acceptodds
Withdrawn from ICLR 2027

Unified Motion Deblurring via Fourier Convolution Learning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Image deblurring is a fundamental task in computer vision, yet existing methods mainly specialize in isolated scenarios. Consequently, the deblurring performance of these models drops significantly across scenarios with different cameras and image contents. This motivates the development of a unified motion deblurring framework that can handle diverse blur patterns. To this end, we construct a multi-source dataset and develop a data pruning and quality control pipeline. Then, to restore blur images with various motion blur kernels across different scenarios, we propose a unified motion deblurring framework via Fourier Convolution Learning. Existing deblurring networks usually rely on pixel-level mappings and overlook the convolutional nature of blur. Thus, the network mostly fits the image content itself rather than learning the convolution process, making it difficult to handle multi-source blur images. Instead of relying on spatial mapping, we perform blur-adaptive global filter learning in Fourier space, which avoids the limited spatial support of kernel estimation. Extensive experiments demonstrate that our unified model effectively handles motion blurs across diverse sources and scenes, achieving competitive performance across both in-domain and OOD benchmarks. Code will be released.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Image Deblur, Low-level Vision

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_jC4nOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary
  1. This paper try to solve the image deblurring task using the proposed method called UniDeblur.
  2. The main idea of the paper is replacing the residual connection in the Unet with the Fourier Convolution Learning (FCL).
  3. Multiple datasets are used to train the network to avoid reduce the domain gap between datasets, the OOD testing proved the effectiveness of this training. Thus, authors also claim that the proposed method is a "unified motion deblurring framework"
Critical strengths
  1. The kernel shown in the Figure 6 correlated with what adaptive filter learned is convincing.
  2. simple and easy-to-understand ideas and methods. Fourier transform convert the convolution to element-wise multiplication, which avoid the OOM when the convolution kernel size is too large.
  3. The adaptive filter mimic the dynamic convolution that adapt the generated kernels for different images, which is well-motivated.
  4. The experiments are broad. They cover a lot of datasets with OOD testing.
  5. All compared methods are retrained on new datasets instead of naively copying values from the original papers.
  6. FCL is general, it can be plug-and-play into different architectures.
Critical weaknesses
  1. One of my biggest concern is that only the convolution (spatial-invariant blur) can be represent with the multiplication in the Fourier domain, a lot of the classic image deblurring datasets are handling spatial-variant motion blur like GoPro, HIDE, REDS, etc. I think the key assumption is not correct in my opinion. Spatial-variant blur can only be represented as matrix multiplication instead of the convolution.
  2. Though Figure 6 shows the learned kernels, but a more convincing experiments could be: with known blur kernels, what will the learned filters look like. Will they look similar?
  3. I know it makes sense to use no-reference image quality metrics to evaluate the OOD because the gt is imperfect, a user study will be preferred since these no-reference image evaluators are far from perfect, sometimes they prefer noisy than clean images.
  4. The dataset pruning is a bit unconvincing. I know it makes sense somehow, but pruning could also removed some hard cases that cannot be resolved using the proposed methods, for example spatial-variant blur. If so, we cannot say the proposed method is "unified".
  5. Contribution 2: reformulates complex spatial kernel prediction into efficient frequency-domain multiplication. I feel this has been done in tons of previous works, for example: @inproceedings{xintm2025DeepRFTv2, title = {DeepRFTv2: Fourier Kernel Estimation for Image Deblurring}, author = {Xintian Mao, JIansheng Wang, Xingran Xie, Qingli Li and Yan Wang}, booktitle = {arXiv}, year = {2025} }
Questions for authors
  1. How did authors obtain and visualize the kernels in Figure 6 and 7.
  2. How does FCL handle spatial-variant blur? This is critical!!!
  3. Any details about the data prunning?
  4. How sensitive are the retrained models on the pruned data.
  5. Have you done any user study?
2 comments
  1. My spatial-variant kernel concern is not resolved. I know "neural networks inherently possess the ability to handle spatially varying motion blur", but multiplying by a fixed Fourier filter still corresponds to spatially invariant convolution within the region represented by that filter. How do different patches learn different motion kernels? It would make more sense if the patch-based FCL is used instead of image-based. And kernels shown in Figure 6 and 7 are still global kernels instead of local kernels.
  2. I do not think user study is useless. It can complement the quantitative metrics instead of replace them.
  3. How sensitive the results are to pruning is not explained. "Nearly all models can effectively handle motion blur, but fail to process defocus blur." does not have experimental supports.
  4. I am not saying the authors plagiarized DeepRFTv2, I just meant the Fourier kernels idea is widely used in this topic. So the distinguish of this work and other works is still unclear to me.
AuthorsOpenReview ↗

