acceptodds
Withdrawn from ICLR 2027

DS-DeBlur: Decoupled Supervision for Diffusion-based Image Deblurring

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Real-world image deblurring remains challenging due to a fundamental mismatch between blur realism and ground-truth fidelity. Real datasets capture authentically blurred images but contain noisy references, whereas synthetic datasets offer clean supervision but fail to reflect realistic motion patterns. This discrepancy leads to a prior conflict with pretrained diffusion models, whose generative priors are learned from clean images, thereby limiting their effectiveness in real motion blur removal. To address this issue, we propose DS-DeBlur, a diffusion-based image deblurring model built upon a new Real-Synthetic Decoupled Supervision (ReSyDS) paradigm. The paradigm explicitly disentangles the learning of blur realism and supervision fidelity through two complementary paths. The synthetic path leverages clean paired data to activate diffusion priors and preserve structural fidelity, while the real path directly learns from authentically blurred images to promote perceptual realism and improve generalization without relying on imperfect references. To further support ReSyDS, we introduce an instance-driven, realistic blur synthesis method, a physically-grounded approach to generate high-quality and realistic blur–sharp pairs for effective supervision. Extensive experiments demonstrate that the proposed method performs favorably against state-of-the-art methods.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Diffuison Model, Image Deblur

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_tLKoOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the challenge of deblurring by tackling the domain gap and prior conflicts between real and synthetic datasets. Real datasets capture complex, realistic motion blur but often suffer from noisy ground-truth references, whereas synthetic datasets provide clean ground truth but fail to faithfully simulate real-world blur dynamics. When fine-tuning pre-trained diffusion models, which are typically trained on high-quality, sharp images, this mismatch leads to prior conflict. To address this, the authors propose Real-Synthetic Decoupled Supervision (ReSyDS), equipped with Instance-Driven Realistic Blur (InstaBlur) and Motion-Strength Modeling (MSM). ReSyDS decouples the learning of fidelity and blur realism through two different paths: the synthetic path aims to preserve the fidelity, while the real path learns realistic blur patterns. InstaBlur generates diverse global, local, and mixed blur patterns using an instance-based synthesis strategy, while the MSM module captures the motion of pixels, repositioning the detail in the latent space. Experiments demonstrate that the proposed framework achieves strong performance across multiple benchmark datasets under no-reference quality metrics.

Critical strengths

S1. Promising No-Reference Metric Performance: As shown in Table 1, the proposed DS-DeBlur consistently and significantly outperforms state-of-the-art methods across all six benchmarks on five no-reference metrics. The paper effectively leverages pre-trained diffusion models while successfully mitigating prior conflicts.

S2. A Reasonably Designed Method with Clear Motivation: The paper clearly identifies the issue of prior conflicts (Figs. 2a and 2b), highlighting the quality gap and distribution mismatch between real and synthetic datasets that pose challenges for diffusion models. The authors thoroughly analyze the dilemma inherent in deblurring datasets, specifically, the trade-off between sharp-but-noisy images and clean-but-blur images. To address this, the proposed method decouples real and synthetic processing paths. This allows the model to learn fidelity from synthetic data and realistic blur degradation from real data without cross-domain interference. Table 3 empirically confirms these complementary benefits, showing that training on real data favors full-reference metrics, whereas synthetic data yields better no-reference metrics.

Critical weaknesses

W1. Limited Technical Novelty: The technical novelty of the proposed method appears somewhat incremental, as several key modules rely heavily on existing techniques. Specifically, the one-step SD-based generator is adapted from OSEDiff, the fidelity discriminator leverages DINOv2, and InstaBlur combines motion trajectories from DeblurGAN with second-order blur synthesis from Real-ESRGAN. Additionally, the proposed MSM integrates spatial-transformer warping with dynamic convolution. While the concept of decoupling real and synthetic paths provides a meaningful insight, the overall methodological novelty remains limited.

W2. Suboptimal Performance in Full-Reference Metrics: The proposed method fails to improve no-reference metrics while maintaining performance in full-reference metrics. Furthermore, performance on full-reference metrics is concerning. As shown in Table 3, the synthetic-only variant performs well on most no-reference metrics (CLIP-IQA, MUSIQ, MANIQ, and LIQE), whereas the real-only path works well on all full-reference metrics (PSNR, SSIM, and LPIPS). However, the adopted ReSyDS framework fails to achieve the best performance on almost any metric except NIQE, and even falls behind by noticeable margins on full-reference metrics. A similar trend is observed for the MSM in Table 5. Overall, comparing the performance on no-reference metrics, the performance on full-reference metrics is concerning. This issue is reflected in Table 2 of the supplementary material, where DS-Deblur yields sub-optimal full-reference results.

Questions for authors

Q1. Suggestion for Evaluation on Datasets with Clean Ground Truth: The authors are suggested to evaluate the proposed method on standard synthetic benchmarks with clean ground-truth images. In the supplementary material, the suboptimal full-reference metric performance is attributed to potential low quality in the ground truth. Evaluating the method on synthetic datasets with clean ground-truth targets would eliminate this confounding factor and can provide a clearer assessment of the model's true capability in preserving full-reference fidelity.

