Pith. sign in

REVIEW 3 major objections 4 minor 14 references

A single 0.82M-parameter, exactly scale-homogeneous U-Net matches or beats the classic sub-1M denoisers on all four standard color sets, and keeps denoising at more than three times its training noise ceiling.

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

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

T0 review · deepseek-v4-flash

2026-08-01 05:12 UTC pith:MKFK7JBF

load-bearing objection Honest engineering report whose abstract overstates the main result; the bias-free machinery is real, but the 'every set and level' claim does not survive the unclipped protocol. the 3 major comments →

arxiv 2607.22793 v1 pith:MKFK7JBF submitted 2026-07-24 eess.IV cs.CV

Small, Bias-Free, Blind and Convolutional Denoiser: A compact ConvNeXt U-Net for blind Gaussian color-image denoising

Nikolas Markou This is my paper
classification eess.IV cs.CV
keywords blind image denoisingbias-free CNNdegree-1 homogeneityscale equivarianceMiyasawa-Tweedie identityConvNeXt U-NetGabor filter bankLaplacian pyramid
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper claims that a compact, 0.82-million-parameter U-Net can act as a single blind color-image denoiser across a wide range of noise levels, provided every operation in the network is bias-free so that the whole map is exactly scale-homogeneous at inference: D(alpha y) = alpha D(y). One model, trained on a noise curriculum up to sigma≈64, is claimed to match or beat the classic sub-million-parameter denoisers DnCNN and FFDNet on all twelve cells of four standard color sets at sigma 15/25/50, averaging about +0.7 dB over DnCNN, and to extrapolate without a cliff to sigma 200, more than three times its training ceiling. The paper is explicit about the boundaries: the [0,1] input clip it uses at training and inference is worth up to +2.23 dB on CBSD68 at sigma 50, so that set is a match rather than a win, and the learned residual is a local, non-conservative score, not a global prior, so plug-and-play-style guarantees do not transfer.

Core claim

The central discovery is that end-to-end bias-freedom is a workable route to blind denoising in a modern compact architecture, not just a theoretical nicety. By removing every additive constant (no bias in convolutions, linear head, LeakyReLU, variance-only batch norm), the trained operator satisfies D(alpha y)=alpha D(y) to floating-point precision at inference, which makes one model behave like an input-adaptive linear filter whose action depends only on the direction of the noisy image. The paper reads the residual through the classical empirical-Bayes score identity, treats it as a local score estimate, and shows empirically that this single checkpoint denoises gracefully far outside its

What carries the argument

Degree-1 homogeneity: the network is built so that every constituent operation is linear through the origin or positively homogeneous (bias-free convolutions, linear head, LeakyReLU, variance-only batch norm, Laplacian pyramid skips), yielding the exact identity D(alpha y)=alpha D(y) at inference. This scale equivariance is the mechanism that carries blind generalization and off-range extrapolation; the empirical-Bayes score identity is the lens that converts the denoiser residual into a local score usable in sampling and inverse problems.

Load-bearing premise

The entire leaderboard claim rests on the assumption that clipping the noisy input to [0,1] is a harmless protocol detail; the paper measured that detail only on CBSD68, where it is worth more than the margin it is meant to protect.

What would settle it

Take the released checkpoint, disable the [0,1] clip, and evaluate on Kodak24, McMaster and Urban100 at sigma 50 under the same full-image protocol used by the published baselines. If any of those PSNR values drops below the corresponding DnCNN/FFDNet numbers, the 'matches or beats on every set and level' claim is false; the paper's own CBSD68 measurement predicts the clip is worth +2.23 dB at sigma 50.

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

If this is right

  • One blind 0.82M checkpoint covers the standard benchmark levels and keeps working at 2-3x the maximum training noise, so per-noise-level training is unnecessary in this regime.
  • The same frozen model can drive stochastic sampling and linear inverse problems (inpainting, super-resolution, deblurring, compressive sensing) because those loops need only a local score, not a symmetric Jacobian.
  • On the four standard color sets at sigma 15/25/50 the model averages about +0.7 dB over DnCNN and +0.3 dB over FFDNet, and comes within roughly 0.3-1.7 dB of 12-32M-parameter state-of-the-art models.
  • The negative result is equally concrete: because the measured Jacobian is non-symmetric (asymmetry ~0.93), plug-and-play and RED convergence guarantees do not transfer to this denoiser.

