acceptodds
Under review as a conference paper at ICLR 2027

Recorrupted-to-Noise: A Unifying Framework for Recorruption-Based Self-Supervised Denoising

11 views

Abstract

Self-supervised learning has emerged as an effective paradigm for image denoising when clean ground-truth data are expensive or unavailable. Existing approaches often rely on restrictive assumptions regarding noise independence or distribution, limiting their applicability across imaging scenarios. In this paper, we introduce a Recorrupted-to-Noise (R2N) self-supervised denoising framework, which offers a unified theoretical interpretation of recorruption-based denoising methods. R2N recorrupts noisy observations and trains a network to predict the recorruption noise. The validity of R2N rests on conditional projective equivalence (CPE) between the observation noise and the recorruption noise, a weaker condition than those required by many existing self-supervised denoising methods. This condition supports a broad range of noise models, including general additive noise, natural exponential-family models, and certain multiplicative noise models. We also develop a test-time strategy for non-uniform noise with unknown parameters. We further analyze the effect of the recorruption coefficient on denoising error and establish a robustness bound under observation-model mismatch. Extensive experiments demonstrate that R2N outperforms existing self-supervised methods in challenging scenarios involving broader noise models and unknown noise parameters.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
self-supervised learning, image restoration, denoising

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_uyeuOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes Recorrupted to Noise (R2N), a framework for self supervised image denoising. Instead of predicting a clean image from a recorrupted observation, R2N trains a network to estimate the added recorruption noise. The method is built on the Conditional Projective Equivalence (CPE) condition, under which the estimated recorruption noise can be rescaled to recover the posterior mean of the clean image. The authors show that the framework covers general additive noise, natural exponential family models, and certain multiplicative noise models. They also connect R2N with existing recorruption methods such as R2R and GR2R, analyze the effect of the recorruption coefficient, and introduce a strategy to estimate this coefficient when the exact noise level is unknown. Experiments on synthetic noise settings and microscopy data are used to validate the theoretical analysis and denoising performance.

Critical strengths

1.The paper provides a clear theoretical formulation for recorruption based denoising through the CPE condition. This formulation covers several noise models within a common framework and gives a useful interpretation of existing methods such as R2R and GR2R.

  1. The theoretical discussion is supported by targeted experiments. In particular, the experiments on the recorruption coefficient agree with the predicted error trends, and the additional analysis examines coefficient estimation and robustness under observation model mismatch.
Critical weaknesses
  1. Limited methodological novelty relative to existing recorruption methods. The proposed framework is closely related to R2R and especially GR2R. The paper itself shows that the GR2R objective can be rewritten as the R2N objective up to a constant scaling factor. Therefore, in the settings already supported by GR2R, the practical distinction between the two formulations appears limited. The main novelty lies in the CPE interpretation and its extension to a broader set of noise models. I am not fully convinced that this represents a sufficiently large methodological advance for ICLR. 25127_Recorrupted_to_Noise_A_U
  2. The experimental evidence is not broad enough to support some of the general claims. The main comparisons rely on PURE, Noise2Score, Neighbor2Neighbor, and GR2R, with GR2R being the main recent baseline. Several experiments involve noise settings where many competing methods are not applicable, so R2N is sometimes mainly compared with Neighbor2Neighbor. In addition, the experiments on unknown noise still assume knowledge of the average noise level or the general noise model. Therefore, the claims about broad applicability and unknown noise conditions appear stronger than what is directly demonstrated by the experiments.
Questions for authors

1.Since the paper shows that the GR2R objective can be transformed into the R2N objective up to a constant factor, could the authors clarify what fundamentally new learning principle R2N introduces beyond this reformulation?

  1. Could the authors provide comparisons with more recent self supervised denoising methods, especially methods published after 2025? This would help determine whether the empirical advantage remains under a broader set of competitive baselines.

  2. For the unknown noise setting, the experiments still assume that the average noise level or the general noise model is available. How would R2N behave when both the noise distribution and its scale are unknown?

  3. The CPE analysis supports general additive noise and certain multiplicative models under specific assumptions. Could the authors clarify the practical failure cases when these assumptions are violated, and provide empirical results under stronger model mismatch?

