Published as a conference paper at ICLR 2025

S CORE - BASED S ELF - SUPERVISED MRI D ENOISING

Jiachen Tu, Yaokun Shi & Fan Lam
Department of Electrical and Computer Engineering
University of Illinois at Urbana-Champaign
Champaign, IL 61820, USA
{jtu9, yaokuns2, fanlam1}@illinois.edu

A BSTRACT

Magnetic resonance imaging (MRI) is a powerful noninvasive diagnostic imaging tool

that provides unparalleled soft tissue contrast and anatomical detail. Noise contamination,
especially in accelerated and/or low-field acquisitions, can significantly degrade image
quality and diagnostic accuracy. Supervised learning based denoising approaches have
achieved impressive performance but require high signal-to-noise ratio (SNR) labels, which
are often unavailable. Self-supervised learning holds promise to address the label scarcity
issue, but existing self-supervised denoising methods tend to oversmooth fine spatial
features and often yield inferior performance than supervised methods. We introduce
Corruption2Self (C2S), a novel score-based self-supervised framework for MRI denois-
ing. At the core of C2S is a generalized denoising score matching (GDSM) loss, which
extends denoising score matching to work directly with noisy observations by modeling
the conditional expectation of higher-SNR images given further corrupted observations.
This allows the model to effectively learn denoising across multiple noise levels directly
from noisy data. Additionally, we incorporate a reparameterization of noise levels to
stabilize training and enhance convergence, and introduce a detail refinement extension
to balance noise reduction with the preservation of fine spatial features. Moreover, C2S
can be extended to multi-contrast denoising by leveraging complementary information
across different MRI contrasts. We demonstrate that our method achieves state-of-the-art
performance among self-supervised methods and competitive results compared to super-
vised counterparts across varying noise conditions and MRI contrasts on the M4Raw and
fastMRI dataset. The project website is available at: https://jiachentu.github.
io/Corruption2Self-Self-Supervised-Denoising/.

1 I NTRODUCTION

Magnetic resonance imaging (MRI) is an invaluable noninvasive imaging modality that provides exceptional
soft tissue contrast and anatomical detail, playing a crucial role in clinical diagnosis and research. However,
the inherent sensitivity of MRI to noise, particularly in accelerated acquisitions and/or low-field settings, can
impair diagnostic accuracy and subsequent computational analysis. Unlike natural images, MRI data often
contains Rician or non-central chi-distributed noise in magnitude images. Denoising has become an important
processing step in the MRI data analysis pipeline. Higher signal-to-noise ratios (SNR) enable better trade-offs
for reduced scanning times or increased spatial resolution, both of which can improve patient experience
and diagnostic accuracy. Traditional denoising methods, such as Non-Local Means (NLM) Froment (2014)
and BM3D Mäkinen et al. (2020), often rely on handcrafted priors such as Gaussianity, self-similarity,
or low-rankness. The performance of these methods is limited by the accuracy of their assumptions and
often requires prior knowledge specific to the acquisition methods used. With the advent of deep learning,

1
Published as a conference paper at ICLR 2025

supervised learning based denoising methods Zhang et al. (2017); Liang et al. (2021); Zamir et al. (2022)
have demonstrated impressive performance by learning complex mappings from noisy inputs to clean targets.
However, they rely heavily on the availability of high-quality, high-SNR “ground truth” data for training—a
resource that is not always readily available or feasible to acquire in MRI. Acquiring such clean labels
necessitates longer scan times, leading to increased costs and patient discomfort. This limitation underscores
the need for self-supervised denoising algorithms, which remove or reduce the dependency on annotated
datasets by leveraging self-generated pseudo-labels as supervisory signals. Moreover, supervised approaches
often face generalization issues due to distribution shifts caused by differences in imaging instruments,
protocols, and noise levels Xiang et al. (2023), limiting their utility in real-world clinical settings. Techniques
such as Noise2Noise Lehtinen et al. (2018), Noise2Void Krull et al. (2019), Noise2Self Batson & Royer
(2019), and their extensions Xie et al. (2020); Huang et al. (2021); Pang et al. (2021); Jang et al. (2024); Wang
et al. (2023) have demonstrated the potential to learn effective denoising models without explicit clean targets.
Recent innovations tailored for MRI, like Patch2Self Fadnavis et al. (2020) and Coil2Coil Park et al. (2022),
further exploit domain-specific characteristics to enhance performance. However, these self-supervised
methods often have limitations. For example, Noise2Noise requires repeated noisy measurements, which
may not be practical. Methods like Noise2Void and Noise2Self rely on masking strategies that can limit the
receptive field or introduce artifacts, potentially leading to oversmoothing and loss of fine details. Approaches
such as Pfaff et al. (2023) and Park et al. (2022) require additional data processing steps, restricting their
applicability. In this paper, we introduce Corruption2Self, a score-based self-supervised framework for MRI
denoising. Building upon the principles of denoising score matching (DSM) Vincent (2011), we extend
DSM to the ambient noise setting, where only noisy observations are available, enabling effective learning in
practical MRI settings where high-SNR data are scarce or impractical to obtain. An overview of the C2S
workflow is illustrated in Figure 1. Additionally, we incorporate a reparameterization of noise levels for
a consistent coverage of the noise level range during training, leading to enhanced training stability and
convergence. In medical imaging, the visual quality of the output and the preservation of diagnostically
relevant features are often more critical than achieving high scores on standard metrics. To address this
priority, a detail refinement extension is introduced to balance noise reduction with the preservation of fine
spatial feature. Furthermore, we extend C2S to multi-contrast denoising by integrating data from multiple
MRI contrasts. This approach leverages complementary information across contrasts, indicating the potential
of C2S to better exploit the rich information available in multi-contrast MRI acquisitions.

Figure 1: Overview of the Corruption2Self (C2S) workflow for MRI denoising. Starting from a noisy MRI
image Xtdata , the forward corruption process adds additional Gaussian noise to create progressively noisier
versions Xt . During training, the model learns to reverse this process by estimating the clean image X0
from these corrupted observations, despite having access only to noisy data. The denoising function hθ
approximates the conditional expectation E[X0 | Xt ], effectively learning to denoise without clean targets. A
reparameterized function Dθ (Xt , t), which shares parameters with hθ , is used to compute the loss.

We conduct extensive experiments on publicly available datasets, including a low-field MRI dataset, to
evaluate the performance of C2S. Our results demonstrate that C2S achieves state-of-the-art performance

2
Published as a conference paper at ICLR 2025

among self-supervised methods and, after extending to multi-contrast on the M4Raw dataset, shows state-of-
the-art performance among both self-supervised and supervised methods. Notably, we are among the first to
comprehensively analyze and compare self-supervised and supervised learning approaches in MRI denoising.
Our findings reveal that C2S not only bridges the performance gap but also offers robust performance under
varying noise conditions and MRI contrasts. This indicates the potential of self-supervised learning to achieve
competitive performance with supervised approaches when the latter are trained on practically obtainable
higher-SNR labels, particularly in scenarios where perfectly clean ground truth is unavailable, offering a
practical and robust solution adaptable to broader clinical settings.

2 BACKGROUND

2.1 L EARNING - BASED DENOISING WITHOUT CLEAN TARGETS

A central concept in many of the self-supervised denoising methods is J-invariance Batson & Royer (2019),
where the denoising function is designed to be invariant to certain subsets of pixel values. Techniques like
Noise2Void Krull et al. (2019) and Noise2Self Batson & Royer (2019) utilize this property by masking pixels
and predicting their values based on their surroundings, approximating supervised learning objectives without
the need for clean data. Noise2Same Xie et al. (2020) extends these concepts by implicitly enforcing J -
invariance via optimizing a self-supervised upper bound. Approaches such as Neighbor2Neighbor Huang et al.
(2021), Noisier2Noise Moran et al. (2020), and Recorrupted2Recorrupted Pang et al. (2021) create pairs of
images from a single noisy observation to mimic the effect of supervised training objectives with independent
noisy pairs. Noise2Score Kim & Ye (2021) exploits the assumption that noise follows an exponential family
distribution and utilizes Tweedie’s formula to obtain the posterior expectation of clean images using the
estimated score function via the AR-DAE Lim et al. (2020) approach. This approach highlights a connection
between Stein’s Unbiased Risk Estimator (SURE)-based denoising methods Soltanayev & Chun (2018); Kim
et al. (2020) and score matching objectives Hyvärinen & Dayan (2005). Specifically, under additive Gaussian
noise, the SURE cost function can be reformulated as an implicit score matching objective, differing only
by a scaling factor and a constant term Kim & Ye (2021). In the context of MRI, recent innovations have
capitalized on the intrinsic properties of data within a self-supervised framework Moreno López et al. (2021)
to enhance denoising performance without clean labels. For standard MRI, Pfaff et al. (2023) utilizes SURE
Kim et al. (2020) and spatially resolved noise maps to improve denoising performance, while Coil2Coil Park
et al. (2022) leverages coil sensitivity to exploit multi-coil information effectively. In the realm of diffusion
MRI (dMRI), specialized approaches have emerged to address the unique challenges of 4D data. Patch2Self
Fadnavis et al. (2020) employs a J-invariant approach that preserves critical anatomical details by selectively
excluding target volume data from its training inputs, though it requires a minimum of ten additional diffusion
volumes. DDM2 Xiang et al. (2023) introduces a three-stage framework incorporating statistical image
denoising into diffusion models. More recently, Wu et al. Wu et al. (2025) proposed Di-Fusion, a single-stage
self-supervised approach that performs dMRI denoising by integrating statistical techniques with diffusion
models through a Fusion process that prevents drift in results and a ”Di-” process that better characterizes
real-world noise distributions.

2.2 D ENOISING S CORE M ATCHING

Denoising Score Matching (DSM) Vincent (2011) is a framework for learning the score function (i.e., the
gradient of the log-density) of the data distribution by training a model to reverse the corruption process
of adding Gaussian noise. In DSM, given a clean data sample X0 ∈ Rd and a noise level σt > 0, a noisy
observation Xt is generated by adding Gaussian noise:Xt = X0 + σt Z, Z ∼ N (0, Id ). The objective
is to train a denoising function hθ : Rd × R → Rd , parameterized by θ, that estimates the clean image
X0 from its noisy counterpart Xt . This is achieved by minimizing the expected mean squared error (MSE)

3
Published as a conference paper at ICLR 2025

h i
2
loss EX0 ,Xt ∥hθ (Xt , t) − X0 ∥ . Vincent (2011) demonstrated that optimizing the denoising function hθ
using the MSE loss is equivalent to learning the score function of the noisy data distribution pt (xt ) up to a
scaling factor. Specifically, using Tweedie’s formula, the relationship between the denoising function and
the score function is given by: ∇xt log pt (xt ) = σ12 (hθ (xt , t) − xt ) . By learning hθ , we implicitly learn
t
∇xt log pt (xt ), which forms the foundation of score-based generative models Song & Ermon (2019); Ho
et al. (2020). However, DSM relies on access to clean data during training, limiting its applicability in
situations where only noisy observations are available. Ambient Denoising Score Matching (ADSM) Daras
et al. (2024) leverages a double application of Tweedie’s formula to relate noisy and clean distributions,
enabling score-based learning from noisy observations. While ADSM was originally designed to mitigate
memorization in large diffusion models by treating noisy data as a form of regularization Daras et al. (2024), its
potential for self-supervised denoising remains underexplored. In our work, we bridge the gap of score-based
self-supervised denoising in practical imaging applications.

