Blind Denoising Diffusion Models: Theory & Applications

• BDDMs are diffusion-based models that omit explicit noise amplitude inputs by marginalizing over unknown nuisance parameters to infer implicit noise schedules.
• They unify schedule-free generative formulations with practical conditioning strategies for ill-posed tasks such as denoising, deblurring, and super-resolution.
• Empirical results show BDDMs achieve competitive performance in medical imaging, face restoration, and remote sensing via guided sampling and stabilization techniques.

Searching arXiv for papers on Blind Denoising Diffusion Models and closely related blind diffusion formulations. Blind Denoising Diffusion Models (BDDMs) are diffusion-based models for settings in which a critical nuisance variable is unavailable at training time, sampling time, or both. In the narrow theoretical sense, a BDDM is a generative diffusion model whose denoiser is blind to the noise amplitude $\sigma$, so the network $f_\theta:\mathbb R^d\to\mathbb R^d$ never receives $\sigma$ as input yet is trained by marginalizing over $\sigma$ and can still follow an implicit reverse-time noise schedule (Kadkhodaie et al., 10 Feb 2026). In a broader applied sense found across recent arXiv literature, the term encompasses diffusion formulations for blind denoising and blind inverse problems in which the clean image, degradation parameters, or measurement operator are unknown, and the diffusion process is conditioned instead on auxiliary structure such as blind-spot predictions, degradation representations, or learned operator priors (Demir et al., 31 Mar 2025). This dual usage makes BDDMs both a theoretical object in schedule-free generative modeling and a practical design pattern for ill-posed restoration problems.

1. Terminological scope and problem settings

Within the cited literature, blindness refers to several distinct unknowns. In the theoretical formulation of “Blind denoising diffusion models and the blessings of dimensionality” (Kadkhodaie et al., 10 Feb 2026), the blind variable is the noise amplitude: the denoiser never sees $\sigma$ in either training or sampling. In medical self-supervised denoising, the blind variable is the unknown corruption process in observations of the form $x' = x + n$, where $n$ may come from an unknown, possibly non-i.i.d. distribution such as signal-dependent Gaussian, Poisson, or spatially correlated noise (Demir et al., 31 Mar 2025). In blind inverse problems, both the target image and the measurement operator can be unknown, as in $y = H_\theta(x) + \varepsilon$ with unknown $\theta$ (Li et al., 28 May 2025), or $y = Ax + z$ with unknown linear operator $f_\theta:\mathbb R^d\to\mathbb R^d$0 (Murata et al., 2023). In blind face restoration, the unknown is the degradation type and degree affecting a low-quality face image (Qiu et al., 2024). In blind super-resolution for remote sensing images, the challenge is the unknown degradation kernel (Xu et al., 2023).

This suggests that BDDMs are better understood as a family of diffusion methods for latent nuisance inference under uncertainty rather than a single standardized architecture. A plausible implication is that comparisons across papers require care, because “blind” may mean unknown noise amplitude, unknown corruption law, unknown kernel, or unknown operator, and the conditioning strategy changes accordingly.

A compact view of the problem classes represented in the literature is given below.

Setting	Unknown quantity	Representative paper
Schedule-free generative diffusion	Noise amplitude $f_\theta:\mathbb R^d\to\mathbb R^d$1	(Kadkhodaie et al., 10 Feb 2026)
Self-supervised medical denoising	Clean reference and noise law	(Demir et al., 31 Mar 2025)
Blind deblurring / inverse problems	Measurement operator parameters	(Li et al., 28 May 2025, Murata et al., 2023)
Blind restoration	Degradation representation	(Qiu et al., 2024)
Blind super-resolution	Blur kernel	(Xu et al., 2023)
Feature-space blind denoising	Noise in transferred quality-aware features	(Li et al., 2024)