Where Pith is reading between the lines

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

  • The sigma=50 margins on Kodak24 and McMaster are smaller than the +2.23 dB clip advantage measured on CBSD68, so an unclipped like-for-like comparison would likely erase or reverse some of the claimed wins at the highest level.
  • The reported high-sigma extrapolation (22.8 dB at sigma 150, 20.0 dB at sigma 200) is measured on clipped input, which raises the noisy-input PSNR; part of the smooth 'no cliff' behavior may be a clipping effect rather than pure homogeneity.
  • Since the paper reports no ablations, a natural next experiment is to train the same network with biases/LayerNorm and with the Gabor stem and pyramid skips removed, to attribute the gain to bias-freedom versus architecture.
  • If the clip protocol were removed, the calibration of the score interpretation changes with noise level; an unclipped variant would be a cleaner test of the empirical-Bayes score reading at high sigma.

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 / 4 minor

Summary. This paper presents BF-ConvUNeXt, a 0.82M-parameter bias-free ConvNeXt U-Net for blind color AWGN denoising. The architecture combines a frozen Gabor stem, Laplacian-pyramid skips, ConvNeXt-V1 blocks, and a bias-free construction that makes the network exactly degree-1 homogeneous at inference. A single checkpoint is trained on a noise-sigma curriculum (sigma_255 roughly 6.4 to 64) and evaluated on CBSD68, Kodak24, McMaster, and Urban100 at sigma in {15,25,50} under the full-image protocol. The paper claims it matches or beats DnCNN and FFDNet on every set and level, trails larger state-of-the-art models by roughly 0.3–1.7 dB, extrapolates smoothly to sigma=150/200, and can be used as a local score for sampling and linear inverse problems. The paper is explicit that the [0,1] input clip is a protocol difference and measures its effect on CBSD68.

Significance. If the main comparison were protocol-matched, the paper would be a useful data point: a compact blind denoiser with a clearly stated architectural mechanism, verified exact homogeneity, open code/checkpoint, and unusually candid limitation statements (clip asymmetry, single checkpoint, no ablations, non-conservative Jacobian). The paper does not claim to beat the state of the art and correctly identifies where its theory does not apply. However, the central 'matches or beats ... every set and level' claim currently exceeds what the paper's own measurements support, so the main quantitative significance is conditional on either re-running the comparison under a matched protocol or revising the claims.

major comments (3)
  1. [Abstract, §5.2 Table 1, §6] The headline 'matches or beats DnCNN and FFDNet on every set and level' is not supported by the paper's own protocol analysis. Section 6 measures the [0,1] input clip to be worth +0.28/+0.71/+2.23 dB on CBSD68 at sigma=15/25/50, states it was not measured on the other sets, and explicitly declines to claim that the sigma=50 margins on Kodak24/McMaster/Urban100 survive it. Yet Table 1's Kodak24 sigma=50 margin is +0.40 dB over DnCNN and +0.37 dB over FFDNet, both smaller than the sigma=25 clip advantage measured on CBSD68. Under the unclipped protocol used by the collated baselines, that cell is expected to flip, and the abstract's 'every set and level' formulation is not justified. Please re-run all four sets at all three levels with the clip disabled (or re-run baselines under the clipped protocol) and revise the abstract/conclusion claims accordingly.
  2. [§5.1, Figs. 2–3; §2.5 Eq. (1)] The extrapolation claim (22.8 dB at sigma=150, 20.0 dB at sigma=200) is computed on inputs clipped to [0,1], and the 'no cliff' curve in Fig. 3 is plotted against the clipped noisy-input PSNR (7.4 dB at sigma=200). Clipping raises the noisy-input PSNR substantially at high sigma and changes the corruption from additive to clipped Gaussian, so the smooth degradation is partly an artifact of the input domain. Moreover, Eq. (1) holds for D, but the deployed composite y -> clip(y) -> D(y) is not degree-1 homogeneous in y, so the Mohan et al. mechanism does not by itself justify the clipped-input extrapolation curve. Please report the unclipped-input curve and, at minimum, restate the extrapolation claim as specific to the clipped protocol.
  3. [§6, Table 1] The statement that 'McMaster, Urban100, and Kodak24 margins are wide enough to absorb an effect of the sigma=15–25 magnitude' is contradicted by the Kodak24 sigma=50 row: the margins are +0.40 dB (DnCNN) and +0.37 dB (FFDNet), both below the +0.71 dB clip advantage measured on CBSD68 at sigma=25 and far below the +2.23 dB at sigma=50. Without measuring the clip advantage on each set/level, the paper cannot assert that those margins survive. This is not a presentation issue; it directly affects every 'clear win' statement on the three non-CBSD68 sets.
minor comments (4)
  1. [§2.2] The Gabor stem parameter ranges are attributed to 'Özbulak–Ekenel Table-I' but no citation is given. Please add the reference or state the exact ranges inline.
  2. [§5.2, Table 1] The bold-face convention marks every cell where BF-ConvUNeXt exceeds a baseline, but the text in §6 says several CBSD68 cells should be read as protocol-dependent matches. Consider marking protocol-sensitive cells (e.g., with an asterisk) so that the table does not visually overstate the claim.
  3. [§2.5, §4] The homogeneity proof applies only at inference, while the variance-only batch norm uses per-batch statistics during training. Please clarify in §4 how the running variance is obtained and used in the single-image full-image evaluation, so readers can reproduce the inference-time normalization exactly.
  4. [§8] The repository URL is useful; adding a commit hash or version tag for the released checkpoint would strengthen reproducibility further.

