Pith. sign in

REVIEW 3 major objections 5 minor 48 references

Low-light cameras go color-blind mainly from black-level error; a small global estimator corrects it before denoising and restores natural colors without camera calibration.

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 · grok-4.5

2026-07-14 07:07 UTC pith:VFIEICSC

load-bearing objection Solid systems paper that isolates black-level error as a fixable global color-bias source in blind low-light raw denoising and ships a corrected SIDD variant. the 3 major comments →

arxiv 2607.11090 v1 pith:VFIEICSC submitted 2026-07-13 cs.CV

Why Low-Light Cameras Go Color Blind: Removing Color Bias in Raw Denoising

Mohammad Mohammadi , Sina Honari , Stavros Tsogkas , Tristan Aumentado-Armstrong , Michael S. Brown , Iqbal Mohomed , Konstantinos G. Derpanis , Alex Levinshtein
show 1 more author
Igor Gilitschenski
This is my paper
classification cs.CV
keywords low-light raw denoisingblack-level errorcolor biasblind denoisingcamera-agnosticSIDD-CCdark-shading 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.

Low-light raw photographs suffer from severe color shifts that make dark regions look purple or hazy even after modern denoisers. The paper shows that the dominant cause is black-level error: a small mismatch between the true sensor black level and the value written in the file metadata. Because low-light pipelines amplify the image by 100× or more, that tiny offset becomes a large, global color cast. The authors introduce a lightweight bias-estimator network that reads the noisy raw image, predicts a single per-channel offset, subtracts it, and only then hands the corrected image to a standard denoiser. Trained end-to-end on synthetic noise with no camera-specific calibration, the method outperforms other blind denoisers on ELD, SID and LRID and often matches or exceeds calibrated or paired-data baselines on color accuracy. The same insight also reveals that the popular SIDD ground-truth images themselves contain the identical bias; a corrected SIDD-CC release removes it and changes how existing models should be scored.

Core claim

Black-level error is the primary source of the color bias that appears in low-light raw denoising. Modeling that error as a single global per-channel offset, predicting it from the noisy image alone, and correcting it before any denoising step yields state-of-the-art color fidelity among calibration-free methods and frequently matches methods that require sensor-specific dark frames or paired data.

What carries the argument

The black-level bias estimator (BLBE): a shared-encoder network that outputs one residual offset per RGBG channel. The predicted offset is added to the metadata black level, the image is re-normalized, and only then is the result passed to the denoiser.

Load-bearing premise

That a single global per-channel offset estimated from one noisy image is enough to capture the main dark-shading artifact, and that residual fixed-pattern noise can be left unmodeled without spoiling the offset estimate.

What would settle it

On a sensor whose dark-shading is dominated by strong spatially varying fixed-pattern noise (for example LRID at 1024 imes amplification), the method should lose its color-accuracy advantage over plain PGRQ baselines; residual green or purple casts should remain after BLE correction.

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

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 argues that black-level error (BLE) is a primary driver of color bias (unnatural chromatic shifts, especially in dark regions) in low-light raw denoising. It proposes a camera-agnostic, calibration-free pipeline: a black-level bias estimator (BLBE) network (shared U-Net encoder + MLP head) predicts a global per-channel offset E_bl ~ U(-M_bl, M_bl) from the noisy input; this corrects the metadata black-level before a standard U-Net denoiser. Training uses only clean SID-Sony images plus synthetic PGRQB noise (shot/read/row/quantization + BLE) with broad, uncorrelated parameter ranges. Evaluations on real ELD (Sony/Nikon), SID, and LRID report superior PSNR/SSIM/CIEDE2000 among blind methods and competitiveness with calibrated/paired baselines (Tables 1, 6, 7), with qualitative color recovery. Separately, the authors diagnose clipping-before-averaging bias in SIDD ground-truth construction, release a corrected SIDD-CC pipeline and split, and benchmark existing methods on it.