2 comments

Thank you very much for your invaluable comments and constructive suggestions. Here are answers to your questions.

(1) Our framework makes three contributions. First, we formulate a conditional proportionality criterion and construct recorruption models beyond the NEF settings of GR2R, including composite additive, clipped Cauchy, and multiplicative Beta noise. Second, we estimate decoding coefficients from noisy observations, reducing reliance on known noise parameters within the assumed model. We will clarify the additional assumptions this requires. Third, our error decomposition separates recorruption uncertainty from prediction-error amplification, clarifying the role of the recorruption parameter and guiding its selection.

(2) We have conducted additional comparisons with ZS-HPD (CVPR 2026), AdaReNet-N2V (CVPR 2025), TTAD (ICCV 2025), and TBSN (AAAI 2025). These cover zero-shot denoising, test-time adaptation, masking-based learning, and Transformer-based blind-spot networks.

  • AdaReNet-N2V: We implement N2V training with the official architecture and our Table 1 budget, retaining both auxiliary-loss weights of 0.1.
  • TBSN: We evaluate three configurations: the official training recipe, an L2 variant with our Table 1 training budget, and the official recipe with multiplicative mean normalization. We consider both direct and R3 inference.
  • ZS-HPD and TTAD: We use their released implementations. We evaluate both 10-step and 100-step adaptation of TTAD.

For multiplicative noise, AdaReNet-N2V, ZS-HPD, TTAD, and TBSN receive known-mean normalization of their noisy inputs and targets. For each noise condition, we report the highest measured PSNR among the tested configurations of each baseline.

MethodMul. BetaMul. BernoulliCompositedCauchy
TBSN26.46 / 0.78725.08 / 0.74125.45 / 0.72825.50 / 0.741
AdaReNet-N2V32.04 / 0.91730.14 / 0.86929.91 / 0.87328.36 / 0.808
TTAD30.63 / 0.87920.77 / 0.51929.66 / 0.84524.60 / 0.633
ZS-HPD30.39 / 0.89523.46 / 0.66428.79 / 0.83725.42 / 0.688
ours31.95 / 0.9033.57 / 0.9530.08 / 0.8729.26 / 0.85

These results support R2N’s competitiveness against recent methods, with advantages depending on the noise model.

(3) Our "unknown-noise" experiments address uncertainty within an assumed noise family or with approximate scale information.

When both the noise distribution and its scale are unknown, R2N can still be trained using a chosen surrogate recorruption process, such as additive Gaussian noise. The decoding coefficient can then be estimated from noisy observations. However, successful prediction of the added noise alone does not guarantee accurate recovery of the underlying signal: the conditional proportionality relation may no longer hold, and empirical coefficient calibration cannot eliminate arbitrary distribution mismatch. Recorruption strength remains a hyperparameter. Thus, R2N is not an assumption-free solution to jointly unknown distributions and scales.

The real fluorescence microscopy experiment in the appendix provides one practical example. We apply Gaussian recorruption to observations with mixed Poisson–Gaussian noise and obtain 31.15 dB PSNR and 0.821 SSIM (Table 8). Here, the chosen Gaussian perturbation level is a hyperparameter, rather than a measurement of the true acquisition noise. This result illustrates that useful denoising is possible without an exact observation-noise model. It does not establish automatic noise-model identification.

(4) When CPE is violated, minimizing noise-prediction loss can yield biased reconstruction. Matching the noise mean and variance alone does not guarantee the required conditional relationship.

We isolated this failure in a controlled numerical experiment with multiplicative Beta observations. We replaced the compatible Beta recorruption multiplier with a two-point distribution having exactly the same mean and variance. We computed the optimal conditional noise predictors analytically and evaluated their reconstruction risks by numerical integration.

Recorr​uption and decodingBayes MSEExcess MSETotal MSE
Compatible Beta kernel0.002773approximately 00.002773
Moment-matched two-point kernel, original coefficient0.0027150.0027580.005473
Moment-matched two-point kernel, clean-oracle optimal scalar coefficient0.0027150.0026190.005334
Here, excess MSE measures the reconstruction’s deviation from the posterior mean under each setting’s own observation model. CPE-based decoding of the mismatched kernel introduces substantial additional error. Even allowing the scalar coefficient to be optimized using clean information reduces total MSE by only about 2.5%, leaving a substantial gap to the Bayes risk. Thus, this failure cannot be explained solely by an incorrectly chosen scalar coefficient.

