Pith. sign in

REVIEW 4 major objections 6 minor 6 references

Towards Robust Image Denoising with Scale Equivariance

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Scale equivariance is the architectural bias that makes denoisers robust to spatially varying noise.

desk verdict The simple equivariant U-Net beating sophisticated models on OOD noise is a genuine empirical hook, but Eq. (3) overreaches: the proven global homogeneity does not imply the per-pixel equivariance the argument needs. read the letter →

arxiv 2508.02967 v1 pith:HDMO5IPO submitted 2025-08-05 cs.CV

Dawei Zhang , Xiaojie Guo This is my paper
classification cs.CV
keywords scaleequivariancefirst-orderhomogeneityimagedenoisingout-of-distributiongeneralizationspatiallyvariantnoiseHeterogeneousNormalizationModuleInteractiveGatingblind
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

Deep denoisers trained on uniform Gaussian noise fail on spatially varying noise because their features get entangled with the noise magnitude. This paper argues that the fix can be built into the architecture: if every operation in the network is first-order homogeneous, scaling the input by $k$ scales the output by $k$, then the noise magnitude factors out of the computation. Under that property, training on uniform noise and testing on per-pixel-varying noise become the same normalized task, removing the main distribution shift. The authors build a denoiser from such components and show it outperforms sophisticated but non-equivariant models on Poisson, speckle, and mixed noise while staying competitive in-distribution. If scale equivariance does the causal work, out-of-distribution robustness in denoising becomes a design property rather than a training-data property.

What carries the argument

The load-bearing object is scale equivariance, first-order homogeneity: a function $H$ with $H(0)=0$ satisfying $H(kI)=kH(I)$ for any scalar $k>0$. The paper shows convolution and ReLU satisfy this, residual blocks preserve it, and designs three components that do too: Constant Scaling divides features by $\sqrt{c}$ to stabilize magnitude without a scale-dependent normalization; the Normalized Self-Modulator applies per-token affine parameters $\gamma,\beta$ obtained from a $1\times1$ convolution, whose homogeneity cancels the scaling of the normalized vector; and the Interactive Gating Module splits features into value and gate streams and divides their Hadamard product by $\sqrt{\sigma^2(F_v)+\sigma^2(F_g)}$, converting an otherwise second-order numerator into a first-order output. This denominator is the central mechanism: it makes the gate's effect scale-invariant while letting the gate's strength be co-determined by both signals. The argument then rests on Eq. (3), which extends the property from a global scalar $\lambda$ to a pixel-wise noise-level map $\Lambda$.

What would settle it

One concrete test: feed the network an image multiplied by a noise-level map $\Lambda$ made of a sharp step or checkerboard, and compare the output to $\Lambda\odot F(I/\Lambda+\tilde{N})$ predicted by Eq. (3). If the relative error grows with the spatial contrast or gradient of $\Lambda$, for example, doubling the step height more than doubles the mismatch, then per-pixel equivariance does not hold and the central claim is falsified. A simpler screening check is to measure $F(\Lambda\odot I)$ versus $\Lambda\odot F(I)$ on a clean image: any systematic deviation as $\Lambda$ becomes more non-uniform contradicts the assumed equivariance.

Watch

Extended reading notes

Core claim

The paper's central claim is that enforcing scale equivariance, defined as first-order homogeneity, $H(kI)=kH(I)$, inside a denoising network suppresses the impact of spatially varying noise levels and closes the train-test gap between uniform and non-uniform noise. With this property, the network's effective task becomes denoising an input normalized to unit noise variance, whether the noise is global or pixel-wise, so the learned computation transfers across noise types. The paper validates the principle with a minimal U-Net built only from convolutions and ReLU, which beats sophisticated non-equivariant models on out-of-distribution noise, and then introduces SEVNet, whose Heterogeneous Normalization Module and Interactive Gating Module preserve first-order homogeneity by construction. The authors report consistent PSNR and SSIM gains over state-of-the-art rivals on synthetic out-of-distribution benchmarks and, after coupling the equivariant backbone with a lightweight noise translator, on real-world noise benchmarks.

Load-bearing premise

The load-bearing premise is that the network's behavior under per-pixel scaling by the spatially varying noise-level map $\Lambda$ is the same as its behavior under global scaling, a per-pixel equivariance strictly stronger than the first-order homogeneity the paper defines and proves; if Eq. (3) fails for spatially varying $\Lambda$, the theoretical case collapses and only the empirical correlation remains.