Significance. If the scoped claim holds, the work meaningfully advances practical blind low-light raw denoising by isolating a simple, globally correctable artifact (BLE) that prior physics-based and learning-based synthesizers under-address without per-sensor dark frames or paired data. The progressive noise-model ablations (PG o PGRQ o PGRQB), the PGRQ+BLBE isolation experiment, capacity/resolution/clipping/per-channel controls (Table 2), and multi-metric multi-sensor results constitute solid, reproducible evidence. Explicit FPN limitation discussion (Appendix G) and the SIDD-CC contribution (new GT extraction + initial benchmark) are community assets. Strengths include end-to-end training with auxiliary L1 on BLE, broad noise sampling for generalization, and honest residual-error analysis on high-ratio LRID.

major comments (3)
  1. [Tables 1, 6, 7 / Sec. 5.3] Tables 1, 6 and 7 report single-run point estimates with no standard deviations, multiple random seeds, or bootstrap intervals, despite stochastic noise synthesis and random cropping. The headline gains (e.g., +2.62 dB PSNR over PGRQB on ELD-Nikon imes200, +2.57 dB on SID imes100) are therefore difficult to judge for statistical reliability; at minimum a 3-seed mean/std or paired significance test on the same test images is needed to support the 'superior among blind / competitive with calibrated' claims.
  2. [Sec. 5.5 / Appendix D] Appendix D states that only five of the original 40 SIDD validation scenes retain the full 150-frame RAW stacks required for the corrected GT pipeline, yielding a 29-scene / 928-patch validation set after filtering. This is too small and device-skewed (S6 Edge heavily pruned) to serve as a stable community benchmark; either expand the set (or release the raw stacks if recoverable) or clearly qualify SIDD-CC results as preliminary and report per-device breakdowns.
  3. [Eq. (7) / Sec. 3.2 / Appendix C, G] The central modeling assumption (Eq. 7 + Sec. 3.2) treats BLE as a single global per-channel offset while leaving fixed-pattern noise (FPN) unmodeled. Appendix C and G correctly flag residual color errors when FPN dominates (high-ratio LRID, high-ISO SID), yet the main-text claim that BLE is 'a primary source' is not accompanied by a quantitative breakdown of BLE vs. FPN variance across the evaluated exposure ratios; without it the reader cannot verify that the reported gains are not partly confounded by residual FPN that the BLBE inadvertently absorbs.
minor comments (5)
  1. [Figs. 3, 5] Fig. 3 and Fig. 5 captions claim 'most closely matches the ground-truth color' but do not report per-image CIEDE2000 or channel-wise mean offsets; adding these numbers would make the visual claim quantitative.
  2. [Appendix A] The BLE predictor head (Appendix A) uses Tanh followed by rescaling; the exact mapping from normalized output to physical E_bl units is not stated, which hinders re-implementation.
  3. [Table 2f] Table 2f 'sensor adaptation' mixes ELD-Sony calibrated PGRQ ranges with blind BLE; clarify whether the same ranges were used for the PGRQB baseline or only for the adapted variant.
  4. [Sec. 3.4 / Fig. 2] Minor notation inconsistency: E_bl vs. Ebl' vs. E_bl' appears in text, equations and Fig. 2; standardize.
  5. [Sec. 5.2] The illumination-correction step (footnote 1) is applied post-denoising for all methods; confirm it does not interact with residual BLE offsets and report uncorrected PSNR as a sensitivity check.

Circularity Check

0 steps flagged

No significant circularity: BLE estimator is standard supervised recovery of a synthetically injected offset, validated externally on real multi-sensor benchmarks.

full rationale

The paper's central chain is empirical, not a first-principles derivation that collapses to its inputs. Clean SID-Sony images receive synthetic noise (shot/read/row/quantization + independent per-channel BLE drawn from a broad uniform range, Sec. 3.1–3.2 and Tab. 5) with no target-sensor calibration. The BLBE head is trained with an ordinary L1 loss against the known injected E_bl (Eq. 9); the corrected black-level is then fed to a U-Net whose L1 image loss (Eq. 10) is the only other term. At test time the same network is applied, without re-fitting, to real noisy captures from ELD-Sony, ELD-Nikon, SID and LRID; gains are measured by PSNR/SSIM/CIEDE2000 against the datasets’ own ground truth (Tabs. 1, 6, 7). Progressive ablations (PG → PGRQ → PGRQB → full model) and the isolation experiment that attaches a frozen BLBE to an off-the-shelf PGRQ baseline further show that the performance lift is not forced by construction. The SIDD-CC reprocessing merely changes the order of averaging versus clipping and is likewise externally falsifiable. No self-citation supplies a uniqueness theorem, no fitted constant is renamed a prediction, and no ansatz is smuggled in. The only residual limitation (unmodeled FPN) is explicitly scoped as future work (App. G) and does not underwrite the reported BLE claim. Hence the derivation is self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 2 invented entities

The method rests on standard CMOS noise decomposition plus the modeling choice that BLE is a uniform per-channel offset drawable from a fixed uniform range; all other quantities are either taken from prior calibrated ranges or set by ordinary training hyper-parameters.

