Frequency-Decoupled Diffusion Guidance for Non-Blind Image Deblurring

Sihan Wang; Jinshu Huang; Haibin Su; Yunhua Xue

cs.CVOct 5, 2026

Frequency-Decoupled Diffusion Guidance for Non-Blind Image Deblurring

Authors: Sihan Wang, Jinshu Huang, Haibin Su, Yunhua Xue

Organizations: School of Mathematical Sciences, Nankai University, Tianjin, China. · Yau Mathematical Sciences Center, Tsinghua University, Beijing, China.

Abstract

Pretrained diffusion models provide powerful image priors for training-free posterior sampling in image restoration. To guide this sampling process, frequency-aware methods progressively incorporate measurement information across frequency bands, facilitating coarse-to-fine reconstruction. However, existing methods typically do not explicitly separate frequency activation from degradation-induced attenuation, leaving attenuation differences among inactive frequencies insufficiently modeled. In this work, we propose frequency-decoupled posterior guidance to separate frequency activation from attenuation-aware spectral regularization. Specifically, a progressive low-to-high frequency schedule determines the active measurement band, while a kernel-derived attenuation map defines a selective spectral prior over inactive components. To stabilize the sampling process, we also introduce a local trajectory regularizer that suppresses spatially irregular state-to-clean deviations. For a fixed endpoint energy, we provide a KL-regularized path-space interpretation. In practice, we construct time-dependent guidance through local energy corrections using a Tweedie plug-in approximation. Experiments on natural-image benchmarks demonstrate strong PSNR and SSIM performance across challenging non-blind deblurring settings, even at higher measurement noise levels.

Figures & tables

	Gaussian Blur		Motion Blur		Uniform Blur
Method	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$
FFHQ Dataset
DPS ( Chung et al., 2023 )	25.253	0.702	22.582	0.619	23.697	0.645
DSG ( Yang et al., 2024 )	27.299	0.778	25.960	0.749	26.548	0.746
FGPS ( Thaker et al., 2025 )	27.380	0.780	26.990	0.770	26.961	0.757
DPS-MAP ( Li and Wang, 2025 )	29.274	0.836	29.260	0.834	28.858	0.817

Table 1: Main fidelity comparison on FFHQ and ImageNet under three deblurring settings. Higher PSNR and SSIM are better. Bold and underlined values indicate the best and second-best results, respectively.

Figure 1: Qualitative comparison on FFHQ and ImageNet under Gaussian, motion, and uniform deblurring. Columns show the ground truth, measurement, DPS, FGPS, DPS-MAP, and our method. Red dashed boxes highlight matched local regions for visual comparison. Additional examples and enlarged crops are provided in the supplementary material.

Variant	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Full	29.133	0.8368	0.1322
Full $-\mathcal{P}_{t}$	29.084	0.8366	0.1342
Full w/ $G_{t}\equiv 1$	29.084	0.8365	0.1343
Full $-\mathcal{R}_{t}$	24.446	0.5659	0.3243

Table 2: Ablation of the main guidance components on the fixed 50-image FFHQ motion-deblurring subset using a single DDPM branch.

	Gaussian		Motion		Uniform
Method	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$
DPS	24.540	0.6758	21.784	0.5893	23.010	0.6204
DSG	26.893	0.7596	25.432	0.7215	26.077	0.7260
FGPS	26.736	0.7528	25.626	0.7216	25.916	0.7197
DPS-MAP	28.364	0.8095	27.396	0.7836	27.414	0.7768
DMPS	28.444	0.8120	27.367	0.7842	27.483	0.7792

Table 3: Quantitative comparison on FFHQ under stronger measurement noise

\sigma_{y}=0.10

. Each result is averaged over

1{,}000

FFHQ evaluation images, with identical degraded measurements used across methods within each task.