2. Core mathematical formulation of blind denoising diffusion

The canonical theoretical BDDM replaces the standard denoiser $f_\theta:\mathbb R^d\to\mathbb R^d$2, which is trained with explicit access to the true noise level $f_\theta:\mathbb R^d\to\mathbb R^d$3, by a blind denoiser $f_\theta:\mathbb R^d\to\mathbb R^d$4 trained by marginalizing over $f_\theta:\mathbb R^d\to\mathbb R^d$5 drawn from a prior $f_\theta:\mathbb R^d\to\mathbb R^d$6 on $f_\theta:\mathbb R^d\to\mathbb R^d$7 (Kadkhodaie et al., 10 Feb 2026). The population objective is

$f_\theta:\mathbb R^d\to\mathbb R^d$8

Proposition 3.1 in that work shows that the population minimizer admits the Bayesian form

$f_\theta:\mathbb R^d\to\mathbb R^d$9

so the blind residual

$\sigma$0

is a mixture of non-blind score functions (Kadkhodaie et al., 10 Feb 2026). This is the defining mathematical distinction: the score surrogate is not indexed by a provided noise amplitude, but by a posterior over amplitudes induced by the noisy sample itself.

The forward process in that paper is a variance-exploding SDE,

$\sigma$1

with $\sigma$2 and $\sigma$3 (Kadkhodaie et al., 10 Feb 2026). The blind reverse process replaces the true $\sigma$4-conditioned score by the blind estimator $\sigma$5 and uses

$\sigma$6

No $\sigma$7 input is required at sampling time (Kadkhodaie et al., 10 Feb 2026).

This differs materially from conventional DDPM and EDM parameterizations in the applied papers, which usually retain timestep conditioning $\sigma$8 and sometimes additional blind-side information. For example, DiffDenoise adopts the standard DDPM forward process

$\sigma$9

with reverse kernel

$\sigma$0

where $\sigma$1 is the blind-spot conditioning image (Demir et al., 31 Mar 2025). Blind-Spot Guided Diffusion uses two reverse kernels, one unconditional branch and one blind-spot branch, and fuses their noise estimates in a classifier-free-style guidance (Cheng et al., 19 Sep 2025). DR-BFR works in latent space and conditions on both low-quality content features and a degradation representation learned by a Degradation Representation Module (Qiu et al., 2024).

3. Implicit schedules, dimensionality, and theoretical guarantees

The central theoretical claim of the strict BDDM formulation is that blind denoisers automatically track an implicit noise schedule along the reverse process (Kadkhodaie et al., 10 Feb 2026). Under the heuristic ansatz that $\sigma$2 and replacing the posterior on $\sigma$3 by $\sigma$4, the ideal SDE becomes

$\sigma$5

A Fokker–Planck calculation shows that consistency $\sigma$6 holds exactly if

$\sigma$7

equivalently

$\sigma$8

(Kadkhodaie et al., 10 Feb 2026). Lemma 3.3 and 3.4 in that paper show that $\sigma$9 decreases to $\sigma$0 if $\sigma$1.

The main rigorous error bound depends on bounded support, low intrinsic dimension $\sigma$2, and a score-estimation error assumption. Theorem 3.6 shows

$\sigma$3

(Kadkhodaie et al., 10 Feb 2026). The last term comes from Bayesian noise-level estimation error, and the paper’s posterior concentration argument yields

$\sigma$4

(Kadkhodaie et al., 10 Feb 2026). The phrase “blessings of dimensionality” refers precisely to this regime: when the ambient dimension is sufficiently large relative to intrinsic dimension, blind estimation of the residual noise level becomes accurate.

For the special schedule $\sigma$5 with $\sigma$6, the paper derives $\sigma$7 and gives a discretization result for the exponential–Euler integrator. Theorem 3.11 states

$\sigma$8

with $\sigma$9, and notes that if $x' = x + n$0 the $x' = x + n$1 term vanishes (Kadkhodaie et al., 10 Feb 2026). Corollary 3.14 then gives a sampling guarantee in the perceptual metric $x' = x + n$2: with

$x' = x + n$3

one obtains $x' = x + n$4 provided $x' = x + n$5 up to logs (Kadkhodaie et al., 10 Feb 2026).