free parameters (4)
  • alpha (BLE loss weight) = 1.0
    Set to 1.0 by hand; balances L1 image loss against L1 BLE loss (Eq. 11).
  • M_bl (max BLE magnitude) = 2
    Constant 2 used for uniform sampling of synthetic BLE (Eq. 7, Table 5); chosen to cover observed sensor ranges after widening.
  • noise-parameter ranges (K, sigma_read, sigma_row, exposure ratio) = K in [0.05,30], log-sigma_read in [-2,3], etc.
    Broadened by >2x from the four-sensor calibrations of Wei et al. 2021; sampled independently per patch (Table 5).
  • training crop size / epochs / lr = 1536^2 / 750 / 2e-4
    1536^2 crops, 750 epochs, 2e-4 Adam with cosine schedule; selected via ablation on ELD-Nikon.
axioms (4)
  • domain assumption Recorded raw intensity equals clean signal plus additive shot/read/banding/quantization noise (Eqs. 1-6).
    Standard CMOS noise formation model used throughout Sec. 3.1; taken from prior literature.
  • ad hoc to paper Black-level error can be modeled as a single global offset drawn from U(-M_bl, M_bl) and applied uniformly per RGBG channel (Eq. 7).
    Core modeling choice of Sec. 3.2; justified by metadata fluctuations but not derived from first principles.
  • ad hoc to paper Fixed-pattern noise is secondary to BLE for the evaluated exposure ratios and can be left unmodeled.
    Stated limitation in Appendix G; heat-maps in Appendix C show FPN is non-negligible at high ISO/ratio.
  • domain assumption Illumination mismatch between noisy and clean captures can be corrected by a single multiplicative factor after denoising.
    Standard practice cited from prior ELD/SID papers; applied uniformly in Sec. 5.2.
invented entities (2)
  • Black-level bias estimator (BLBE) network no independent evidence
    purpose: Predicts the four per-channel BLE offsets from a single noisy raw image so they can be subtracted before denoising.
    New architectural component (shared U-Net encoder + MLP head) introduced in Sec. 3.4; no independent physical existence outside the trained model.
  • SIDD-CC ground-truth extraction pipeline no independent evidence
    purpose: Averages aligned noisy frames before black-level subtraction and clipping to avoid the purple bias of the original SIDD GT.
    New processing order defined in Sec. 4; produces a corrected dataset whose existence is internal to the paper.

pith-pipeline@v1.1.0-grok45 · 31750 in / 3094 out tokens · 30924 ms · 2026-07-14T07:07:39.292964+00:00 · methodology

0 comments
read the original abstract

Raw images inherently suffer from noise due to the stochastic nature of light and sensor hardware imperfections. As real photon counts fall, the ratio of this noise to the signal degrades; consequently, for low-light conditions, robust denoising is especially vital for high-quality results. While recent data-driven methods achieve strong performance, they typically rely on large-scale noisy-clean image pairs that are costly and difficult to collect. Alternatively, parametric noise models can generate synthetic training data, but this necessitates precise camera calibration, which is often impractical for unknown devices. In this work, we propose a camera-agnostic, calibration-free paradigm for low-light raw denoising. We identify that color bias from black-level error is a primary source of performance degradation and causes severe color shifts. To mitigate this, we introduce a bias estimator network that predicts the black-level error as a global feature of the noisy input. We evaluate our approach across the ELD, SID, and LRID datasets, demonstrating superior performance among blind denoisers, particularly in terms of color correction. In many cases, we are competitive with-or can even surpass-methods with stronger supervision. Furthermore, we reveal that the widely used SIDD dataset contains significant color bias in its ground-truth images, which yields unrealistic color reproduction in trained models. We introduce a new ground-truth extraction framework to resolve this issue and provide a benchmark of existing methods on the corrected dataset.

Figures

Figures reproduced from arXiv: 2607.11090 by Alex Levinshtein, Igor Gilitschenski, Iqbal Mohomed, Konstantinos G. Derpanis, Michael S. Brown, Mohammad Mohammadi, Sina Honari, Stavros Tsogkas, Tristan Aumentado-Armstrong.

