Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

Towards Controllable Real Image Denoising with Camera Parameters

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Camera metadata can both improve and steer a real-image denoiser.

desk verdict A sensible but under-evidenced ICIP paper: camera-conditioned denoising via adaLN shows a real SID gain, but the authors never rule out that extra parameters do the work, and the SIDD gain is negligible. read the letter →

arxiv 2507.01587 v2 pith:3TBG7CYF submitted 2025-07-02 cs.CV eess.IV

Youngjin Oh , Junhyeong Kwon , Keuntek Lee , Nam Ik Cho This is my paper · ORCID
classification cs.CVeess.IV
keywords controllableimagedenoisingcameraparametersadaptivelayernormalizationrealISOshutterspeedF-numberEXIFmetadata
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper sets out to show that the noise level of a real photograph is encoded well enough in three camera settings—ISO, shutter speed, and F-number—that a denoising network given those values as a condition outperforms the same network without them and can be adjusted by the user at inference time. The authors turn the three metadata values into a single conditioning vector and inject it into every block of a U-shaped denoising network through adaptive layer normalization. On the SID and SIDD real-image benchmarks the conditioned network, CPADNet, raises PSNR and SSIM over its baseline, and the same conditioning also improves a Transformer-based restorer. If the claim holds, metadata that already accompanies ordinary photo files becomes a free knob for choosing how much noise to remove.

What carries the argument

The camera parameter vector $v \in \mathbb{R}^{27}$. For each of ISO, shutter speed, and F-number, nine non-linear functions—$x$, $1/x$, $\sqrt{x}$, $x^{-1/2}$, $x^{1/4}$, $x^{-1/4}$, $\log x$, $\sin(\log x)$, $\cos(\log x)$—are applied and normalized to $[0,1]$, then concatenated. This vector is fed to adaptive layer normalization (adaLN) layers, which predict channel-wise affine parameters $\gamma$ and $\beta$, replacing the plain layer normalization of NAFBlock with a conditioning mechanism; dropout at rate 0.2 is applied before the affine prediction. The vector encodes an approximate noise level and conditions every block of the U-shaped network, so the network's denoising strength responds to the supplied camera settings. For smartphone images lacking F-number, the F-number branch is replaced by a one-hot encoding of camera type passed through an embedding layer.

What would settle it

Train a capacity-matched baseline (NAFBlock enlarged to 9.45M parameters, no camera conditioning) on the same SID patches; if its PSNR is not below CPADNet's 25.57 dB, the central claim that camera parameters drive the improvement is falsified.

Watch

Extended reading notes

Core claim

The central claim is that camera parameters are valuable ground-truth information about real noise level, so conditioning a denoising network on them during training improves accuracy and grants interactive control at inference. Concretely, ISO, shutter speed, and F-number are mapped through nine non-linear functions, normalized to [0,1], and concatenated into a camera parameter vector $v \in \mathbb{R}^{27}$. The vector is consumed by adaptive layer normalization layers that produce per-channel scale and shift, replacing the plain layer normalization of the NAFBlock baseline in every block. At inference the user can replace the true metadata with a milder or harsher camera setting to lighten or strengthen denoising. The paper reports gains on SID (24.58 to 25.57 dB patch PSNR) and on SIDD (39.41 to 39.43 dB patch PSNR) over the baseline, and similar gains when the conditioning is added to Restormer.

Load-bearing premise

The load-bearing premise is that the reported gains come from the camera-parameter conditioning itself, not from the larger network: CPADNet has 9.45M parameters versus 6.67M for the baseline, and no capacity-matched ablation is included.

Editorial extensions

If this is right

  • A single trained CPADNet can produce lighter or stronger denoising at inference by substituting lower or higher ISO, shutter speed, or F-number values, with no retraining.
  • The conditioning also improves Restormer, a Transformer architecture, so it is not tied to the specific NAFBlock design.
  • Because the three parameters come from standard EXIF metadata, the method does not require extra capture-time instrumentation.
  • The improvement is larger on SID (about 1 dB) than on SIDD (about 0.02 dB), which the paper ties to the closer match between training and test ISO distributions on SID.

