) 1, 4, 5, 10

    Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: CVPR (2022) 1, 4, 5, 10

  33. [33]

    Machine Intelligence Research (2023).https://doi.org/10.1007/s11633- 023-1466-010

    Zhang, K., Li, Y., Liang, J., Cao, J., Zhang, Y., Tang, H., Fan, D.P., Timofte, R., Gool, L.V.: Practical blind image denoising via Swin-Conv-UNet and data syn- thesis. Machine Intelligence Research (2023).https://doi.org/10.1007/s11633- 023-1466-010

  34. [34]

    IEEE TPAMI (2021).https://doi

    Zhang, K., Li, Y., Zuo, W., Zhang, L., Gool, L.V., Timofte, R.: Plug-and-play image restoration with deep denoiser prior. IEEE TPAMI (2021).https://doi. org/10.1109/TPAMI.2021.30889141, 4, 10

  35. [35]

    IEEE Transactions on Im- age Processing26(7), 3142–3155 (2017).https://doi.org/10.1109/TIP.2017

    Zhang, K., Zuo, W., Chen, Y., Meng, D., Zhang, L.: Beyond a Gaussian denoiser: Residual learning of deep CNN for image denoising. IEEE Transactions on Im- age Processing26(7), 3142–3155 (2017).https://doi.org/10.1109/TIP.2017. 26622061, 4 Brightness Bias in Learned Image Denoising 17

  36. [36]

    arXiv preprint arXiv:2303.02881 (2023) 10

    Zhang, Y., Li, D., Shi, X., He, D., Song, K., Wang, X., Qin, H., Li, H.: KBNet: Kernel basis network for image restoration. arXiv preprint arXiv:2303.02881 (2023) 10