A common misconception is that omitting the explicit noise amplitude necessarily destroys control of the reverse process. The theoretical analysis argues the opposite under low-intrinsic-dimensional structure: the blind denoiser can infer the residual noise variance from the noisy image itself and thereby recover an implicit schedule (Kadkhodaie et al., 10 Feb 2026).

4. Conditioning mechanisms in practical BDDMs

Applied BDDMs rarely remain fully unconditional. Instead, they compensate for blind uncertainty by injecting auxiliary signals that are informative about structure, degradation, or operator state.

DiffDenoise conditions a DDPM-style denoiser on the output of a pretrained Blind-Spot Network, $x' = x + n$6, obtained under $x' = x + n$7-invariance (Demir et al., 31 Mar 2025). The conditioning image is concatenated channel-wise with $x' = x + n$8 and injected at multiple resolutions via simple concatenation in the encoder and decoder of a U-Net–style backbone; the paper notes that cross-attention or FiLM layers are also possible, but concatenation sufficed in its experiments (Demir et al., 31 Mar 2025). The training objective is the standard simple denoising loss

$x' = x + n$9

with $n$0 (Demir et al., 31 Mar 2025).

Blind-Spot Guided Diffusion likewise combines blind-spot structure with a conventional diffusion branch. It learns two Gaussian reverse kernels, an unconditional branch $n$1 and a blind-spot branch $n$2, and fuses their noise predictions as

$n$3

(Cheng et al., 19 Sep 2025). The blind-spot branch is based on the PUCA blind-spot network modified to be time-dependent, while the non-blind branch is a standard U-shaped architecture as in Ho et al. (Cheng et al., 19 Sep 2025).

DR-BFR uses a different conditioning logic. Its Degradation Representation Module extracts a degradation prompt $n$4 from low-quality faces via unsupervised contrastive learning, LQ reconstruction, and a distribution-matching loss, with total loss

$n$5

(Qiu et al., 2024). The Latent Diffusion Restoration Module then conditions on both latent low-quality content features $n$6 and the degradation representation $n$7. The former is injected by simple channel-wise concatenation at the U-Net input, while the latter is injected via cross-attention layers, with time-adaptive prompt strength $n$8 where $n$9 (Qiu et al., 2024).

Blind super-resolution in DDSR splits conditioning across two DDPMs: a kernel-estimation DDPM conditioned on an LR encoding $y = H_\theta(x) + \varepsilon$0, and a reconstruction DDPM conditioned on both $y = H_\theta(x) + \varepsilon$1 and the predicted kernel $y = H_\theta(x) + \varepsilon$2 (Xu et al., 2023). Blind inverse-problem samplers use yet another strategy. Blind-PnPDM alternates between image and operator updates, each formulated as a Gaussian denoising task solved by a separate diffusion prior, one for clean images and one for operator parameters (Li et al., 28 May 2025). GibbsDDRM keeps a pretrained unconditional diffusion model for the signal prior and couples it with explicit online updates of the operator through a partially collapsed Gibbs sampler and Langevin dynamics (Murata et al., 2023).

Taken together, these papers suggest that practical BDDMs often replace direct access to the unknown nuisance with a surrogate observable: a blind-spot estimate, a latent degradation code, a current operator sample, or a pseudo-clean conditioning image.

5. Sampling, stabilization, and blind posterior inference

Sampling behavior is a major differentiator within the BDDM literature. The theoretical schedule-free BDDM trains a blind denoiser with Algorithm 1 and samples with Algorithm 2 by initializing $y = H_\theta(x) + \varepsilon$3, iterating

$y = H_\theta(x) + \varepsilon$4

with $y = H_\theta(x) + \varepsilon$5, and stopping when $y = H_\theta(x) + \varepsilon$6 (Kadkhodaie et al., 10 Feb 2026). The stopping rule itself reflects the model’s ability to estimate residual noise from the sample.

