Diffusion Model-Based Denoising

• Diffusion model-based denoising is a generative framework that leverages a learned Markov chain to add noise and then iteratively reverse it using neural approximations.
• The approach employs a forward process that corrupts data and a reverse process that predicts noise via score functions, ensuring robust signal recovery.
• Training minimizes a regression loss between predicted and true noise, enabling state-of-the-art performance in applications like image, medical, and signal denoising.

Diffusion model-based denoising refers to a paradigm in which a learned stochastic process, typically a Markov chain that incrementally adds noise to data samples (the forward process), is equipped with a neural approximation of the time-reversed, denoising process. This "score-based" or "denoising diffusion" model learns to iteratively invert the noising process and recover the signal or posterior distribution of interest from a corrupted observation. The resulting denoising procedures exhibit remarkable sample quality, robustness to various noise models, and strong theoretical links to optimal estimation in high-dimensional spaces.

1. Foundational Principles of Diffusion Model-Based Denoising

Diffusion model-based denoising systems are built on the construction of a parameterized Markov process that progressively adds noise to structured data—images, signals, embeddings—until the samples are nearly or exactly i.i.d. noise. The reverse process, parameterized by a neural network, seeks to iteratively recover clean data from a noisy input by predicting either the underlying “score” $\nabla_x \log q_t(x)$ or the explicit added noise $\epsilon$ at each step. Most modern approaches follow the Denoising Diffusion Probabilistic Model (DDPM) or its variants, with forward dynamics

$$q(x_t | x_{t-1}) = \mathcal{N}\bigl(x_t; \sqrt{\alpha_t} x_{t-1}, \beta_t I\bigr)$$

and associated closed-form marginals

$$q(x_t | x_0) = \mathcal{N}\bigl(x_t; \sqrt{\bar\alpha_t} x_0, (1-\bar\alpha_t) I\bigr)$$

where $\{\beta_t\}$ is a variational noise-variance schedule and $\bar\alpha_t = \prod_{s=1}^t \alpha_s$ with $\alpha_t = 1-\beta_t$. Denoising proceeds by learning a parameterized reverse kernel

$$p_\theta(x_{t-1} | x_t) = \mathcal{N}\bigl(x_{t-1}; \mu_\theta(x_t, t), \Sigma_t I\bigr)$$

with $\mu_\theta$ computed either by noise-prediction, $x_0$-prediction, or a dynamically mixed combination of both (Benny et al., 2022).

This framework generalizes beyond image generation and denoising, admitting extensions to arbitrary state spaces (Euclidean, discrete, manifold), domain-conditioned reverse steps, and arbitrary noise models (Gaussian, Poisson, Gamma) (Xie et al., 2023, Benton et al., 2022).

2. Forward and Reverse Processes: Variants and Generalizations

While standard DDPMs use additive Gaussian forward noise, application-specific denoising often requires adapting the forward process to the empirical noise model. For example, generative image denoising considers both synthetic (AWGN) and non-Gaussian (Gamma, Poisson) models, constructing the forward Markov chain such that its terminal distribution matches the observed noisy measurement. The reverse Markov chain is then parameterized to stochastically sample from the conditional posterior $\epsilon$0 (Xie et al., 2023). For spatially-variant real-world denoising, per-pixel time-variable maps are introduced to faithfully match signal-dependent noise (Pearl et al., 2023).

The general diffusion Markov model (DMM) framework extends this principle even further, showing that time-reversed Markov processes equipped with generalized score networks and implicit/explicit score-matching objectives can encompass inference in arbitrary noising dynamics, including discrete and manifold-valued spaces (Benton et al., 2022).

The reverse chain may be stochastic (as in DDPM, with residual noise injection) or deterministic (as in DDIM), permitting, at inference, tradeoffs between sample diversity and mean-square error optimality (Fesl et al., 2024).

3. Training and Inference Algorithms

Training diffusion denoisers consists in minimizing a (weighted) regression loss between the true noise and the predicted noise, or between true $\epsilon$1 and its prediction at each time step. For instance,

$\epsilon$2

where $\epsilon$3. More robust training can be attained by replacing the squared loss with robust alternatives (Huber, least-trimmed squares) to accommodate contaminated data, as in RDDPM for unsupervised anomaly segmentation (Moradi et al., 4 Aug 2025). For conditional denoising (e.g., posterior estimation), the reverse network may be conditioned on an observation $\epsilon$4, as described in the DMM framework (Benton et al., 2022) or for model-based imaging applications (Xu et al., 24 Jul 2025).

Inference proceeds by initializing the Markov chain at the observed noisy point $\epsilon$5 (or a suitably transformed variant) and iteratively applying the learned reverse dynamics for $\epsilon$6 steps. For truncated/accelerated denoising, step-skipping or shortest-path schemes have been formulated (e.g., ShortDF), achieving high-fidelity denoising in a fraction of the steps by optimizing a graph-theoretic relaxation (Chen et al., 5 Mar 2025).

