Single-Step Diffusion Regularization (SSD-Reg)

• SSD-Reg is a learned regularization strategy for microwave tomography that integrates a single reverse-diffusion step per iteration to provide a diffusion-based prior gradient.
• It couples a physics-based electromagnetic forward model with a pretrained diffusion prior within a variational plug-and-play framework using automatic differentiation.
• Empirical results demonstrate that SSD-Reg achieves faster convergence, higher SSIM/PSNR, and improved noise robustness compared to classical and recent reconstruction methods.

Searching arXiv for the named method and closely related diffusion-regularization literature. arXiv search query: "Single-Step Diffusion Regularization microwave tomography SSD-Reg" Single-Step Diffusion Regularization (SSD-Reg) is a learned regularization strategy that embeds a diffusion-model prior into a physics-based iterative reconstruction algorithm for microwave tomography (MWT). In the formulation introduced in "Learned Regularization for Microwave Tomography" (Tong et al., 11 Aug 2025), SSD-Reg uses a single reverse-diffusion step per reconstruction iteration to provide a prior gradient, rather than running a full diffusion sampling chain. The method is positioned as a hybrid, physics–generative framework in which an accurate Fréchet-differentiable forward model for electromagnetic scattering is combined with a pretrained diffusion prior on images of shapes, and the two are integrated through a Plug-and-Play (PnP) variational scheme.

1. Definition, scope, and interpretive boundaries

The explicit term Single-Step Diffusion Regularization (SSD-Reg) is introduced for microwave tomography, where the objective is to reconstruct the complex-valued dielectric contrast field $\chi(x)$ from measured scattered electromagnetic fields. The inverse problem is described as nonlinear and severely ill-posed. Within this setting, SSD-Reg denotes a learned regularizer that acts through a single-step denoising or score correction inside each iteration of a conventional optimization procedure, rather than through an end-to-end predictor or a multi-step reverse diffusion trajectory (Tong et al., 11 Aug 2025).

Formally, after normalizing the discrete complex contrast $\underline{\chi}$ to an image $\mathbf{x}_0$ and applying random flips, the regularizer is defined as

$$R(\mathbf{x}_0) = \lambda_t \ \operatorname{sg}\big[\xi_\phi(\mathbf{x}_t, t) - \xi_t\big]^\mathsf{T} \mathbf{x}_0, \qquad \lambda_t = \frac{\lambda}{\mathrm{SNR}_t},$$

where $\xi_\phi(\mathbf{x}_t,t)$ is the noise predicted by a pretrained diffusion model, $\xi_t \sim \mathcal{N}(0,I)$ is the actual injected noise, and

$$\mathbf{x}_t = \sqrt{\alpha_t}\,\mathbf{x}_0 + \sqrt{1-\alpha_t}\,\xi_t .$$

Because of the stop-gradient operator, the gradient with respect to $\mathbf{x}_0$ is

$$\nabla_{\mathbf{x}_0} R(\mathbf{x}_0) = \lambda_t \ \operatorname{sg}[\xi_\phi(\mathbf{x}_t, t) - \xi_t]^\mathsf{T},$$

which the paper interprets as a single-step denoising or score correction provided by the diffusion model (Tong et al., 11 Aug 2025).

A central interpretive point is that single-step does not mean a single-shot inversion from measurements to reconstruction. In SSD-Reg, the main optimization trajectory is still iterative and data-consistency-driven; the single-step designation refers specifically to the use of one denoiser call per outer reconstruction iteration. A related misconception is that SSD-Reg is a diffusion sampler. The paper states the opposite: SSD-Reg does not sample from the diffusion reverse process, but instead uses the diffusion model only to compute a single gradient term per iteration in Adam on $\underline{\chi}$ (Tong et al., 11 Aug 2025).