Figure 1
Figure 1. Figure 1: Identification and mitigation of color bias (CB) in low-light denoising. Left: CB induced by black-level error (BLE) in the denoising pipeline. Right: CB due to raw processing of GT images, based on multi-frame fusion. Bottom: Diagram of our proposed CB-correcting denoising method, including the raw BLE estimator. Abstract—Raw images inherently suffer from noise due to the stochastic nature of light and se… view at source ↗
Figure 2
Figure 2. Figure 2: Training and inference pipeline. During training (top), synthetic noise and black-level error (BLE), Ebl, are added to the clean image, which is then passed to the BLE predictor. The estimated BLE, Ebl’, is then used to update the black-level, ℓbl, in the preprocessing before the denoising step (right). The model is trained end-to-end. For inference (bottom) on real noisy images, the BLE predictor is used … view at source ↗
Figure 3
Figure 3. Figure 3: Black-level-induced color bias handling of different noise models on the ELD-Sony [18] dataset. Progressive noise modeling with PG (Poisson-Gaussian), PGRQ (+Row, Quantization), and PGRQB (+Black-Level Error) show incremental improvement, yet color bias persists. The calibrated ELD noise model, and the paired-data diffusion-based noise synthesis model of NoiseDiff also exhibit color bias. Our noise modelin… view at source ↗
Figure 4
Figure 4. Figure 4: SIDD vs. SIDD-CC. Black-level (BL) denotes the lowest intensity recorded by a camera sensor. In practice, however, raw image values do not zero-out below BL, due to dark noise (top-left histogram). As a result, when generating GT from noisy bursts, the order of processing steps, such as black-level subtraction (BLS), clipping, and averaging, can impact color quality. The original SIDD pipeline (top flow di… view at source ↗
Figure 5
Figure 5. Figure 5: Comparison on ELD-Nikon images [18]. Progressive noise modeling with PG (Poisson-Gaussian), PGRQ (+Row, Quantization), and PGRQB (+Black-Level Error) all exhibit color bias. The calibrated ELD and 2-Shots noise models also show it. In contrast, our noise modeling approach has much more accurate colors (see zoomed inset in particular). suffer a noticeable drop on some sensors. We consistently outperform PGR… view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative comparison on SIDD [36] and SIDD￾CC. (d), (e), and (h) contain the purple tint seen in (b), while (f) and (g) match the dark regions of (c). inputs are identical between SIDD and SIDD-CC; only the ground truth differs. As shown in Figs. 6b and 6c, the original SIDD ground truth exhibits a noticeable purple tint compared to SIDD-CC. Comparing two supervised NAFNet models trained on SIDD and SIDD… view at source ↗
Figure 7
Figure 7. Figure 7: Stage-by-stage visualization of our denoising pipeline on an ELD-Sony [18] ×200 sample. (a) The cap￾tured low-light raw input is severely underexposed and barely visible. (b) Multiplying by the dataset-provided ×200 gain matches the ground-truth exposure but reveals strong noise and a pronounced color shift. (c) The BLBE mod￾ule predicts the four per-channel black-level error offsets and subtracts each fro… view at source ↗
Figure 8
Figure 8. Figure 8: Additional qualitative comparisons on ELD-Sony [18]. Two scenes are shown across four rows (blind baselines on top, calibrated baselines and ours on the bottom of each scene). PG and PGRQ exhibit residual color bias, and PGRQB introduces local color distortions, whereas our method closely matches the ground truth without target-sensor data (see zoomed insets). C.1 Dark-shading Noise in ELD-Sony, SID, and L… view at source ↗
Figure 9
Figure 9. Figure 9: Additional qualitative comparisons on ELD-Nikon [18]. Two scenes are shown across four rows (blind baselines on top, calibrated baselines and ours on the bottom of each scene). PG and PGRQ exhibit residual color bias, and PGRQB introduces local color distortions, whereas our method closely matches the ground truth without target-sensor data (see zoomed insets). of dark shading is substantially smaller, the… view at source ↗
Figure 10
Figure 10. Figure 10: Additional qualitative comparisons on SID [42]. Two scenes across four rows (blind baselines on top, calibrated baselines and ours on the bottom of each scene). PG and PGRQ exhibit pronounced color shifts due to uncompensated black-level error, while our method recovers colors that closely match the ground truth (see zoomed insets). A first observation is that the dark shading characteris￾tics of the SID … view at source ↗
Figure 11
Figure 11. Figure 11: Qualitative comparisons on LRID [19] at the 256× exposure ratio. As discussed in Sec. C.1, the dark shading noise—and hence the BLE—is weak in this setting. Our method performs on par with the blind baselines, while PGRQB and ELD introduce noticeable color artifacts. 0.03 0.02 0.01 0.00 0.01 0.02 Black-Level Error 0.0 2.5 5.0 Frequency R 0.03 0.02 0.01 0.00 0.01 0.02 Black-Level Error 0.0 2.5 5.0 Frequenc… view at source ↗
Figure 12
Figure 12. Figure 12: Distribution of predicted black-level error (BLE) on SID ×250 [42]. Per-channel histograms of the BLE offsets estimated by our BLBE module across the RGBG channels of the test set. ISO and exposure time, and the resulting predictions are averaged to correct all corresponding noisy inputs. This aggregation stabilizes the estimate when individual images lack sufficient visual content for accurate per-image … view at source ↗
Figure 13
Figure 13. Figure 13: Dark-shading noise heatmaps for the SID and LRID sensors at ISO 6400. Each subfigure displays the spatial distribution of dark shading noise for the four Bayer channels (R, G1, B, G2), computed after black-level subtraction and normalization by the white level. The 1st and 99th percentile values of each channel are annotated in the corresponding panel (denoted p1 and p99). The LRID sensor exhibits dark sh… view at source ↗
Figure 14
Figure 14. Figure 14: Synthetic dark-shading denoising across sensors and ISO settings. A clean image from SID is corrupted with dark shading noise (amplified by 200×) drawn from four different sources: the SID sensor at ISO 1600, 6400, and 12800 (rows 1–3), and the LRID sensor at ISO 6400 (row 4). PGRQ fails to suppress the dark-shading across all configurations. PGRQB attenuates the overall dark-shading magnitude but introdu… view at source ↗
Figure 15
Figure 15. Figure 15: Qualitative comparison on LRID [19] at the 1024× exposure ratio. A representative failure case for our approach. PG and PGRQ exhibit residual dark shading noise. PGRQB alters the colors but introduces local inconsistencies. ELD, despite applying dark shading correction prior to denoising, fails to recover accurate colors, whereas 2-Shots produces the result closest to the ground truth. Our model partially… view at source ↗
Figure 16
Figure 16. Figure 16: Effect of pixel clipping on SIDD ground-truth quality. Patches cropped from SIDD ground-truth images of a scene containing a color checker, captured at three different ISO levels. As the fraction of initially clipped pixels increases, dark regions develop a pronounced purple tint due to the biased mean estimation. Below our 1% threshold (a), the bias is negligible. APPENDIX E MEMORY AND RUNTIME EFFICIENCY… view at source ↗
Figure 18
Figure 18. Figure 18: Qualitative results on real-world low-light captures from an iPhone 16 Pro. All images are captured under the same lighting condition, and noisy inputs are amplified by 100× for visualization. As ISO increases, dark shading noise becomes increasingly dominant. PGRQ exhibits a yellowish tint caused by uncompensated black-level error, while PGRQB introduces local color artifacts, most notably in the ISO 160… view at source ↗
Figure 19
Figure 19. Figure 19: Failure case on a well-exposed capture from an iPhone 16 Pro. The scene is captured under normal lighting with a low ISO setting, so the input is essentially noise-free and free of black-level error. PGRQ preserves the true colors. PGRQB introduces local color distortions, while our model predicts a spurious black-level error because the image content does not provide a sufficient signal for accurate BLE … 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