3 M ETHODOLOGY
Consider a clean image X0 ∈ Rd and its corresponding noisy observation Xtdata ∈ Rd . We formulate the
self-supervised denoising problem as estimating the conditional expectation E[X0 | Xtdata ], which constitutes
the optimal estimator of X0 in the minimum mean square error (MMSE) sense. While estimating this
typically requires clean or high-SNR reference images in a supervised learning setting, we adopt an ambient
score-matching perspective inspired by Daras et al. (2024), circumventing the need for clean labels. The
noisy image can be modeled as:
Xtdata = X0 + σtdata N, (1)
where σtdata > 0 denotes the noise standard deviation at level tdata , and N ∼ N (0, Id ) represents the noise
component. In MRI, the noise N is typically assumed to be additive, ergodic, stationary, uncorrelated, and
white in k-space Liang & Lauterbur (2000). When the signal-to-noise ratio (SNR) exceeds two, the noise
in the image domain can be well-approximated as Gaussian distributed Gudbjartsson & Patz (1995). While
this Gaussian assumption underpins our theoretical framework, our empirical results suggest robustness even
when this condition is not strictly satisfied.
The noise level σtdata scales the noise to match the observed noise level in the MRI data. To facilitate self-
supervised learning, we introduce a forward corruption process that systematically adds additional Gaussian
noise to Xtdata , defining a continuum of increasingly noisy versions of the data:
q
Xt = Xtdata + σt2 − σt2data Z, Z ∼ N (0, Id ), t > tdata , (2)

where σt is a strictly monotonically increasing noise schedule function for t ∈ (tdata , T ], with T being the
maximum noise level. This process allows us to model the distribution of the noisy data at different noise
levels and forms the foundation for our generalized denoising score matching approach. In scenarios where
the noise deviates from Gaussianity (e.g., Rician noise in low-SNR regions), pre-processing techniques such as
the Variance Stabilizing Transform (VST) Foi (2011) can be applied to better approximate Gaussian statistics.
Our goal is to train a denoising function hθ : Rd × R → Rd , parameterized by θ, which maps a noisy input
Xt at noise level t to an estimate of either the clean image X0 or a less noisy version corresponding to a
target noise level σttarget ≤ σtdata , where Xttarget = X0 + σttarget N0 , with N0 ∼ N (0, Id ).

3.1 G ENERALIZED D ENOISING S CORE M ATCHING

We introduce the Generalized Denoising Score Matching (GDSM) loss, which enables learning a denoising
function directly from noisy observations by modeling the conditional expectation of a higher-SNR image
given a further corrupted version of the noisy data (proof provided in Appendix 4).

4
Published as a conference paper at ICLR 2025

Theorem 1 (Generalized Denoising Score Matching). Let X0 ∈ Rd be a clean data sample drawn from the
distribution p0 (x0 ). Suppose that the noisy observation at a given data noise level tdata is
Xtdata = X0 + σtdata N, N ∼ N (0, Id ),
and that for any t > tdata the observation Xt is generated according to the forward process described in
Equation equation 2. Let hθ : Rd × (tdata , T ] → Rd be a denoising function parameterized by θ, and fix a
target noise level σttarget satisfying 0 ≤ σttarget ≤ σtdata . Define the loss function
h i


2
J(θ) = EXtdata , t, Xt γ t, σttarget hθ (Xt , t) + δ t, σttarget Xt − Xtdata , (3)
where t is sampled uniformly from (tdata , T ] and the coefficients are defined by
σ 2 − σt2data σt2 − σt2target
γ t, σttarget := 2t



and δ t, σttarget := data .
σt − σt2target σt2 − σt2target
Then any minimizer θ∗ of J(θ) satisfies
h i
hθ∗ (Xt , t) = E Xttarget Xt .
In other words, the optimal denoising function recovers the conditional expectation of the image with noise
level σttarget given the more heavily corrupted observation Xt .
Remark 1. The proposed GDSM framework generalizes several existing methods: when σttarget = σtdata ,
GDSM reduces to the standard denoising score matching (DSM) Vincent (2011); when σttarget = 0, it recovers
the ambient denoising score matching (ADSM) formulation Daras et al. (2024). Moreover, GDSM subsumes
Noisier2Noise Moran et al. (2020) as a special case. By setting σttarget = 0 and fixing the noise level to
α2 1
σt2 = (1 + α2 )σt2data , the coefficients simplify to γ(t, 0) = 1+α 2 and δ(t, 0) = 1+α 2 , thereby generalizing

Noisier2Noise to a continuous range of noise levels. For further details, see Section B.2.
To enhance training stability and improve convergence, we introduce a reparameterization of the noise
levels. Let τ ∈ (0, T ′ ] be a new variable defined by
q
στ2 = σt2 − σt2data , T ′ = σT2 − σt2data . (4)
The original t can be recovered via the inverse of σt , as:
q 
t = σt−1 στ2 + σt2data . (5)

Since σt is strictly increasing, σt−1 is well-defined. In practice, as T ′ ≫ tdata , we approximate T ′ ≈ T to

allow uniform sampling over τ and consistent coverage of the noise level range during training, leading to
smoother and faster convergence. To further improve training stability, we combine our reparameterization
strategy with Exponential Moving Average (EMA) of model parameters. Details of the training dynamics
with and without reparameterization (and the effect of EMA) are provided in Appendix I (see Figure 12).
Under this reparameterization, the loss function in Equation equation 12 becomes:
h i
2
J ′ (θ) = EXtdata ,τ,Xt γ ′ (τ, σttarget ) hθ (Xt , t) + δ ′ (τ, σttarget ) Xt − Xtdata , (6)
where the coefficients are:
στ2 σt2data − σt2target
γ ′ (τ, σttarget ) = , δ ′ (τ, σttarget ) = . (7)
στ2 + σt2data − σt2target στ2 + σt2data − σt2target
Corollary 2 (Reparameterized Generalized Denoising Score Matching). With our reparameterization strategy,
the minimizer θ∗ of the objective J ′ (θ) satisfies (proof provided in Appendix 5):
hθ∗ (Xt , t) = E Xttarget | Xt , ∀Xt ∈ Rd , t > tdata .
 

5
Published as a conference paper at ICLR 2025

3.2 C ORRUPTION 2S ELF : A S CORE - BASED S ELF - SUPERVISED MRI D ENOISING F RAMEWORK

Building upon the Reparameterized Generalized Denoising Score Matching introduced earlier, we propose
Corruption2Self
 (C2S) where  the objective is to train a denoising function hθ that approximates the conditional
expectation E Xttarget | Xt , where Xttarget is a less noisy version of Xtdata with noise level σttarget ≤ σtdata .
Denote Dθ (Xτ , τ ) as the weighted combination of the network output with the input through skip connections:

Dθ (Xτ , τ ) = λout (τ, σttarget ) hθ (Xt , t) + λskip (τ, σttarget ) Xt , (8)

q 
where Xt = Xtdata + στ Z with Z ∼ N (0, Id ), and t relates to τ through t = σt−1 στ2 + σt2data . The
blending coefficients are defined as:

στ2 σt2data − σt2target

λout (τ, σttarget ) = , λskip (τ, σttarget ) = . (9)
στ2 + σt2data − σt2target στ2 + σt2data − σt2target

When σttarget = 0 (maximum denoising), the goal is to predict E[X0 | Xt ] and the coefficients simplify to:
λout (τ, 0) = στ2 /(στ2 + σt2data ), λskip (τ, 0) = σt2data /(στ2 + σt2data ). Our loss function is expressed as:

1 h
2
i
LC2S (θ) = EXtdata ∼ptdata (x),τ ∼U (0,T ] w(τ ) ∥Dθ (Xτ , τ ) − Xtdata ∥2 , (10)
2
where Xτ is constructed by adding noise to Xtdata according to the reparameterized noise level τ , and w(τ ) is a
weighting function designed to balance the contributions from different noise levels. Following practices from

α
prior works Song et al. (2020); Kingma et al. (2021); Karras et al. (2022), w(τ ) can be set to στ2 + σt2data ,
with α being a hyperparameter controlling the weighting.
During inference, given a noisy observation Xtdata , the denoised output is obtained by:

X̂ = hθ∗ (Xtdata , tdata ), (11)

where the trained model hθ∗ approximates E[X0 | Xtdata ], providing a clean estimate of the image.
While the C2S training procedure effectively approximates E[X0 | Xt ] when σttarget = 0, it can lead to
oversmoothing. To maintain a balance between noise reduction and feature preservation, we introduce a detail
refinement extension where the network is trained to predict E[Xttarget | Xt ] with a non-zero target noise level
σttarget > 0, allowing the network to retain a controlled amount of noise for preserving finer image textures
(details are provided in Appendix G). As shown in Table 1, incorporating the detail refinement extension
leads to a statistically significant improvement in image quality across contrasts. Our denoising model builds
upon the U-Net architecture employed in DDPM Ho et al. (2020), enhanced with time conditioning and the
Noise Variance Conditioned Multi-Head Self-Attention (NVC-MSA) module Hatamizadeh et al. (2023),
which enables the self-attention layers to dynamically adapt to varying noise scales. Empirically, we found
that setting the noise schedule function στ equal to the noise level τ yields good performance. More details
regarding the architecture and implementation are provided in Appendix C.