Q2. Request for Evidence on Adaptive Restoration Claims: In Lines 350–352, the authors claim several advantages of the adaptive restoration scheme. To support these claims, could the authors provide qualitative and quantitative evidence (e.g., visual comparisons and ablation metrics) demonstrating the specific contributions and effectiveness of the adaptive restoration mechanism?

Q3. Insight into NIQE Performance Advantage: In Table 3, NIQE is the sole no-reference metric where the adopted ReSyDS outperforms both the real-only and synthetic-only variants. Could the authors provide further insight or analysis into why ReSyDS specifically excels on NIQE compared to the other no-reference metrics?

Reviewer_eaQiOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DS-DeBlur, a one-step diffusion-based image deblurring framework designed to address the mismatch between realistic blur degradation and the quality of available ground-truth images. The central contribution is a Real-Synthetic Decoupled Supervision (ReSyDS) paradigm that separates the learning of structural fidelity and real-world blur characteristics into two complementary training paths. The synthetic path employs Instance-Driven Realistic Blur Synthesis (InstaBlur) to generate clean blur–sharp pairs and introduces a fidelity discriminator to encourage high-quality reconstruction. The real path utilizes authentic blurred images without relying on imperfect ground-truth references, employing a wavelet-based blur assessor to provide additional supervision. Furthermore, a Motion-Strength Modeling (MSM) module combines motion alignment and dynamic convolution to handle spatially varying blur. Experiments across multiple deblurring benchmarks demonstrate substantial improvements in no-reference perceptual quality metrics, while ablation studies investigate the contributions of the proposed components.

Critical strengths

The paper addresses an important challenge in real-world image deblurring by considering the mismatch between realistic blur degradation and ground-truth fidelity. The proposed Real-Synthetic Decoupled Supervision paradigm provides a reasonable approach to combining the advantages of clean synthetic supervision and authentic real-world blur observations. In addition, the Instance-Driven Realistic Blur Synthesis pipeline incorporates instance-level motion, global blur, and noise simulation to generate diverse training samples, while the Motion-Strength Modeling module introduces adaptive spatial alignment and restoration. Experimental results demonstrate substantial improvements in no-reference perceptual quality metrics across multiple benchmarks, suggesting promising generalization to different blur distributions. Furthermore, the one-step diffusion architecture achieves relatively efficient inference compared with the evaluated diffusion-based restoration baselines, and the component-level ablations provide useful insights into the proposed framework.

Critical weaknesses

Despite the substantial improvements in no-reference perceptual quality metrics, several limitations warrant further investigation. First, the experimental evidence does not fully establish the effectiveness of the proposed decoupled supervision strategy. In Table 3, ReSyDS underperforms real-only training on PSNR, SSIM, and LPIPS, while synthetic-only training achieves better results on several perceptual metrics, including CLIP-IQA, MUSIQ, MANIQA, and LIQE. Although ReSyDS provides a trade-off between these objectives, its advantages over simpler joint training or alternative supervision strategies remain insufficiently demonstrated. Second, the evaluation primarily emphasizes no-reference image quality metrics, which may favor visually plausible textures without necessarily guaranteeing accurate recovery of the original scene content. More comprehensive full-reference evaluations and analyses of structural fidelity would strengthen the conclusions. Third, the blur-aware real path relies on a learned assessor that distinguishes restored outputs from blurred inputs without independent sharp-image supervision. Consequently, it remains unclear whether this objective consistently promotes genuine blur removal rather than excessive sharpening or the generation of artificial high-frequency details. Furthermore, although InstaBlur is described as physically grounded, its instance-level PSF-based synthesis does not explicitly model all the geometric and occlusion effects associated with complex object motion, and the realism of the synthesized degradation remains insufficiently validated. Finally, the proposed method benefits from additional synthetic training data and a pretrained diffusion backbone, while the comparison baselines do not appear to receive fully equivalent augmentation resources, making it difficult to isolate the contributions of the proposed supervision strategy from those of the training data and generative prior.

Questions for authors
  1. Can the authors provide a direct comparison between ReSyDS and simpler joint real-synthetic training strategies to better demonstrate the necessity and effectiveness of the proposed decoupled supervision?

  2. Given that ReSyDS underperforms real-only or synthetic-only training on several metrics in Table 3, how do the authors justify its claimed superiority in balancing reconstruction fidelity and perceptual quality?

  3. How does the wavelet-based blur assessor distinguish genuine blur removal from excessive sharpening or hallucinated high-frequency details without access to reliable sharp-image references?

  4. Can the authors provide more comprehensive quantitative evaluations of the physical realism of InstaBlur, particularly regarding complex object motion, occlusion boundaries, and spatially varying blur?

  5. Can the authors clarify whether the comparison methods receive equivalent synthetic training data and pretrained model resources, and provide additional full-reference evaluations to establish the structural fidelity of the proposed method?

1 comment

As the deadline approaches and the authors have not submitted a rebuttal, I am unable to assess whether my concerns have been addressed. I therefore retain my original score.