Circularity Check

0 steps flagged

No circular derivation: the homogeneity property is derived from bias-free construction and verified by direct measurement; external theory is imported as published background, not as a self-citation; the [0,1]-clip caveat is a protocol-fairness issue, not circularity.

full rationale

BF-ConvUNeXt's central claim, D(αy)=αD(y), is not fitted from the benchmarks it predicts: it follows from the stated bias-free construction (use_bias=False, linear head, LeakyReLU, variance-only BatchNorm) and is verified by direct numerical measurement, with exact-zero dyadic checks and sub-1e-4 non-dyadic error reported in Section 3. No parameter is fit to the target comparisons and then reported as a prediction; the noise-curriculum blind generalization is an empirical sweep over a single trained checkpoint, and the paper explicitly measures and discloses the [0,1]-clip advantage on CBSD68 (+0.28/+0.71/+2.23 dB) and declines to claim the unmeasured sigma=50 margins survive it. The Miyasawa/Tweedie score reading is imported from external, properly cited theory (Miyasawa 1961, Robbins 1956, Mohan et al. 2020), and the paper itself limits its force by reporting the non-conservative Jacobian (asymmetry 0.93). There are no self-citations by the author, no uniqueness theorem invoked from the authors' prior work, and no renamed empirical pattern presented as derivation. The leaderboard comparison is an engineering measurement with explicit protocol caveats; those caveats affect correctness/fairness, not circularity.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 0 invented entities

The paper introduces no new physical entities; the 'local score' is a property of the trained operator, not a new postulated object. The central claims rest on standard math (Tweedie, homogeneity, PSNR scaling) and prior-paper domain assumptions, plus several hand-chosen architectural and training hyperparameters.