2. Variational structure and electromagnetic forward model

SSD-Reg is embedded in a variational formulation for the 2D transverse-magnetic electromagnetic scattering problem. The contrast is $\underline{\chi}$0, supported in a domain $\underline{\chi}$1. For an incident field $\underline{\chi}$2, the scattered field $\underline{\chi}$3 satisfies the Lippmann–Schwinger equation

$\underline{\chi}$4

with integral operator

$\underline{\chi}$5

The forward operator mapping contrast to measured scattered field is written as

$\underline{\chi}$6

where $\underline{\chi}$7 denotes the Lippmann–Schwinger solution operator (Tong et al., 11 Aug 2025).

The inverse problem is posed as

$\underline{\chi}$8

The data-consistency loss is

$\underline{\chi}$9

and the total loss used in optimization is

$\mathbf{x}_0$0

The gradient is correspondingly decomposed as

$\mathbf{x}_0$1

with the regularization gradient obtained by automatic differentiation through the preprocessing map $\mathbf{x}_0$2 (Tong et al., 11 Aug 2025).

This construction makes SSD-Reg a variationally integrated prior rather than a post-processing heuristic. The forward model supplies the measurement-space geometry and explicit Jacobian structure; the diffusion prior supplies an image-space correction toward a learned distribution of plausible shapes. A plausible implication is that the method is designed to preserve fidelity to the governing physics while replacing handcrafted analytic priors, such as total variation, with a learned structural prior.

3. Diffusion prior and the meaning of the single-step correction

The diffusion prior used in SSD-Reg is of the VP-SDE / DDPM-style family. The noising process is written as

$\mathbf{x}_0$3

with discrete-time training implemented through

$\mathbf{x}_0$4

The model $\mathbf{x}_0$5 predicts the noise at time $\mathbf{x}_0$6, and training uses the denoising score-matching loss

$\mathbf{x}_0$7

The paper states that the diffusion model is trained in an unsupervised fashion on images sampled from a distribution of complex shapes and that there is no need for paired data consisting of measurements and ground truth (Tong et al., 11 Aug 2025).

At reconstruction time, the prior is used only as a regularizer. For each iteration, SSD-Reg selects a diffusion timestep $\mathbf{x}_0$8, samples one Gaussian noise vector $\mathbf{x}_0$9, perturbs the current reconstruction estimate once to obtain $$R(\mathbf{x}_0) = \lambda_t \ \operatorname{sg}\big[\xi_\phi(\mathbf{x}_t, t) - \xi_t\big]^\mathsf{T} \mathbf{x}_0, \qquad \lambda_t = \frac{\lambda}{\mathrm{SNR}_t},$$0, evaluates the pretrained diffusion model once, and computes a regularization gradient proportional to $$R(\mathbf{x}_0) = \lambda_t \ \operatorname{sg}\big[\xi_\phi(\mathbf{x}_t, t) - \xi_t\big]^\mathsf{T} \mathbf{x}_0, \qquad \lambda_t = \frac{\lambda}{\mathrm{SNR}_t},$$1. There is no reverse chain, no repeated denoising trajectory, and no need to backpropagate through time (Tong et al., 11 Aug 2025).

This single-step mechanism is closely related, in spirit, to other low-NFE diffusion adaptations, but the operational role is different. In "Single and Few-step Diffusion for Generative Speech Enhancement" (Lay et al., 2023), a diffusion model is adapted to coarse reverse integration through a second-stage corrective loss on the final solver output. In "One-Step Effective Diffusion Network for Real-World Image Super-Resolution" (Wu et al., 2024), a one-step generator is regularized through latent-space variational score distillation and a KL-divergence surrogate. These comparisons suggest that SSD-Reg belongs to a broader family of methods that use diffusion models as structural constraints in extreme low-step regimes, but in the MWT formulation the diffusion model functions explicitly as a gradient-based regularizer rather than as a one-step generator or a distilled sampler.

4. Plug-and-Play realization and optimization algorithm

SSD-Reg is implemented within a PnP-style reconstruction procedure, although the paper emphasizes that it does not split the update into two separate operators. Instead, both the data-consistency gradient and the diffusion-regularization gradient are combined in a single Adam update (Tong et al., 11 Aug 2025).