Reading between the lines

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

  • An implication left implicit is that existing photo libraries could be re-rendered at different noise strengths by editing the metadata values fed to the network, since EXIF data already exists for most images.
  • A capacity-matched control—enlarging the baseline to CPADNet's 9.45M parameters without camera conditioning—would separate the contribution of the metadata from the contribution of extra parameters; the paper does not include such an experiment.
  • If the conditioning vector faithfully represents noise level, feeding an ISO between two training values should yield a denoised result whose strength lies between the two endpoints; the paper demonstrates the endpoints but not this interpolation curve.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes CPADNet, a denoising network that conditions a NAFBlock-based U-Net on camera parameters (ISO, shutter speed, F-number) via adaptive layer normalization. The camera parameters are non-linearly equalized into a 27-dimensional vector, and a learned embedding is used for smartphone cameras that lack F-number metadata. Experiments on the SID Sony subset and SIDD compare CPADNet with the baseline NAFNet and CPA-Restormer with Restormer, reporting PSNR/SSIM gains and qualitative inference-time controllability demonstrations in which the user changes the conditioning vector to adjust denoising strength.

Significance. If the reported performance gains are attributable to the camera-parameter conditioning, the method is practically attractive: it uses ground-truth EXIF metadata rather than estimated noise levels, is end-to-end trainable, adds almost no MACs, and is presented as a plug-in module for existing restoration networks. The code is promised to be released. However, the empirical evidence is currently insufficient to separate the effect of the camera-conditioning prior from the effect of added model capacity, and no comparison with prior noise-level-conditioning methods is provided. The core idea is plausible and worth pursuing, but the central quantitative claim needs stronger experimental support.

major comments (3)
  1. [Section 3.3, Table 1] The central quantitative claim is confounded by model capacity. CPADNet has 9.45M parameters versus 6.67M for the baseline, a 42% increase, while MACs are identical (18.53G). On SIDD valid and test, the gain is only +0.02 dB (39.41 to 39.43 and 39.57 to 39.59), which is within ordinary run-to-run variation for a single training run. The paper acknowledges the parameter increase but provides no capacity-matched baseline, no ablation that removes or scrambles the camera-parameter information while keeping the architecture fixed, and no repeated-seed statistics. The same issue affects CPA-Restormer (11.72M to 15.06M parameters). Please add a capacity-matched baseline (e.g., a wider NAFNet with the same parameter count as CPADNet) and a condition-ablated control (e.g., a constant or shuffled camera vector), and report mean and standard deviation over at least three training runs.
  2. [Section 1 and Section 3.3] The claim that the method improves state-of-the-art denoising networks is not supported by comparison with prior noise-level-conditioning or controllable denoising methods. CBDNet, DUDN, and variational approaches [3]-[6] also provide inference-time control of denoising strength, but no quantitative comparison is given on the same datasets. Without such baselines, the paper should either add a comparison table or delimit the claim to 'adds controllability via camera metadata' rather than 'improves state-of-the-art denoising networks.'
  3. [Section 3.4, Figs. 4 and 5] The controllability demonstration is qualitative only, and some of the manipulated camera parameters may lie outside the training distribution (e.g., ISO 50 and F/22 for SID, and ISO 10000 for SIDD, whose training distribution in Fig. 3 has little support at such values). If these extreme conditioning vectors are extrapolations, the visual changes may reflect out-of-distribution behavior rather than a learned, meaningful denoising-strength axis. Please provide a quantitative evaluation of controllability, such as PSNR/SSIM or a noise-reduction metric as a function of the conditioning vector on held-out data, and clarify whether the shown parameter combinations occur in the training set.
minor comments (4)
  1. [Overall] The title in the arXiv header reads 'TOW ARDS' with an extra space, and there are a few other spacing artifacts in the text; a copyedit pass would improve readability.
  2. [Section 2.2] The choice of the nine nonlinear functions is inherited from [15], but no analysis or ablation is provided for the specific function set. A sentence explaining why this set is appropriate for the three camera parameters, or a small ablation, would strengthen the method section.
  3. [Fig. 2(a)] The legend for 'Input feature from previous block' and 'Output feature from current block' is unclear because both arrows in the CPA-NAFBlock diagram use the same style, making the dataflow difficult to follow. Please use distinct arrow styles or labels.
  4. [Section 3.1] The evaluation protocol for SID and SIDD should state explicitly how the patch-level PSNR/SSIM are computed (e.g., average over non-overlapping test patches) and how the full-frame results are produced; currently this information must be inferred from the note about Restormer.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: camera parameters are external metadata, the conditioning is learned end-to-end, and the nonlinear encoding is imported from an external prior work; the main caveat is a capacity confound, which is a validity issue, not circularity.