48 extracted references · 1 linked inside Pith

  1. [1]

    Ma ˆıtre,From photon to pixel: the digital camera hand- book

    H. Ma ˆıtre,From photon to pixel: the digital camera hand- book. John Wiley & Sons, 2017

    2017

  2. [2]

    Deep joint demosaicking and denoising,

    M. Gharbi, G. Chaurasia, S. Paris, and F. Durand, “Deep joint demosaicking and denoising,”ACM Trans- actions on Graphics (TOG), vol. 35, no. 6, pp. 1–12, 2016

    2016

  3. [3]

    Practical blind image denoising via Swin-Conv-UNet and data synthesis,

    K. Zhang, Y. Li, J. Liang, J. Cao, Y. Zhang, H. Tang, D.-P . Fan, R. Timofte, and L. V . Gool, “Practical blind image denoising via Swin-Conv-UNet and data synthesis,” Machine Intelligence Research, vol. 20, no. 6, pp. 822–836, 2023

    2023

  4. [4]

    Joint HDR denoising and fusion: A real- world mobile HDR image dataset,

    S. Liu, X. Zhang, L. Sun, Z. Liang, H. Zeng, and L. Zhang, “Joint HDR denoising and fusion: A real- world mobile HDR image dataset,” inIEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023

    2023

  5. [5]

    Rethinking learning- based demosaicing, denoising, and super-resolution pipeline,

    G. Qian, Y. Wang, J. Gu, C. Dong, W. Heidrich, B. Ghanem, and J. S. Ren, “Rethinking learning- based demosaicing, denoising, and super-resolution pipeline,” inProceedings of the IEEE International Con- ference on Computational Photography (ICCP), 2022

    2022

  6. [6]