DiffDenoise uses DDIM sampling for speed, but introduces Symmetric-Noise Stabilization (SRDS). Two DDIM trajectories are run from initial noises $y = H_\theta(x) + \varepsilon$7 and $y = H_\theta(x) + \varepsilon$8:

$y = H_\theta(x) + \varepsilon$9

and the final output is averaged,

$\theta$0

(Demir et al., 31 Mar 2025). The stated rationale is that score-model errors are roughly odd functions in $\theta$1, so averaging cancels leading-order bias terms (Demir et al., 31 Mar 2025).

Blind-Spot Guided Diffusion uses Random & Complementary Replacement during sampling. At each step it forms

$\theta$2

replaces pixels of $\theta$3 by $\theta$4 with probability $\theta$5 in Base Replacement, repeats sampling for $\theta$6 rounds, and averages the resulting estimates (Cheng et al., 19 Sep 2025). The paper states that random and complementary pixel replacement reintroduces noisy-pixel information in controlled fashion, preventing over-smoothing and preserving spatial continuity (Cheng et al., 19 Sep 2025).

Blind inverse-problem samplers use alternating posterior updates rather than a single reverse chain. Blind-PnPDM employs a split-Gibbs sampler: an image-update alternates a Gaussian likelihood step with a prior denoising step implemented by an EDM image denoiser $\theta$7, and an operator-update alternates an analogous likelihood step with a prior denoising step implemented by a second diffusion denoiser $\theta$8 (Li et al., 28 May 2025). GibbsDDRM similarly samples from $\theta$9 by interleaving modified DDRM updates for the latent signal states with Langevin updates for the unknown operator, using SVD spectral coordinates for efficient Gaussian conditionals (Murata et al., 2023).

A plausible implication is that “sampling” in BDDMs spans three distinct computational idioms: schedule-free reverse SDE integration (Kadkhodaie et al., 10 Feb 2026), guided diffusion with stabilization heuristics (Demir et al., 31 Mar 2025, Cheng et al., 19 Sep 2025), and MCMC-style alternating posterior simulation for blind inverse problems (Li et al., 28 May 2025, Murata et al., 2023).

6. Empirical domains and reported performance

The empirical record of BDDMs is diverse because the blind variable differs by domain.

In self-supervised medical image denoising, DiffDenoise was tested on FastMRI knee MRI with Gaussian, Poisson, Gamma pixel noise, SIIM chest X-rays, spatially correlated variants, and real low-field brain MRI (M4Raw) (Demir et al., 31 Mar 2025). The paper reports that on independent Gaussian pixel noise, DiffDenoise achieves $y = Ax + z$0 PSNR gain over the corrupted input and sits only $y = Ax + z$1 below a fully supervised NAF-Net while using no clean training images, and that SSIM matches or exceeds supervised baselines (Demir et al., 31 Mar 2025). Against Neighbor2Neighbor, PUCA, LG-BPN, and Noise2Self, it is reported as $y = Ax + z$2–$y = Ax + z$3 PSNR better and as preserving vessel and cartilage edges that other methods blur (Demir et al., 31 Mar 2025). Under spatially correlated noise, competing BSN-based approaches degrade by $y = Ax + z$4–$y = Ax + z$5 as noise correlation grows, whereas DiffDenoise stays within $y = Ax + z$6 (Demir et al., 31 Mar 2025). On real M4Raw brain MRI, it outperforms PUCA by $y = Ax + z$7–$y = Ax + z$8 PSNR and surpasses NAF-Net in SSIM (Demir et al., 31 Mar 2025).

