Scale-Aware Degradation Kernels

• Scale-aware degradation kernels are operators that adapt their form based on scale factors, linking HR-to-LR downsampling, kernel support, and restoration behavior.
• They integrate classical image formation models with modern techniques like conditioned convolutions and diffusion modules to handle various degradations.
• They mitigate issues such as the larger-kernel effect by using structural priors like low-rank regularization and support continuous adaptation across different degradation scenarios.

Searching arXiv for the cited papers to ground the article in current records. Scale-aware degradation kernels are degradation operators whose form depends on scale, where “scale” may denote the HR-to-LR downsampling factor, the spatial support chosen for a blur kernel, or the resolution level at which a restoration network senses and inverts degradation. In the strict image-formation sense, the kernel appears in models such as $I^{LR} = (I^{HR} \otimes k)\downarrow_s + n$; in a broader contemporary sense, conditioned convolutions, attention blocks, and diffusion modules realize families of effective operators indexed by scale and related degradation descriptors. Across blind super-resolution, variable-rate compression-aware super-resolution, and unified restoration, the common premise is that degradation modeling should not treat scale as incidental: the same kernel is generally not appropriate across different super-resolution factors, and a fixed operator can be mismatched even when the nominal task remains unchanged (Do et al., 18 Jul 2025, Chai et al., 18 Mar 2026).

1. Formal scope of scale awareness

In the classical blind SR formulation, the degradation process is written as

$I^{LR} = (I^{HR} \otimes k)\downarrow_s + n,$

where $k$ is a blur kernel, $\downarrow_s$ is downsampling by scale factor $s$, and $n$ is additive noise. Within this formulation, scale awareness means that the effective degradation should depend jointly on $k$ and $s$, rather than treating $k$ as reusable across $\times 2$, $I^{LR} = (I^{HR} \otimes k)\downarrow_s + n,$0, or other settings. The reference-based blind SR work makes this point explicit, stating that degradation kernels “should account for not only the degradation process but also the downscaling factor,” and that applying the same degradation kernel across varying super-resolution scales may be impractical (Do et al., 18 Jul 2025).

A second meaning of scale awareness concerns kernel support itself. In blind deconvolution, the chosen kernel size $I^{LR} = (I^{HR} \otimes k)\downarrow_s + n,$1 defines the support domain over which blur is estimated. In that setting, “scale” is the spatial extent of the kernel, and the central question is whether an oversized support introduces spurious degrees of freedom. The blind deconvolution analysis shows that it does, even for noise-free blurry images, so size-awareness is not merely a modeling convenience but a structural requirement (Si-Yao et al., 2017).

A third meaning appears in recent conditional restoration networks, where scale-aware behavior is implemented without explicit PSF estimation. In DACG-IR, degradation characteristics are extracted by multi-scale branches with different depth-wise kernel sizes $I^{LR} = (I^{HR} \otimes k)\downarrow_s + n,$2, and these responses are turned into layer-wise prompts that modulate attention, frequency-domain filtering, and feature aggregation across the hierarchy (He et al., 2 May 2026). In ASSR-EIC, the only explicit kernel is bicubic downsampling, while the decoder learns implicit, learned, scale-aware degradation operators by conditioning on a continuous rescaling factor $I^{LR} = (I^{HR} \otimes k)\downarrow_s + n,$3, a codec quality indicator $I^{LR} = (I^{HR} \otimes k)\downarrow_s + n,$4, and a codec type indicator $I^{LR} = (I^{HR} \otimes k)\downarrow_s + n,$5 (Chai et al., 18 Mar 2026).

Taken together, the literature uses “scale-aware degradation kernel” in two compatible senses: an explicit image-formation kernel whose shape depends on scale, and an implicit effective operator whose behavior changes with scale-conditioned latent variables.

