Pith. sign in

REVIEW 3 major objections 5 minor 41 references

Motion-Aware Adaptive Pixel Pruning for Efficient Local Motion Deblurring

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read M2AENet prunes sharp pixels and follows motion trajectories to deblur locally blurry images better and cheaper than prior methods.

desk verdict Pixel-level pruning for local deblurring is a genuinely useful step; the reblur-loss temporal alignment is the one load-bearing assumption to press on. read the letter →

arxiv 2507.07708 v1 pith:SJ5XDFV6 submitted 2025-07-10 cs.CV

Wei Shang , Dongwei Ren , Wanying Zhang , Pengfei Zhu , Qinghua Hu , Wangmeng Zuo This is my paper
classification cs.CV
keywords imagedeblurringlocalmotionblurpixelpruningintra-motionanalyzermaskpredictiondeformableconvolutionstructuralreparameterizationefficientinference
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Local motion blur affects only parts of an image, yet most deblurring networks process every pixel equally, wasting computation on sharp areas and sometimes distorting them. This paper proposes a network that predicts a per-pixel blur mask, prunes sharp pixels during inference, and estimates each pixel's displacement across the exposure interval so a deformable convolution can sample along the actual motion trajectory. On the local-blur dataset ReLoBlur it reports gains of 0.25 dB in PSNR and nearly 0.8 dB in mask-weighted PSNR over prior local deblurring methods, plus a 0.55 dB gain on the global-blur GoPro dataset. Those gains come with 49% fewer FLOPs, 38.5% shorter inference time, and about one-fifth the GPU memory of the strongest compared transformer. The paper's claim is that pixel-level computation allocation plus explicit motion-trajectory guidance is sufficient to beat patch- and window-level pruning strategies.

What carries the argument

The load-bearing object is the mask- and motion-aware simple (M2AS) block, a modified NAFNet block with two inserted mechanisms. First, mask-aware convolution multiplies features elementwise by the predicted blur mask so sharp pixels are excluded while gradients still flow; at test time the network samples only the $Q$ blurred pixels and replaces each $3\times3$ convolution by the equivalent combination of an unfold operation and a reparameterized $1\times1$ convolution whose weights are the original kernel reshaped from $\mathbb{R}^{C_{\text{out}}\times C_{\text{in}}\times3\times3}$ to $\mathbb{R}^{C_{\text{out}}\times(C_{\text{in}}\times9)\times1\times1}$, cutting that layer's work by the factor $Q/(HW)$. Second, motion-aware deformable convolution takes the predicted per-pixel displacements $\{\boldsymbol{o}(t_n)\}_{n=0}^{N-1}$, converts them into relative offsets for a $3\times3$ deformable kernel by adding them to the standard kernel positions, and performs depthwise deformable convolution so the receptive field follows the motion trajectory. A quadratic interpolation formula generates the intermediate displacements from the two predicted endpoints, and a reblur loss warps the sharp ground truth along those displacements to supervise them.

What would settle it

Build a synthetic local-blur test set from a high-frame-rate video so the true exposure midpoint is known, then retrain M2AENet twice: once with the sharp reference at the true midpoint and once with it shifted to a known non-midpoint frame. If the second model's PSNR_w does not fall or its estimated displacement fields do not rotate or translate to compensate, the midpoint assumption is not actually carrying the result; if it degrades, the assumption is confirmed as load-bearing.

Watch

Extended reading notes

Core claim

The central claim is that local motion deblurring can be cast as two coupled subproblems: locating the blurred pixels and following their motion. A mask predictor assigns each pixel a blur probability, converted to a binary mask through Gumbel-Softmax during training and a threshold at test time; an intra-motion analyzer predicts each pixel's displacement from the exposure midpoint to the start and end of the exposure, with intermediate positions filled by a quadratic trajectory. The mask gates feature processing so sharp pixels contribute nothing, and the trajectory offsets are fed into a depthwise deformable convolution so that restoration samples pixels along the direction and magnitude of motion. The paper reports that these two mechanisms together, not either alone, produce the best results: ablations show the mask mainly lifts weighted PSNR on local blur while the motion convolution mainly lifts global PSNR. The result is a single end-to-end network that outperforms existing local and global deblurring methods on ReLoBlur and GoPro while spending computation almost exclusively on blurred pixels.