In real-world image denoising, Blind-Spot Guided Diffusion was evaluated on SIDD and DND. On SIDD, the paper reports 37.98 dB / 0.944, compared with approximately 37.69 dB / 0.937 for the best prior BSN, SelfFormer; on DND it reports 38.99 dB / 0.943, compared with approximately 38.92 dB / 0.943 for the best prior method (Cheng et al., 19 Sep 2025). Ablation results indicate guidance weight $y = Ax + z$9 performs best around $f_\theta:\mathbb R^d\to\mathbb R^d$00–$f_\theta:\mathbb R^d\to\mathbb R^d$01 for PSNR, replacement probability $f_\theta:\mathbb R^d\to\mathbb R^d$02 is optimal at approximately $f_\theta:\mathbb R^d\to\mathbb R^d$03, sampling steps are best at $f_\theta:\mathbb R^d\to\mathbb R^d$04–$f_\theta:\mathbb R^d\to\mathbb R^d$05, and $f_\theta:\mathbb R^d\to\mathbb R^d$06 rounds $f_\theta:\mathbb R^d\to\mathbb R^d$07 $f_\theta:\mathbb R^d\to\mathbb R^d$08 steps are recommended (Cheng et al., 19 Sep 2025).

In blind face restoration, DR-BFR reports on CelebA-Test that DR-BFR attains FID $f_\theta:\mathbb R^d\to\mathbb R^d$09 and NIQE $f_\theta:\mathbb R^d\to\mathbb R^d$10, both best, with LPIPS $f_\theta:\mathbb R^d\to\mathbb R^d$11 and PSNR $f_\theta:\mathbb R^d\to\mathbb R^d$12, SSIM $f_\theta:\mathbb R^d\to\mathbb R^d$13 (Qiu et al., 2024). The paper states that the next best SOTA methods, CodeFormer, RestoreFormer, and DiffBIR, obtain FID approximately $f_\theta:\mathbb R^d\to\mathbb R^d$14–$f_\theta:\mathbb R^d\to\mathbb R^d$15 and NIQE approximately $f_\theta:\mathbb R^d\to\mathbb R^d$16–$f_\theta:\mathbb R^d\to\mathbb R^d$17, and that on real-world sets DR-BFR consistently attains top-2 NIQE and FID (Qiu et al., 2024). Its ablation indicates that removing degradation representation worsens FID and NIQE, with DR-None at FID $f_\theta:\mathbb R^d\to\mathbb R^d$18, NIQE $f_\theta:\mathbb R^d\to\mathbb R^d$19, and the full DR-ALL at FID $f_\theta:\mathbb R^d\to\mathbb R^d$20, NIQE $f_\theta:\mathbb R^d\to\mathbb R^d$21 (Qiu et al., 2024).

In blind deblurring, Blind-PnPDM was tested on 100 FFHQ images under Gaussian kernels with random sigma in $f_\theta:\mathbb R^d\to\mathbb R^d$22 px and motion kernels with random length in $f_\theta:\mathbb R^d\to\mathbb R^d$23 px (Li et al., 28 May 2025). For Gaussian blur it reports 27.13 dB PSNR, 0.802 SSIM, and 0.180 LPIPS; for motion blur it reports 27.42 dB PSNR, 0.795 SSIM, and 0.176 LPIPS (Li et al., 28 May 2025). These exceed the listed baselines Pan-DCP, DeblurGAN-v2, BlindDPS, and GibbsDDRM on those test conditions (Li et al., 28 May 2025). GibbsDDRM itself, on FFHQ $f_\theta:\mathbb R^d\to\mathbb R^d$24 blind image deblurring with random motion kernels and additive Gaussian noise $f_\theta:\mathbb R^d\to\mathbb R^d$25, reports FID 38.71, LPIPS 0.115, and PSNR 25.80 (Murata et al., 2023). On vocal dereverberation it reports FAD 4.21, SI-SDR improvement $f_\theta:\mathbb R^d\to\mathbb R^d$26, and SRMR 8.40 (Murata et al., 2023).

