Progressive Guidance Denoising Techniques

Updated 6 July 2026

Progressive Guidance Denoising is a staged restoration approach that decomposes image corruption into sequential, interpretable components for targeted removal.
It employs distinct denoisers and guidance signals—ranging from noise classifiers to frequency-based modules—to accurately refine image quality over iterative steps.
The framework improves recovery metrics such as PSNR and SSIM by aligning intermediate restoration states with progressive noise removal strategies.

Searching arXiv for the primary paper and closely related guidance-based denoising work. Progressive Guidance Denoising denotes a class of denoising strategies in which restoration is performed through an explicitly staged process, and each stage is guided by auxiliary information about what should be removed, refined, or preserved next. In its clearest formulation, the framework alternates between estimating the dominant corruption remaining in the current signal and applying a denoiser specialized to that corruption, rather than attempting to invert a complex degradation in a single pass. The term is directly exemplified by NoiseBreaker, described as a “Progressive Guidance Denoising framework,” which assumes an observed image corrupted by a sequential mixture of known primary noises and removes them iteratively under the control of a noise-analysis module (Lemarchand et al., 2020). Subsequent work extends the same coarse-to-fine logic to inter-frequency guidance, diffusion trajectory regularization, BEV feature denoising, low-dose PET and CT restoration, gradient-managed inverse problems, and even pseudo-label denoising in source-free adaptation (Jia, 2021, Ye et al., 27 Feb 2025, Jing et al., 2 Mar 2026, Liu et al., 2 Jul 2026, Wu et al., 9 Jul 2025, Bui-Tran et al., 29 Oct 2025).

1. Core definition and mathematical structure

In NoiseBreaker, the clean image is denoted by $x$ , and the observed image $y$ is modeled as a sequentially composed mixture of $m$ known primary noises (Lemarchand et al., 2020): $y \;=\; h(x) \;=\; x \;+\; n_1 \;+\; n_2 \;+\;\cdots+\;n_m\,,$ where each $n_i$ is drawn from a known primary distribution $\eta_i$ . The decomposition notation makes the iterative intent explicit: $y \;=\; x + \sum_{i=1}^m n_i \quad\Longrightarrow\quad u_k \;=\; x + \sum_{i=k+1}^m n_i \quad (k=0,1,\dots,m)\,,$ so that $u_0=y$ and, ideally, $u_m=x$ .

The denoising rule is equally explicit. Starting from $u_0=y$ , each iteration $y$ 0 computes

$y$ 1

where $y$ 2 is a Noise Analysis module that predicts the dominating remaining noise class, and $y$ 3 is the corresponding primary denoiser (Lemarchand et al., 2020). The process repeats until $y$ 4 predicts “clean,” or until a maximum number of steps.

This formulation distinguishes progressive guidance denoising from monolithic denoisers. The guidance signal is not merely an auxiliary feature map or a conditioning vector; it determines the next restoration operator. A plausible implication is that the method is especially suited to settings in which the degradation can be factorized into interpretable components, since the algorithmic sequence mirrors that factorization.

2. NoiseBreaker as the canonical progressive pipeline

NoiseBreaker alternates between two modules: a Noise Analysis module $y$ 5 and a Denoising module $y$ 6 (Lemarchand et al., 2020). The operational sequence is:

$y$ 7, the predicted noise class on the current image $y$ 8.
$y$ 9, the output of the primary denoiser for class $m$ 0.

The Noise Analysis module is designed to predict which primary noise type—and in some variants, which intensity range—is currently dominant. Its architecture uses MobileNetV2 pre-trained on ImageNet with input size $m$ 1, followed by two fully-connected layers, specifically “1024 units + ReLU, then $m$ 2 units + softmax,” where $m$ 3 is the total number of noise classes plus one “clean” class (Lemarchand et al., 2020). The output is a probability vector $m$ 4, with predicted label $m$ 5.

The classification loss is categorical cross-entropy,

$m$ 6

with $m$ 7 the one-hot ground truth. Training uses Adam, initial learning rate $m$ 8, halved every 50 epochs, 200 epochs total, and batch size 64. The training set contains 9 600 images from the ImageNet validation set, each synthetically noised according to one of the $m$ 9 classes, with validation and test sets of 200 images each (Lemarchand et al., 2020).