The reconstruction algorithm proceeds as follows. The contrast $$R(\mathbf{x}_0) = \lambda_t \ \operatorname{sg}\big[\xi_\phi(\mathbf{x}_t, t) - \xi_t\big]^\mathsf{T} \mathbf{x}_0, \qquad \lambda_t = \frac{\lambda}{\mathrm{SNR}_t},$$2 is initialized, for example by zeros. At each iteration, the predicted scattered field is computed through the forward model, the residual is formed, and the data-consistency loss is evaluated. The method uses early stopping if $$R(\mathbf{x}_0) = \lambda_t \ \operatorname{sg}\big[\xi_\phi(\mathbf{x}_t, t) - \xi_t\big]^\mathsf{T} \mathbf{x}_0, \qquad \lambda_t = \frac{\lambda}{\mathrm{SNR}_t},$$3 and $$R(\mathbf{x}_0) = \lambda_t \ \operatorname{sg}\big[\xi_\phi(\mathbf{x}_t, t) - \xi_t\big]^\mathsf{T} \mathbf{x}_0, \qquad \lambda_t = \frac{\lambda}{\mathrm{SNR}_t},$$4. The Jacobian blocks are updated asymmetrically: $$R(\mathbf{x}_0) = \lambda_t \ \operatorname{sg}\big[\xi_\phi(\mathbf{x}_t, t) - \xi_t\big]^\mathsf{T} \mathbf{x}_0, \qquad \lambda_t = \frac{\lambda}{\mathrm{SNR}_t},$$5 is always updated, and $$R(\mathbf{x}_0) = \lambda_t \ \operatorname{sg}\big[\xi_\phi(\mathbf{x}_t, t) - \xi_t\big]^\mathsf{T} \mathbf{x}_0, \qquad \lambda_t = \frac{\lambda}{\mathrm{SNR}_t},$$6 is updated every 5 iterations. The data-consistency gradient is

$$R(\mathbf{x}_0) = \lambda_t \ \operatorname{sg}\big[\xi_\phi(\mathbf{x}_t, t) - \xi_t\big]^\mathsf{T} \mathbf{x}_0, \qquad \lambda_t = \frac{\lambda}{\mathrm{SNR}_t},$$7

A diffusion time step $$R(\mathbf{x}_0) = \lambda_t \ \operatorname{sg}\big[\xi_\phi(\mathbf{x}_t, t) - \xi_t\big]^\mathsf{T} \mathbf{x}_0, \qquad \lambda_t = \frac{\lambda}{\mathrm{SNR}_t},$$8 and noise $$R(\mathbf{x}_0) = \lambda_t \ \operatorname{sg}\big[\xi_\phi(\mathbf{x}_t, t) - \xi_t\big]^\mathsf{T} \mathbf{x}_0, \qquad \lambda_t = \frac{\lambda}{\mathrm{SNR}_t},$$9 are then sampled; $\xi_\phi(\mathbf{x}_t,t)$0 is normalized and randomly flipped to obtain $\xi_\phi(\mathbf{x}_t,t)$1; the perturbation

$\xi_\phi(\mathbf{x}_t,t)$2

is formed; the SSD loss is computed; $\xi_\phi(\mathbf{x}_t,t)$3 is obtained by autodiff; and the two gradients are combined before an Adam update with learning rate $\xi_\phi(\mathbf{x}_t,t)$4 (Tong et al., 11 Aug 2025).

The random flip augmentation is not an incidental detail. The paper states that horizontal and vertical flipping make the effective prior distribution more symmetric and reduce overfitting to orientation biases in the training dataset. In the reported ablation, removing random flipping reduces SSIM from $\xi_\phi(\mathbf{x}_t,t)$5 to $\xi_\phi(\mathbf{x}_t,t)$6 and PSNR from $\xi_\phi(\mathbf{x}_t,t)$7 to $\xi_\phi(\mathbf{x}_t,t)$8, with visible geometric distortions such as circular structures becoming polygonal and shapes deforming (Tong et al., 11 Aug 2025).