4 R ESULTS AND D ISCUSSION

We first evaluate C2S on the M4Raw dataset, which includes in-vivo MRI data with real noise. The training
and validation data use three-repetition-averaged images, while the test data comprises higher-SNR labels,

6
Published as a conference paper at ICLR 2025

Algorithm 1 Corruption2Self Training Procedure Table 1: Effectiveness of detail refinement module

on the M4Raw validation dataset. Improvements are
Require: Noisy dataset Xtdata N
i=1 , hθ , max noise level statistically significant (∗p < 0.05) using paired t-tests.
T , batch size m, total iterations K, noise schedule
function στ Additional details are provided in Appendix G.
1: for k = 1 to K do
2: Sample minibatch Xtdata m i=1 , τ ∼ U(0, T ], Z ∼
Contrast PSNR p-value
N (0, Id ) T1 34.56±0.062 → 34.89±0.038 0.001∗
2
στ
3: Compute λout (τ, 0) = 2 +σ 2
στ t
, λskip (τ, 0) = T2 33.84±0.090 → 34.07±0.121 0.001∗
σt2
data
FLAIR 32.44±0.073 → 32.58±0.089 0.016∗
data
2 +σ 2
στ t
data q SSIM p-value
4: Recover t = σt−1 ( στ2 + σt2data )
T1 0.885±0.002 → 0.892±0.003 0.007∗
5: Compute Xt ← Xtdata + στ ZP T2 0.860±0.003 → 0.867±0.002 0.003∗
1 m
6: Compute loss: LC2S (θ) = 2m i=1 w(τ )∥ · ∥ FLAIR 0.812±0.004 → 0.818±0.001 0.005∗
7: Update θ using Adam optimizer Kingma (2014)

created by averaging six repetitions for T1 and T2, and four for FLAIR. This setup allows us to assess how well
denoising methods perform when evaluated on cleaner test data. Table 2 compares the performance of C2S
against classical methods (NLM, BM3D), supervised learning (SwinIR, Restormer, Noise2Noise), and self-
supervised approaches (Noise2Void, Noise2Self, PUCA, LG-BPN, Noisier2Noise, Recorrupted2Recorrupted).
C2S consistently outperforms other self-supervised methods, achieving the highest PSNR and SSIM across
all contrasts. Our base C2S model significantly outperforms existing self-supervised methods across all
contrasts, achieving PSNRs of 32.59dB, 32.28dB, and 32.43dB for T1, T2, and FLAIR respectively. With
our detail refinement extension, C2S further improves performance to 32.77dB/0.919, 32.33dB/0.890, and
32.51dB/0.876 for T1, T2, and FLAIR contrasts respectively. Notably, recent self-supervised methods like
PUCA and LG-BPN, demonstrate lower performance on MRI data. This performance gap can be attributed
to the blind-spot architecture design, often leading to information loss and oversmoothed results.

Figure 2: Comparison of different denoising methods for T1 contrast from the M4Raw dataset.

An important observation is that supervised methods such as SwinIR and Restormer, trained on three-
repetition-averaged labels, do not significantly outperform self-supervised methods on higher-SNR test data.
Supervised models typically learn E[Xttarget | Xtdata ], where tdata > ttarget > 0 when multi-repetition averaged
samples are used as labels. This makes supervised methods less effective at handling shifts to higher-SNR test
data. In contrast, C2S approximates E[X0 | Xt ], allowing it to achieve competitive performance on test data.
Empirical results on test labels (three-repetition-average) matching the SNR of the training data (presented

7
Published as a conference paper at ICLR 2025

Methods T1 T2 FLAIR
PSNR / SSIM ↑ PSNR / SSIM ↑ PSNR / SSIM ↑
Classical Non-Learning-Based Methods
NLM Froment (2014) 31.90 / 0.898 31.17 / 0.876 32.01 / 0.870
BM3D Mäkinen et al. (2020) 32.07 / 0.903 31.20 / 0.877 32.14 / 0.873
Supervised Learning Methods
SwinIR Liang et al. (2021) 32.53 / 0.913 31.90 / 0.891 32.15 / 0.885
Restormer Zamir et al. (2022) 32.35 / 0.912 31.79 / 0.890 32.31 / 0.886
Noise2Noise Lehtinen et al. (2018) 32.59 / 0.911 32.37 / 0.886 32.70 / 0.871
Self-Supervised Single-Contrast Methods
Noise2Void Krull et al. (2019) 31.46 / 0.870 30.93 / 0.857 31.17 / 0.851
Noise2Self Batson & Royer (2019) 31.72 / 0.887 31.18 / 0.873 31.72 / 0.870
PUCA Jang et al. (2024) 30.52 / 0.870 29.11 / 0.827 29.57 / 0.807
LG-BPN Wang et al. (2023) 31.15 / 0.890 30.66 / 0.868 30.82 / 0.862
Noisier2Noise Moran et al. (2020) 31.60 / 0.876 31.45 / 0.871 31.59 / 0.861
Recorrupted2Recorrupted Pang et al. (2021) 31.67 / 0.876 31.33 / 0.870 31.57 / 0.863
C2S 32.59 / 0.915 32.28 / 0.888 32.43 / 0.872
C2S (w/ Detail Refinement) 32.77 / 0.919 32.33 / 0.890 32.51 / 0.876

Table 2: Quantitative comparison of denoising performance on the M4Raw test dataset. Results show PSNR
(dB) and SSIM metrics across T1, T2, and FLAIR contrasts. C2S with detail refinement (bold) achieves the
best performance among all self-supervised methods across all contrasts, and even outperforms supervised
approaches in some cases. Second-best results among self-supervised methods are underlined.

in Appendix F) show that supervised methods like SwinIR and Restormer perform better when the noise
characteristics of the training and test data are similar.

Methods PD, σ = 13/255 PD, σ = 25/255 PDFS, σ = 13/255 PDFS, σ = 25/255

PSNR / SSIM ↑ PSNR / SSIM ↑ PSNR / SSIM ↑ PSNR / SSIM ↑
Classical Non-Learning-Based Methods
NLM 30.40 / 0.772 21.63 / 0.327 28.82 / 0.726 21.12 / 0.350
BM3D 33.16 / 0.829 30.58 / 0.755 30.64 / 0.705 28.49 / 0.592
Supervised Learning Methods
Noise2True (SwinIR) 34.44 / 0.868 32.39 / 0.820 31.35 / 0.774 29.55 / 0.665
Noise2True (U-Net) 34.54 / 0.870 32.61 / 0.825 31.39 / 0.775 29.62 / 0.669
Noise2Noise Lehtinen et al. (2018) 34.06 / 0.854 30.83 / 0.769 31.33 / 0.773 29.12 / 0.654
Self-Supervised Methods
Noise2Void Krull et al. (2019) 32.19 / 0.804 29.79 / 0.706 29.50 / 0.629 27.99 / 0.558
Noise2Self Batson & Royer (2019) 32.47 / 0.808 30.60 / 0.757 29.32 / 0.613 28.31 / 0.563
PUCA Jang et al. (2024) 31.03 / 0.771 29.88 / 0.740 28.65 / 0.594 27.54 / 0.527
LG-BPN Wang et al. (2023) 31.15 / 0.776 30.32 / 0.751 29.14 / 0.603 27.77 / 0.535
Noisier2Noise Moran et al. (2020) 33.18 / 0.807 30.35 / 0.741 30.39 / 0.683 27.83 / 0.559
Recorrupted2Recorrupted Pang et al. (2021) 33.29 / 0.810 30.52 / 0.745 30.95 / 0.752 28.32 / 0.561
C2S 33.36 / 0.831 30.62 / 0.753 30.72 / 0.750 28.58 / 0.592
C2S (w/ Detail Refinement) 33.48 / 0.832 30.67 / 0.761 30.91 / 0.756 28.62 / 0.601

Table 3: Quantitative evaluation on the fastMRI test dataset with simulated noise at two different levels
(σ = 13/255 and σ = 25/255) across PD and PDFS contrasts.

8
Published as a conference paper at ICLR 2025

Figure 3: Comparison of denoising methods for the PD contrast (σ = 13/255) from the fastMRI dataset.

To further evaluate the robustness of C2S under different noise levels, we conducted experiments on the
fastMRI dataset Zbontar et al. (2018), simulating Gaussian noise with σ = 13/255 and σ = 25/255.
As shown in Table 3, the same baseline methods are analyzed and C2S consistently achieves the best or
comparable results among self-supervised methods. On PDFS with σ = 13/255, Recorrupted2Recorrupted
achieves a slightly higher PSNR (30.95 dB vs. 30.91 dB); however, C2S records the highest SSIM (0.756),
indicating better detail preservation. It is worth noting that although the labels in this simulated dataset do not
have added synthetic noise, they still contain inherent noise typical in MRI, albeit with higher SNR. Figure 3
demonstrates that our method balances feature preservation and noise removal, resulting in much cleaner
visual representations compared to other methods. For additional results on fastMRI, refer to Appendix E.
We assessed the effect of reparameterization on training stability and performance by mapping the noise
levels τ ∈ (0, T ] to a new scale for more uniform sampling. As shown in Table 4a, reparameterization
improves PSNR and SSIM across all contrasts. The training dynamics, illustrated in Figure 12, confirm the
stabilizing effect of reparameterization. The model with reparameterization (blue) shows smoother and faster
convergence than the model without it (orange), which fluctuates more and converges slower.

M4Raw fastMRI
Architecture
T1 T2 FLAIR PSNR ↑ SSIM ↑ PSNR ↑ SSIM ↑
Method
PSNR ↑ SSIM ↑ PSNR ↑ SSIM ↑ PSNR ↑ SSIM ↑
U-Net 33.11 0.865 32.32 0.807
Without Reparam. 31.14 0.837 30.53 0.807 30.43 0.771 DDPM 34.82 0.886 33.48 0.835
With Reparam. 34.43 0.882 33.82 0.860 32.56 0.814 Ours 34.91 0.890 33.63 0.837

(a) Impact of reparameterization of noise levels on the M4Raw dataset. (b) Influence of model architecture on the
Results are validation results obtained after training for 200 epochs. M4Raw dataset (T1) and fastMRI dataset (PD).

Table 4: Ablation studies on the impact of reparameterization and model architecture.

We evaluated the impact of different architectural choices on the performance of our denoising model. As
shown in Table 4b, incorporating time conditioning significantly improves both PSNR and SSIM across
the M4Raw (T1 contrast) and fastMRI (PD contrast, noise level 13/255) datasets. The best performance is
achieved by further incorporating the NVC-MSA module in our model (Appendix C), which allows the model
to dynamically adapt to varying noise levels by integrating noise variance into the self-attention mechanism.
Our approach demonstrates strong robustness to noise level estimation errors, making it suitable for prac-
tical applications where exact noise levels may be unknown. Through extensive experiments detailed in
Appendix H, we show that C2S maintains stable performance even with significant estimation errors (±50%
of the true noise level). This robustness, combined with the incorporation of standard noise estimation

9
Published as a conference paper at ICLR 2025

techniques (e.g., from the skimage package Van der Walt et al. (2014)), enables our method to effectively
function as a blind denoising model. In practice, we find such noise estimation tools provide sufficiently
accurate estimates for optimal model performance, alleviating the need for precise noise level knowledge.
More quantitative results and analysis of the effect of noise estimation error are provided in Appendix H.

