Masked Denoising Score-Matching (MDSM)

• Masked Denoising Score-Matching is a technique that uses masks to restrict supervision, enforcing J-invariance and mitigating trivial identity mappings.
• It leverages a partial corruption kernel where losses are computed only on masked coordinates, linking score-based objectives with selective denoising.
• MDSM bridges concepts from DSM and Soft Diffusion, offering practical benefits in self-supervised denoising while posing challenges in sampling speed and artifact control.

Masked Denoising Score-Matching (MDSM) denotes denoising score-matching objectives in which supervision is restricted by a mask, typically so that a model predicts a denoised value or a score at masked locations from unmasked context. In blind-spot and J-invariant denoising, the mask prevents trivial identity mapping; in more formal score-based treatments, masking appears either as an explicit elementwise weighting of the loss or as a corruption operator or Markov kernel that acts only on selected coordinates (Tu et al., 8 May 2025, Daras et al., 2022, Benton et al., 2022). A common source of confusion is that the acronym MDSM explicitly expands to “Multi-scale Denoising Score Matching” in a different energy-based modeling paper; that work concerns multi-scale Gaussian corruption, and masked corruption variants are not part of that work (Li et al., 2019).

1. Terminology and core definition

In the MRI denoising literature discussed by Corruption2Self (C2S), masked or blind-spot denoising strategies from the Noise2Void/Noise2Self family enforce J-invariance by masking pixels or voxels so the model cannot trivially copy the input. In DSM parlance, this corresponds to combining score or denoiser regression with a mask and computing the loss only on masked positions, using the unmasked neighborhood as context (Tu et al., 8 May 2025).

A prototypical masked denoiser objective under AWGN is

$$\mathcal{L}_{\mathrm{MDSM}}(\theta)=\mathbb{E}_{\mathbf{y},\,\sigma,\,\boldsymbol{\varepsilon},\,\mathbf{m}} \big\|\,\mathbf{m}\odot\big[\mathbf{h}_\theta(\mathbf{y}+\sigma\boldsymbol{\varepsilon},\sigma)-\mathbf{y}\big]\big\|_2^2,$$

with $\mathbf{m}\in\{0,1\}^d$ selecting target locations. An equivalent score-regression form is

$$\mathcal{L}_{\mathrm{MDSM}^{\mathrm{score}}}(\theta)=\mathbb{E}\big\|\mathbf{m}\odot\big[\mathbf{s}_\theta(\mathbf{y}+\sigma\boldsymbol{\varepsilon},\sigma)+\boldsymbol{\varepsilon}/\sigma\big]\big\|_2^2,$$

together with the blind-spot requirement that the target pixel is not observed by the receptive field (Tu et al., 8 May 2025).

Usage of “MDSM”	Setting	Source
Masked Denoising Score-Matching	Blind-spot or masking-based score/denoiser regression	(Tu et al., 8 May 2025, Daras et al., 2022, Benton et al., 2022)
Multi-scale Denoising Score Matching	Multi-scale Gaussian corruption in EBMs; masked corruption variants are not part of this work	(Li et al., 2019)

This terminological split matters because the masked and multi-scale usages share denoising score-matching ancestry but correspond to different corruption models, objectives, and sampling procedures.

2. Score-matching interpretation of masking

Within the Denoising Markov Model framework, masking can be written as a corruption kernel on $\mathbb{R}^d$. Let $M\in\{0,1\}^d$ be a random mask, let $\epsilon\sim\mathcal{N}(0,\sigma^2 I_d)$, and define

$y = x + M \odot \epsilon.$

Equivalently, coordinates with $M_j=1$ receive Gaussian corruption, while coordinates with $M_j=0$ are left unchanged. The corresponding kernel is

$$K_{M,\sigma}(y|x) = \prod_{j: M_j=1} \mathcal{N}(y_j; x_j, \sigma^2)\cdot \prod_{j: M_j=0}\delta(y_j-x_j),$$

and its gradient with respect to $\mathbf{m}\in\{0,1\}^d$0 on masked coordinates is

$\mathbf{m}\in\{0,1\}^d$1

The resulting masked DSM loss is

$\mathbf{m}\in\{0,1\}^d$2

Under the standard DSM assumptions stated in the DMM treatment, the minimizer satisfies

$\mathbf{m}\in\{0,1\}^d$3

so the masked objective still targets the score of a corrupted marginal distribution rather than merely a heuristic reconstruction map (Benton et al., 2022).