The authors have added comparisons with recent methods and conducted additional experiments on CPE violations, addressing some of my previous concerns and strengthening the experimental evidence. I acknowledge these improvements and am willing to raise my score accordingly. However, concerns remain regarding the methodological novelty of R2N over GR2R and its applicability under unknown noise conditions.

Reviewer_omuBOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a self-supervised denoising method which learns to predict added corruption and rescales this prediction to recover a clean-image posterior mean. Its core idea, conditional projective equivalence, leads to constructions for additive and selected multiplicative noise models. The paper also studies corruption strength, test-time coefficient estimation, and model mismatch, with synthetic experiments and evaluation on a real microscopy dataset

Critical strengths
  • CPE provides a compact way to organize recorruption constructions. The explicit Beta and Bernoulli examples make the framework more concrete than a reformulation

  • The evaluation covers different corruption models, unknown noise levels, and real microscopy data

Critical weaknesses
  • The novelty over related methods is not clearly explained enough. Eq. (14) establishes an exact loss reparameterization of GR2R in shared settings. Moreover , Noisier2Noise already uses conditional-expectation identities and output correction for additive and Bernoulli noise. The paper should distinguish its new ideas from this existing methods.
  • The Cauchy noise setting is not fully consistent with the MSE theory. Since untruncated Cauchy noise has infinite variance, the usual finite-second-moment assumption does not hold. Although appendix mentions the integrability requirement, it does not clarify how this requirement is met in the experiments or whether clipping the noise affects the claimed CPE relation
  • Test-time calibration needs an additional signal assumption. Eq.18 assumes approximate orthogonality between predicted noise and the clean image, which does not follow from CPE or independent corruption.
  • The coefficient analysis is partly heuristic. The U-shaped approximation error and preferred coefficient range depend on how prediction errors change with α\alpha. They do not follow from the displayed scaling factor alone. The masking connection is also intuitive rather than a formal equivalence.
  • The experimental positioning and reproducibility need strengthening. Relevant unknown-noise approaches such as UNSURE (ICLR 2025) and CANC (CVPR 2026) are missing. Dataset splits, training schedules, and whether inference uses the original or freshly recorrupted observation are also insufficiently specified
Questions for authors
  • Were Cauchy observations or training targets clipped? If so, how does this affect the CPE identity and the population-MSE interpretation?
  • Was Neighbor2Neighbor corrected for the non-unit mean of the multiplicative noise? For the unknown-noise comparisons, what prevents applying the same coefficient calibration to the equivalent GR2R residual predictor?
1 comment

Thank you very much for your invaluable comments and constructive suggestions. Below, we address the key questions as well as the limitations of our work.

Response to the Key Questions

We sincerely thank the reviewers for these insightful questions, which are highly valuable for strengthening our work.

(1) We agree that the algebraic reformulation alone does not establish a fundamentally different learning principle. In settings where both formulations apply, their objectives are equivalent up to a constant factor, so we do not claim that predicting noise inherently yields a better optimum than GR2R.

Our contribution lies in developing this formulation into a broader framework with three components. First, we formulate a conditional proportionality criterion and provide explicit recorruption constructions for noise models beyond the NEF settings considered by GR2R, including composite additive noise, clipped Cauchy noise, and multiplicative Beta noise. Second, we develop decoding-coefficient estimation from noisy observations, reducing reliance on known noise parameters within the assumed recorruption model. This estimation requires additional assumptions, which we will state more clearly. Third, we derive an error decomposition that separates the uncertainty introduced by recorruption from the amplification of prediction errors. This analysis explains the role of the recorruption parameter and provides guidance for its selection.

(2) In the standard Cauchy experiment, both observation and additional noise samples were clipped to [−1,1][-1,1] before addition. The targets therefore have finite second moments. We will explicitly describe this setting as clipped Cauchy-derived noise.