Load-bearing premise

The load-bearing premise is that the paired sharp ground-truth frame sits exactly at the midpoint of the exposure interval, because the offset-prediction loss uses that midpoint as the anchor for all warped sharp frames; the paper says the ground truth only 'typically represents' the midpoint and does not verify this for the beam-splitter ReLoBlur captures.

Editorial extensions

If this is right

  • Because pruning happens per pixel, the computational saving grows as the blurred region shrinks, which is exactly the regime of high-resolution images with small moving objects.
  • Mask and motion information are complementary: removing the mask collapses weighted PSNR on local blur, while removing the motion convolution collapses global PSNR, so the paper's accuracy gain depends on both mechanisms being present.
  • The same trained model handles global blur by treating the whole image as blurred with a mask of ones on GoPro, meaning pixel pruning does not prevent the network from behaving like a global deblurrer when needed.
  • Applying mask-aware convolution to additional positions inside the block lowers FLOPs further but hurts accuracy, so the paper places it at only the first convolution of the block to balance speed and quality.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The midpoint-exposure assumption behind the offset loss is testable: on a synthetic blur benchmark with known shutter timing, shifting the reference sharp frame away from the midpoint should measurably bias the learned displacements if the assumption is load-bearing.
  • The predicted displacement fields are an interpretable by-product that could serve as self-supervised motion estimates for related tasks such as motion segmentation or optical-flow initialization, since the reblur loss provides supervision without manual annotations.
  • The flat-region failure mode noted in the appendix suggests that reblur supervision alone cannot distinguish a sharp flat area from a blurred one; adding edge-aware or temporal constraints is a natural next step the paper leaves open.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes M2AENet, a CNN-based local motion deblurring method built on NAFNet, adding a trainable mask predictor for blurred-region localization and an intra-motion analyzer that estimates per-pixel displacements during exposure. The method uses mask-aware convolution with a pixel-pruning inference strategy that reparameterizes 3×3 convolutions into 1×1 convolutions, and a motion-aware deformable convolution guided by the estimated motion trajectories. The network is trained end-to-end with reconstruction, mask, and reblur losses. Experiments on ReLoBlur and GoPro report state-of-the-art PSNR/SSIM (and weighted metrics on ReLoBlur) while reducing FLOPs, inference time, and GPU memory compared with LMD-ViT. Ablations study the contribution of each component and the placement of mask-aware convolution.

Significance. If the claims hold, M2AENet offers a practically valuable balance of quality and efficiency for local motion deblurring: pixel-level computation allocation is a natural next step beyond patch/window pruning, and the reported 0.25–0.8 dB gains on ReLoBlur and 0.55 dB on GoPro over strong baselines, together with a 49% FLOPs reduction versus LMD-ViT, are substantial. The paper's experimental protocol is more rigorous than many in the field, since all compared methods are retrained on the same combined dataset and cropping strategy. The authors also provide code, which supports reproducibility. However, several load-bearing technical descriptions are inconsistent or incomplete, and one physical assumption about the training data is unverified; these issues must be resolved before the contributions can be fully trusted.