The same framework also yields a Fisher-divergence interpretation. If $\mathbf{m}\in\{0,1\}^d$4, then for each $\mathbf{m}\in\{0,1\}^d$5 the objective equals the Fisher divergence between $\mathbf{m}\in\{0,1\}^d$6 and $\mathbf{m}\in\{0,1\}^d$7, up to constants (Benton et al., 2022). This places MDSM within the same score-matching lineage as Vincent-style DSM, but on a masked subspace.

A notable implication is that masking is not merely a training trick. In the DMM formulation it is a partial corruption Markov transition, and the denoising target is the logarithmic gradient of the corresponding corrupted marginal. This suggests that properly formulated MDSM inherits the consistency properties of DSM on the coordinates where the corruption is smooth.

3. Linear-corruption formulations and measurement-space objectives

Soft Diffusion generalizes DSM to arbitrary linear corruption operators $\mathbf{m}\in\{0,1\}^d$8 through

$\mathbf{m}\in\{0,1\}^d$9

with target score

$$\mathcal{L}_{\mathrm{MDSM}^{\mathrm{score}}}(\theta)=\mathbb{E}\big\|\mathbf{m}\odot\big[\mathbf{s}_\theta(\mathbf{y}+\sigma\boldsymbol{\varepsilon},\sigma)+\boldsymbol{\varepsilon}/\sigma\big]\big\|_2^2,$$0

The generalized Tweedie relation in measurement space is

$$\mathcal{L}_{\mathrm{MDSM}^{\mathrm{score}}}(\theta)=\mathbb{E}\big\|\mathbf{m}\odot\big[\mathbf{s}_\theta(\mathbf{y}+\sigma\boldsymbol{\varepsilon},\sigma)+\boldsymbol{\varepsilon}/\sigma\big]\big\|_2^2,$$1

Specializing to masking sets $$\mathcal{L}_{\mathrm{MDSM}^{\mathrm{score}}}(\theta)=\mathbb{E}\big\|\mathbf{m}\odot\big[\mathbf{s}_\theta(\mathbf{y}+\sigma\boldsymbol{\varepsilon},\sigma)+\boldsymbol{\varepsilon}/\sigma\big]\big\|_2^2,$$2, where $$\mathcal{L}_{\mathrm{MDSM}^{\mathrm{score}}}(\theta)=\mathbb{E}\big\|\mathbf{m}\odot\big[\mathbf{s}_\theta(\mathbf{y}+\sigma\boldsymbol{\varepsilon},\sigma)+\boldsymbol{\varepsilon}/\sigma\big]\big\|_2^2,$$3 and $$\mathcal{L}_{\mathrm{MDSM}^{\mathrm{score}}}(\theta)=\mathbb{E}\big\|\mathbf{m}\odot\big[\mathbf{s}_\theta(\mathbf{y}+\sigma\boldsymbol{\varepsilon},\sigma)+\boldsymbol{\varepsilon}/\sigma\big]\big\|_2^2,$$4 indicate observed or unobserved dimensions. The corruption becomes

$$\mathcal{L}_{\mathrm{MDSM}^{\mathrm{score}}}(\theta)=\mathbb{E}\big\|\mathbf{m}\odot\big[\mathbf{s}_\theta(\mathbf{y}+\sigma\boldsymbol{\varepsilon},\sigma)+\boldsymbol{\varepsilon}/\sigma\big]\big\|_2^2,$$5

and the Soft Score Matching loss reduces to

$$\mathcal{L}_{\mathrm{MDSM}^{\mathrm{score}}}(\theta)=\mathbb{E}\big\|\mathbf{m}\odot\big[\mathbf{s}_\theta(\mathbf{y}+\sigma\boldsymbol{\varepsilon},\sigma)+\boldsymbol{\varepsilon}/\sigma\big]\big\|_2^2,$$6