(3) Equation (18) is the exact least-squares solution to γ^=arg⁡min⁡γ∥Y~−γR(Y~)∥2.\widehat\gamma=\arg\min_\gamma\|\widetilde Y-\gamma R(\widetilde Y)\|^2. Its derivation does not require orthogonality. Approximate orthogonality is needed to interpret it as a denoising coefficient: weak signal alignment limits unintended signal removal. This approximation does not follow from CPE alone, and the appendix provides empirical support rather than a universal guarantee.

(4) The error decomposition is exact, whereas the U-shaped trend additionally assumes that the numerator in Eq. (17) remains relatively stable as α\alpha changes. Since network prediction errors also change with α\alpha, the scaling factor alone cannot guarantee this shape. We will make this assumption explicit.

For the masking connection, we provide a proof under Gaussian noise and bounded clean images: when increasingly strong additional noise is applied only at selected locations, the correctly decoded predictor converges in mean square to the conditional mean based solely on the remaining observations. We will state these assumptions explicitly instead of presenting the connection only through intuition.

(5) We have completed additional UNSURE comparisons and will report them. CANC has not been evaluated because its official repository currently provides no implementation.

We will also clarify the experimental protocols. The original DIV2K synthetic experiments use the official 800 image training and 100 image validation split, with Adam, a learning rate of 5e−4, batch size 16, and 500 epochs. The validation set is used for checkpoint selection. At inference, R2N receives a recorrupted version of the original noisy observation.

(6) The original Neighbor2Neighbor comparison did not apply mean correction: the true multiplicative-noise mean was not supplied for preprocessing. R2N likewise operates on the unnormalized observations without explicitly dividing by the true multiplier mean. We have additionally evaluated a mean-corrected version of the same frozen Neighbor2Neighbor models. Using the known multiplier mean to correct the outputs of the same frozen Neighbor2Neighbor models improves PSNR from 16.19 to 31.87 dB for Beta noise and from 15.98 to 34.68 dB for Bernoulli noise, confirming that intensity bias substantially affected the original comparison.

We also agree that the proposed calibration can be applied to the equivalent GR2R residual predictor. We have evaluated this control, and calibration benefits GR2R as well. We do not claim that calibration is exclusive to R2N. Our contributions additionally include the CPE formulation, compatible constructions for additional noise models, and the analysis of reconstruction error and recorruption strength.

Reviewer_9i34OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes R2N, a self-supervised denoising framework that recorrupts a noisy image and trains a network to predict the added recorruption noise, then rescales that estimate to recover the posterior mean of the clean image. The whole thing rests on a single condition they call conditional projective equivalence (CPE), which is weaker than the independence/distribution assumptions most prior methods need, and they use it to fold mask-based and statistic-based (R2R/GR2R/SURE) approaches into one umbrella, plus a test-time trick for estimating the coefficient when the noise level is unknown or non-uniform.

Critical strengths

The unification angle is the real contribution here and the U-shaped reducible-error analysis for the coefficient α is a clean result that the experiments actually reproduce.

Critical weaknesses

"Harness a broad range of noise models" oversells it a bit — the admissible constructions still boil down to additive / NEF / multiplicative, so the unification is more elegant bookkeeping than a genuinely general theory. On standard noise the gains over GR2R and supervised are basically nil (Table 1, Poisson/Gamma are within rounding), and the test-time estimator quietly depends on predicted-noise ⊥ clean-image, which the appendix only verifies empirically with a non-zero cosine similarity. The mask-method connection is also just an informal α→1 limiting argument, not a real proof.

Questions for authors

"Harness a broad range of noise models" oversells it a bit — the admissible constructions still boil down to additive / NEF / multiplicative, so the unification is more elegant bookkeeping than a genuinely general theory. On standard noise the gains over GR2R and supervised are basically nil (Table 1, Poisson/Gamma are within rounding), and the test-time estimator quietly depends on predicted-noise ⊥ clean-image, which the appendix only verifies empirically with a non-zero cosine similarity. The mask-method connection is also just an informal α→1 limiting argument, not a real proof.

1 comment

Thank you very much for your invaluable comments and constructive suggestions. Below, we address the key questions as well as the limitations of our work.

Response to the Key Questions

We sincerely thank the reviewers for these insightful questions, which are highly valuable for strengthening our work.