The pipeline terminates when the classifier predicts “clean,” yielding a stopping criterion that is itself learned rather than externally specified. In practice there are at most as many iterations as there are noise classes $y \;=\; h(x) \;=\; x \;+\; n_1 \;+\; n_2 \;+\;\cdots+\;n_m\,,$ 0, and termination often occurs earlier once the classifier labels the image “clean” (Lemarchand et al., 2020). This gives the method a diagnosis-like character: the framework not only restores but also exposes the sequence of inferred corruptions.

3. Tailored denoisers, empirical behavior, and limitations

For each class $y \;=\; h(x) \;=\; x \;+\; n_1 \;+\; n_2 \;+\;\cdots+\;n_m\,,$ 1, NoiseBreaker trains a denoiser $y \;=\; h(x) \;=\; x \;+\; n_1 \;+\; n_2 \;+\;\cdots+\;n_m\,,$ 2 on clean–noisy pairs $y \;=\; h(x) \;=\; x \;+\; n_1 \;+\; n_2 \;+\;\cdots+\;n_m\,,$ 3, where $y \;=\; h(x) \;=\; x \;+\; n_1 \;+\; n_2 \;+\;\cdots+\;n_m\,,$ 4 is $y \;=\; h(x) \;=\; x \;+\; n_1 \;+\; n_2 \;+\;\cdots+\;n_m\,,$ 5 corrupted only by that single primary noise (Lemarchand et al., 2020). The chosen architectures come “from an OpenDenoising benchmark”:

Noise class	Primary denoiser
Gaussian classes ( $y \;=\; h(x) \;=\; x \;+\; n_1 \;+\; n_2 \;+\;\cdots+\;n_m\,,$ 6 in three ranges)	MWCNN
Speckle classes ( $y \;=\; h(x) \;=\; x \;+\; n_1 \;+\; n_2 \;+\;\cdots+\;n_m\,,$ 7 in three ranges) and Uniform classes (two ranges)	SGN
Bernoulli and Poisson	SRResNet

Each primary denoiser is trained exactly as in its original paper, except that the training images come from the ImageNet subset and the noise parameters are sampled uniformly within the class’s predefined interval (Lemarchand et al., 2020). This modularity is central: the framework does not require a universal denoiser to learn all mixtures at once.

On 2 200 held-out images, Top-1 accuracy for the Noise Analysis module was 0.93 for grayscale and 0.91 for RGB; confusion matrices show that most errors occur between neighboring intensity ranges or low-noise vs. clean (Lemarchand et al., 2020). On BSD68 corrupted by six two-stage mixtures $y \;=\; h(x) \;=\; x \;+\; n_1 \;+\; n_2 \;+\;\cdots+\;n_m\,,$ 8, the reported quantitative results are given in average PSNR and SSIM over the 68 test images. In grayscale, NoiseBreaker is 2 dB above the classification-guided denoiser from [38], and 5 dB above BM3D and Noise2Void on average, with SSIM 0.85 versus 0.75 for the best competitor. In RGB, NoiseBreaker exceeds [38] by approximately 4.8 dB and BM3D by approximately 3 dB, with SSIM 0.81 versus 0.58 for BM3D and 0.59 for 38. Qualitative comparisons show that NoiseBreaker preserves edges and textures better than blind or monolithic methods.

The framework’s adaptability is also explicit. To integrate a new primary noise $y \;=\; h(x) \;=\; x \;+\; n_1 \;+\; n_2 \;+\;\cdots+\;n_m\,,$ 9, one adds a new output neuron in $n_i$ 0 and fine-tunes it, or re-trains the heads, on images corrupted by $n_i$ 1, and trains a dedicated denoiser $n_i$ 2 on synthetic pairs for that noise (Lemarchand et al., 2020). At the same time, the limitations are clear: the method assumes a sequential mixture model rather than spatially varying mixtures, requires knowledge of primary noise families and the ability to synthetically sample them, and uses a patch-based classifier input of $n_i$ 3, so very large or very small images may require tiling or resizing (Lemarchand et al., 2020).