full rationale

The paper's derivation chain does not define any quantity in terms of its own target. ISO, shutter speed, and F-number are external EXIF inputs, not fitted constants or outputs of the denoiser; the camera-parameter vector v is formed by fixed nonlinear functions (x, 1/x, sqrt(x), etc.) following [15] and then normalized, and the denoising network is trained end-to-end with L1 loss against clean images. There is no step where a fitted value is relabeled as a prediction, no uniqueness theorem imported from the authors' prior work, and no equation in which the claimed improvement reduces by construction to the input. The self-citations [5,6] are background descriptions of prior adjustable-denoising methods and do not support the central claim. The main quantitative weakness--Table 1 shows CPADNet has 9.45M parameters vs 6.67M for the baseline and gives only +0.02 dB on SIDD--is a capacity/ablation confound that bears on external validity, not on circularity. The nonlinear equalization is adopted from an external paper ([15]) rather than from the authors' own prior work, so no ansatz is smuggled in via self-citation. Overall, the method is self-contained against standard benchmarks; no circular step is present.

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

The method introduces no new physical entities. Free parameters are design choices, not fitted constants in a physical model. The main axioms are standard noise modeling and metadata assumptions.

free parameters (4)
  • Non-linear equalization function set = 9 functions: x, 1/x, sqrt(x), x^-1/2, x^1/4, x^-1/4, log(x), sin(log(x)), cos(log(x))
    Chosen from [15]; not fitted, but a design choice that shapes the camera parameter vector.
  • Smartphone type embedding = Learned embedding for SIDD phone types
    Used in place of F-number for smartphone data; trained on SIDD.
  • Normalization range = [0,1]
    Camera parameter mapped values are normalized to [0,1] before concatenation.
  • Dropout probability = 0.2
    Applied before affine parameter computation to reduce overfitting to parameter combinations.
assumptions (4)
  • domain assumption Noise follows a heteroscedastic Gaussian model: y ~ N(x, lambda_read + lambda_shot * x)
    Equation (1), cited from [9-11]; used to justify selecting ISO, shutter speed, F-number.
  • domain assumption ISO, shutter speed, and F-number are sufficient metadata to represent the noise level
    The paper argues from the noise model and Fig. 1, but the sufficiency is assumed and not quantitatively established.
  • ad hoc to paper The nonlinear function set from ParamISP [15] captures the mapping from camera parameters to noise-relevant features
    The functions are taken from prior work without analysis specific to denoising; their adequacy is assumed.
  • domain assumption Rawpy postprocess on SID short/long exposure pairs yields clean/noisy RGB training pairs
    The pipeline for forming training data is standard but introduces processing choices (demosaicing, white balance) that affect the task.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Towards Controllable Real Image Denoising with Camera Parameters." pith.science (2026). https://pith.science/paper/3TBG7CYF

@misc{pith2026250701587,
  author       = {Pith},
  title        = {Pith review of: Towards Controllable Real Image Denoising with Camera Parameters},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3TBG7CYF}},
  note         = {Machine review of arXiv:2507.01587}
}
read the original abstract

Recent deep learning-based image denoising methods have shown impressive performance; however, many lack the flexibility to adjust the denoising strength based on the noise levels, camera settings, and user preferences. In this paper, we introduce a new controllable denoising framework that adaptively removes noise from images by utilizing information from camera parameters. Specifically, we focus on ISO, shutter speed, and F-number, which are closely related to noise levels. We convert these selected parameters into a vector to control and enhance the performance of the denoising network. Experimental results show that our method seamlessly adds controllability to standard denoising neural networks and improves their performance. Code is available at https://github.com/OBAKSA/CPADNet.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. DarkVRAI: Capture-Condition Conditioning and Burst-Order Selective Scan for Low-light RAW Video Denoising

    cs.CV 2025-08 conditional novelty 5.0 of 10

    DarkVRAI combines capture-condition conditioning and a burst-order selective scan to win the AIM 2025 low-light RAW video denoising challenge.