Extending C2S to Multi-Contrast Settings MRI typically involves acquiring multiple contrasts to provide
comprehensive diagnostic information. By leveraging complementary information from different contrasts,
denoising performance can be further enhanced. To capitalize on this, we extend the C2S framework to
multi-contrast settings by incorporating additional MRI contrasts as inputs. Figure 4 demonstrates the visual
comparison of different denoising methods for the T1 contrast on the M4Raw dataset. It is evident that
using multi-contrast inputs (T1 & T2, T1 & FLAIR) allows for better structural preservation and more
detailed reconstructions compared to single-contrast denoising techniques. Quantitatively, Table 5 shows that
multi-contrast C2S consistently outperforms classical BM3D, supervised Noise2Noise, and single-contrast
C2S in terms of PSNR and SSIM. More details can be found in Appendix D.

Target Contrast Best Classical Best Supervised Best Self-Supervised Multi-Contrast C2S
PSNR / SSIM ↑ BM3D Noise2Noise C2S T1 & T2 FLAIR & T1 T2 & FLAIR
T1 32.07 / 0.903 32.59 / 0.911 32.77 / 0.919 33.57 / 0.921 33.89 / 0.922 N/A
T2 31.20 / 0.877 32.37 / 0.886 32.33 / 0.890 33.01 / 0.895 N/A 33.36 / 0.901
FLAIR 32.14 / 0.873 32.70 / 0.871 32.51 / 0.876 N/A 32.71 / 0.879 32.62 / 0.877

Table 5: Multi-contrast denoising results on the M4Raw dataset. For the multi-contrast C2S results, entries
such as ”T1 & T2” indicate that T1 and T2 contrasts were used as inputs for denoising the target contrast.

Figure 4: Comparison of different denoising methods for T1 contrast in the M4Raw dataset. The figure
showcases Noisy, BM3D, Noise2Noise, and C2S, along with multi-contrast C2S variants (T1 & T2, T1 &
FLAIR). Multi-contrast C2S preserves more structural details and produces sharper reconstructions.

5 C ONCLUSION
We have introduced Corruption2Self, a score-based self-supervised denoising framework tailored for MRI
applications. By extending denoising score matching to the ambient noise setting through our Generalized
Denoising Score Matching approach, C2S enables effective learning directly from noisy observations without
the need for clean labels. Our method incorporates a reparameterization of noise levels to stabilize training and
enhance convergence, as well as a detail refinement extension to balance noise reduction with the preservation
of fine spatial features. By extending C2S to multi-contrast settings, we further leverage complementary
information across different MRI contrasts, leading to enhanced denoising performance. Notably, C2S
exhibits superior robustness across varying noise conditions and MRI contrasts, highlighting its potential for
broader applicability in clinical settings.

10
Published as a conference paper at ICLR 2025

ACKNOWLEDGEMENTS
This research was partially supported by the following fundings: NSF-CBET-1944249 and NIH-
R35GM142969. The authors would also like to thank Ruiyang Zhao and Yizun Wang for the helpful
discussions.

R EFERENCES
Joshua Batson and Loı̈c Royer. Noise2self: Blind denoising by self-supervision. In Proceedings of the 36th
International Conference on Machine Learning, volume 97, pp. 524–533, 2019.
Giannis Daras, Alexandros G Dimakis, and Constantinos Daskalakis. Consistent diffusion meets tweedie:
Training exact ambient diffusion models with noisy data. arXiv preprint arXiv:2404.10177, 2024.
Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Un-
terthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth
16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.
Shreyas Fadnavis, Joshua Batson, and Eleftherios Garyfallidis. Patch2self: Denoising diffusion mri with
self-supervised learning. Advances in Neural Information Processing Systems, 33:16293–16303, 2020.
Chun-Mei Feng, Huazhu Fu, Shuhao Yuan, and Yong Xu. Multi-contrast mri super-resolution via a multi-
stage integration network. In Medical Image Computing and Computer Assisted Intervention–MICCAI
2021: 24th International Conference, Strasbourg, France, September 27–October 1, 2021, Proceedings,
Part VI 24, pp. 140–149. Springer, 2021.
Alessandro Foi. Noise estimation and removal in mr imaging: The variance-stabilization approach. In 2011
IEEE International symposium on biomedical imaging: from nano to macro, pp. 1809–1814. IEEE, 2011.
Jacques Froment. Parameter-free fast pixelwise non-local means denoising. Image Processing On Line, 4:
300–326, 2014.
Hákon Gudbjartsson and Samuel Patz. The rician distribution of noisy mri data. Magnetic resonance in
medicine, 34(6):910–914, 1995.
Ali Hatamizadeh, Jiaming Song, Guilin Liu, Jan Kautz, and Arash Vahdat. Diffit: Diffusion vision transform-
ers for image generation. arXiv preprint arXiv:2312.02139, 2023.
Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural
information processing systems, 33:6840–6851, 2020.
Tao Huang, Songjiang Li, Xu Jia, Huchuan Lu, and Jianzhuang Liu. Neighbor2neighbor: Self-supervised
denoising from single noisy images. In Proceedings of the IEEE/CVF conference on computer vision and
pattern recognition, pp. 14781–14790, 2021.
Aapo Hyvärinen and Peter Dayan. Estimation of non-normalized statistical models by score matching.
Journal of Machine Learning Research, 6(4), 2005.
Hyemi Jang, Junsung Park, Dahuin Jung, Jaihyun Lew, Ho Bae, and Sungroh Yoon. Puca: patch-unshuffle
and channel attention for enhanced self-supervised image denoising. Advances in Neural Information
Processing Systems, 36, 2024.
Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based
generative models. Advances in neural information processing systems, 35:26565–26577, 2022.

11
Published as a conference paper at ICLR 2025

Kwanyoung Kim and Jong Chul Ye. Noise2score: tweedie’s approach to self-supervised image denoising
without clean images. Advances in Neural Information Processing Systems, 34:864–874, 2021.

Kwanyoung Kim, Shakarim Soltanayev, and Se Young Chun. Unsupervised training of denoisers for low-dose
ct reconstruction without full-dose ground truth. IEEE Journal of Selected Topics in Signal Processing, 14
(6):1112–1125, 2020.

Diederik Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. Advances in
neural information processing systems, 34:21696–21707, 2021.

Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.

Alexander Krull, Tim-Oliver Buchholz, and Florian Jug. Noise2void-learning denoising from single noisy
images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.
2129–2137, 2019.

Jaakko Lehtinen, Jacob Munkberg, Jon Hasselgren, Samuli Laine, Tero Karras, Miika Aittala, and Timo Aila.
Noise2noise: Learning image restoration without clean data. In International Conference on Machine
Learning, pp. 2965–2974. PMLR, 2018.

Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image
restoration using swin transformer. In Proceedings of the IEEE/CVF international conference on computer
vision, pp. 1833–1844, 2021.

Zhi-Pei Liang and Paul C Lauterbur. Principles of magnetic resonance imaging. SPIE Optical Engineering
Press Bellingham, 2000.

Jae Hyun Lim, Aaron Courville, Christopher Pal, and Chin-Wei Huang. Ar-dae: towards unbiased neural
entropy gradient estimation. In International Conference on Machine Learning, pp. 6061–6071. PMLR,
2020.

Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin
transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF
international conference on computer vision, pp. 10012–10022, 2021.

Mengye Lyu, Lifeng Mei, Shoujin Huang, Sixing Liu, Yi Li, Kexin Yang, Yilong Liu, Yu Dong, Linzheng
Dong, and Ed X Wu. M4raw: A multi-contrast, multi-repetition, multi-channel mri k-space dataset for
low-field mri research. Scientific Data, 10(1):264, 2023.

Ymir Mäkinen, Lucio Azzari, and Alessandro Foi. Collaborative filtering of correlated noise: Exact transform-
domain variance for improved shrinkage and patch matching. IEEE Transactions on Image Processing, 29:
8339–8354, 2020.

Nick Moran, Dan Schmidt, Yu Zhong, and Patrick Coady. Noisier2noise: Learning to denoise from unpaired
noisy data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.
12064–12072, 2020.

Marc Moreno López, Joshua M Frederick, and Jonathan Ventura. Evaluation of mri denoising methods using
unsupervised learning. Frontiers in Artificial Intelligence, 4:642731, 2021.

Tongyao Pang, Huan Zheng, Yuhui Quan, and Hui Ji. Recorrupted-to-recorrupted: unsupervised deep
learning for image denoising. In Proceedings of the IEEE/CVF conference on computer vision and pattern
recognition, pp. 2043–2052, 2021.

12
Published as a conference paper at ICLR 2025

Juhyung Park, Dongwon Park, Hyeong-Geol Shin, Eun-Jung Choi, Hongjun An, Minjun Kim, Dongmyung
Shin, Se Young Chun, and Jongho Lee. Coil2coil: Self-supervised mr image denoising using phased-array
coil images. arXiv preprint arXiv:2208.07552, 2022.
Laura Pfaff, Julian Hossbach, Elisabeth Preuhs, Fabian Wagner, Silvia Arroyo Camejo, Stephan Kan-
nengiesser, Dominik Nickel, Tobias Wuerfl, and Andreas Maier. Self-supervised mri denoising: leveraging
stein’s unbiased risk estimator and spatially resolved noise maps. Scientific Reports, 13(1):22629, 2023.
Shakarim Soltanayev and Se Young Chun. Training deep learning based denoisers without ground truth data.
Advances in neural information processing systems, 31, 2018.
Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances
in neural information processing systems, 32, 2019.
Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben
Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint
arXiv:2011.13456, 2020.
Stefan Van der Walt, Johannes L Schönberger, Juan Nunez-Iglesias, François Boulogne, Joshua D Warner,
Neil Yager, Emmanuelle Gouillart, and Tony Yu. scikit-image: image processing in python. PeerJ, 2:e453,
2014.
Pascal Vincent. A connection between score matching and denoising autoencoders. Neural computation, 23
(7):1661–1674, 2011.
Zichun Wang, Ying Fu, Ji Liu, and Yulun Zhang. Lg-bpn: Local and global blind-patch network for self-
supervised real-world denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and
Pattern Recognition, pp. 18156–18165, 2023.
Chenxu Wu, Qingpeng Kong, Zihang Jiang, and S Kevin Zhou. Self-supervised diffusion MRI denoising via
iterative and stable refinement. In The Thirteenth International Conference on Learning Representations,
2025. URL https://openreview.net/forum?id=wxPnuFp8fZ.
Tiange Xiang, Mahmut Yurt, Ali B Syed, Kawin Setsompop, and Akshay Chaudhari. ddm2 : Self-supervised
diffusion mri denoising with generative diffusion models. arXiv preprint arXiv:2302.03018, 2023.
Yaochen Xie, Zhengyang Wang, and Shuiwang Ji. Noise2same: Optimizing a self-supervised bound for
image denoising. Advances in neural information processing systems, 33:20320–20330, 2020.
Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar 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, pp. 5728–5739, 2022.
Jure Zbontar, Florian Knoll, Anuroop Sriram, Tullie Murrell, Zhengnan Huang, Matthew J Muckley, Aaron
Defazio, Ruben Stern, Patricia Johnson, Mary Bruno, et al. fastmri: An open dataset and benchmarks for
accelerated mri. arXiv preprint arXiv:1811.08839, 2018.
Kai Zhang, Wangmeng Zuo, Yunjin Chen, Deyu Meng, and Lei Zhang. Beyond a gaussian denoiser: Residual
learning of deep cnn for image denoising. IEEE Transactions on Image Processing, 26(7):3142–3155,
2017.