A common misconception is that progressive guidance denoising is simply repeated application of the same denoiser. NoiseBreaker contradicts that interpretation: the procedure is class-selective, the denoisers are class-specific, and the controller can terminate early once “clean” is detected (Lemarchand et al., 2020).

4. Inter-frequency progressive guidance

A distinct formulation of progressive guidance appears in IGNet, which addresses AGWN denoising by exploiting the observation that low-frequency bands tend to preserve a higher PSNR than high-frequency bands under additive Gaussian white noise (Jia, 2021). The network therefore uses low-frequency content as guidance to refine more contaminated high-frequency bands.

IGNet first decomposes encoder features using iterative 2D DWT into one low-frequency subband $n_i$ 4 and three high-frequency subbands $n_i$ 5, $n_i$ 6, and $n_i$ 7. The Frequency-Progressive Feature Refiner then repeats this split for three levels, and at each level refines the current high-frequency tensor under the guidance of its corresponding low-frequency partner via a Low-frequency Guided Refiner (LGR) (Jia, 2021). The LGR uses two $n_i$ 8 convolutions, one on $n_i$ 9 and one on $\eta_i$ 0, channel-wise concatenation, another $\eta_i$ 1 convolution, and a residual addition back to $\eta_i$ 2. No additional attention mechanism is used; fusion is purely convolutional plus residual.

The progressive process is explicitly staged. After three levels of splitting and guided refinements, the subbands are merged back by inverse DWT in reverse order, yielding a reconstructed full-band feature $\eta_i$ 3, which is then decoded to a denoised image $\eta_i$ 4 (Jia, 2021). Training uses an $\eta_i$ 5 loss between $\eta_i$ 6 and clean ground truth, Adam with $\eta_i$ 7, $\eta_i$ 8, a learning rate starting at $\eta_i$ 9 with cosine annealing, and 120 epochs per noise level for $y \;=\; x + \sum_{i=1}^m n_i \quad\Longrightarrow\quad u_k \;=\; x + \sum_{i=k+1}^m n_i \quad (k=0,1,\dots,m)\,,$ 0.

Quantitatively, IGNet with $y \;=\; x + \sum_{i=1}^m n_i \quad\Longrightarrow\quad u_k \;=\; x + \sum_{i=k+1}^m n_i \quad (k=0,1,\dots,m)\,,$ 1 has approximately 0.87 MB of parameters and reports average PSNR and SSIM over Set12, BSD68, and Urban100 of $y \;=\; x + \sum_{i=1}^m n_i \quad\Longrightarrow\quad u_k \;=\; x + \sum_{i=k+1}^m n_i \quad (k=0,1,\dots,m)\,,$ 2 at $y \;=\; x + \sum_{i=1}^m n_i \quad\Longrightarrow\quad u_k \;=\; x + \sum_{i=k+1}^m n_i \quad (k=0,1,\dots,m)\,,$ 3, $y \;=\; x + \sum_{i=1}^m n_i \quad\Longrightarrow\quad u_k \;=\; x + \sum_{i=k+1}^m n_i \quad (k=0,1,\dots,m)\,,$ 4 at $y \;=\; x + \sum_{i=1}^m n_i \quad\Longrightarrow\quad u_k \;=\; x + \sum_{i=k+1}^m n_i \quad (k=0,1,\dots,m)\,,$ 5, and $y \;=\; x + \sum_{i=1}^m n_i \quad\Longrightarrow\quad u_k \;=\; x + \sum_{i=k+1}^m n_i \quad (k=0,1,\dots,m)\,,$ 6 at $y \;=\; x + \sum_{i=1}^m n_i \quad\Longrightarrow\quad u_k \;=\; x + \sum_{i=k+1}^m n_i \quad (k=0,1,\dots,m)\,,$ 7. IGNet+ with $y \;=\; x + \sum_{i=1}^m n_i \quad\Longrightarrow\quad u_k \;=\; x + \sum_{i=k+1}^m n_i \quad (k=0,1,\dots,m)\,,$ 8 has approximately 3.49 MB and reports $y \;=\; x + \sum_{i=1}^m n_i \quad\Longrightarrow\quad u_k \;=\; x + \sum_{i=k+1}^m n_i \quad (k=0,1,\dots,m)\,,$ 9, $u_0=y$ 0, and $u_0=y$ 1, while MWCNN with 23.8 MB achieves $u_0=y$ 2, $u_0=y$ 3, and $u_0=y$ 4 (Jia, 2021). Ablation shows that each extra wavelet level adds approximately 0.1 dB on BSD68 at $u_0=y$ 5, and separating $u_0=y$ 6 versus $u_0=y$ 7 plus LGR adds approximately 0.4 dB.