Within this optimization view, SSD-Reg can be understood as a PnP module that contributes a learned prior gradient while leaving the main inverse-problem variable, the forward operator, and the optimizer unchanged. This differs from posterior-sampling-style diffusion inverse solvers, which evolve the reconstruction along a diffusion trajectory. SSD-Reg instead preserves a conventional iterative inverse-problem backbone and inserts the diffusion model as a single-step regularization term.

5. Empirical behavior in microwave tomography

The paper reports results on simulated data, a high-contrast breast dataset, and real-world Fresnel data. The simulated scenarios include the Austria benchmark, 2D shapes, and MNIST-based permittivity distributions. The high-contrast breast experiments use a $\xi_\phi(\mathbf{x}_t,t)$9 domain with coupling medium $\xi_t \sim \mathcal{N}(0,I)$0 and multi-frequency measurements at $\xi_t \sim \mathcal{N}(0,I)$1, $\xi_t \sim \mathcal{N}(0,I)$2, $\xi_t \sim \mathcal{N}(0,I)$3, and $\xi_t \sim \mathcal{N}(0,I)$4. The real-world experiments involve nylon and foam objects in $\xi_t \sim \mathcal{N}(0,I)$5 TM-mode with $\xi_t \sim \mathcal{N}(0,I)$6–$\xi_t \sim \mathcal{N}(0,I)$7 transmitters and $\xi_t \sim \mathcal{N}(0,I)$8 receivers (Tong et al., 11 Aug 2025).

The baselines are BP, PDA, BPS, INR+TV, and DPS. All except BPS are unsupervised. According to the reported comparisons, SSD-Reg significantly improves PSNR and SSIM compared to classical and recent methods on simulated and real data, and it yields the best PSNR, SSIM, and lowest LPIPS in most simulated cases. Qualitatively, BP and PDA are described as blurry or speckled; INR+TV and PDA show periodic artifacts; DPS suppresses artifacts but can distort permittivity values and boundaries; SSD-Reg recovers shapes close to ground truth with realistic permittivity and minimal artifacts (Tong et al., 11 Aug 2025).

A concise summary of several reported performance properties is given below.

Property	Reported behavior	Context
Convergence	About 200–300 iterations	Faster than $\xi_t \sim \mathcal{N}(0,I)$9 for INR+TV
Reconstruction time	SSD $$\mathbf{x}_t = \sqrt{\alpha_t}\,\mathbf{x}_0 + \sqrt{1-\alpha_t}\,\xi_t .$$0	DPS $$\mathbf{x}_t = \sqrt{\alpha_t}\,\mathbf{x}_0 + \sqrt{1-\alpha_t}\,\xi_t .$$1, INR+TV $$\mathbf{x}_t = \sqrt{\alpha_t}\,\mathbf{x}_0 + \sqrt{1-\alpha_t}\,\xi_t .$$2
Noise robustness	SSIM/PSNR remain high at 5%, 15%, 30% noise	SSD outperforms DPS and INR+TV in SSIM and PSNR across all noise levels

For additive Gaussian noise, the paper reports approximate SSD values of $$\mathbf{x}_t = \sqrt{\alpha_t}\,\mathbf{x}_0 + \sqrt{1-\alpha_t}\,\xi_t .$$3, $$\mathbf{x}_t = \sqrt{\alpha_t}\,\mathbf{x}_0 + \sqrt{1-\alpha_t}\,\xi_t .$$4, and $$\mathbf{x}_t = \sqrt{\alpha_t}\,\mathbf{x}_0 + \sqrt{1-\alpha_t}\,\xi_t .$$5 in SSIM/PSNR at $$\mathbf{x}_t = \sqrt{\alpha_t}\,\mathbf{x}_0 + \sqrt{1-\alpha_t}\,\xi_t .$$6, $$\mathbf{x}_t = \sqrt{\alpha_t}\,\mathbf{x}_0 + \sqrt{1-\alpha_t}\,\xi_t .$$7, and $$\mathbf{x}_t = \sqrt{\alpha_t}\,\mathbf{x}_0 + \sqrt{1-\alpha_t}\,\xi_t .$$8 noise, respectively. It further states that SSD-Reg maintains high structural fidelity even at $$\mathbf{x}_t = \sqrt{\alpha_t}\,\mathbf{x}_0 + \sqrt{1-\alpha_t}\,\xi_t .$$9 additive Gaussian noise and remains stable in high-contrast scenarios with breast phantoms of contrast greater than $\mathbf{x}_0$0 (Tong et al., 11 Aug 2025).