13
Published as a conference paper at ICLR 2025

A A DDITIONAL E XPERIMENTAL D ETAILS

A.1 DATASETS

We evaluated our Corruption2Self (C2S) method on two distinct datasets:

In-vivo Dataset (M4Raw): The M4Raw dataset Lyu et al. (2023) contains multi-channel k-space images
across three contrasts (T1-weighted, T2-weighted, and FLAIR) from 183 participants. The dataset was
partitioned into training (128 individuals, 6,912 slices), validation (30 individuals, 1,620 slices), and testing
(25 individuals, 1,350 slices) subsets. For training and validation, pseudo-ground truth labels were generated
by averaging three repetitions for T1/T2-weighted contrasts and two repetitions for FLAIR. Test data utilized
higher-SNR labels derived from averaging six repetitions (T1/T2-weighted) and four repetitions (FLAIR),
enabling assessment of model generalization to cleaner data. The dataset (20.7GB) is distributed under
CC-BY-4.0 license.
Simulated Dataset (fastMRI): We utilized single-coil knee data from the fastMRI dataset Zbontar et al.
(2018), selecting patient entries matching those in MINet Feng et al. (2021) to ensure contrast correspondence.
The dataset was structured into training (180 pairs, 6,648 slices), validation (47 pairs, 1,684 slices), and
testing (45 pairs, 1,665 slices) subsets. White Gaussian noise with σ ∈ [13, 25] was introduced to simulate
real-world MRI noise patterns. The dataset (31.5GB) is distributed under MIT license.

A.2 T RAINING PARAMETERS

All experiments were conducted on NVIDIA A6000 GPUs. For the M4Raw dataset, optimization was
performed using Adam with learning rate 1 × 10−4 and weight decay 1 × 10−4 . For the fastMRI dataset,
Adam was configured with learning rate 1 × 10−4 and weight decay 5 × 10−2 . Critical hyperparameters
(learning rate, weight decay, batch size, maximum noise level T ) were optimized based on validation
performance. Early stopping was implemented to prevent overfitting, and final models were selected based on
optimal validation metrics before test set evaluation.

A.3 E VALUATION P ROTOCOL

To assess the efficacy and robustness of the Corruption2Self (C2S) framework, we employed the following
evaluation protocol:
Pseudo-Ground Truth Generation: For the in-vivo dataset (M4Raw), pseudo-ground truth labels were
generated by averaging the multi-repetition images.
Image Quality Metrics: The quality of the denoised images was quantified using two metrics:

• Peak Signal-to-Noise Ratio (PSNR): This metric measures the ratio between the maximum possible
power of a signal and the power of corrupting noise, assessing the quality of the denoised image
against the reference label.
• Structural Similarity Index Measure (SSIM): This metric evaluates the structural similarity
between the denoised image and the reference, focusing on luminance, contrast, and structure.

B T HEORETICAL R ESULTS
Lemma 3. Given the objective function J(θ):
h h h iii
2
J(θ) = EXtdata Et EXt |Xtdata ∥gθ (Xt , t) − Xtdata ∥ ,

14
Published as a conference paper at ICLR 2025

where Xt ∼ N (Xtdata , σ 2 (t)I), the function gθ∗ (Xt , t) that minimizes J(θ) is the conditional expectation:
gθ∗ (Xt , t) = E [Xtdata | Xt ] .

Proof. Our goal is to find gθ∗ (Xt , t) that minimizes the objective function J(θ). Since the outer expectations
over Xtdata and t do not affect the minimization with respect to θ, we focus on minimizing the inner expectation:
h i
2
EXt |Xtdata ∥gθ (Xt , t) − Xtdata ∥ .

For fixed Xt and t, the optimal function gθ∗ (Xt , t) minimizes the expected squared error:
h i
2
min EXtdata |Xt ∥gθ (Xt , t) − Xtdata ∥ .
gθ

According to estimation theory, the function that minimizes this expected squared error is the conditional
expectation of Xtdata given Xt :
gθ∗ (Xt , t) = E [Xtdata | Xt ] .
Therefore, the function gθ∗ (Xt , t) = E [Xtdata | Xt ] minimizes the objective function J(θ).
Theorem 4 (Generalized Denoising Score Matching; restated Theorem 1). Let the following assumptions
hold:
1. Data Distribution: The clean data vector X0 ∈ Rd is distributed according to p0 (x0 ).
2. Noise Level Functions: The noise level or schedule functions σt are strictly positive, scalar-valued
functions of time t, and are monotonically increasing with respect to t.
3. Noisy Observations at Data Noise Level tdata : The observed data Xtdata is given by Xtdata =
X0 + σtdata Ztdata , where Ztdata ∼ N (0, Id ).
4. Target Noise Level ttarget ≤ tdata : The target noisy data Xttarget is defined as Xttarget = X0 +
σttarget Zttarget , where Zttarget ∼ N (0, Id ).
5. Higher Noise Levels t ≥ tdata : For any t ≥ tdata , the noisy data Xt is given by Xt = X0 + σt Zt ,
where Zt ∼ N (0, Id ).
6. Function Class Expressiveness: The neural network class {hθ } is sufficiently expressive, satisfying
the universal approximation property.
Define the objective function:
h i
2
J(θ) = EXtdata ,t,Xt γ(t, σttarget ) hθ (Xt , t) + δ(t, σttarget ) Xt − Xtdata , (12)
where t is uniformly sampled from (tdata , T ] and
σt2 − σt2data σt2data − σt2target
γ(t, σttarget ) = , δ(t, σttarget ) = .
σt2 − σt2target σt2 − σt2target

Given Xtdata and t ≥ tdata , Xt is sampled as:

q
Xt = Xtdata + σt2 − σt2data Z, where Z ∼ N (0, Id ). (13)

Then, the minimizer θ∗ of J(θ) satisfies:

hθ∗ (Xt , t) = E[Xttarget | Xt ], ∀Xt ∈ Rd , t ≥ tdata . (14)

15
Published as a conference paper at ICLR 2025

Proof. Our goal is to find θ∗ that minimizes J(θ). Note that Xt can be expressed in terms of both Xtdata and
Xttarget :
q
Xt = Xtdata + σt2 − σt2data Z1 , Z1 ∼ N (0, Id ),
q
Xt = Xttarget + σt2 − σt2target Z2 , Z2 ∼ N (0, Id ).

Using properties of Gaussian distributions, the score function ∇Xt log pt (Xt ) can be written in two ways:
 
E [Xtdata | Xt ] − Xt E Xttarget | Xt − Xt
∇Xt log pt (Xt ) = = .
σt2 − σt2data σt2 − σt2target
Equating these expressions and rearranging terms:
σt2 − σt2data   σt2data − σt2target
E [Xtdata | Xt ] = E X ttarget | X t + Xt .
σt2 − σt2target σt2 − σt2target

Recognizing the coefficients as γ(t, σttarget ) and δ(t, σttarget ), respectively:

 
E [Xtdata | Xt ] = γ(t, σttarget )E Xttarget | Xt + δ(t, σttarget )Xt .
Define the auxiliary function:
gθ (Xt , t) = γ(t, σttarget )hθ (Xt , t) + δ(t, σttarget )Xt .
Substituting into the loss function J(θ), we have:
h i
2
J(θ) = E ∥gθ (Xt , t) − Xtdata ∥ .

This is a mean squared error (MSE) objective between gθ (Xt , t) and Xtdata .
The objective function becomes:
h i
2
J(θ) = EXtdata Et EXt |Xtdata ∥gθ (Xt , t) − Xtdata ∥ .

By the property of MSE minimization (Lemma 3), the function gθ∗ (Xt , t) that minimizes J(θ) satisfies:
gθ∗ (Xt , t) = E [Xtdata | Xt ] .
Substituting the earlier expression for E [Xtdata | Xt ]:
 
γ(t, σttarget )hθ∗ (Xt , t) + δ(t, σttarget )Xt = γ(t, σttarget )E Xttarget | Xt + δ(t, σttarget )Xt .
Subtracting δ(t, σttarget )Xt from both sides:
 
γ(t, σttarget )hθ∗ (Xt , t) = γ(t, σttarget )E Xttarget | Xt .
Since γ(t, σttarget ) > 0 (due to σt being strictly increasing and t ≥ tdata ), we can divide both sides by
γ(t, σttarget ):  
hθ∗ (Xt , t) = E Xttarget | Xt .
This completes the proof.
Corollary 5 (Reparameterized Generalized Denoising Score Matching; restated Corollary 2). Let the
assumptions of the Generalized Denoising Score Matching theorem 4 hold. Additionally, define:

16
Published as a conference paper at ICLR 2025

1. Reparameterization: For t ≥ tdata , define τ such that

στ2 = σt2 − σt2data , (15)
′ ′
where τ ∈ (0, T ] and T is determined by = σT2 ′ σT2 − σt2data .
Note that given τ and tdata , we can
recover t using the inverse function of σt , denoted as σt−1 :
q
t = σt−1 ( στ2 + σt2data ). (16)
This inverse function exists and is well-defined due to the strictly monotonically increasing property
of σt .

Define the new objective function:

h i
2
J ′ (θ) = EXtdata ,τ,Xt γ ′ (τ, σttarget ) hθ (Xt , t) + δ ′ (τ, σttarget ) Xt − Xtdata , (17)
where τ is uniformly sampled from [0, T ′ ], and the coefficients are defined as:
στ2 σt2data − σt2target
γ ′ (τ, σttarget ) = , δ ′ (τ, σttarget ) = .
στ2 + σt2data − σt2target στ2 + σt2data − σt2target

Given Xtdata and τ , Xt is sampled as:

Xt = Xtdata + στ Z′ , where Z′ ∼ N (0, Id ). (18)

Then, the minimizer θ∗ of J ′ (θ) satisfies:

hθ∗ (Xt , t) = E[Xttarget | Xt ], ∀Xt ∈ Rd , t ≥ tdata . (19)

Proof. Substituting σt2 into γ(t, σttarget ) and δ(t, σttarget ) from Theorem 1, we obtain:
στ2 σt2data − σt2target
γ ′ (τ, σttarget ) = γ(t, σttarget ) = , δ ′ (τ, σttarget ) = δ(t, σttarget ) = .
στ2 + σt2data − σt2target στ2 + σt2data − σt2target
Define gθ (Xt , t) = γ ′ (τ, σttarget )hθ (Xt , t) + δ ′ (τ, σttarget )Xt . The objective function becomes:
h i
2
J ′ (θ) = EXtdata Eτ EXt |Xtdata ∥gθ (Xt , t) − Xtdata ∥ .
By Lemma 3, the function minimizing J ′ (θ) is:
gθ∗ (Xt , t) = E [Xtdata | Xt ] .
From proof of Theorem 1, we have:
E [Xtdata | Xt ] = γ ′ (τ, σttarget )E Xttarget | Xt + δ ′ (τ, σttarget )Xt .
 