2. Kernel support, oversized models, and the larger-kernel effect

The most direct treatment of scale as kernel size appears in blind deconvolution. The observation model is

$I^{LR} = (I^{HR} \otimes k)\downarrow_s + n,$6

with MAP estimation

$I^{LR} = (I^{HR} \otimes k)\downarrow_s + n,$7

When convolution is rewritten as $I^{LR} = (I^{HR} \otimes k)\downarrow_s + n,$8, increasing kernel size enlarges the number of columns of $I^{LR} = (I^{HR} \otimes k)\downarrow_s + n,$9. The analysis identifies an “inflating effect”: adding linearly independent columns strictly enlarges the solution subspace, so the least-squares residual decreases while the estimate acquires extra components that correspond to noise directions rather than meaningful blur structure (Si-Yao et al., 2017).

This mechanism explains why oversized kernels degrade estimation quality. The effect persists in two regimes analyzed in the paper. In the first, the sharp image estimate is assumed perfect and kernel error is driven by explicit noise through $k$0; empirical singular-value analysis shows that the error grows sharply with kernel size. In the second, the blurry image is noise-free but the current sharp estimate contains error, so $k$1 acts as implicit noise; again, the kernel error operator worsens as support grows (Si-Yao et al., 2017).

To suppress this larger-kernel effect, the paper replaces simple element-wise priors with a low-rank kernel regularizer based on the log-determinant surrogate

$k$2

where $k$3 are singular values. The motivation is structural rather than merely sparse: real blur kernels are smooth and highly correlated across rows and columns, whereas noisy oversized estimates exhibit a broad singular-value tail. The resulting kernel update penalizes high-rank contamination in the extended support and remains robust when kernel size is substantially larger than the true support (Si-Yao et al., 2017).

This line of work establishes a basic principle that later SR literature generalizes: scale awareness is not only about the upsampling factor, but also about choosing or regularizing the support on which degradation is represented.

3. Explicit scale-specific kernels in blind super-resolution

The most explicit argument that degradation kernels should vary with super-resolution scale is given by the reference-based blind SR framework. Its image-formation model is

$k$4

The paper argues that “The degradation processes for an LR image derived from HR images with different resolutions should be distinct. This suggests that the corresponding blur kernel from HR images downsampled to varying resolutions should also differ.” On this view, a single kernel $k$5 shared across $k$6 and $k$7 is generally mismatched (Do et al., 18 Jul 2025).

The framework operationalizes this claim by learning a scale-specific HR-LR cycle. The upsampling branch $k$8 uses DASR and produces an implicit degradation representation $k$9, while the downsampling branch $\downarrow_s$0 is a KernelGAN-like deep linear convolutional network whose weights define an explicit kernel $\downarrow_s$1 with effective receptive field $\downarrow_s$2. Scale awareness is enforced structurally: different target scales use different $\downarrow_s$3, the downsampler is trained to be the inverse of that specific $\downarrow_s$4, and HR reference images are chosen at the target HR scale (Do et al., 18 Jul 2025).

A key regularizer is

$\downarrow_s$5

which aligns the implicit degradation code of the target LR image with the code of synthetic LR images obtained by downsampling content-irrelevant HR references. Because the references share degradation but not content, the model is forced to isolate the scale-specific degradation process rather than transfer textures or semantic correspondences (Do et al., 18 Jul 2025).

The empirical behavior reinforces the scale-specific interpretation. The paper reports that replacing the learned downsampler with the ground-truth kernel in the $\downarrow_s$6 anisotropic setting raises performance from 31.161 dB / 0.8683 to 33.144 dB / 0.8994, indicating that accurate scale-aware kernels are a major performance bottleneck. The contribution of the method is therefore not only better super-resolution, but also a reframing of kernel estimation: not one kernel per LR image in isolation, but a scale-tied inverse pair $\downarrow_s$7 learned through cycle consistency and reference-based degradation alignment (Do et al., 18 Jul 2025).