Reference graph

Works this paper leans on

24 extracted references · 20 canonical work pages · cited by 1 Pith paper

  1. [3]

    Dataset We train and evaluate our method using the Sony subset of the SID [2] and SIDD-medium datasets [1]

    EXPERIMENTAL RESULTS 3.1. Dataset We train and evaluate our method using the Sony subset of the SID [2] and SIDD-medium datasets [1]. SID. The Sony subset of SID dataset is taken with Sony α7S II camera, and from the total of 2697 short-/long-exposure RAW image pairs, we use 2099 pairs as training data and Table 1. Evaluation of our proposed method applie...

    work page 2000

  2. [6]

    Learning to see in the dark,

    Chen Chen, Qifeng Chen, Jia Xu, and Vladlen Koltun, “Learning to see in the dark,” in Proceedings of the IEEE conference on computer vision and pattern recog- nition, 2018, pp. 3291–3300

    work page 2018

  3. [1]

    INTRODUCTION Image denoising is an essential process in photography that aims to restore a clear image from a noisy one. The primary source of noise in optical camera images is the sensor, which inevitably produces highly noisy images when sensor size and light intensity are reduced, as is common in many smartphone cameras [1]. Even high-end DSLR cameras,...

    work page

  4. [2]

    Camera Parameters Related to Noise Level Noise from optical sensors can generally be characterized us- ing a heteroscedastic Gaussian model [9–11]

    PROPOSED METHOD 2.1. Camera Parameters Related to Noise Level Noise from optical sensors can generally be characterized us- ing a heteroscedastic Gaussian model [9–11]. When we have the intensity of a noise-free clean raw image represented asx, the noisy raw image y is modeled as: y ∼ N(µ = x, σ2 = λread + λshotx), (1) where the parameter λread represents...

    arXiv 2025

  5. [4]

    To achieve this, we have created a vector containing information about the noise level and utilize adaptive layer normalization to empower the network with this vector

    CONCLUSION We have developed an effective method to enhance and con- trol the performance of a denoising network by using noise- related camera parameters. To achieve this, we have created a vector containing information about the noise level and utilize adaptive layer normalization to empower the network with this vector. Experiments demonstrate that our...

    work page 2025

  6. [5]

    A high-quality denoising dataset for smart- phone cameras,

    Abdelrahman Abdelhamed, Stephen Lin, and Michael S Brown, “A high-quality denoising dataset for smart- phone cameras,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2018, pp. 1692–1700

    2018

  7. [7]

    Ffdnet: Toward a fast and flexible solution for cnn-based image denoising,

    Kai Zhang, Wangmeng Zuo, and Lei Zhang, “Ffdnet: Toward a fast and flexible solution for cnn-based image denoising,” IEEE Transactions on Image Processing , vol. 27, no. 9, pp. 4608–4622, 2018

    work page 2018

  8. [8]

    Toward convolutional blind denoising of real photographs,

    Shi Guo, Zifei Yan, Kai Zhang, Wangmeng Zuo, and Lei Zhang, “Toward convolutional blind denoising of real photographs,” in Proceedings of the IEEE/CVF confer- ence on computer vision and pattern recognition , 2019, pp. 1712–1722

    work page 2019

Show all 24 references
  1. [9]

    Deep universal blind image denoising,

    Jae Woong Soh and Nam Ik Cho, “Deep universal blind image denoising,” in 2020 25th International Confer- ence on Pattern Recognition (ICPR) . IEEE, 2021, pp. 747–754

    2020

  2. [10]

    Variational deep im- age restoration,

    Jae Woong Soh and Nam Ik Cho, “Variational deep im- age restoration,” IEEE Transactions on Image Process- ing, vol. 31, pp. 4363–4376, 2022

    2022

  3. [11]

    Simple baselines for image restoration,

    Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun, “Simple baselines for image restoration,” in Eu- ropean Conference on Computer Vision. Springer, 2022, pp. 17–33

    2022

  4. [12]

    Scalable diffu- sion models with transformers,

    William Peebles and Saining Xie, “Scalable diffu- sion models with transformers,” in Proceedings of the IEEE/CVF International Conference on Computer Vi- sion, 2023, pp. 4195–4205