free parameters (6)
  • Noise sigma curriculum maximum = 0.025 to 0.25 linearly over 100 epochs (sigma255 6.4 to 64)
    Hand-chosen schedule defines the training noise range; the blind generalization and extrapolation claims are measured relative to this range.
  • Architecture size (width, depth, blocks per scope) = flat width 66, depth 2, 3 blocks per scope + 3 per high-frequency skip band
    Hand-chosen; affects parameter count, capacity, and the efficiency claims.
  • Gabor stem parameter ranges = s in [2,21], theta in [0,360], lambda in [8,100], gamma in [0,300%], psi in [0,360] from Ozbulak-Ekenel Table-I
    Frozen, hand-swept filter bank drawn from an uncited prior table; contributes no trainable parameters but shapes the input representation.
  • Gaussian blur kernel for Laplacian split = 5x5
    Fixed, hand-chosen; defines the lossless split and the high-frequency residual routing.
  • Block hyperparameters = LayerScale init 1e-4 (floored 1e-6), LeakyReLU slope 0.1, dropout 0.1
    Hand-chosen; LayerScale scaling is linear and preserves homogeneity.
  • Training hyperparameters = AdamW lr 1e-3, cosine decay, 10 warmup epochs, weight decay 4e-3, batch size 4, 100 epochs
    Hand-chosen; the checkpoint's behavior depends on the training recipe.
axioms (6)
  • standard math Miyasawa/Tweedie identity: for y=x+eps, eps~N(0,sigma^2 I), the MMSE estimator is y + sigma^2 grad log p_sigma(y).
    Section 3 motivates the residual-as-score reading; assumed to apply to the trained network as a local score.
  • domain assumption Mohan et al.'s blind-generalization argument: a degree-1 homogeneous denoiser has a degree-0 Jacobian, so it generalizes across noise levels.
    Section 3 imports this from Mohan et al. (2020); the paper verifies homogeneity empirically but does not re-prove the generalization claim.
  • domain assumption Training-set representativeness: COCO/DIV2K patches generalize to CBSD68/Kodak24/McMaster/Urban100.
    Sections 4 and 5 assume standard benchmark transfer; no distribution-shift analysis is provided.
  • standard math Lossless Laplacian split: up(l) + h = z.
    Section 2.3 defines the split and relies on it for the skip routing.
  • standard math PSNR comparability: max_val=1.0 PSNR is directly comparable to max_val=255 numbers.
    Section 4 states scale invariance of PSNR under a matched range.
  • standard math Non-conservativeness implies no global prior: a non-symmetric Jacobian means the residual is not the gradient of any scalar potential.
    Section 3 uses the measured Jacobian asymmetry (0.93) to conclude plug-and-play/RED guarantees do not transfer.

pith-pipeline@v1.3.0-alltime-deepseek · 12632 in / 17236 out tokens · 129234 ms · 2026-08-01T05:12:06.573071+00:00 · methodology

0 comments
read the original abstract

We describe and evaluate BF-ConvUNeXt, a compact bias-free ConvNeXt U-Net for blind additive-white-Gaussian-noise color image denoising, combining four existing ingredients so a single property survives end to end: a frozen depthwise Gabor stem (oriented band-pass, zero trainable parameters), a Laplacian-pyramid encoder routing the high-frequency residual into each skip connection, a ConvNeXt-V1 U-Net body, and bias-free construction throughout (no additive bias, linear head, LeakyReLU, variance-only batch norm). Together these make the 0.82M-parameter network exactly degree-1 homogeneous at inference, D(alpha y) = alpha D(y), licensing a Miyasawa/Tweedie score reading of the residual and blind generalization across noise levels from one model. We train a single blind model on a noise-sigma curriculum (sigma approximately 6.4 to 64, 0-255 scale); it extrapolates past that ceiling without a cliff, degrading smoothly to 22.8 dB at sigma=150 and 20.0 dB at sigma=200. Evaluated unchanged on four standard color sets (CBSD68, Kodak24, McMaster, Urban100) at sigma in {15,25,50}, it matches or beats DnCNN and FFDNet on every set and level, averaging about +0.7 dB over DnCNN. Against heavyweight CNN/transformer state of the art it trails by a small margin (roughly 0.3-1.7 dB depending on set) at 1/15 to 1/39 of their parameters. The homogeneity is inference-only and checkpoint-specific, and the learned residual is a local, not global, score (non-conservative Jacobian), so plug-and-play/RED guarantees do not transfer; it still drives stochastic sampling and linear inverse problems (inpainting, super-resolution, deblurring, compressive sensing).

Figures

Figures reproduced from arXiv: 2607.22793 by Nikolas Markou.

Figure 1
Figure 1. Figure 1: BF-ConvUNeXt data flow (channel widths shown). The encoder (left) descends, the decoder spine (right) ascends, and the bottleneck sits at the base. A frozen Gabor stem, followed by a trainable bias￾free 1×1 projection (gabor_stem_projection) that remixes its 66 responses, feeds a two-level ConvNeXt￾V1 U-Net. Downsampling is a Laplacian-pyramid split: the low-frequency band descends (solid arrows) while the… view at source ↗
Figure 2
Figure 2. Figure 2: Blind PSNR vs. noise on DIV2K-validation ( [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Denoising gain: denoised PSNR (solid blue) vs. the corrupted noisy-input PSNR (dotted red) across [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: A clean reference followed by (noisy, denoised) pairs at [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Four-set average PSNR at σ=25 vs. (approximate) parameter count. BF-ConvUNeXt (blue) sits on the efficient side of the trade-off: above the sub-1M CNNs and within ∼ 0.71 dB of the 12–32M state of the art at ∼ 1/15–1/39 of their size. Parameter counts are approximate and collated from the respective papers. The parameter count is only half the efficiency story; [PITH_FULL_IMAGE:figures/full_fig_p011_5.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