Editorial extensions

If this is right

  • A denoiser made of scale-equivariant components can be trained on uniform Gaussian noise and applied directly to spatially varying Poisson, speckle, and mixed noise without adaptation.
  • Standard normalization layers and exponential activations such as GELU and Swish are architectural liabilities for out-of-distribution robustness; replacing them with equivariant stabilizers is predicted to help any denoising architecture.
  • In-distribution performance does not have to be sacrificed for robustness: the authors report competitive PSNR on Gaussian noise while leading on out-of-distribution benchmarks.
  • For real-world sensor noise, the equivariant backbone plus a lightweight noise translator outperforms much larger end-to-end models, suggesting a modular path from synthetic to real denoising.
  • The principle generalizes beyond the specific network: robustness to degradation magnitude is an inductive bias of network design, not only of data or loss.

Reading between the lines

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

  • The paper proves first-order homogeneity for global scaling; Eq. (3) implicitly demands per-pixel equivariance under the level map $\Lambda$, a stronger condition that standard convolutions do not satisfy. A direct test would be measuring how much $F(\Lambda\odot I)$ deviates from $\Lambda\odot F(I)$ as the spatial gradient of $\Lambda$ increases.
  • If the per-pixel version holds only approximately, the theory predicts that out-of-distribution gains should shrink as noise-level maps become more high-frequency; experiments with step-function or checkerboard $\Lambda$ would isolate this boundary.
  • The same first-order-homogeneity argument should transfer to other inverse problems where the degradation strength varies spatially, such as non-uniform deblurring or spatially varying super-resolution, though each operator would need its own equivariant modules.
  • The Interactive Gating denominator is a form of energy normalization; this mechanism could be reused outside denoising as a scale-equivariant nonlinearity for any gated network.
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

4 major / 6 minor

Summary. The paper proposes that enforcing scale equivariance (first-order homogeneity) in a denoising network improves robustness to out-of-distribution (OOD) noise, especially spatially varying noise. The authors derive two theoretical relations, Eqs. (2) and (3), from first-order homogeneity, and on this basis design two modules: the Heterogeneous Normalization Module (HNM), composed of Constant Scaling (CS) and a Normalized Self-Modulator (NSM), and the Interactive Gating Module (IGM). They validate the approach with a simple equivariant U-Net baseline and with the full SEVNet on synthetic OOD benchmarks (Poisson, Speckle, mixtures, and spatially varying variants) and on real-world noise benchmarks, reporting consistent PSNR/SSIM improvements over prior methods.

Significance. If the theoretical claim were established, the paper would make a useful contribution by showing that an architecture-level inductive bias—first-order homogeneity—can be imposed by construction and yields OOD robustness without sacrificing in-distribution performance. The manuscript has notable strengths: the algebraic lemmas in the appendix are explicit, the first-principles baseline U-Net is a genuine independent test, and the proposed modules are simple and clearly described. The promised public code would also aid reproducibility. However, the central theoretical bridge from global scale equivariance to spatially varying noise, Eq. (3), is not justified and is in fact false for standard convolutions. The empirical results are suggestive, but the paper does not provide a controlled comparison isolating scale equivariance as the causal factor. The contribution is therefore currently supported only partially, and the main mechanism remains unproven.