    2023

  5. [13]

    Practical poissonian-gaussian noise modeling and fitting for single-image raw-data,

    Alessandro Foi, Mejdi Trimeche, Vladimir Katkovnik, and Karen Egiazarian, “Practical poissonian-gaussian noise modeling and fitting for single-image raw-data,” IEEE transactions on image processing, vol. 17, no. 10, pp. 1737–1754, 2008

    2008

  6. [14]

    Unpro- cessing images for learned raw denoising,

    Tim Brooks, Ben Mildenhall, Tianfan Xue, Jiawen Chen, Dillon Sharlet, and Jonathan T Barron, “Unpro- cessing images for learned raw denoising,” in Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019, pp. 11036–11045

    2019

  7. [15]

    Toward efficient deep blind raw image restoration,

    Marcos V Conde, Florin Vasluianu, and Radu Timofte, “Toward efficient deep blind raw image restoration,” in 2024 IEEE International Conference on Image Process- ing (ICIP). IEEE, 2024, pp. 1725–1731

    2024

  8. [16]

    Practical deep raw image denoising on mobile devices,

    Yuzhi Wang, Haibin Huang, Qin Xu, Jiaming Liu, Yiqun Liu, and Jue Wang, “Practical deep raw image denoising on mobile devices,” in European Conference on Computer Vision. Springer, 2020, pp. 1–16

    2020

  9. [17]

    Uformer: A general u-shaped transformer for image restoration,

    Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wen- gang Zhou, Jianzhuang Liu, and Houqiang Li, “Uformer: A general u-shaped transformer for image restoration,” in Proceedings of the IEEE/CVF confer- ence on computer vision and pattern recognition , 2022, pp. 17683–17693

    2022

  10. [18]

    Defocus deblurring using dual-pixel data,

    Abdullah Abuolaim and Michael S Brown, “Defocus deblurring using dual-pixel data,” in Computer Vision– ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part X 16 . Springer, 2020, pp. 111–126

    2020

  11. [19]

    Paramisp: Learned forward and inverse isps using cam- era parameters,

    Woohyeok Kim, Geonu Kim, Junyong Lee, Seungy- ong Lee, Seung-Hwan Baek, and Sunghyun Cho, “Paramisp: Learned forward and inverse isps using cam- era parameters,” in Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition , 2024, pp. 26067–26076

    2024

  12. [20]

    Sigmoid- weighted linear units for neural network function ap- proximation in reinforcement learning,

    Stefan Elfwing, Eiji Uchibe, and Kenji Doya, “Sigmoid- weighted linear units for neural network function ap- proximation in reinforcement learning,” Neural net- works, vol. 107, pp. 3–11, 2018

    2018

  13. [21]

    Restormer: Efficient transformer for high- resolution image restoration,

    Syed Waqas Zamir, Aditya Arora, Salman Khan, Mu- nawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang, “Restormer: Efficient transformer for high- resolution image restoration,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 5728–5739

    2022

  14. [22]

    Retinexformer: One- stage retinex-based transformer for low-light image en- hancement,

    Yuanhao Cai, Hao Bian, Jing Lin, Haoqian Wang, Radu Timofte, and Yulun Zhang, “Retinexformer: One- stage retinex-based transformer for low-light image en- hancement,” in Proceedings of the IEEE/CVF Inter- national Conference on Computer Vision , 2023, pp. 12504–12513

    2023

  15. [23]

    Mipi 2024 challenge on few-shot raw image denoising: Methods and results,

    Xin Jin, Chunle Guo, Xiaoming Li, Zongsheng Yue, Chongyi Li, Shangchen Zhou, Ruicheng Feng, Yuekun Dai, Peiqing Yang, Chen Change Loy, et al., “Mipi 2024 challenge on few-shot raw image denoising: Methods and results,” in Proceedings of the IEEE/CVF Con- ference on Computer Vi...

    2024

  16. [24]

    Noise2noiseflow: Realistic cam- era noise modeling without clean images,

    Ali Maleky, Shayan Kousha, Michael S Brown, and Marcus A Brubaker, “Noise2noiseflow: Realistic cam- era noise modeling without clean images,” in Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 17632–17641

    2022

Pith tools

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