Therefore,
gθ∗ (Xt , t) = γ ′ (τ, σttarget )hθ∗ (Xt , t) + δ ′ (τ, σttarget )Xt = γ ′ (τ, σttarget )E Xttarget | Xt + δ ′ (τ, σttarget )Xt .
 

Comparing both expressions, we conclude:

γ ′ (τ, σttarget )hθ∗ (Xt , t) = γ ′ (τ, σttarget )E Xttarget | Xt .
 

Since γ ′ (τ, σttarget ) > 0, we divide both sides by γ ′ (τ, σttarget ):

 
hθ∗ (Xt , t) = E Xttarget | Xt .
This completes the proof.

17
Published as a conference paper at ICLR 2025

B.1 E XTENSION TO VARIANCE P RESERVING C ASE

Our GDSM framework can be naturally extended to the variance preserving (VP) Song et al. (2020); Ho et al.
(2020) case. For example, when σttarget = 0, which corresponds to the VP formulation in ADSM Daras et al.
(2024). In this case, the data model follows:
q
Xtdata = 1 − σt2data X0 + σtdata Z, 0 < σtdata < 1 (20)

Let X0 ∈ Rd represent clean data, and for any σtdata < σt < 1, the forward corrupted Xt is given by:
q
Xt = 1 − σt2 X0 + σt Zt , Zt ∼ N (0, Id ) (21)

Define the objective function:

 
p 2
σt2 q
2 2 1 − σt2
LVP (θ) = EXtdata ,t,Xt  1 − σ tdata hθ (X t , t) − σ tdata 2 Xt − Xtdata  (22)
σt2 − σt2data σt − σt2data

Then, the minimizer θ∗ of JVP (θ) satisfies:

hθ∗ (Xt , t) = E[X0 | Xt ], ∀Xt ∈ Rd , t ≥ tdata . (23)

B.2 C ONNECTION TO N OISIER 2N OISE

We demonstrate that Noisier2Noise Moran et al. (2020) emerges as a special case of our Generalized Denoising
Score Matching (GDSM) framework under specific conditions. Let us establish the correspondence between
notations: in Noisier2Noise, X represents the clean image, Y = X + N represents the noisy observation
with noise N , and Z = Y + M represents the doubly-noisy image with additional synthetic noise M . These
correspond to our formulation where X is X0 , Y is Xtdata , and Z is Xt .
From proof of Theorem 1, when setting σttarget = 0, we obtain:
σt2 − σt2data σ2
E[Xtdata | Xt ] = 2 E[X0 | Xt ] + tdata Xt (24)
σt σt2

In the improved variant of Noisier2Noise, a parameter α controls the magnitude of synthetic noise M relative
to the original noise N . We can establish that this parameter corresponds to our noise schedule through:

σt2 − σt2data
α2 = (25)
σt2data

Under this relationship, Equation equation 24 becomes equivalent to the Noisier2Noise formulation:
α2 1
E[Y |Z] = 2
E[X|Z] + Z (26)
1+α 1 + α2

This equivalence leads to the characteristic Noisier2Noise correction formula:

(1 + α2 )E[Y |Z] − Z
E[X|Z] = (27)
α2

18
Published as a conference paper at ICLR 2025

In the standard case where α = 1, this reduces to E[X|Z] = 2E[Y |Z] − Z, which corresponds to our
framework with σt2 = 2σt2data .
Our GDSM framework offers several advances over Noisier2Noise. First, it provides a continuous noise
schedule through σt , allowing the model to learn from a spectrum of noise levels rather than a fixed ratio
determined by α. Second, it introduces explicit time conditioning in the network architecture, enabling better
adaptation to different noise magnitudes. Third, and perhaps most importantly, it eliminates the need to tune
the α parameter, which according to Moran et al. (2020) is ”difficult or impossible to derive in the absence of
clean validation data.” Instead, our approach automatically learns to handle different noise levels through the
continuous schedule and time conditioning. Furthermore, GDSM extends beyond the clean image prediction
task by supporting arbitrary target noise levels through σttarget , providing a unified framework for various
denoising objectives.

C M ODEL A RCHITECTURE

In this appendix, we provide a comprehensive description of the architectures employed for both single-
contrast and multi-contrast MRI denoising. Our designs build upon the U-Net structure utilized in Denoising
Diffusion Probabilistic Models (DDPM) Ho et al. (2020), incorporating advanced conditioning and attention
mechanisms to enhance performance. For detailed implementation of the Noise Variance Conditioned
Multi-Head Self-Attention (NVC-MSA) module, please refer to Appendix C.2.

C.1 S INGLE -C ONTRAST M ODEL A RCHITECTURE

Our single-contrast denoising model employs a U-Net backbone augmented with time conditioning and the
NVC-MSA module Hatamizadeh et al. (2023).
Time Conditioning: The model adapts its processing based on the noise level t by integrating time em-
beddings into the convolutional layers. This is achieved through adaptive normalization (e.g., instance
normalization followed by an affine transformation conditioned on the time embedding), as introduced in
DDPM.
NVC-MSA Module: To enable the network to adjust to varying noise levels, we incorporate the NVC-MSA
module into the self-attention mechanisms of the U-Net. The module conditions the attention on the current
noise variance, allowing the network to effectively capture long-range dependencies and adapt to different
noise scales. Mathematically, the queries, keys, and values are computed as:
Q = WQ (X) + bQ (t), (28)
K = WK (X) + bK (t), (29)
V = WV (X) + bV (t), (30)
where bQ (t), bK (t), and bV (t) are learned affine transformations of the time embedding. This NVC-MSA
mechanism allows the attention modules to be aware of the noise level and adjust their focus accordingly,
effectively capturing long-range dependencies. The implementation details and pseudo-code for the NVC-
MSA module are provided in Appendix C.2.

C.2 N OISE VARIANCE C ONDITIONED M ULTI -H EAD S ELF -ATTENTION (NVC-MSA) P SEUDO C ODE

Below is the code snippet of the NVC-MSA module, which is integral to both single-contrast and multi-
contrast model architectures. This module conditions the self-attention mechanism on the noise variance
level, enabling the model to adapt its attention based on the current noise level.

19
Published as a conference paper at ICLR 2025

The code snippet encapsulates the core functionality of the NVC-MSA module. The module first normalizes
the input tensor and generates queries, keys, and values for spatial tokens. It then reshapes and projects the
noise variance embeddings using 1x1 convolutions. These noise-conditioned components are added to the
queries, keys, and values before applying the attention mechanism. Finally, the output is rearranged and
projected to produce the final feature map.

20
Published as a conference paper at ICLR 2025

def forward(self, x, noise_emb):

# Get shape of input tensor

b, c, h, w = x.shape
n = h * w

# Normalize the input tensor

x = self.norm(x)

# Generate queries, keys, and values for spatial tokens

qkv = self.to_qkv(x).chunk(3, dim=1)
q, k, v = map(lambda t: rearrange(t, ’b (h d) x y -> b (x y) h d’, h=self.
heads), qkv)

# Reshape and project noise variance embeddings using 1x1 convolutions

noise_emb = noise_emb.view(b, -1, 1, 1)
noise_q = self.noise_query_conv(noise_emb)
noise_k = self.noise_key_conv(noise_emb)
noise_v = self.noise_value_conv(noise_emb)

# Rearrange the projected noise variance embeddings

noise_q = rearrange(noise_q, ’b (h d) x y -> b (x y) h d’, h=self.heads)
noise_k = rearrange(noise_k, ’b (h d) x y -> b (x y) h d’, h=self.heads)
noise_v = rearrange(noise_v, ’b (h d) x y -> b (x y) h d’, h=self.heads)

# Add noise variance-dependent components to queries, keys, and values

q = q + noise_q
k = k + noise_k
v = v + noise_v

# Apply attention mechanism

out = self.attend(q, k, v)

# Rearrange and project the output

out = rearrange(out, ’b (x y) h d -> b (h d) x y’, x=h, y=w)

return self.to_out(out)

Listing 1: Code snippet for NVC-MSA implementation

D M ULTI - CONTRAST C2S

D.1 M ULTI -C ONTRAST M ODEL A RCHITECTURE

The multi-contrast denoising model extends the single-contrast architecture to handle multiple input contrasts,
thereby enhancing denoising performance by leveraging complementary information.
Multicontrast Fusion: The model accepts multiple contrast inputs by concatenating the primary and
complementary contrast images (e.g., (T1, T2)). An initial convolution layer extracts feature embeddings
from the fused contrasts, which are then processed through the U-Net architecture.
NVC-MSA Module: Similar to the single-contrast model, the multi-contrast model integrates the NVC-MSA
module into its self-attention mechanisms. By conditioning the attention on the noise variance level σ, the

21
Published as a conference paper at ICLR 2025

model can adaptively adjust its focus based on the current noise level, as detailed in the pseudo-code provided
in Appendix C.2.
Output Head: Following the U-Net processing, the output head generates a single-channel image representing
the denoised primary contrast image.
Flexibility and Extensions: The architecture can dynamically adjust to accommodate any number of input
contrasts by modifying the input layer accordingly. Although our current implementation is based on U-Net,
the NVC-MSA mechanism is compatible with other architectures, such as Vision Transformers Dosovitskiy
et al. (2020); Liu et al. (2021), where it can replace standard multi-head self-attention modules to enhance
model complexity and performance. This extension remains an avenue for future research.