4. Implicit degradation representations and dynamic kernel families

A large part of the recent literature does not regress an explicit blur kernel at inference time. Instead, it learns latent degradation representations that control dynamic convolutions, attention, or forward/inverse degradation networks. In this setting, “kernel” refers to the effective operator induced by conditioned layers.

MRDA is an early formulation of this idea. It adopts the same classical degradation model but introduces a Meta-Learning Network, a Degradation Extraction Network, and a Region Degradation Aware SR Network. The central object is an implicit degradation representation (IDR), denoted by $\downarrow_s$8 in spatial form and $\downarrow_s$9 in compressed form. RDAN then uses $s$0 to generate dynamic convolution weights $s$1 and combines them with region-wise modulation: $s$2 Because the degradation process $s$3 includes blur, noise, JPEG compression, downsampling scheme, and implicitly the scale factor, the IDR functions as a latent scale-aware kernel descriptor rather than an explicit PSF (Xia et al., 2022).

DSAT pushes the same idea into a CNN-Transformer hybrid. It uses contrastive learning to obtain a content-invariant degradation representation $s$4, then injects $s$5 into a degradation-aware CNN layer and a degradation-aware Swin self-attention layer. The DCL generates depth-wise convolution kernels $s$6 and channel coefficients $s$7, while the attention layer modulates the value pathway via degradation-conditioned weights: $s$8 Separate models are trained for $s$9, $n$0, and $n$1, with different kernel-width ranges per scale, so the learned degradation manifold is scale-specific even though no explicit scale token is concatenated to the input. The paper reports 32.43 dB on Urban100 at $n$2, 0.94 dB higher than DASR, and 26.62 dB on Urban100 at $n$3, a 0.26 dB improvement over KDSR (Liu et al., 2023).

DADiff generalizes the notion of a kernel even further by replacing the known linear degradation operator $n$4 and its pseudo-inverse $n$5 with learned networks. An encoder $n$6 extracts a spatially varying degradation representation $n$7; a degradation model $n$8 approximates the forward operator; and a restoration model $n$9 approximates the pseudo-inverse. These are integrated into DDNM-style diffusion guidance through a neural correction of the form

$k$0

The paper does not estimate explicit blur kernels, but it directly compares explicit and implicit operator models on DIV2K-Val at $k$1: the fully implicit setting reaches PSNR 24.47, SSIM 0.621, and LPIPS 0.307, whereas the explicit Gaussian-kernel setting gives PSNR 20.96, SSIM 0.400, and LPIPS 0.599 (Lu et al., 15 Jan 2025).

A common misconception is that scale-aware degradation kernels must be explicit, low-dimensional blur maps. These works show a different possibility: scale can be encoded in task-specific training, degradation manifolds, dynamic convolutions, or diffusion guidance, and the resulting effective operators can be more expressive than a single global kernel.

5. Continuous-scale and rate-adaptive degradation operators

The most direct extension from discrete scales to continuous scale conditioning appears in extreme image compression with arbitrary-scale super-resolution. ASSR-EIC models degradation as HR downsampling by an arbitrary factor $k$2, followed by codec compression parameterized by codec type $k$3 and bitrate-based quality parameter $k$4. The encoder side uses explicit bicubic interpolation for $k$5, but the decoder side uses a diffusion-based joint degradation-aware ASSR decoder conditioned on $k$6, $k$7, $k$8, and $k$9 (Chai et al., 18 Mar 2026).

The paper’s conditioning mechanism makes the scale dependence explicit. Each scalar parameter is embedded by positional encoding plus MLP, and the resulting vectors are summed: $s$0 This global encoding embedding is injected into ResNet blocks in the latent UNet backbone and in the fidelity module. In parallel, a local compression-rescaling modulator applies FiLM-style channel gains and biases to multiscale fidelity features: $s$1 with $s$2 and $s$3 derived from embeddings of codec type, codec quality, and scale (Chai et al., 18 Mar 2026).