major comments (4)
  1. [Problem Analysis, Eq. (3)] Equation (3) claims that for a scale-equivariant network F, F(X + Λ⊙Ñ) = Λ⊙F(X/Λ + Ñ) for a spatially varying noise-level map Λ. Writing Z = X/Λ + Ñ, this is equivalent to F(Λ⊙Z) = Λ⊙F(Z) for every positive pixel-wise map Λ, i.e., equivariance under per-pixel scaling. Definition 1 and Lemmas 3–5 establish only homogeneity under a global scalar k. A 3×3 convolution mixes neighboring pixels with different scaling factors: (Conv(Λ⊙Z))_{i,j} = Σ_{m,n} W_{m,n} Λ_{i-m,j-n} Z_{i-m,j-n}, which equals Λ_{i,j}(Conv Z)_{i,j} only when Λ is constant on the kernel support. Since SEVNet uses 3×3 convolutions in IGM and standard convolutional layers, Eq. (3) fails exactly in the spatially varying regime it is meant to explain. The paper needs either a proof of per-pixel equivariance for the proposed modules, or a substantially revised theoretical argument (e.g., an approximate or asymptotic statement), or a clear restriction of the OOD claim to settings that Eq. (3) can actually support.
  2. [Lemma 1 and network implementation] Lemma 1 states that convolution is first-order homogeneous, but this holds only for convolutions without bias terms. With a bias b, Conv(k·I) = k·W*I + b, which is not generally equal to k·Conv(I) unless b = 0. The paper never states that all convolutional layers in SEVNet and the baseline U-Net are bias-free, and PyTorch's default is bias=True. Since Eqs. (2) and (3) and the baseline validation depend on strict first-order homogeneity, the paper must either specify and justify bias-free convolutions or account for the bias contribution in the homogeneity analysis.
  3. [Experiments, Table 1 and Baseline] The first-principles validation in Table 1 compares a simple equivariant U-Net baseline against sophisticated non-equivariant models, but this is not a controlled experiment. To support the claim that scale equivariance is the causal factor behind the OOD gains, the paper should compare the equivariant baseline against a non-equivariant U-Net with the same architecture, capacity, training data, and training schedule. Without such an ablation, the superior OOD performance could be due to other differences such as architecture simplicity, regularization, or training details, rather than the equivariance property itself.
  4. [Experiments, real-world extension] The real-world results in Table 4 are obtained with 'Ours+NT', where NT is a noise translator borrowed from Ha et al. (2024). The paper credits the gains to the scale-equivariant denoiser, but it reports no comparison of the same noise translator with a non-equivariant denoiser, nor a comparison of SEVNet without NT. As a result, the contribution of the scale-equivariant architecture to real-world performance is not isolated, and the claim that 'the applicability of our scale-equivariant network... can be successfully extended to real-world data' is not directly supported by the presented experiments.
minor comments (6)
  1. [Eq. (5)] The NSM formula uses X in the expression (X - μ(F))/σ(F), but the input to NSM was denoted F; please unify the notation.
  2. [Eq. (6) and surrounding text] The text describes the gating denominator as a relative importance weight σ(F_v)/(σ(F_v)+σ(F_g)), but Eq. (6) uses the square root of the sum of variances; the relationship between these expressions should be clarified.
  3. [Experiments, Table 2 caption] The caption states that average results across three noise levels are reported, but the specific level values for Speckle and Poisson are only given later in the appendix; it would help to state them in the table caption.
  4. [Experiments, appendix] The main text promises efficiency comparisons in the appendix, but the appendix contains only qualitative figures and no efficiency table; please either add the table or remove the promise.
  5. [Introduction and Related Work] The reference 'Chen et al. 2023' is used for both adversarial perturbations and MaskedDenoising; please disambiguate the two citations.
  6. [Figure 2] The figure mentions 'DSS' (dual-signal scaling) but the term is not defined in the main text; please define it where IGM is introduced.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; Eq. (3) is a derivation gap, not a circular reduction.

full rationale

The principal derivation, Eq. (2), is a valid consequence of Definition 1: global homogeneity gives F(X+λN)=λF(X/λ+N), and the target X leads to the normalized denoising task. Lemmas 3-5 are independent algebraic checks of the modules' global homogeneity, not assumptions of the result; the modules were designed to satisfy the constraint, which is construction rather than circularity. The baseline U-Net experiment (Table 1) is an external test that does not train on the OOD target and is therefore a genuine non-fitted validation. No fitted parameter is relabeled as a prediction, and no load-bearing self-citation or imported uniqueness theorem occurs; the only self-citation (Zhang et al. 2023a, MM-BSN) is contextual. The serious issue is Eq. (3), which silently replaces global homogeneity with per-pixel scaling equivariance F(Λ⊙Z)=Λ⊙F(Z); standard 3×3 convolutions violate this, and no proof is provided. This is an unsupported generalization and a correctness risk for the stated mechanism, not a case where the conclusion is identical to the input by construction or where an output is forced by a fitted constant. Hence the circularity score is 0.

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

The theory is parameter-free aside from ordinary trainable network weights (e.g., the per-channel η in CS), so no hand-fitted constants enter the central claim. The main burden is the unproven pointwise-equivariance assumption in Eq. (3), which is recorded as an axiom above.

assumptions (3)
  • domain assumption Spatially varying noise can be modeled as Y = X + Λ⊙Ñ with pixel-wise independent noise Ñ.
    Used in Eq. (3) and in the synthetic OOD noise generation (Poisson, Speckle, Variant 1-3). The real-world section notes the network assumes independent noise, so this is an explicit domain restriction.
  • ad hoc to paper First-order homogeneity under a global scalar k implies the pointwise scaling relation F(X + Λ⊙Ñ) = Λ⊙F(X/Λ + Ñ).
    This is asserted in Problem Analysis, Eq. (3), but not derived. It is generally false for networks with standard convolutions when Λ varies within the receptive field, so it is a load-bearing unproven premise.
  • domain assumption The training objective for the synthetic experiments is such that the re-parameterization in Eq. (2) correctly describes the effective learning task.
    The loss function is never specified in the provided text; the argument that the network no longer learns a direct mapping from noisy to clean implicitly assumes a loss (e.g., L2) whose minimizer is the scaled target, and it ignores the λ² weighting of the loss under scaling.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Towards Robust Image Denoising with Scale Equivariance." pith.science (2026). https://pith.science/paper/HDMO5IPO