Appendix figures & tables19 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	Task and branch	$\tau_{\mathrm{start}}$	$\tau_{\mathrm{end}}$	$\lambda_{\mathrm{lp}}$	$\lambda_{\mathrm{prior}}$	$\lambda_{\mathrm{reg}}$	$(s_{\mathrm{early}},s_{\mathrm{late}})$	$w$
FFHQ	Gaussian DDPM/DDIM	10	75	1.30	0.0030	0.0050	$(2.5,0.60)$	0.50
FFHQ	Motion DDPM/DDIM	8	125	2.20	0.0050	0.0050	$(2.5,0.60)$	0.50
FFHQ	Uniform CGJ-DDPM	8	115	1.20	0.0005	0.0030	$(2.5,0.60)$	0.50
FFHQ	Uniform DDIM	8	115	1.20	0.0005	0.0030	$(2.5,0.60)$	0.50
ImageNet	Gaussian DDPM/DDIM	10	75	1.00	0.0010	0.0030	$(2.0,0.50)$	0.70
ImageNet	Motion DDPM	8	105	2.20	0.00125	0.0050	$(2.5,0.60)$	0.50

Appendix

Table 4: Task-specific configurations. A combined DDPM/DDIM row means that both branches use the listed guidance parameters. For uniform deblurring, CGJ is enabled only in the CGJ-DDPM branch. The weight

w

applies to the DDPM or CGJ-DDPM output; the DDIM output receives

1-w

. It is repeated on separate branch rows for clarity.

	Gaussian Blur		Motion Blur		Uniform Blur
Method	LPIPS $\downarrow$	FID $\downarrow$	LPIPS $\downarrow$	FID $\downarrow$	LPIPS $\downarrow$	FID $\downarrow$
FFHQ Dataset
DPS ( Chung et al., 2023 )	0.150	31.161	0.197	34.278	0.175	32.164
DSG ( Yang et al., 2024 )	0.104	26.155	0.130	28.557	0.134	28.609
FGPS ( Thaker et al., 2025 )	0.093	23.593	0.103	24.670	0.116	26.136
DPS-MAP ( Li and Wang, 2025 )	0.236	61.684	0.195	67.177	0.241	69.985

Appendix

Table 5: Complementary perceptual comparison on FFHQ and ImageNet under three deblurring settings. Lower LPIPS and FID are better. Bold and underlined values indicate the best and second-best results, respectively.

Variant	PSNR $\uparrow$	$\Delta$ PSNR	SSIM $\uparrow$	LPIPS $\downarrow$
Full	29.133	–	0.8368	0.1322
w/o attenuation weighting	29.138	+0.005	0.8367	0.1318
Fixed $G_{t}$	29.133	+0.000	0.8368	0.1325
w/o spectral prior	29.084	-0.049	0.8366	0.1342
w/o trajectory regularization	24.446	-4.687	0.5659	0.3243
w/o frequency decoupling	29.084	-0.049	0.8365	0.1343

Appendix

Table 6: Complete component ablation on the fixed 50-image FFHQ motion-deblurring subset using a single DDPM branch.

\Delta

PSNR is measured relative to the full configuration.

Dataset	Configuration	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
FFHQ	Previous	29.404	0.8380	0.1569
FFHQ	Unified	29.447	0.8392	0.1534
ImageNet	Previous	26.382	0.7606	0.2948
ImageNet	Unified	26.415	0.7627	0.2913

Appendix

Table 7: Effect of the unified spectral prior and simplified motion configuration on the full 1000-image evaluation sets.

Variant	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Base two-branch	29.108	0.8252	0.1836
CGJ two-branch	29.160	0.8268	0.1757

Appendix

Table 8: Effect of the operator-specific CGJ correction on FFHQ uniform deblurring over the full 1000-image evaluation set.

Dataset	Task	Ours-DDPM	Ours-DDIM	Ours	Fusion weights
FFHQ	Gaussian	29.095 / 0.8308 / 0.1741	29.236 / 0.8340 / 0.2127	29.528 / 0.8423 / 0.1943	0.50 / 0.50
FFHQ	Motion	28.954 / 0.8319 / 0.1378	28.799 / 0.8168 / 0.1893	29.447 / 0.8392 / 0.1534	0.50 / 0.50
FFHQ	Uniform	28.775 / 0.8187 / 0.1621	28.216 / 0.8025 / 0.2011	29.160 / 0.8268 / 0.1757	0.50 / 0.50
ImageNet	Gaussian	25.463 / 0.7234 / 0.3867	25.333 / 0.7148 / 0.4103	25.500 / 0.7249 / 0.3936	0.70 / 0.30
ImageNet	Motion	26.256 / 0.7570 / 0.2889	26.133 / 0.7509 / 0.2995	26.415 / 0.7627 / 0.2913	0.50 / 0.50
ImageNet	Uniform	25.593 / 0.7165 / 0.3204	25.193 / 0.6959 / 0.3972	25.869 / 0.7240 / 0.3324	0.55 / 0.45

Appendix

Table 9: Single-branch and two-branch reconstruction results on the full 1000-image evaluation sets. Each entry reports PSNR / SSIM / LPIPS. The two fusion weights correspond to the first and second branch, respectively. For uniform deblurring, the Ours-DDPM column corresponds to the CGJ-DDPM branch.