Reviewer_r64fOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DS-DeBlur, a one-step diffusion model for image deblurring. Its Real-Synthetic Decoupled Supervision (ReSyDS) strategy combines paired supervision from clean synthetic data with unpaired supervision from real blurry images. The method further introduces an instance-level blur synthesis pipeline, InstaBlur, and a Motion-Strength Modeling (MSM) module that combines latent-space motion alignment with dynamic convolution to improve restoration quality in real-world scenarios.

Critical strengths
  1. The problem is practically relevant, and the overall design is clearly motivated. The paper addresses the difficulty of obtaining training data that simultaneously provides realistic blur and high-quality sharp references. Using synthetic data to provide reconstruction supervision while exploiting real blurry inputs for adaptation is a reasonable approach.
  2. The reported no-reference perceptual results are strong, and the evaluation covers a broad range of benchmarks. Table 1 includes six datasets, and the proposed method achieves the best results on all reported no-reference metrics. For example, its MUSIQ score on FoundIR reaches 70.34, compared with 55.61 for the strongest competing method on this metric. The qualitative examples also exhibit sharper textures, and one-step inference offers practical value.
Critical weaknesses
  1. The proposed method additionally uses clean images and InstaBlur-generated training pairs, whereas the main baselines use GoPro training data. Consequently, the current comparisons do not clearly distinguish improvements attributable to additional high-quality training data from those attributable to the proposed decoupled supervision strategy. A controlled comparison using the same training data, backbone, and training budget is needed to isolate the contribution of ReSyDS.

  2. Equation (5) directly treats the current restoration outputs as “sharp” positive samples and real blurry inputs as negative samples. The assessor may therefore learn to distinguish the distributions of generated outputs and input images, rather than specifically assess residual blur. Although reconstruction supervision on the synthetic branch provides some constraints, further evidence is needed to establish that the real branch does not exploit shortcuts associated with generated textures or sharpening artifacts.

  3. The background compositing term in Equation (1) uses an image that still contains the original foreground. When object motion reveals previously occluded background regions, this formulation may retain foreground content at its original location. The main text also does not explain how occlusion relationships between multiple instances are handled. While the method is reasonable as an instance-aware data augmentation strategy, its claims of physical realism require explicit approximation assumptions and supporting analysis of motion boundaries, occluded regions, and real blur data.

  4. Some interpretations of the ablations are stronger than the results support, and the trade-offs across metrics need clearer discussion. In Table 3, moving from Real-only training to ReSyDS decreases PSNR from 21.70 to 21.51 dB and increases LPIPS from 0.2152 to 0.2490, indicating that the improvements in no-reference perceptual quality come with a reduction in reference-based reconstruction quality. In Table 4, removing instance-level blur yields an LPIPS of 0.2330, which is better than the full model’s 0.2490. This does not support the broad statement that removing components consistently degrades performance. Table 5 similarly reveals trade-offs across metrics. These trade-offs are not inherently problematic, but the intended contribution of each module and the criterion for selecting the final configuration remain insufficiently explained. A clearer discussion is needed to establish how the perceptual benefits are balanced against changes in reconstruction fidelity.

  5. The main evaluation provides an incomplete picture of fidelity and does not sufficiently cover directly related diffusion-based methods. Since structural fidelity is a central objective of the paper, the key PSNR, SSIM, and LPIPS comparisons with competing methods should be presented and discussed alongside the no-reference results in the main text. Deferring these comparisons to the supplementary material makes it difficult to assess the reconstruction trade-offs underlying the reported perceptual improvements. In addition, while Diff-Plugin and the retrained OSEDiff are informative baselines, the main comparison omits directly related diffusion-based deblurring methods such as DeblurDiff. Including such methods, or explaining why they cannot be meaningfully compared, would strengthen the evaluation.

Questions for authors
  1. What losses and data sampling ratios are used for the Real-only, Synthetic-only, and ReSyDS variants in Table 3? When using the same real and synthetic training data, backbone, and training budget, does ReSyDS still provide a clear advantage over conventional mixed-data training?

  2. Equation (5) treats the model’s own restoration outputs as sharp positive samples. Is there independent experimental evidence that its scores distinguish blur from noise and excessive sharpening, rather than primarily identify distributional differences between the model outputs and the original blurry inputs?

  3. The background compositing term in Equation (1) still contains the foreground at its original location. How does the formulation avoid retaining this foreground content or producing boundaries inconsistent with the exposure process? When multiple moving instances overlap, how are their occlusion relationships determined?

  4. The full configuration does not dominate all metrics in Tables 4 and 5. How do the authors weigh the perceptual benefits of each component against its effects on reconstruction quality? Are there cross-dataset results or analyses of the loss weights that support the final configuration?

  5. Please report PSNR, SSIM, and LPIPS alongside the no-reference metrics on the main paired benchmarks, and discuss how the perceptual improvements affect reconstruction similarity. Additionally, why was DeblurDiff not included in the main comparison? Could this comparison be added, or could the reason for its omission be clarified?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.