This suggests that progressive guidance denoising need not rely on explicit noise-class decisions. In IGNet, the guidance variable is spectral reliability: low-frequency components with more fidelity guide high-frequency recovery (Jia, 2021).

5. Guided diffusion as a progressive denoising paradigm

Diffusion-based restoration has generalized progressive guidance denoising from discrete denoiser selection to timestep-conditioned trajectory control. In BEVDiffuser, the object of denoising is a BEV feature map $u_0=y$ 8, and the guidance signal is a ground-truth object layout $u_0=y$ 9, embedded by a transformer-based Layout Fusion Module (Ye et al., 27 Feb 2025). Global conditioning adds a scene-summary token to the noisy BEV representation, while local object-aware conditioning injects cross-attention at each U-Net layer. The forward process follows standard Gaussian diffusion, and the reverse process uses classifier-free guidance with weight $u_m=x$ 0, so that layout-conditioned and unconditional predictions are combined at every denoising step. BEVDiffuser is trained to regress back to the original noisy BEV prediction $u_m=x$ 1 from a pretrained BEV encoder because the true clean BEV $u_m=x$ 2 is unknown. After being trained once on the feature maps of any base BEV model, it can be plugged in during training of a new BEV pipeline exactly as a supervisory head; at inference it is simply removed, preserving original latency (Ye et al., 27 Feb 2025).

On the nuScenes 3D-detection benchmark, BEVDiffuser improves four backbones: BEVFormer-tiny from NDS 35.5 to 39.1 and mAP 25.2 to 28.3; BEVFormer-base from 51.8 to 53.7 and 41.7 to 43.0; BEVFormerV2 from 41.1 to 44.7 and 32.7 to 37.1; and BEVFusion from 70.9 to 71.9 and 67.6 to 69.2 (Ye et al., 27 Feb 2025). The paper also reports notable improvements of 12.3\% in mAP and 10.1\% in NDS for 3D object detection without introducing additional computational complexity, with especially strong gains for long-tail objects and low-light or rainy scenarios.

MAP-Diff introduces a different mechanism: clinically observed intermediate-dose scans act as trajectory anchors in progressive 3D whole-body PET denoising (Jing et al., 2 Mar 2026). The reverse process is conditioned on the ultra-low-dose input $u_m=x$ 3, but training also imposes a timestep-weighted anchor loss on $u_m=x$ 4, where the anchor assigned to $u_m=x$ 5 is chosen from the intermediate-dose volumes $u_m=x$ 6. Anchor timesteps are calibrated by degradation matching between simulated diffusion corruption and real multi-dose PET pairs using NMAE and $u_m=x$ 7 (Jing et al., 2 Mar 2026). On the internal dataset, MAP-Diff improves PSNR from 42.48 dB to 43.71 dB, increases SSIM to 0.986, and reduces NMAE from 0.115 to 0.103 compared to 3D DDPM; on the external cohort it achieves 34.42 dB PSNR and 0.141 NMAE (Jing et al., 2 Mar 2026). Here, progressive guidance is imposed on the denoising path itself rather than only on the endpoint.