    DNF: Decouple and feedback network for seeing in the dark,

    X. Jin, L.-H. Han, Z. Li, C.-L. Guo, Z. Chai, and C. Li, “DNF: Decouple and feedback network for seeing in the dark,” inIEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023

    2023

  7. [7]

    IDF: Iterative dynamic filtering networks for generalizable image de- noising,

    D. Kim, J. Ko, M. K. Ali, and T. H. Kim, “IDF: Iterative dynamic filtering networks for generalizable image de- noising,” inInternational Conference on Computer Vision (ICCV), 2025

    2025

  8. [8]

    Fast and high quality image denoising via malleable convolution,

    Y. Jiang, B. Wronski, B. Mildenhall, J. T. Barron, Z. Wang, and T. Xue, “Fast and high quality image denoising via malleable convolution,” inEuropean Con- ference on Computer Vision (ECCV), 2022

    2022

  9. [9]

    Plug-and-play image restoration with deep denoiser prior,

    K. Zhang, Y. Li, W. Zuo, L. Zhang, L. Van Gool, and R. Timofte, “Plug-and-play image restoration with deep denoiser prior,”IEEE Transactions on Pattern Anal- ysis and Machine Intelligence (TP AMI), 2021

    2021

  10. [10]

    Restormer: Efficient transformer for high-resolution image restoration,

    S. W. Zamir, A. Arora, S. Khan, M. Hayat, F. S. Khan, and M.-H. Yang, “Restormer: Efficient transformer for high-resolution image restoration,” inIEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022

    2022

  11. [11]

    Simple base- lines for image restoration,

    L. Chen, X. Chu, X. Zhang, and J. Sun, “Simple base- lines for image restoration,” inEuropean Conference on Computer Vision (ECCV), 2022

    2022

  12. [12]

    Noise flow: Noise modeling with conditional normal- izing flows,

    A. Abdelhamed, M. A. Brubaker, and M. S. Brown, “Noise flow: Noise modeling with conditional normal- izing flows,” inInternational Conference on Computer Vision (ICCV), 2019

    2019

  13. [13]

    Dark noise diffusion: Noise synthesis for low-light image denois- ing,

    L. Lu, R. Achddou, and S. Susstrunk, “Dark noise diffusion: Noise synthesis for low-light image denois- ing,”IEEE Transactions on Pattern Analysis and Machine Intelligence (TP AMI), 2025

    2025

  14. [14]

    Physics-guided ISO-dependent sensor noise modeling for extreme low-light photography,

    Y. Cao, M. Liu, S. Liu, X. Wang, L. Lei, and W. Zuo, “Physics-guided ISO-dependent sensor noise modeling for extreme low-light photography,” inIEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023

    2023

  15. [15]

    Towards general low-light raw noise syn- thesis and modeling,

    F. Zhang, B. Xu, Z. Li, X. Liu, Q. Lu, C. Gao, and N. Sang, “Towards general low-light raw noise syn- thesis and modeling,” inInternational Conference on Computer Vision (ICCV), 2023

    2023

  16. [16]

    Calibration-free raw image denoising via fine- grained noise estimation,

    Y. Zou, Y. Fu, Y. Zhang, T. Zhang, C. Yan, and R. Tim- ofte, “Calibration-free raw image denoising via fine- grained noise estimation,”IEEE Transactions on Pattern Analysis and Machine Intelligence (TP AMI), 2025

    2025

  17. [17]

    Rethinking noise synthesis and modeling in raw denoising,

    Y. Zhang, H. Qin, X. Wang, and H. Li, “Rethinking noise synthesis and modeling in raw denoising,” in International Conference on Computer Vision (ICCV), 2021

    2021

  18. [18]

    Physics-based noise modeling for extreme low-light photography,

    K. Wei, Y. Fu, Y. Zheng, and J. Yang, “Physics-based noise modeling for extreme low-light photography,” IEEE Transactions on Pattern Analysis and Machine In- telligence (TP AMI), 2021

    2021

  19. [19]

    Learnabil- ity enhancement for low-light raw denoising: Where paired real data meets noise modeling,

    H. Feng, L. Wang, Y. Wang, and H. Huang, “Learnabil- ity enhancement for low-light raw denoising: Where paired real data meets noise modeling,” inProceedings of the ACM International Conference on Multimedia, 2022

    2022

  20. [20]

    MSFA image denoising using physics-based noise model and noise- decoupled network,

    Y. Jiang, Y. Fu, Q. Liu, and J. Zhang, “MSFA image denoising using physics-based noise model and noise- decoupled network,”IEEE Transactions on Pattern Anal- ysis and Machine Intelligence (TP AMI), 2025