D.2 M ULTI -C ONTRAST C2S A LGORITHM

The multi-contrast C2S framework extends the single-contrast algorithm to leverage complementary informa-
tion from auxiliary contrast images. Given a collection of noisy multi-contrast training images

{(Xtdata ,i , Ci )}N
i=1 ,

where Xtdata ,i ∈ Rd is the noisy target contrast image and Ci ∈ Rd×c represents c auxiliary noisy contrast
images, our goal is to estimate the clean target contrast image X0 using both Xtdata and C. In the multi-
contrast setting, we focus on the case where ttarget = 0 (i.e., σttarget = 0), aiming to directly estimate the
MMSE estimator E[X0 | Xtdata , C]. While the single-contrast C2S incorporates a detail refinement extension
with a non-zero target noise level, we leave the exploration of such extensions and more advanced contrast
fusion architectures for multi-contrast C2S as future work. For the implementation of the denoising function
Dθ (Xτ , τ | C), the conditioning on auxiliary contrasts C is achieved through a CNN encoder architecture
that extracts features from each auxiliary contrast image. These extracted features are then concatenated
with the features from the target contrast image in the feature space, allowing the model to effectively
integrate complementary information from all available contrasts. The concatenated features are subsequently
processed through the U-Net backbone with NVC-MSA modules, as in the single-contrast case. Following the
reparameterization strategy introduced in Section 3.1, we define a reparameterized function Dθ (Xτ , τ | C)
as:

Dθ (Xτ , τ | C) = λout (τ, σttarget ) hθ (Xt , t | C) + λskip (τ, σttarget ) Xt , (31)

where Xt = Xtdata + στ Z, with Z ∼ N (0, Id ), and the coefficients remain consistent with the single-contrast
case:

στ2 σt2data − σt2target

λout (τ, σttarget ) = , λskip (τ, σttarget ) = (32)
στ2 + σt2data − σt2target στ2 + σt2data − σt2target

The multi-contrast C2S loss function is then formulated as:

1 h
2
i
LMC-C2S (θ) = EXtdata ∼ptdata (x) w(τ ) ∥Dθ (Xτ , τ | C) − Xtdata ∥2 (33)
2 C∼p(c)
τ ∼U [0,T ]
Z∼N (0,Id )

The complete training procedure for multi-contrast C2S is outlined in Algorithm 2.

22
Published as a conference paper at ICLR 2025

Algorithm 2 Multi-Contrast Corruption2Self Training Procedure

Require: Noisy multi-contrast dataset {(Xitdata , Ci )}N
i=1 , hθ , max noise level T , batch size m, total iterations K, noise
schedule function στ
1: for k = 1 to K do
2: Sample minibatch {(Xitdata , Ci )}m
i=1 , τ ∼ U (0, T ], Z ∼ N (0, Id )
σ2 σt2
3: Compute λout (τ, 0) = σ2 +στ2 , λskip (τ, 0) = σ2 +σdata2
q τ tdata  τ tdata

4: Recover t = σt−1 στ2 + σt2data

5: Compute Xt ← Xtdata + στ Z
1
Pm i i i i 2
6: Compute loss: L = 2m i=1 w(τ ) λout (τ, 0) hθ (Xt , t | C ) + λskip (τ, 0) Xt − Xtdata
7: Update θ using Adam optimizer to minimize L

During inference, given a noisy target contrast observation Xtdata and auxiliary contrasts C, the denoised
output is obtained by:

X̂ = hθ∗ (Xtdata , tdata | C) (34)

where the trained model hθ∗ approximates E[X0 | Xtdata , C], providing a clean estimate of the target contrast
image that benefits from the complementary information in the auxiliary contrasts.

E A DDITIONAL R ESULTS ON FAST MRI

In this section, we present additional results on the fastMRI dataset, evaluating the performance of various
denoising methods across different noise levels and contrasts.
Table 6 summarizes the performance comparison between the baseline method (Without Reparameterization)
and our proposed Corruption2Self (C2S) approach, under four configurations: PD with σ = 13/255, PDFS
with σ = 13/255, PD with σ = 25/255, and PDFS with σ = 25/255.

Method PD, σ = 13/255 PDFS, σ = 13/255 PD, σ = 25/255 PDFS, σ = 25/255

PSNR ↑ / SSIM ↑ PSNR ↑ / SSIM ↑ PSNR ↑ / SSIM ↑ PSNR ↑ / SSIM ↑
Without Reparam. 32.65 / 0.821 30.08 / 0.676 30.16 / 0.747 28.24 / 0.570
With Reparam. 33.48 / 0.832 30.67 / 0.761 30.91 / 0.756 28.62 / 0.601

Table 6: Impact of reparameterization of noise levels on the fastMRI dataset. The ”Without Reparam.”
row contains estimated baseline results, while ”With Reparam.” represents the proposed method with
reparameterization.

F A DDITIONAL R ESULTS ON M4R AW

F.1 V ISUAL C OMPARISON OF D ENOISING M ETHODS

In this section, we provide a visual comparison of several denoising methods applied to T1, T2, and FLAIR
contrast images in the M4Raw dataset. The denoising methods evaluated include Noise2Noise, BM3D,
SwinIR, R2R, Noise2Self, and C2S, with Multi-repetition Averaged Label serving as the ground truth
reference for comparison.

23
Published as a conference paper at ICLR 2025

Figure 5: Comparison of different denoising methods for PD contrast (noise level 13/255) in fastMRI.

Figure 6: Comparison of different denoising methods for PDFS contrast (noise level 25/255) in fastMRI.

24
Published as a conference paper at ICLR 2025

Figure 7: Comparison of different denoising methods for T1 contrast in M4Raw. The top row shows the
original noisy image and results from Noise2Noise, BM3D, and SwinIR. The bottom row includes the
multi-repetition averaged label, R2R, Noise2Self, and C2S methods. A zoomed-in section of each image is
presented below each corresponding brain image for detailed comparison.

Figure 8: Top: Comparison of different denoising methods for T1 contrast in M4Raw. Bottom: Comparison
of different denoising methods for T2 contrast in M4Raw.

25
Published as a conference paper at ICLR 2025

Figure 9: Comparison of different denoising methods for FLAIR contrast in M4Raw.

Figure 10: Comparison of different denoising methods for FLAIR contrast in M4Raw.

F.2 E VALUATION ON M ATCHING -SNR T EST L ABELS

When evaluated on test data matching the SNR of the training data, supervised methods like SwinIR and
Restormer achieve the best performance, with PSNR improvements over self-supervised methods. This
indicates that supervised approaches excel when the training and testing conditions are similar. However, the
performance gap narrows when considering C2S, which attains PSNR and SSIM values comparable to those
of the supervised methods.
These observations highlight a limitation of supervised methods: they may not generalize well to scenarios
where the test data has different noise characteristics or higher SNR than the training data. In contrast,
C2S demonstrates robust performance across different SNR levels, achieving competitive results on both
matching-SNR and higher-SNR test labels. This suggests that our self-supervised approach is more resilient
to variations in noise levels and better generalizes to cleaner images without relying on clean ground truth
during training.

26
Published as a conference paper at ICLR 2025

Methods T1 T2 FLAIR
PSNR / SSIM ↑ PSNR / SSIM ↑ PSNR / SSIM ↑
Classical Non-Learning-Based Methods
NLM Froment (2014) 34.65 / 0.897 33.72 / 0.873 32.83 / 0.830
BM3D Mäkinen et al. (2020) 35.27 / 0.900 34.01 / 0.875 33.22 / 0.841
Supervised Learning Methods
SwinIR Liang et al. (2021) 36.09 / 0.926 34.57 / 0.902 34.34 / 0.909
Restormer Zamir et al. (2022) 35.96 / 0.926 34.13 / 0.898 34.21 / 0.908
Noise2Noise Lehtinen et al. (2018) 34.82 / 0.892 33.92 / 0.861 33.73 / 0.879
Self-Supervised Single-Contrast Methods
Noise2Void Krull et al. (2019) 32.83 / 0.870 31.73 / 0.857 30.90 / 0.821
Noise2Self Batson & Royer (2019) 34.17 / 0.883 32.64 / 0.847 31.96 / 0.823
Recorrupted2Recorrupted Pang et al. (2021) 33.60 / 0.801 32.93 / 0.820 32.42 / 0.794
C2S (Ours) 36.11 / 0.925 34.87 / 0.904 34.15 / 0.898

Table 7: Quantitative results on the M4Raw dataset evaluated on three-repetition-averaged test labels
(matching training and validation SNR). Mean PSNR and SSIM metrics are reported. The best results
among self-supervised methods are in bold.

G D ETAIL R EFINEMENT A LGORITHM

In this appendix, we present the Detail Refinement algorithm as an extension to the primary Corruption2Self
(C2S) training procedure. While the primary stage focuses on learning the conditional expectation E[X0 | Xt ]
for maximum denoising, this aggressive approach may lead to the loss of fine details and important features.
The Detail Refinement stage addresses this issue by training the network to predict E[Xttarget | Xt ] with a non-
zero target noise level σttarget > 0, allowing the preservation of intricate structures and textures. Empirically,
we discovered that uniformly sampling ttarget from the interval [0, tdata ] during training already yields good
results and leave more advanced tuning process to future work. This sampling strategy provides a good
balance between noise reduction and detail preservation, allowing the network to learn a range of refinement
levels adaptively.

G.1 L OSS F UNCTION FOR D ETAIL R EFINEMENT

The loss function for the Detail Refinement stage is similar to the primary C2S stage but introduces a target
noise level σttarget . The denoised output Dθ (Xτ , τ, σttarget ) is defined as:
Dθ (Xτ , τ, σttarget ) = λout (τ, σttarget ) hθ (Xt , t) + λskip (τ, σttarget ) Xt ,
where hθ (Xt , t) is the denoising network parameterized by θ, with input Xt and noise level t, and
λout (τ, σttarget ) and λskip (τ, σttarget ) are blending factors between the network output and the noisy input.
The loss function is given by:
1 h
2
i
Lrefine (θ) = E Xtdata ∼ptdata (x) w(τ ) Dθ (Xτ , τ, σttarget ) − Xtdata 2
, (35)
2 τ ∼U [0,T ]
σttarget ∼U (0,σtdata ]
Z∼N (0,Id )

where:

27
Published as a conference paper at ICLR 2025

• Xtdata ∼ ptdata (x) is the noisy observed data.

• τ ∼ U[0, T ] is the reparameterized noise level uniformly sampled from [0, T ], where T is the
maximum noise level.
• σttarget ∼ U(0, σtdata ] is the target noise level, sampled uniformly from the interval (0, σtdata ].
• Xt = Xtdata + στ Z, where Z ∼ N (0, Id ) is Gaussian noise.
• λout (τ, σttarget ) and λskip (τ, σttarget ) are defined as:

στ2 σt2data − σt2target

λout (τ, σttarget ) = , λskip (τ, σttarget ) = . (36)
στ2 + σt2data − σt2target στ2 + σt2data − σt2target

The goal of this refinement stage is to retain a controlled amount of noise, preventing the loss of fine details
and features while still performing denoising.