ProSAC-CT also makes the reverse process explicitly stage-aware (Liu et al., 2 Jul 2026). It formulates LDCT denoising as a residual conditional diffusion process on the RDDM backbone, with three specialized elements: anatomical-prior-guided conditioning (APGC), residual frequency-domain decoupling stage (RFDDS), and time-step-decoupling denoising decoder (TD $u_m=x$ 8). The 1 000 reverse steps are partitioned into early, middle, and late regimes, with each decoder specializing respectively in global anatomy, boundary finesse, and fine-detail cleanup (Liu et al., 2 Jul 2026). The details describe this as a smooth, coarse-to-fine trajectory that preserves global CT intensities, sharpens anatomical boundaries, and recovers fine detail without oversmoothing or hallucination.

SPGD targets a different failure mode in diffusion restoration: instability in the interaction between the prior score and likelihood guidance (Wu et al., 9 Jul 2025). The method identifies conflicts between prior and likelihood gradient directions and temporal fluctuations in the likelihood gradient, then introduces a progressive likelihood warm-up strategy and adaptive directional momentum smoothing. Instead of one large likelihood-gradient step after each denoising update, SPGD splits the step into $u_m=x$ 9 smaller steps before denoising, with effective likelihood weight progressively ramped from $u_0=y$ 0 to $u_0=y$ 1 in $u_0=y$ 2 increments of size $u_0=y$ 3 (Wu et al., 9 Jul 2025). On FFHQ 256×256 and ImageNet 256×256 under random inpainting, Gaussian deblur, motion deblur, and $u_0=y$ 4 super-resolution, SPGD achieves the best or second-best metrics across tasks, and on FFHQ random inpainting reports PSNR 30.87 dB versus the next best 28.64 dB (Wu et al., 9 Jul 2025).

Taken together, these diffusion-based works broaden progressive guidance denoising from “what noise should be removed next” to “what state should the reverse trajectory approach next,” “what structure should condition each step,” and “how should guidance strength evolve across timesteps.”

The underlying principle of progressive guidance has also been extended beyond conventional image denoising. UP2D treats pseudo-label corruption in source-free domain adaptive medical image segmentation as a denoising problem and progressively improves pseudo-label quality through Refined Prototype Filtering, Uncertainty-Guided EMA, and quantile-based entropy minimization (Bui-Tran et al., 29 Oct 2025). The framework is explicitly described as Uncertainty-aware Progressive Pseudo-label Denoising. In the SFDA setting, it reports, for example, on Drishti-GS, optic-disc Dice $u_0=y$ 5 and optic-cup Dice $u_0=y$ 6, and on RIM-ONE-r3, disc Dice $u_0=y$ 7 and cup Dice $u_0=y$ 8 (Bui-Tran et al., 29 Oct 2025). This suggests that progressive guidance denoising has become a broader methodological motif for suppressing unreliable signal components under staged control.

Several misconceptions recur across this literature. One is that guidance necessarily means external labels or classifiers. The record is more heterogeneous. NoiseBreaker uses a noise classifier (Lemarchand et al., 2020); IGNet uses low-frequency bands as internal guidance (Jia, 2021); BEVDiffuser uses ground-truth object layout (Ye et al., 27 Feb 2025); MAP-Diff uses clinically observed intermediate-dose anchors (Jing et al., 2 Mar 2026); ProSAC-CT uses anatomical priors and spectral decomposition (Liu et al., 2 Jul 2026); SPGD uses managed likelihood gradients (Wu et al., 9 Jul 2025). Another misconception is that progressive methods are only relevant when the final endpoint is poor. In MAP-Diff and SPGD, the intermediate trajectory itself is a primary object of design, regularization, or stabilization (Jing et al., 2 Mar 2026, Wu et al., 9 Jul 2025).

The broader significance of progressive guidance denoising lies in its replacement of single-shot inversion by structured restoration paths. In the original NoiseBreaker formulation, that path is a sequence of primary denoisers selected by noise analysis (Lemarchand et al., 2020). In later systems, it becomes a sequence of guided frequency refinements, trajectory anchors, stage-specific decoders, or controlled gradient updates (Jia, 2021, Jing et al., 2 Mar 2026, Liu et al., 2 Jul 2026, Wu et al., 9 Jul 2025). A plausible implication is that the enduring research question is not only how to denoise, but how to specify the intermediate states that make denoising reliable, interpretable, and compatible with downstream objectives.