    2025

  21. [21]

    Learning physics-informed noise models from dark frames for low-light raw image denoising,

    H. Feng, L. Wang, Y. Huang, Y. Wang, L. Zhu, and H. Huang, “Learning physics-informed noise models from dark frames for low-light raw image denoising,” IEEE Transactions on Pattern Analysis and Machine Intel- ligence (TP AMI), 2026

    2026

  22. [22]

    2-shots in the dark: Low-light denoising with minimal data acquisi- tion,

    L. Lu, R. Achddou, and S. S ¨usstrunk, “2-shots in the dark: Low-light denoising with minimal data acquisi- tion,”IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2026

    2026

  23. [23]

    Noise2Noise: Learning image restoration without clean data,

    J. Lehtinen, J. Munkberg, J. Hasselgren, S. Laine, T. Kar- ras, M. Aittala, and T. Aila, “Noise2Noise: Learning image restoration without clean data,”International Conference on Machine Learning (ICML), 2018

    2018

  24. [24]

    Noise2Self: Blind denoising by self-supervision,

    J. Batson and L. Royer, “Noise2Self: Blind denoising by self-supervision,” inInternational Conference on Machine Learning (ICML), 2019

    2019

  25. [25]

    Noise2void- learning denoising from single noisy images,

    A. Krull, T.-O. Buchholz, and F. Jug, “Noise2void- learning denoising from single noisy images,” inIEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019

    2019

  26. [26]

    Noise2Score: Tweedie’s approach to self-supervised image denoising without clean im- ages,

    K. Kim and J. C. Ye, “Noise2Score: Tweedie’s approach to self-supervised image denoising without clean im- ages,”Neural Information Processing Systems (NeurIPS), 2021

    2021

  27. [27]

    Noise2Same: Optimizing a self-supervised bound for image denoising,

    Y. Xie, Z. Wang, and S. Ji, “Noise2Same: Optimizing a self-supervised bound for image denoising,”Neural Information Processing Systems (NeurIPS), 2020

    2020

  28. [28]

    Noise2Info: Noisy image to information of noise for self-supervised image denoising,

    J. Wang, S. Di, L. Chen, and C. W. W. Ng, “Noise2Info: Noisy image to information of noise for self-supervised image denoising,” inInternational Conference on Com- puter Vision (ICCV), 2023

    2023

  29. [29]

    Self2Self with dropout: Learning self-supervised denoising from sin- gle image,

    Y. Quan, M. Chen, T. Pang, and H. Ji, “Self2Self with dropout: Learning self-supervised denoising from sin- gle image,” inIEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020

    2020

  30. [30]

    Blind2Unblind: Self-supervised image denoising with visible blind spots,

    Z. Wang, J. Liu, G. Li, and H. Han, “Blind2Unblind: Self-supervised image denoising with visible blind spots,” inIEEE Conference on Computer Vision and Pat- 10 tern Recognition (CVPR), 2022

    2022

  31. [31]

    Zero-shot Noise2Mean: Gap minimization for efficient denoising from a single noisy image,

    D. Liu, Y. Shi, G. Zhang, S. Li, and L. Zhang, “Zero-shot Noise2Mean: Gap minimization for efficient denoising from a single noisy image,” inProceedings of the National Conference on Artificial Intelligence (AAAI), 2025

    2025

  32. [32]

    Zero-shot Noise2Noise: Efficient image denoising without any data,

    Y. Mansour and R. Heckel, “Zero-shot Noise2Noise: Efficient image denoising without any data,” inIEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023

    2023

  33. [33]

    AP-BSN: Self- supervised denoising for real-world images via asym- metric pd and blind-spot network,

    W. Lee, S. Son, and K. M. Lee, “AP-BSN: Self- supervised denoising for real-world images via asym- metric pd and blind-spot network,” inIEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022

    2022

  34. [34]

    CVF-SID: Cyclic multi-variate function for self- supervised image denoising by disentangling noise from image,

    R. Neshatavar, M. Yavartanoo, S. Son, and K. M. Lee, “CVF-SID: Cyclic multi-variate function for self- supervised image denoising by disentangling noise from image,” inIEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022

    2022

  35. [35]

    YOND: Practical blind raw image denois- ing free from camera-specific data dependency,

    H. Feng, L. Wang, Y. Huang, T. Li, L. Zhu, and H. Huang, “YOND: Practical blind raw image denois- ing free from camera-specific data dependency,”arXiv preprint arXiv:2506.03645, 2025