@misc{pith2026250802967,
  author       = {Pith},
  title        = {Pith review of: Towards Robust Image Denoising with Scale Equivariance},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HDMO5IPO}},
  note         = {Machine review of arXiv:2508.02967}
}
read the original abstract

Despite notable advances in image denoising, existing models often struggle to generalize beyond in-distribution noise patterns, particularly when confronted with out-of-distribution (OOD) conditions characterized by spatially variant noise. This generalization gap remains a fundamental yet underexplored challenge. In this work, we investigate \emph{scale equivariance} as a core inductive bias for improving OOD robustness. We argue that incorporating scale-equivariant structures enables models to better adapt from training on spatially uniform noise to inference on spatially non-uniform degradations. Building on this insight, we propose a robust blind denoising framework equipped with two key components: a Heterogeneous Normalization Module (HNM) and an Interactive Gating Module (IGM). HNM stabilizes feature distributions and dynamically corrects features under varying noise intensities, while IGM facilitates effective information modulation via gated interactions between signal and feature paths. Extensive evaluations demonstrate that our model consistently outperforms state-of-the-art methods on both synthetic and real-world benchmarks, especially under spatially heterogeneous noise. Code will be made publicly available.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

6 extracted references · 6 canonical work pages

  1. [2]

    The noise level at each position is affected by the corresponding pixel, while σ ∈ {60, 80, 90} controls the global damage intensity

    Speckle noise: Y = X + σ · √ X ⊙ N, where N is derived from the standard Gaussian distribution. The noise level at each position is affected by the corresponding pixel, while σ ∈ {60, 80, 90} controls the global damage intensity. 3) Figure 6: Qualitative comparisons on mixture noise of σ = 90 and α = 6. Figure 7: Qualitative comparisons on Poisson noise o...

    work page

  2. [33]

    Transfer CLIP for Generalizable Image Denoising

    Springer. Chen, X.; Li, Z.; Pu, Y .; Liu, Y .; Zhou, J.; Qiao, Y .; and Dong, C. 2024. A Comparative Study of Image Restora- tion Networks for General Backbone Network Design. In Leonardis, A.; Ricci, E.; Roth, S.; Russakovsky, O.; Sattler, T.; and Varol, G., eds.,Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4...

    work page Pith review arXiv 2024

  3. [1703]

    Chen, L.; Chu, X.; Zhang, X.; and Sun, J

    IEEE. Chen, L.; Chu, X.; Zhang, X.; and Sun, J. 2022. Simple Baselines for Image Restoration. In Avidan, S.; Brostow, G. J.; Ciss ´e, M.; Farinella, G. M.; and Hassner, T., eds., Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part VII, volume 13667 of Lecture Notes in Computer Science , 17–

    work page 2022

  4. [2019]

    In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 1712–1722

    Toward Convolutional Blind Denoising of Real Pho- tographs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 1712–1722. Ha, I.; Ryou, D.; Seo, S.; and Han, B. 2024. Learning to Translate Noise for Robust Image Denoising. CoRR, abs/2412.04727. Kim, C.; Kim, T. H.; and Baik, S. 2024. LAN: Learning to Adapt Noise for Imag...

    work page arXiv 2024

  5. [2022]

    In IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, 17662–17672

    Uformer: A General U-Shaped Transformer for Im- age Restoration. In IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, 17662–17672. IEEE. Yue, Z.; Yong, H.; Zhao, Q.; Zhang, L.; Meng, D.; and Wong, K. K. 2024. Deep Variational Network Toward Blind Image Restoration. IEEE Trans. Pattern Anal...

    work page 2022

  6. [2025]

    IEEE Trans

    Pixel2Pixel: A Pixelwise Approach for Zero-Shot Sin- gle Image Denoising. IEEE Trans. Pattern Anal. Mach. In- tell., 47(6): 4614–4629. Ma, X.; Dai, X.; Bai, Y .; Wang, Y .; and Fu, Y . 2024. Rewrite the Stars. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, 5694–5703. IEEE. Mansour, Y .; an...

    work page 2024

Pith tools

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