Dataset	Task	Baseline	$\Delta$ PSNR	$\Delta$ SSIM	$\Delta$ LPIPS
FFHQ	Gaussian	DPS-MAP	$+0.253\,[+0.240,+0.266]$	$+0.0058\,[+0.0055,+0.0062]$	$-0.0418\,[-0.0428,-0.0407]$
FFHQ	Gaussian	DMPS	$+0.149\,[+0.133,+0.165]$	$+0.0030\,[+0.0027,+0.0033]$	$-0.0253\,[-0.0264,-0.0241]$
FFHQ	Motion	DPS-MAP	$+0.182\,[+0.169,+0.196]$	$+0.0051\,[+0.0046,+0.0057]$	$-0.0418\,[-0.0430,-0.0407]$
FFHQ	Motion	DMPS	$+0.221\,[+0.199,+0.243]$	$+0.0043\,[+0.0037,+0.0048]$	$-0.0350\,[-0.0361,-0.0338]$
FFHQ	Uniform	DPS-MAP	$+0.302\,[+0.286,+0.318]$	$+0.0101\,[+0.0096,+0.0106]$	$-0.0652\,[-0.0666,-0.0637]$
FFHQ	Uniform	DMPS	$+0.224\,[+0.209,+0.239]$	$+0.0079\,[+0.0074,+0.0083]$	$-0.0551\,[-0.0564,-0.0537]$

Appendix

Table 10: Paired statistical comparison with DPS-MAP and DMPS on the full 1000-image evaluation sets. Each entry reports the mean per-image difference and its Student-

t

95% confidence interval. We define

\Delta=\mathrm{Ours}-\mathrm{Baseline}

; positive values favor our method for PSNR and SSIM, whereas negative values favor our method for LPIPS.

Figure 2: Hyperparameter sensitivity on the FFHQ motion-deblurring Dev50 set. Left: sensitivity to the attenuation parameter

\rho

, with

\rho=10^{-3}

as the default. Right: sensitivity to the spectral-prior strength, shown as a multiplier of the default

\lambda_{\mathrm{prior}}=0.005

. In both panels, all other parameters are fixed to the default configuration. Error bars denote paired Student-

t

95% confidence intervals of the PSNR difference relative to the default setting, shifted to the corresponding absolute PSNR levels.

Figure 3: Additional qualitative comparisons on FFHQ under Gaussian, motion, and uniform deblurring at $\sigma_{\mathrm{y}}=0.05$ . Two additional Full1000 examples are shown for each degradation setting. Columns show the ground truth, measurement, DPS, FGPS, DPS-MAP, and our method.

Figure 4: Additional qualitative comparisons on ImageNet under Gaussian, motion, and uniform deblurring at $\sigma_{\mathrm{y}}=0.05$ . Two additional Full1000 examples are shown for each degradation setting. Columns show the ground truth, measurement, DPS, FGPS, DPS-MAP, and our method.

Figure 5: Zoomed comparisons on FFHQ at $\sigma_{\mathrm{y}}=0.05$ . The examples cover Gaussian, motion, and uniform deblurring. For each row, all methods are shown using the same spatial crop. The uniform-deblurring comparison uses CGJ-DDPM as the first Ours branch.

Figure 6: Zoomed comparisons on ImageNet at $\sigma_{\mathrm{y}}=0.05$ . The examples cover Gaussian, motion, and uniform deblurring. For each row, all methods are shown using the same spatial crop. For uniform deblurring, the first Ours branch is CGJ-DDPM.

Figure 7: Absolute reconstruction error on representative FFHQ examples. Brighter regions indicate larger pixel-wise deviation from the ground truth. All methods are displayed using the same error scale within each example.

Figure 8: Absolute reconstruction error on representative ImageNet examples. Brighter regions indicate larger pixel-wise deviation from the ground truth. All methods are displayed using the same error scale within each example.

Figure 9: Kernel-dependent spectral attenuation weights $W_{\kappa}$ for Gaussian, motion, and uniform blur with $\rho=10^{-3}$ . Larger values correspond to frequencies that are more strongly attenuated by the blur kernel. The Gaussian kernel produces a radially structured attenuation pattern, the motion kernel produces a directional pattern, and the uniform kernel produces a periodic spectral structure.