    Pith/arXiv arXiv 2025

  36. [36]

    A high- quality denoising dataset for smartphone cameras,

    A. Abdelhamed, S. Lin, and M. S. Brown, “A high- quality denoising dataset for smartphone cameras,” in IEEE Conference on Computer Vision and Pattern Recogni- tion (CVPR), 2018

    2018

  37. [37]

    Noise modeling in one hour: Minimizing preparation efforts for self-supervised low- light raw image denoising,

    F. Li, H. Jiang, and D. Iso, “Noise modeling in one hour: Minimizing preparation efforts for self-supervised low- light raw image denoising,” inIEEE Conference on Com- puter Vision and Pattern Recognition (CVPR), 2025

    2025

  38. [38]

    Lighting every darkness in two pairs: A calibration-free pipeline for raw denoising,

    X. Jin, J.-W. Xiao, L.-H. Han, C. Guo, R. Zhang, X. Liu, and C. Li, “Lighting every darkness in two pairs: A calibration-free pipeline for raw denoising,” inInterna- tional Conference on Computer Vision (ICCV), 2023

    2023

  39. [39]

    Image denoising in mixed Poisson–Gaussian noise,

    F. Luisier, T. Blu, and M. Unser, “Image denoising in mixed Poisson–Gaussian noise,”IEEE Transactions on Image Processing (TIP), vol. 20, no. 3, pp. 696–708, 2010

    2010

  40. [40]

    Clipped noisy images: Heteroskedastic model- ing and practical denoising,

    A. Foi, “Clipped noisy images: Heteroskedastic model- ing and practical denoising,”Signal Processing, vol. 89, no. 12, pp. 2609–2629, 2009

    2009

  41. [41]

    Practical Poissonian-Gaussian noise modeling and fit- ting for single-image raw-data,

    A. Foi, M. Trimeche, V . Katkovnik, and K. Egiazarian, “Practical Poissonian-Gaussian noise modeling and fit- ting for single-image raw-data,”IEEE Transactions on Image Processing (TIP), vol. 17, no. 10, pp. 1737–1754, 2008

    2008

  42. [42]

    Learning to see in the dark,

    C. Chen, Q. Chen, J. Xu, and V . Koltun, “Learning to see in the dark,” inIEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018

    2018

  43. [43]

    U-Net: Convo- lutional networks for biomedical image segmentation,

    O. Ronneberger, P . Fischer, and T. Brox, “U-Net: Convo- lutional networks for biomedical image segmentation,” inProceedings of the International Conference on Medi- cal Image Computing and Computer Assisted Intervention (MICCAI), 2015

    2015

  44. [44]

    Adam: A method for stochastic optimization,

    D. P . Kingma and J. Ba, “Adam: A method for stochastic optimization,” inInternational Conference on Learning Representations (ICLR), 2015

    2015

  45. [45]

    The ciede2000 color-difference formula: Implementation notes, sup- plementary test data, and mathematical observations,

    G. Sharma, W. Wu, and E. N. Dalal, “The ciede2000 color-difference formula: Implementation notes, sup- plementary test data, and mathematical observations,” Color Research & Application, vol. 30, no. 1, pp. 21–30, 2005

    2005

  46. [46]

    Im- age denoising by sparse 3-D transform-domain collab- orative filtering,

    K. Dabov, A. Foi, V . Katkovnik, and K. Egiazarian, “Im- age denoising by sparse 3-D transform-domain collab- orative filtering,”IEEE Transactions on Image Processing (TIP), vol. 16, no. 8, pp. 2080–2095, 2007

    2080

  47. [47]

    Exploring efficient asymmetric blind-spots for self-supervised de- noising in real-world scenarios,

    S. Chen, J. Zhang, Z. Yu, and T. Huang, “Exploring efficient asymmetric blind-spots for self-supervised de- noising in real-world scenarios,” inIEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1 APPENDIXA IMPLEMENTATIONDETAILS We provide additional implementation details for our model and the baselines—PG (Poisson–Gaussian), PGRQ (...

    arXiv 2024

  48. [48]

    Combined (c), our method reaches a footprint comparable to the patch-based PGRQB (d) while keeping a∼2.8 dB PSNR advantage

    Downsampling the BLBE input cuts its cost by an order of magnitude with no loss in accuracy, and tiling the denoiser trades runtime for memory. Combined (c), our method reaches a footprint comparable to the patch-based PGRQB (d) while keeping a∼2.8 dB PSNR advantage. with default settings; results are in Sec. 5.5. In raw, we again use BM3D as a non-learni...