14 extracted references · 10 linked inside Pith

  1. [1]

    J., & Adelson, E

    Burt, P. J., & Adelson, E. H. (1983). The Laplacian pyramid as a compact image code. IEEE Transactions on Communications, 31(4), 532--540

    1983

  2. [2]

    Gabor, D. (1946). Theory of communication. Journal of the Institution of Electrical Engineers, 93(26), 429--457

    1946

  3. [3]

    Kadkhodaie, Z., & Simoncelli, E. P. (2021). Stochastic solutions for linear inverse problems using the prior implicit in a denoiser. Advances in Neural Information Processing Systems (NeurIPS). arXiv:2007.13640

    Pith/arXiv arXiv 2021

  4. [4]

    Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., & Timofte, R. (2021). SwinIR: Image restoration using Swin Transformer. ICCV Workshops. arXiv:2108.10257

    Pith/arXiv arXiv 2021

  5. [5]

    Liu, Z., Mao, H., Wu, C.-Y., Feichtenhofer, C., Darrell, T., & Xie, S. (2022). A ConvNet for the 2020s. CVPR. arXiv:2201.03545

    Pith/arXiv arXiv 2022

  6. [6]

    Miyasawa, K. (1961). An empirical Bayes estimator of the mean of a normal population. Bulletin of the International Statistical Institute, 38(4), 181--188

    1961

  7. [7]

    P., & Fernandez-Granda, C

    Mohan, S., Kadkhodaie, Z., Simoncelli, E. P., & Fernandez-Granda, C. (2020). Robust and interpretable blind image denoising via bias-free convolutional neural networks. International Conference on Learning Representations (ICLR). arXiv:1906.05478

    Pith/arXiv arXiv 2020

  8. [8]

    Robbins, H. (1956). An empirical Bayes approach to statistics. Proceedings of the Third Berkeley Symposium on Mathematical Statistics and Probability, 1, 157--163

    1956

  9. [9]

    Ronneberger, O., Fischer, P., & Brox, T. (2015). U-Net: Convolutional networks for biomedical image segmentation. MICCAI. arXiv:1505.04597

    Pith/arXiv arXiv 2015

  10. [10]

    W., Arora, A., Khan, S., Hayat, M., Khan, F

    Zamir, S. W., Arora, A., Khan, S., Hayat, M., Khan, F. S., & Yang, M.-H. (2022). Restormer: Efficient transformer for high-resolution image restoration. CVPR. arXiv:2111.09881

    Pith/arXiv arXiv 2022

  11. [11]

    Zhang, K., Zuo, W., Chen, Y., Meng, D., & Zhang, L. (2017). Beyond a Gaussian denoiser: Residual learning of deep CNN for image denoising. IEEE Transactions on Image Processing, 26(7), 3142--3155. arXiv:1608.03981

    Pith/arXiv arXiv 2017

  12. [12]

    Zhang, K., Zuo, W., & Zhang, L. (2018). FFDNet: Toward a fast and flexible solution for CNN-based image denoising. IEEE Transactions on Image Processing, 27(9), 4608--4622. arXiv:1710.04026

    Pith/arXiv arXiv 2018

  13. [13]

    Zhang, K., Li, Y., Zuo, W., Zhang, L., Van Gool, L., & Timofte, R. (2021). Plug-and-play image restoration with deep denoiser prior. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10), 6360--6376. arXiv:2008.13751

    Pith/arXiv arXiv 2021

  14. [14]

    Zhang, K., Li, Y., Liang, J., Cao, J., Zhang, Y., Tang, H., Fan, D.-P., Timofte, R., & Van Gool, L. (2022). Practical blind image denoising via Swin-Conv-UNet and data synthesis. arXiv:2203.13278

    Pith/arXiv arXiv 2022