Response to Reviewer jC4n

  1. As explained in our previous response, at test time, we apply FCL to patchified features rather than to the full image.
  2. Thanks for this suggestion, and we will consider it.
  3. Because we have not yet devised a suitable experiment to demonstrate how sensitive the results are to pruning.
  4. We accept the your judgment and plan to withdraw this submission. We realize that our writing led to the reviewers' confusion, and we will address these issues in an updated version.
Reviewer_xHqKOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes UniDeblur, a unified motion deblurring framework trained on six datasets. The key component is Fourier Convolution Learning (FCL), which is integrated into the multi-scale encoder-decoder connections of a U-Net architecture. FCL dynamically predicts input-adaptive frequency filters and performs element-wise multiplication in the Fourier domain to achieve global feature convolution.

Critical strengths
  1. The authors retrain all baseline methods under a unified training protocol and validate the effectiveness of FCL across different backbones. For example, incorporating FCL into the EVSBlock backbone improves the GoPro PSNR from 33.26 dB to 34.02 dB, with a relatively small computational increase from 122G to 130G FLOPs.
  2. The experiments cover multiple in-domain and out-of-distribution (OOD) datasets. The ablation studies investigate multi-scale integration, different module architectures, network depth and width, and frequency-domain addition versus multiplication, providing empirical evidence for the effectiveness of the proposed design.
Critical weaknesses
  1. Limited novelty of the core operator and missing closely related work. The proposed FCL operator is highly similar to the input-adaptive frequency filtering mechanism introduced in AFFNet [1]. Both approaches dynamically generate frequency-domain filters conditioned on input features and apply them through element-wise multiplication. Therefore, it is difficult to regard this operator itself as a novel methodological contribution. More importantly, AFFNet [1] is neither cited nor discussed in the paper. Although AFFNet was not originally designed for image deblurring, its substantial methodological similarity makes it an essential related work that should be acknowledged and carefully distinguished from the proposed FCL.
  2. Insufficient transparency in OOD dataset filtering. A substantial portion of the OOD evaluation samples is removed during data filtering. For example, 328 out of 786 samples are removed from MIORe. However, the paper does not provide sufficiently explicit and reproducible filtering criteria or a complete list of excluded samples. Such extensive filtering may alter the evaluation distribution and affect the reliability of the reported generalization results.
  3. Insufficient OOD evaluation and missing diffusion-based baselines. The OOD evaluation relies exclusively on no-reference perceptual quality metrics (CLIPIQA and MUSIQ), despite the availability of ground-truth images in these datasets. While the authors discuss potential GT quality issues, relying solely on no-reference metrics is insufficient to establish restoration fidelity. Moreover, if perceptual quality is the primary evaluation criterion, the comparison should include diffusion-based deblurring methods, such as DeblurDiff [2], which are designed to achieve high perceptual quality.

[1] Zhipeng Huang et al., "Adaptive Frequency Filters As Efficient Global Token Mixers," ICCV, 2023.

[2] Lingshun Kong et al., "DeblurDiff: Real-Word Image Deblurring with Generative Diffusion Models," NeurIPS, 2025.