Under a broad operator view, this means that the decoder does not use one super-resolution kernel. It uses a continuous family of effective operators indexed by $s$4. During training, $s$5 is sampled uniformly from $s$6; at inference, any float $s$7 can be used. The model therefore learns a continuous operator manifold rather than a bank of discrete kernels for $s$8, $s$9, or $k$0 alone (Chai et al., 18 Mar 2026).

This formulation also broadens the meaning of degradation beyond blur. Bitrate reduction is partly controlled by changing $k$1, and partly by codec quality. The resulting “scale-aware degradation kernels” are therefore simultaneously scale-aware and bitrate-aware. The paper reports that one single diffusion model can handle many bitrates, and qualitatively shows a shift from fidelity-oriented behavior at mild degradation to more generative behavior as $k$2 grows and information loss becomes severe (Chai et al., 18 Mar 2026).

ASSR-EIC is important because it shows that scale-aware degradation modeling need not stop at SR scale factors. Scale can be coupled to compression severity, codec family, and semantic guidance, and the decoder can still be interpreted as realizing a degradation-conditioned kernel family.

6. Unified restoration, hierarchical prompts, and open problems

In unified image restoration, scale awareness is tied to network hierarchy rather than only to SR factor. DACG-IR uses a Degradation-Aware Module that extracts shallow features and processes them through $k$3 parallel branches with different depth-wise kernel sizes,

$k$4

followed by fusion, spatial gating, and dual-statistic pooling. The resulting global degradation code $k$5 and layer-wise prompts $k$6 modulate the rest of the architecture (He et al., 2 May 2026).

The prompts control Context Adaptive Gated Attention by changing attention temperature and attention-output gating: $k$7 At the bottleneck, the Context Gated Dual-Domain Modulation module converts $k$8 into a frequency-domain mask $k$9, yielding a degradation-conditioned spectral filter. Skip connections are also filtered by a spatial-channel dual-gated adaptive fusion mechanism. In effect, DACG-IR distributes degradation-aware filtering across local spatial kernels, attention-based aggregation, global spectral masks, and skip-level gating, all aligned to different scales in the encoder-decoder hierarchy (He et al., 2 May 2026).

This suggests a broader taxonomy. Explicit scale-aware kernels remain most interpretable when the task is close to classical image formation, as in blur estimation or reference-based blind SR. Implicit scale-aware kernels become attractive when degradations are heterogeneous, spatially varying, nonlinear, or entangled with codec artifacts, prompting, or adverse weather. The cited works repeatedly trade interpretability for flexibility: MRDA and DSAT do not decode explicit kernels from their latent codes; DADiff replaces $\times 2$0 and $\times 2$1 with learned operators; ASSR-EIC models only bicubic downsampling explicitly and learns the rest end-to-end; DACG-IR extracts degradation cues without kernel supervision (Xia et al., 2022, Lu et al., 15 Jan 2025, Chai et al., 18 Mar 2026, He et al., 2 May 2026).

Several unresolved issues remain visible across the literature. First, explicit disentanglement between scale and other degradation variables is often absent; scale may remain entangled with blur, noise, or codec artifacts in the latent code. Second, many methods are still discrete-scale in practice, either by training separate models per scale or by relying on task-specific backbones, whereas arbitrary continuous scales are explicitly supported only in a subset of frameworks. Third, when explicit kernels are estimated, support selection itself is unstable: oversized supports can corrupt estimation unless structural priors such as low-rank regularization are imposed (Si-Yao et al., 2017). Finally, when scale-aware behavior is implemented through diffusion models, runtime remains a practical limitation, and the learned operator manifold is only implicitly characterized rather than analytically understood (Chai et al., 18 Mar 2026).

Scale-aware degradation kernels therefore designate not one technique but a research direction. Its unifying thesis is that degradation inversion should be conditioned on the scale at which information was lost, whether that scale is the support of a blur PSF, the HR-to-LR sampling factor, the bitrate-induced loss level, or the receptive-field level at which restoration features are processed.