(1) We agree that “harness a broad range of noise models” should be stated more carefully. Our intention was to describe broader coverage across the noise settings considered in comparison with the methods evaluated in the paper. The constructions we establish concern additive noise, natural exponential family models, and selected multiplicative models, subject to the stated assumptions. Applicability to a particular observation model requires a compatible recorruption scheme satisfying CPE; our results do not establish such a scheme for arbitrary noise distributions. We will revise the wording to make these boundaries explicit and describe the scope of each construction more precisely.

(2) We agree that the gains in these standard noise settings are modest, which is consistent with our expectations. In these experiments, GR2R uses the known noise parameters and a correctly specified recorruption scheme. Both GR2R and our known-parameter R2N use α = 0.2, which lies near the broad low-error region in our α analysis. Given the correspondence between their objectives under shared assumptions, similar performance is expected.

Likewise, we do not expect a systematic advantage over a well-trained supervised method: under squared-error loss, the posterior mean is the optimal predictor for a given observation. These comparisons demonstrate that R2N achieves competitive performance through self-supervision. The intended benefits concern its unified formulation, extensions to additional noise settings, and coefficient adaptation when noise parameters are unavailable.

(3) We would like to clarify the rationale behind Eq. (18): it is the exact least-squares solution to minimizing $|\widetilde Y-\gamma R(\widetilde Y)|^2$, and its derivation itself does not require orthogonality. The approximation concerns interpreting this projection coefficient as a denoising coefficient.

Geometrically, this calibration removes the component of the observation along the predicted-noise direction. It is therefore reasonable when the clean image has a small projection onto that direction. More precisely, writing r=R(Y~)r=R(\widetilde Y) and defining the clean-oracle coefficient along the same direction as γX=arg⁡min⁡γ∥Y~−γr−X∥2\gamma_X=\arg\min_\gamma\|\widetilde Y-\gamma r-X\|^2, we obtain γ^−γX=⟨X,r⟩∥r∥2.\widehat\gamma-\gamma_X=\frac{\langle X,r\rangle}{\|r\|^2}. Thus, a non-zero cosine similarity produces a quantifiable discrepancy; exact orthogonality is not necessary for the calibration to be useful. The appendix provides empirical support for this approximation in the evaluated settings, rather than a proof of universal validity.

(4) We prove that the optimally decoded R2N prediction converges to the optimal masked prediction as α→1\alpha \to 1 under Gaussian noise.

Consider one selected pixel, with UU denoting the original noisy observations at the remaining locations. Additional noise is applied only to the selected pixel. Assume a bounded clean pixel and mutually independent Gaussian noises, independent of the clean pixel and UU:

\qquad Z_\tau = X + N + \tau G, \qquad N \sim \mathcal{N}(0,\sigma^2), \quad G \sim \mathcal{N}(0,1).$$ Define $$v_\tau = \sigma^2 + \tau^2, \qquad \alpha_\tau = \frac{\tau^2}{v_\tau} \longrightarrow 1 \quad \text{as } \tau \to \infty.$$ By Bayes' rule, the selected observation reweights the conditional prior of $X$ given $U$ by $$w_\tau(x)=\exp\left(\frac{Z_\tau x}{v_\tau}-\frac{x^2}{2v_\tau}\right).$$ As the additional-noise strength increases, $$\frac{Z_\tau}{v_\tau}=\frac{X+N+\tau G}{\sigma^2+\tau^2}\longrightarrow 0.$$ Since $0 \le x \le 1$, $$\sup_{0 \le x \le 1}\left|\log w_\tau(x)\right|\le\frac{|Z_\tau|}{v_\tau}+\frac{1}{2v_\tau}\longrightarrow 0.$$ Thus, the additional weights converge uniformly to 1. After normalization, the posterior converges to the conditional prior given $U$: the selected observation no longer provides additional information. $$Z_\tau-\alpha_\tau^{-1}R_\tau^*(U,Z_\tau)=\mathbb{E}[X \mid U,Z_\tau]\xrightarrow{}\mathbb{E}[X \mid U].$$ The limit is precisely the optimal prediction of the selected pixel from the remaining observations. This conclusion also extends to a broad class of non-Gaussian additive noises. Under the same boundedness and independence assumptions, finite positive variance and a nonzero absolutely continuous component are sufficient when using the paper’s construction of summing independent noise copies.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.