Questions for authors
  1. Novelty compared with AFFNet: What are the fundamental methodological differences between FCL and the adaptive frequency filtering mechanism in AFFNet [1]? Could the authors provide a detailed comparison and clarify the specific novelty beyond applying a similar operator to motion deblurring?
  2. OOD dataset filtering: What are the exact criteria used to remove OOD test samples? Could the authors release the filtering principle and report whether the relative performance rankings remain consistent before and after filtering?
  3. OOD evaluation protocol: Why are full-reference metrics such as PSNR and SSIM entirely omitted from the OOD evaluation, even for samples with reasonably aligned ground truths? Additionally, could the authors include diffusion-based deblurring methods, such as DeblurDiff [2], to provide a more comprehensive comparison under perceptual quality metrics?
Reviewer_ijTWOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes UniDeblur, a unified motion deblurring framework trained jointly on multiple blur datasets.

The main technical component is Fourier Convolution Learning (FCL). Given a feature map, FCL transforms it into the Fourier domain, predicts an input-adaptive frequency filter, multiplies the predicted filter with the feature spectrum, and then transforms the result back to the spatial domain. The module is applied at multiple encoder-decoder transitions.

The paper also constructs a multi-source training and evaluation protocol using six training datasets and several held-out OOD datasets. Representative deblurring and all-in-one restoration baselines are retrained under the same unified protocol.

Critical strengths
  1. The unified motion-deblurring setting is practically relevant. The training corpus covers different cameras, acquisition pipelines, scenes, lighting conditions, and blur characteristics.

  2. The main comparison is relatively well controlled. The authors retrain the competing methods under the same multi-source training protocol instead of directly comparing against numbers reported under different training settings.

  3. The ablations provide convincing evidence that the proposed FCL module improves restoration performance. FCL improves NAFBlock-, BioBlock-, and EVSBlock-based models. In particular, using the same EVSBlock backbone and depth as EVSSM, UniDeblur-L improves GoPro PSNR from 33.26 dB to 34.02 dB while increasing FLOPs from 122G to 130G.

  4. The paper evaluates a broad range of in-domain and held-out datasets and includes both quantitative and extensive qualitative comparisons.

Critical weaknesses

1. The methodological novelty of FCL is unclear relative to highly similar prior work.

At its core, FCL applies the following procedure: feature → Fourier transform → input-dependent frequency-domain filter prediction → element-wise multiplication with the same feature spectrum → inverse Fourier transform.

A highly relevant prior work that is not cited or discussed is:

  • Xintian Mao, Haofei Song, Yin-Nian Liu, Qingli Li, and Yan Wang. "DeepRFTv2: Kernel-level Learning for Image Deblurring." arXiv:2511.21132, 2025.

DeepRFTv2 introduces a Fourier Kernel Estimator (FKE) with a very similar core motivation and operation: it transforms learned features into Fourier space, predicts a feature-level Fourier kernel with a neural network, directly multiplies the predicted kernel with the feature spectrum, and applies an inverse Fourier transform. It also motivates this design as moving deblurring from pixel-level restoration to kernel-level learning of the blur process.

The two methods are not identical. DeepRFTv2 places FKE in the restoration tail and combines it with a Decoupled Multi-Scale UNet, whereas UniDeblur applies FCL at encoder-decoder transitions, extends it across multiple scales, explores different filter-prediction blocks, and focuses on unified multi-source training.

However, based on the current presentation, the clearest differences appear to lie in module placement, architecture, and training setting, while the central Fourier kernel-learning mechanism remains very similar. The current manuscript instead contrasts FCL mainly with DeepRFT (2023), additive Fourier feature learning, and static global filtering. Without discussing DeepRFTv2 and directly explaining the distinction between FCL and FKE, the novelty of the main technical contribution is difficult to establish.

2. The physical interpretation of the learned Fourier filters is stronger than the current evidence supports.

The manuscript repeatedly describes FCL as explicitly modeling the convolutional nature or the "essence" of motion blur and interprets the learned filters as blur/deconvolution kernels.

The empirical effectiveness of FCL is clear, but this physical interpretation requires more evidence. For a fixed predicted filter, global multiplication in Fourier space corresponds to a global shift-invariant convolution in the spatial domain. Real dynamic motion blur, however, is often spatially varying because different objects and image regions can have different motion trajectories.