Pseudocode, in its archetypal form, is:

$\epsilon$8

4. Architectural Designs and Conditioning Techniques

The diffusion denoiser is customarily realized as a UNet or equivalent convolutional/transformer backbone augmented with explicit time embeddings (often sinusoidal), skip connections, and, where needed, conditioning blocks to accommodate guidance information or domain adaptation. Recent designs include composite UNet backbones that simultaneously predict noise and clean sample, fusing the two via learned per-pixel interpolation (Benny et al., 2022); residual refinement approaches for augmenting reconstructive denoisers with diffusion-based residual synthesis (Wang et al., 2023); specialized modules for absorbing physical measurement models (e.g., DECT forward models in medical imaging (Xu et al., 24 Jul 2025)); and hierarchical/adaptive ensembling or embedding procedures for real/noisy images (Li et al., 2023).

Furthermore, specialized adaptive embedding and ensembling strategies such as DMID first map a real noisy image into the diffusion trajectory at the correct time step via a VAE-style noise transformer, then ensemble multiple denoised trajectories to suppress sample inconsistency and anchor the result to the measurement (Li et al., 2023).

5. Theoretical Guarantees and Trade-Offs

Diffusion-based denoising models possess strong theoretical guarantees under mild assumptions. Deterministic reversal of the learned Markov chain without stochastic re-sampling, when appropriately matched to the observation’s SNR, yields an estimator asymptotically converging to the minimum mean-square error conditional mean (CME), satisfying

$\epsilon$7

with polynomial convergence rates in practice using only a few hundred steps (Fesl et al., 2024). This reveals a deep connection between diffusion models and classical Bayesian estimators.

A key phenomenon is the perception-distortion tradeoff: longer reverse trajectories or increased stochasticity yield more perceptually natural results (often measured by FID or LPIPS) at the expense of pixel fidelity (PSNR/SSIM), while truncated/deterministic samplers yield higher distortion but align more closely with the CME (Dornbusch et al., 18 Mar 2025, Wang et al., 2023). Methods such as the Linear Combination Diffusion Denoiser (LCDD) (Dornbusch et al., 18 Mar 2025) and reconstruct-and-generate paradigms (Wang et al., 2023) exploit this to explicitly interpolate or combine the strengths of low-distortion and high-perception regimes.

6. Applications and Empirical Performance

Diffusion model-based denoising has demonstrated state-of-the-art performance in diverse application domains:

• Image denoising: Outperforming classical CNNs and vision transformers on synthetic (Gaussian, Poisson, Gamma noise) and real-world denoising tasks (Xie et al., 2023, Yang et al., 2023, Pearl et al., 2023, Li et al., 2023).
• Medical image denoising and inversion: Achieving PSNR and SSIM improvements over specialized baselines in low-dose CT (Chen et al., 24 Aug 2025, Xu et al., 24 Jul 2025), with strong generalization across dose and anatomy by integrating domain-aware embeddings or model-based physics constraints.
• Stepwise signal recovery: Denoising single-molecule step transitions and event boundaries more accurately than Hidden Markov or frequency-domain filtering (Tong et al., 9 Feb 2026).
• Recommender systems: Denoising user/item embeddings to improve robustness against noisy implicit feedback (Zhao et al., 2024).
• Semantic communications: Mapping SNRs to diffusion time for optimal denoising of latent codes in end-to-end communications channels (Wang et al., 6 Jun 2025).
• Image compression: Tokenizing images using generalized diffusion codebook models (gDDCM) for improved rate-distortion performance (Kong, 17 Nov 2025).

Empirical results frequently show that diffusion-based approaches surpass or match the best discriminative and self-supervised baselines in both quantitative metrics (PSNR, SSIM) and perceptual quality measures (LPIPS, FID) across varied test scenarios.

Method	Domain	Distortion Metric	Perceptual Metric	Ref.
SVNR	Real image noise	PSNR 24.56 dB	FID 43.1	(Pearl et al., 2023)
RnG	SIDD, DIV2K	LPIPS 0.0719	NIQE 11.85	(Wang et al., 2023)
ShortDF	Gen. synthesis	Step 2 FID 9.08	n/a	(Chen et al., 5 Mar 2025)
DMID-d/DMID-p	CBSD68/ImageNet	PSNR +0.5dB	LPIPS best	(Li et al., 2023)
FoundDiff	LDCT CT	PSNR 44.22 dB	SSIM 0.9731	(Chen et al., 24 Aug 2025)
SSDM	Step signals	MSE 0.0041; F1 0.96	n/a	(Tong et al., 9 Feb 2026)

7. Limitations and Extensions

Although diffusion model-based denoisers offer robust and generalizable performance, they are computationally intensive compared to single-pass discriminative networks, especially at high image resolutions and large diffusion depths. Approximate/truncated and shortest-path-accelerated variants (Chen et al., 5 Mar 2025) alleviate this to a degree. Further, for contaminated or semi-supervised training data, robust losses improve segmentation and anomaly detection but require parameter tuning (Moradi et al., 4 Aug 2025).

Future directions include the development of:

• More expressive and adaptive forward processes for non-Gaussian or structured noise;
• Improved samplers for real-time or resource-constrained environments;
• Explicit uncertainty quantification and diversity guarantees;
• Learning codebooks or quantizers within gDDCMs for better compression-denoising tradeoffs (Kong, 17 Nov 2025).

In summary, diffusion model-based denoising is an active and theoretically grounded research area, driving advances across image, signal, and domain-adaptive denoising applications while providing a unifying generative prior approach that interpolates between Bayesian optimality and perceptual realism (Xie et al., 2023, Fesl et al., 2024, Benton et al., 2022).