G.2 A LGORITHM I MPLEMENTATION

The Detail Refinement training procedure minimizes the loss function Lrefine (θ) over the network parameters
θ. The steps are as follows:

Algorithm 3 Detail Refinement Training Procedure

Require: Noisy dataset {Xitdata }N i=1 , denoising network hθ , max noise level T , batch size m, total iterations Krefine ,
noise schedule function στ
1: for k = 1 to Krefine do
2: Sample minibatch {Xitdata }m i=1 , τ ∼ U (0, T ], Z ∼ N (0, Id )
3: Sample σttarget ∼ U (0, σtdata ]
2
στ
4: Compute λout (τ, σttarget ) = 2 +σ 2
στ −σt2
tdata target
σt2 −σt2
data target
5: Compute λskip (τ, σttarget ) = 2 +σ 2 2
στ tdata −σttarget
q 
6: Recover t = σt−1 στ2 + σt2data
7: Compute Xt ← Xtdata + στ Z
1
Pm i 2
8: Compute loss: L = 2m i=1 w(τ ) Dθ (Xτ , τ, σttarget ) − Xtdata
9: Update θ using Adam optimizer Kingma (2014) to minimize L

H ROBUSTNESS TO N OISE L EVEL E STIMATION E RROR

In practical applications, precise noise level estimation can be challenging, making robustness to estimation
errors a crucial property for denoising models. This section provides a comprehensive analysis of our model’s
performance under various degrees of noise level misestimation, demonstrating its capability to function
effectively as a blind denoising model.

H.1 E XPERIMENTAL S ETUP AND R ESULTS

We evaluate our pretrained model’s robustness by systematically varying the input noise level estimation
tdata from -50% (underestimation) to +50% (overestimation) of the true noise level. Table 8 presents the
quantitative results across T1, T2, and FLAIR contrasts on the M4Raw test set, while Figure 11 visualizes
these trends.

28
Published as a conference paper at ICLR 2025

Noise Level M4Raw Dataset (PSNR / SSIM ↑)

Estimation T1 T2 FLAIR
-50% 32.6004 / 0.9154 32.3210 / 0.8890 32.9496 / 0.8757
-40% 32.5958 / 0.9153 32.3125 / 0.8888 32.9498 / 0.8759
-30% 32.5910 / 0.9151 32.3044 / 0.8886 32.9478 / 0.8761
-20% 32.5861 / 0.9150 32.2964 / 0.8884 32.9436 / 0.8762
-10% 32.5810 / 0.9149 32.2882 / 0.8881 32.9373 / 0.8763
0% 32.5758 / 0.9148 32.2812 / 0.8879 32.9297 / 0.8764
+10% 32.5704 / 0.9147 32.2700 / 0.8876 32.9192 / 0.8763
+20% 32.5649 / 0.9146 32.2593 / 0.8873 32.9076 / 0.8763
+30% 32.5592 / 0.9145 32.2468 / 0.8870 32.8947 / 0.8762
+40% 32.5535 / 0.9143 32.2322 / 0.8866 32.8805 / 0.8760
+50% 32.5475 / 0.9142 32.2151 / 0.8862 32.8655 / 0.8759

Table 8: Performance analysis under varying noise level estimations on the M4Raw dataset. The model
demonstrates remarkable stability across all contrasts, with particularly strong performance under slight
underestimation. The ’-’ and ’+’ indicate underestimation and overestimation of the noise level, respectively.

Figure 11: Visualization of model performance under varying noise level estimations. The plots demonstrate
consistent stability across all contrasts, with minimal performance degradation even under significant estima-
tion errors (±50%).

H.2 A NALYSIS AND D ISCUSSION

Our experiments reveal several key findings: Overall Stability: The model maintains remarkably stable
performance across all tested estimation errors, with maximum PSNR variations of only 0.053 dB, 0.106 dB,
and 0.084 dB for T1, T2, and FLAIR contrasts, respectively. Asymmetric Response: Interestingly, the
model shows slightly better performance under noise level underestimation compared to overestimation. For
instance, with T1 contrast, a 50% underestimation achieves a PSNR of 32.6004 dB, outperforming both
the true noise level (32.5758 dB) and 50% overestimation (32.5475 dB). While all contrasts exhibit robust
performance, the degree of stability varies. T1 shows the most stable response, while T2 demonstrates slightly
higher sensitivity to estimation errors.

29
Published as a conference paper at ICLR 2025

M4Raw T1 fastMRI PD25

T
PSNR ↑ SSIM ↑ PSNR ↑ SSIM ↑
3 33.30 0.869 30.58 0.740
5 33.91 0.879 31.01 0.765
10 34.91 0.890 30.41 0.751
15 34.35 0.882 30.42 0.743
20 34.43 0.882 30.52 0.755

Table 9: Influence of maximum corruption level T on the M4Raw validation dataset.

H.3 P RACTICAL I MPLEMENTATION

In our implementation, we utilize the skimage package Van der Walt et al. (2014) for noise level estimation,
which proves sufficient for optimal performance. The model’s demonstrated robustness suggests that even
relatively simple estimation techniques can provide adequate noise level approximations for effective denois-
ing. Our findings reveal significant practical advantages: the model readily adapts to real-world scenarios
where exact noise levels are unknown, while standard noise estimation tools consistently deliver near-optimal
performance. Furthermore, the observed slight preference for underestimation indicates that conservative
noise level estimates may be advantageous in practice.
While future work could explore more sophisticated noise estimation techniques, particularly for extreme
cases, our current results demonstrate that the model’s inherent robustness already makes it highly practical
for real-world applications. This robustness, combined with the effectiveness of standard noise estimation
tools, enables our approach to function reliably as a blind denoising model, requiring minimal assumptions
about the underlying noise characteristics.

I A DDITIONAL A BLATION S TUDIES

I.1 E FFECTIVENESS OF M AXIMUM CORRUPTION LEVEL T

We also analyzed the impact of the maximum corruption level T . All models were trained for 300 epochs
with the same hyperparameters. As shown in Table 9, performance generally improves with higher T , peaking
at T = 10 for the M4Raw dataset (T1 contrast). For the fastMRI dataset (PD contrast, 25/255 noise level),
the best performance occurs at T = 5. These results indicate that while increasing T generally benefits
performance, excessively high corruption levels may not lead to further improvements within the given
training budget and could require longer training times to converge.

I.2 E FFECTIVENESS OF R EPARAMETERIZATION AND EMA ON T RAINING DYNAMICS

In our framework, we leverage a reparameterization strategy to address the challenges associated with noise
level sampling. The theoretical foundation of GDSM (see Theorem 1) requires sampling noise levels t > tdata .
However, directly sampling t ∼ U(tdata , T ] can be sensitive to errors in noise level estimation. To mitigate
this, we instead sample an auxiliary variable τ ∼ U(0, T ] and define the corresponding variance via

στ2 = σt2 − σt2data .

We now describe two key perspectives that elucidate why this reparameterization improves the training
process:

30
Published as a conference paper at ICLR 2025

Stable Noise Level Sampling. By recovering the original noise level through the transformation
q 
−1 2 2
t = σt στ + σtdata ,

the additive term στ2 ensures that the recovered t remains valid even when σtdata is underestimated. This
robustness contrasts with the direct sampling approach, where inaccuracies in estimating tdata would directly
affect the noise level range.

Consistent Coverage Across Data Samples. Sampling τ from a fixed interval (0, T ] guarantees that
the entire noise level range is consistently covered during training, irrespective of individual sample noise
characteristics. Although we approximate T ≈ T ′ for practical implementation, the fixed-range sampling of
τ avoids the variability inherent in the direct sampling of t (which would otherwise depend on each sample’s
noise level). This consistency is particularly beneficial for datasets with heterogeneous noise levels, leading
to smoother convergence and more stable training dynamics.
In addition to reparameterization, we apply an Exponential Moving Average (EMA) with a decay rate of
0.999 during training. The combination of these techniques not only enhances stability but also accelerates
convergence. Figure 12 presents a comparative analysis of the PSNR and SSIM metrics on the M4Raw
FLAIR dataset over the first 125 training epochs (T = 5 and σtarget = 0). The results demonstrate that the
joint application of reparameterization and EMA stabilizes the training dynamics.

Figure 12: Effectiveness of Reparameterization in Noise Level Adjustment on the M4Raw FLAIR validation
Dataset. Comparison of PSNR and SSIM metrics on the validation set for different model configurations over
the first 125 training epochs (T = 5, σtarget = 0). The combination of reparameterization and EMA (0.999)
yields more stable training dynamics and improved convergence.

J G UIDELINES FOR S ELECTING AND U SING M AXIMUM C ORRUPTION L EVEL

J.1 T HEORETICAL B OUNDS FOR M AXIMUM C ORRUPTION L EVEL

Following insights from score-based generative models Song et al. (2020), the maximum corruption level T
can be theoretically chosen as large as the maximum Euclidean distance between all pairs of training data
points to ensure sufficient coverage for accurate score estimation. Our analysis of MRI datasets reveals the
following maximum pairwise distances:

31
Published as a conference paper at ICLR 2025

• M4Raw Dataset:
– T1 contrast: 56.72
– T2 contrast: 47.99
– FLAIR contrast: 65.20
• FastMRI Dataset:
– PD (σ = 13/255): 114.23
– PDFS (σ = 13/255): 102.59
– PD (σ = 25/255): 115.53
– PDFS (σ = 25/255): 102.94

J.2 P RACTICAL C ONSIDERATIONS AND T RAINING DYNAMICS

While theoretical bounds provide upper limits, our empirical studies show that significantly smaller values of
T can achieve optimal performance while maintaining computational efficiency. Table 10 demonstrates the
relationship between T and training convergence:

Max Corruption Level (T ) Epochs to Converge

20 204
15 183
10 125
5 79
3 42

Table 10: Relationship between maximum corruption level and training convergence on M4Raw dataset (T1).

J.3 A LTERNATIVE : VARIANCE P RESERVING (VP) F ORMULATION

An alternative to the variance exploding (VE) formulation is the variance preserving (VP) formulation
(Appendix B.1). Here, σt is sampled uniformly from (σtdata , 1), eliminating the need to estimate T . This
approach avoids explicit selection of T , as the maximum corruption level is bounded by 1.
The corruption process in this formulation is given by:
q
Xtdata = 1 − σt2data X0 + σtdata Z, 0 < σtdata < 1, (37)
with additional noise levels sampled such that σtdata < σt < 1. This principled approach maintains theoretical
guarantees and achieves comparable performance to VE in our experiments.

J.4 P RACTICAL G UIDELINES

Based on our analysis, we recommend starting with T = 5 for datasets exhibiting moderate noise levels, as
this provides an effective baseline for most applications. For datasets with higher noise levels or more complex
noise patterns, gradually increasing T up to 20 may yield improved results, though practitioners should
note that training time increases approximately linearly with T . Throughout the training process, careful
monitoring of validation metrics is essential to determine optimal stopping points and assess the effectiveness
of the chosen corruption level. In cases where dataset characteristics make the selection of T particularly
challenging, the VP formulation offers a principled alternative that maintains theoretical guarantees while
eliminating the need for explicit T selection.

32