Because $$\mathcal{L}_{\mathrm{MDSM}^{\mathrm{score}}}(\theta)=\mathbb{E}\big\|\mathbf{m}\odot\big[\mathbf{s}_\theta(\mathbf{y}+\sigma\boldsymbol{\varepsilon},\sigma)+\boldsymbol{\varepsilon}/\sigma\big]\big\|_2^2,$$7 is diagonal with zeros on unobserved entries, gradients from the loss propagate only through observed dimensions; the network is incentivized to match measurements only where $$\mathcal{L}_{\mathrm{MDSM}^{\mathrm{score}}}(\theta)=\mathbb{E}\big\|\mathbf{m}\odot\big[\mathbf{s}_\theta(\mathbf{y}+\sigma\boldsymbol{\varepsilon},\sigma)+\boldsymbol{\varepsilon}/\sigma\big]\big\|_2^2,$$8. The paper identifies this as the essence of masked denoising score matching (Daras et al., 2022).

This formulation clarifies an important distinction. In the Soft Diffusion parameterization, the loss is applied in measurement space after the corruption operator acts on the clean-image prediction. In the blind-spot formulation used for self-supervised denoising, the mask is written directly as an elementwise weighting of score or denoiser errors (Tu et al., 8 May 2025). Both fit within DSM, but they expose different aspects of the same principle: only part of the observation should carry direct supervision at each training event.

The masking conventions are not identical across papers. In Soft Diffusion, $$\mathcal{L}_{\mathrm{MDSM}^{\mathrm{score}}}(\theta)=\mathbb{E}\big\|\mathbf{m}\odot\big[\mathbf{s}_\theta(\mathbf{y}+\sigma\boldsymbol{\varepsilon},\sigma)+\boldsymbol{\varepsilon}/\sigma\big]\big\|_2^2,$$9 denotes an observed dimension in $\mathbb{R}^d$0 (Daras et al., 2022), whereas in the DMM specialization $\mathbb{R}^d$1 selects a coordinate that receives Gaussian corruption in $\mathbb{R}^d$2 (Benton et al., 2022). The shared structure is partial corruption plus score learning on the affected coordinates.

4. Denoising identities, sampling, and reverse dynamics

For Gaussian masking, the DMM treatment gives a masked Tweedie identity:

$\mathbb{R}^d$3

and, more specifically, $\mathbb{R}^d$4 for $\mathbb{R}^d$5 (Benton et al., 2022). Thus a masked score directly yields a conditional denoiser on masked coordinates.

Sampling can be performed with annealed Langevin dynamics on masked coordinates:

$\mathbb{R}^d$6

optionally followed by the Tweedie denoising step

$\mathbb{R}^d$7

The DMM framework also allows reverse SDE formulations with time-indexed masks and noise levels, provided the masking schedule is ergodic over coordinates (Benton et al., 2022).

Soft Diffusion proposes a more structured measurement-space reverse update, the Momentum Sampler. Under masking,

$\mathbb{R}^d$8

The first bracket is the “de-masking” momentum term, the second is the denoising term, and the last adjusts the noise level (Daras et al., 2022).

A practical consequence is that masked score-based sampling is often slower than Gaussian diffusion or blur-based linear corruption. Soft Diffusion reports that, under masking on CelebA-64, FID is approximately $\mathbb{R}^d$9 at $M\in\{0,1\}^d$0 steps and approximately $M\in\{0,1\}^d$1 at $M\in\{0,1\}^d$2 steps, and attributes this to the fact that the network is only penalized on observed region at each $M\in\{0,1\}^d$3, so expanding the mask to reveal new pixels requires many small steps for smooth transitions (Daras et al., 2022).