In ultrasound denoising, the DDPM-based unsupervised method of “Deep Ultrasound Denoising Using Diffusion Probabilistic Models” reports at $f_\theta:\mathbb R^d\to\mathbb R^d$27 a PSNR of approximately 23.6 dB versus 22.0 dB for NLM and 22.4 dB for BM3D, with GCNR approximately 92.6% versus 91.6% and 90.4%, while preserving speckles (Asgariandehkordi et al., 2023). In blind BIQA, PFD-IQA reports PLCC 0.937 on KADID and 0.922 on LIVE-C, with overall average SRCC/PLCC of 0.931/0.935 versus previous SOTA 0.905/0.887 (Li et al., 2024).

7. Conceptual implications, misconceptions, and open directions

One recurring misconception is to equate BDDMs exclusively with blind image denoising. The cited literature shows a broader scope: BDDMs appear in generative sampling without explicit noise schedules (Kadkhodaie et al., 10 Feb 2026), self-supervised denoising (Demir et al., 31 Mar 2025, Cheng et al., 19 Sep 2025, Asgariandehkordi et al., 2023), blind restoration (Qiu et al., 2024), blind super-resolution (Xu et al., 2023), blind inverse problems (Li et al., 28 May 2025, Murata et al., 2023), and even feature-space refinement for BIQA (Li et al., 2024). This suggests that the essential ingredient is not denoising in the narrow pixel-space sense, but diffusion-based inference when the corruption descriptor is latent or partially observed.

A second misconception is that blind diffusion must sacrifice fidelity because it lacks explicit conditioning on the true degradation amplitude or operator. The theoretical work argues that schedule-free blind denoisers can outperform non-blind counterparts by correcting mismatch between the true residual noise of the image and the noise assumed by a prescribed schedule (Kadkhodaie et al., 10 Feb 2026). The empirical papers make analogous domain-specific claims: blind-spot guidance is introduced because pure BSNs sacrifice local detail and pure diffusion branches tend to blur (Cheng et al., 19 Sep 2025); degradation prompts in DR-BFR are introduced because diffusion models otherwise lack awareness of specific degradation and may produce unnatural details and inaccurate textures (Qiu et al., 2024); DiffDenoise uses blind-spot conditioning and symmetric-noise averaging to preserve high-frequency structures that self-supervised methods tend to over-smooth (Demir et al., 31 Mar 2025).

The limitations are equally consistent. The theoretical guarantees in (Kadkhodaie et al., 10 Feb 2026) require low intrinsic dimension and large ambient dimension, and the paper notes that quantifying intrinsic dimension for natural images remains an empirical challenge. Blind-PnPDM highlights higher per-sample cost because each outer iteration requires two diffusion runs and depends on well-trained diffusion models (Li et al., 28 May 2025). GibbsDDRM identifies SVD cost, sampling cost, and non-convexity of the joint posterior as practical difficulties (Murata et al., 2023). DR-BFR notes that unseen corruptions such as watermarks or real compression artifacts may not be fully captured because its Degradation Representation Module is trained on synthetic degradations (Qiu et al., 2024).

The trajectory of the field points toward richer blind conditioning and broader task coverage. The papers explicitly mention integrating blind denoisers into latent or text-conditioned models (Kadkhodaie et al., 10 Feb 2026), extending blind diffusion to deblurring, inpainting, and other inverse problems with unknown or spatially varying $f_\theta:\mathbb R^d\to\mathbb R^d$28 (Kadkhodaie et al., 10 Feb 2026), retraining degradation representation modules for deblurring, denoising, and super-resolution (Qiu et al., 2024), and applying blind plug-and-play diffusion to blind compressive sensing or unified joint annealing schemes (Li et al., 28 May 2025). Taken together, these works depict BDDMs not as a fixed model class but as a rapidly diversifying methodology for diffusion-based inference under hidden corruption structure.