The reported convergence behavior also distinguishes SSD-Reg from the baselines. On the Austria benchmark, SSD-Reg reaches a plateau around approximately $\mathbf{x}_0$1 iterations with stable PSNR and SSIM, whereas INR+TV improves more slowly and requires many more iterations. Average reconstruction time per case is reported as approximately $\mathbf{x}_0$2 for SSD, $\mathbf{x}_0$3 for DPS, and $\mathbf{x}_0$4 for INR+TV, which the paper describes as a $\mathbf{x}_0$5 speedup over INR+TV (Tong et al., 11 Aug 2025).

The authors do not present formal convergence theorems. They instead state that the design is inspired by RED and diffusion variational perspectives, and that the diffusion prior approximates a gradient of a log-prior, making SSD-Reg an approximate MAP estimator under a learned diffusion prior. This suggests a variational interpretation, but not a proved one.

6. Relation to adjacent single-step diffusion paradigms

SSD-Reg sits within a wider methodological shift toward reducing the number of diffusion evaluations while retaining the statistical structure of diffusion priors. In the MWT paper, the closest comparisons are to standard DPS and to RED-diff- or SDS-inspired approaches: SSD-Reg is described as conceptually closest to RED-diff and SDS because it uses diffusion as a learned regularizer, employs a single forward call to estimate a gradient direction, and avoids backpropagation through long diffusion chains (Tong et al., 11 Aug 2025).

A distinct but relevant neighboring line is represented by "Single-Step Consistent Diffusion Samplers" (Jutras-Dubé et al., 11 Feb 2025). That work studies unnormalized target distributions and introduces consistent diffusion samplers that generate high-fidelity samples in a single step. Its central mechanisms are consistency and self-consistency losses, used either for data-free distillation from a pretrained diffusion sampler or for fully amortized training from scratch. The paper states that these mechanisms can be interpreted as regularizers that constrain a one-step map to remain aligned with multi-step diffusion dynamics. This suggests a conceptual parallel to SSD-Reg, although the application domain and formal objective are different: (Jutras-Dubé et al., 11 Feb 2025) is concerned with sampling from unnormalized distributions, whereas SSD-Reg is formulated as learned regularization inside a physics-based inverse solver.

The speech-enhancement and Real-ISR papers reinforce the same broader pattern. In speech enhancement, the low-NFE adaptation is achieved through a two-stage procedure in which a predictive loss corrects the coarse reverse process for a chosen schedule, allowing stable performance at $\mathbf{x}_0$6–$\mathbf{x}_0$7 NFEs (Lay et al., 2023). In Real-ISR, OSEDiff uses the degraded observation itself as the diffusion starting point and adds latent-space variational score distillation as a KL-divergence regularizer so that a one-step latent transformation remains consistent with the distribution induced by a full Stable Diffusion model (Wu et al., 2024). A plausible implication is that SSD-Reg belongs to a family of methods in which diffusion models cease to be used primarily as samplers and instead become trainable or frozen sources of regularizing gradients, flow constraints, or distribution-matching penalties in aggressively compressed inference regimes.

Within that broader family, the distinctive features of SSD-Reg are its explicit use of a fully Fréchet-differentiable electromagnetic forward model with Jacobians, its unsupervised diffusion prior on generic shape images rather than paired MWT training data, and its single-step PnP realization for nonlinear inverse scattering. Those characteristics define SSD-Reg not merely as a computational shortcut, but as a particular synthesis of variational inverse problems and learned diffusion regularization (Tong et al., 11 Aug 2025).