5. Relation to generalized DSM and self-supervised MRI denoising

C2S introduces generalized denoising score matching (GDSM) for MRI denoising from noisy-only data. Its key contrast with MDSM is mechanistic. GDSM avoids explicit masking and instead enforces nontrivial supervision by “further corruption”: the input $M\in\{0,1\}^d$4 is stochastically different from the supervisory signal $M\in\{0,1\}^d$5. MDSM enforces nontrivial supervision by hiding pixels and training only on masked locations (Tu et al., 8 May 2025).

The C2S analysis states that MDSM restricts receptive field and can introduce checkerboard artifacts or oversmoothing, whereas GDSM/C2S preserves full context and optimizes an exact conditional-expectation target,

$M\in\{0,1\}^d$6

which yields an unbiased MMSE target under Gaussianity and any $M\in\{0,1\}^d$7 (Tu et al., 8 May 2025). This is presented as the main reason blind-spot self-supervision and generalized DSM behave differently in MRI.

The same paper also proposes an explicit hybrid:

$M\in\{0,1\}^d$8

where the C2S predictor is combined with a blind-spot mask. The stated purpose is to keep the unbiased conditional-expectation target while reducing identity risk even further; this may help when the network has very large receptive fields (Tu et al., 8 May 2025).

In practical terms, the C2S synthesis recommends masked variants when distributional assumptions are weak or when strict J-invariance is desired, and recommends C2S when additive Gaussian-like noise is a reasonable assumption and full-field context, multi-level conditioning, and MMSE targets are important (Tu et al., 8 May 2025). This does not invalidate MDSM; rather, it locates masking among several ways of preventing trivial self-supervision.

6. Empirical behavior, limitations, and recurrent misconceptions

The empirical record summarized in the cited works assigns MDSM both clear advantages and recurring liabilities. Its reported advantages are that it does not need to estimate the noise distribution explicitly, enforces J-invariance, and is widely applicable (Tu et al., 8 May 2025). In the DMM formulation, informative gradients are obtained when every coordinate is masked with positive probability and $M\in\{0,1\}^d$9 remains positive; for iterative masked denoising, irreducibility requires masking schedules that eventually act on all coordinates (Benton et al., 2022).

The main limitations recur across formulations. In the MRI comparison, blind-spot methods such as Noise2Void and Noise2Self oversmooth fine structures and underperform C2S; MDSM-style masking is described as reducing context, increasing variance, and being susceptible to masked-sampling artifacts (Tu et al., 8 May 2025). In Soft Diffusion, masking works but samples more slowly and with worse FID than blur, reflecting the fact that the loss only penalizes observed coordinates and the observed region expands progressively as $\epsilon\sim\mathcal{N}(0,\sigma^2 I_d)$0 decreases (Daras et al., 2022). In the DMM treatment, large mask ratios make the posterior rely heavily on the prior, and the training signal is confined to corrupted coordinates (Benton et al., 2022).

A recurrent misconception is to treat any masked reconstruction loss as interchangeable with masked score matching. The cited score-based formulations are more specific: the target is either the posterior-averaged conditional score under a masked linear corruption (Daras et al., 2022) or the score of the corrupted marginal under a masked Markov kernel (Benton et al., 2022). Another recurrent misconception is terminological. The 2019 energy-based modeling paper titled “Learning Energy-Based Models in High-Dimensional Spaces with Multi-scale Denoising Score Matching” uses MDSM to mean multi-scale Gaussian corruption with a fixed $\epsilon\sim\mathcal{N}(0,\sigma^2 I_d)$1 target score; masked corruption is explicitly outside its scope (Li et al., 2019).

Taken together, these works position Masked Denoising Score-Matching as a principled subclass of DSM in which masks define a partial corruption process and a partial supervision pattern. Its theoretical basis is strongest when the masking operation is incorporated directly into the corruption model or transition kernel, and its practical performance depends on the trade-off between J-invariance and the loss of full receptive-field context.