The nonlinear backbone may still allow FCL to be useful under such conditions, but this does not necessarily mean that the learned feature-space filter corresponds to the physical image-space motion kernel.

Figures 6 and 7 provide qualitative kernel visualizations, but the manuscript does not clearly explain how the displayed "theoretical" blur/deconvolution kernels are obtained for the real datasets or quantitatively measure the agreement between the predicted filters and actual blur kernels. The current visual evidence is therefore insufficient to support the stronger physical interpretation.

3. The OOD evaluation provides useful evidence, but the generalization claim should be interpreted more carefully.

The authors identify quality problems in several OOD datasets and prune samples containing residual GT blur, mixed degradations, or misalignment. This is a reasonable attempt to improve evaluation quality, and importantly all methods are evaluated on the same retained subsets.

However, the exact filtering criteria are not sufficiently specified. The paper reports substantial removal on some datasets, including hundreds of samples from MIORe and VAR-MIORe. Since these curated subsets are central to the OOD claim, deterministic selection criteria and the exact retained sample lists should be released.

The main OOD evaluation also relies on CLIP-IQA and MUSIQ. These no-reference metrics measure perceptual image quality rather than restoration fidelity. A deblurred output can become sharper or more visually pleasing without necessarily being closer to the latent sharp image. The fact that restored images can even obtain no-reference scores above the provided GT illustrates this distinction.

Therefore, the OOD results support improved perceptual quality across domains, but they provide weaker evidence for accurate cross-domain deblurring than full-reference evaluation on reliably aligned subsets would.

4. The multi-source protocol is useful, but it is primarily a training/evaluation protocol rather than a new dataset.

The six training datasets are existing public datasets, and the contribution is mainly their joint use, preprocessing, patch construction, and the proposed quality-control protocol.

This is still useful, especially because the paper retrains all baselines consistently. However, the manuscript should avoid overstating this contribution as dataset construction in the same sense as collecting a new benchmark.

One reproducibility detail is also unclear: the six training datasets contain substantially different numbers of images and extracted patches, but the manuscript does not specify whether batches are sampled uniformly over all patches, uniformly over datasets, or according to another balancing strategy. This may materially affect unified multi-source training.

Questions for authors
  1. What is the precise methodological difference between FCL and the Fourier Kernel Estimator in DeepRFTv2 (2025)?
    Both appear to predict an input-dependent Fourier kernel/filter from learned features, multiply it with the same feature spectrum, and transform the result back to the spatial domain. Please clarify which part of the Fourier kernel-learning mechanism itself is new beyond FKE and provide a direct comparison if possible.

  2. What evidence shows that the learned FCL filters correspond to physical blur/deconvolution kernels rather than useful input-adaptive spectral filters?
    In particular, how are the "theoretical" kblurk_{\mathrm{blur}} and kdeblurk_{\mathrm{deblur}} shown in Figures 6 and 7 obtained for real captured datasets?

  3. How should FCL be interpreted for spatially varying motion blur?
    Since global Fourier multiplication corresponds to a shift-invariant convolution for a fixed filter, how does the proposed formulation represent multiple local motion trajectories within the same image?

  4. What exact criteria are used for OOD sample pruning?
    Please provide reproducible filtering rules and the retained/removed sample lists. Reporting results before and after pruning would also help quantify the effect of this curation.

  5. Can the authors provide full-reference OOD results on subsets with sufficiently reliable alignment?
    CLIP-IQA and MUSIQ are useful perceptual metrics but do not directly measure restoration fidelity. Complementary PSNR/SSIM, LPIPS, DISTS, or other suitable full-reference metrics would strengthen the cross-domain generalization claim.

  6. How are the six training datasets sampled during unified training?
    Are all extracted patches simply concatenated, or is dataset-balanced sampling used? Since the datasets differ substantially in size, this detail is important for reproducibility and interpretation of the multi-source results.

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

Related papers

Discussion (0)

Sign in to comment.