Figure 10: Evolution of $G_{t}$ and $M_{t}$ during reverse sampling for Gaussian deblurring. As the reverse process progresses, the support of $G_{t}$ expands from low to higher frequencies, while the support of $M_{t}=(1-G_{t})W_{\kappa}$ correspondingly contracts. The rows show the frozen branch-specific schedules used for FFHQ and ImageNet.

Figure 11: Evolution of $G_{t}$ and $M_{t}$ during reverse sampling for motion deblurring. The directional structure inherited from $W_{\kappa}$ remains visible in the inactive-band weighting $M_{t}$ , while the active measurement region progressively expands. The rows correspond to the frozen branch-specific schedules used for FFHQ and ImageNet.

Figure 12: Evolution of $G_{t}$ and $M_{t}$ during reverse sampling for uniform deblurring. The periodic attenuation structure of the underlying uniform kernel modulates the inactive-frequency prior through $M_{t}=(1-G_{t})W_{\kappa}$ . For the uniform task, the first Ours branch is CGJ-DDPM and the second branch is DDIM.

Method	FFHQ-G	FFHQ-M	FFHQ-U	ImageNet-G	ImageNet-M	ImageNet-U
DPS	1000	1000	1000	1000	1000	1000
DSG	1000	1000	1000	1000	1000	1000
FGPS	1000	1000	1000	1000	1000	1000
DPS-MAP	1000	1000	1000	100	100	100
DMPS	100	100	100	100	1000	1000
Ours	2000	2000	2150	2000	2000	2150

Appendix

Table 11: Number of diffusion-model UNet evaluations per reconstructed image. The final Ours result combines two independently completed sampling branches. For uniform deblurring, the CGJ-DDPM branch requires additional Tweedie predictions during the operator-specific correction.

Explore similar work

Stop Denoising Your Blurs
May 24, 2026Sasidhar Parvathireddy, Vamsidhar Saraswathula, Rama Krishna GorthiGaussian Deblurring Image Restoration
Indian Institute of Technology Tirupati, India.
FSM-Net: An Efficient Frequency-Spatial Network for Real-World Deblurring
May 29, 2026Vinh-Thuan LyGaussian Deblurring Real-World
University of Science, VNU-HCM, Ho Chi Minh City, Vietnam · Vietnam National University, Ho Chi Minh City, Vietnam
CogSENet: Blind Image Deblurring with Blur-Conditioned Semantic Routing and Explicit Frequency Fusion
Jun 29, 2026Pan Wang, Yihao Hu, Xiujin LiuGaussian Deblurring Motion Blur
University of Science and Technology of China, Hefei, China · Westlake University, Hangzhou, China · University of Michigan, Ann Arbor, MI, USA

May 24, 2026, cs.CVJ/K move · Enter open · S save

Stop Denoising Your Blurs

Sasidhar Parvathireddy, Vamsidhar Saraswathula, Rama Krishna Gorthi

Indian Institute of Technology Tirupati, India.

Gaussian Deblurring Image Restoration Convolutional

Open paper AskPDF arXiv

In recent times, diffusion models have achieved remarkable performance in image restoration tasks. Their core mechanism relies on the restricted presumption of degradation prior to the additive noise operation. However, the blur model, one of the most widely studied degradation formulations, violates this assumption, as it is inherently based on convolution rather than addition. In this paper, we introduce ConvDiff, a novel diffusion based framework that substitutes the additive operation with convolution for the task of image deblurring. In the forward process, we construct a meaningful trajectory from the clean image to its blurred counterpart by exploiting the frequency domain characteristics of convolution, rather than progressively corrupting the image with additive noise. While the current work instantiates this framework for Gaussian blur, where frequency-domain decomposition yields closed-form and physically valid intermediate states, the underlying principle of constructing degradation trajectories from the blur operator extends naturally to other blur families. This formulation bridges the gap between the mathematical principles of blurring and the iterative design of diffusion-based restoration algorithms, enabling more physically grounded and effective image restoration models.

Abstract

Explore similar work

Stop Denoising Your Blurs

FSM-Net: An Efficient Frequency-Spatial Network for Real-World Deblurring

CogSENet: Blind Image Deblurring with Blur-Conditioned Semantic Routing and Explicit Frequency Fusion