major comments (3)
  1. [§3.3.1 and Fig. 2(c)] The reparameterization that converts the masked 3×3 convolution into a 1×1 convolution is described with weight shapes W∈R^{C_out×C_in×3×3} reshaped into W'∈R^{C_out×(C_in×9)×1×1}. However, the layer being reparameterized is the depthwise convolution in Eq. (5), whose actual weight tensor is W∈R^{2C×1×3×3} with groups 2C (as labeled in Fig. 2). A dense 1×1 convolution with input channels C_in×9 and output channels C_out is not equivalent to the original depthwise operation: it would have 4C^2×9 parameters and perform cross-channel mixing, whereas the depthwise convolution has 2C×9 parameters and no cross-channel mixing. The paper must either specify a grouped or block-structured reparameterization that exactly recovers the depthwise operation, or clarify the notation. Without this, the claimed FLOPs reduction and the equivalence of the pixel-pruning inference are not verifiable.
  2. [§3.4.3, Eq. (11)] The reblur loss assumes that the sharp ground-truth y_gt is the sharp frame at the midpoint of the exposure interval, since the model warps y_gt by the estimated displacements o(t_n) and sums over time instants to synthesize the blurred input. The paper states only that y_gt 'typically represents' the midpoint and provides no verification for ReLoBlur, which is captured with a beam-splitter system whose temporal alignment of the sharp frame within the exposure is not specified. If the sharp frame is not at the midpoint, the estimated o(t0) and o(t_{N-1}) are biased by a constant temporal offset, and the 'motion trajectories' used to guide the deformable convolution are systematically incorrect. The ablation in Table 3 shows that the motion-aware component improves PSNR, but that only demonstrates that the offsets carry useful information for the reconstruction loss, not that they are physically correct. The authors should verify the temporal alignment in the dataset or reformulate the loss to be independent of an unknown midpoint, and ideally provide a synthetic experiment where ground-truth motion is known.
  3. [§3.1, §3.4, and Appendix A] The main text repeatedly states that the mask predictor and intra-motion analyzer are applied at each scale and that the losses in Eqs. (9)–(10) sum over S scales. However, Appendix A states: 'To ease training, we added the mask predictor and intra-motion analyzer only to the fourth stage of the encoder.' These statements are directly contradictory. If only one stage uses these components, then the sum over S scales is either incorrect or S=1, and the claimed multi-scale joint optimization is not implemented as described. The authors must correct this inconsistency and specify precisely which stages contain the mask predictor and intra-motion analyzer, because it affects the interpretation of the mask loss, the offset loss, and the overall architecture.
minor comments (5)
  1. [Fig. 2 caption] The caption contains a typo: 'M AS Block (Ours)2' should likely read 'M2AS Block (Ours)'.
  2. [Eq. (8)] The frequency reconstruction loss FFT is not defined in the paper; an explicit equation or a reference would improve reproducibility.
  3. [Abstract and Fig. 1] The abstract claims 'superior performance' over state-of-the-art methods, while Fig. 1 describes 'comparable performance to state-of-the-art methods'; this wording should be unified to avoid undercutting the central claim.
  4. [§3.1, Eq. (3)] The test-time threshold ε is fixed at 0.5 even though the paper describes the pruning as 'adaptive'; the authors should briefly justify this choice or analyze its sensitivity.
  5. [Table 2] The column header 'Times (s)' should be 'Time (s)' for consistency with the caption.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central claims are held-out benchmark comparisons, and the auxiliary losses are externally supervised or forward-model constraints rather than reductions to the method's own outputs.

full rationale

The paper's central claims are benchmark PSNR/SSIM and efficiency comparisons on held-out ReLoBlur and GoPro test sets, with all compared methods retrained under the same protocol. The three training losses are externally supervised or physically constrained: the reconstruction loss uses sharp ground truth; the mask loss uses annotated blur masks; and the reblur loss (Eq. 11) synthesizes a blurred frame by warping the sharp ground truth with predicted displacements and matching the input blurred frame. This is a forward blur-formation self-supervision, not an identity that defines the output in terms of itself. The mask predictor is supervised by ground-truth masks, and the displacement field is constrained by an independent reblur consistency term plus TV smoothness; neither the predicted mask nor the displacements are fitted to the target PSNR/SSIM. Ablations compare component removals against the same losses. The midpoint-exposure assumption in Sec. 3.4.3 is an unverified modeling assumption and therefore a correctness risk, but it is not circular: if the assumption fails, Eq. 11 is biased, yet the derivation does not define any predicted quantity in terms of the final metric. Self-citations ([25], [28]) appear only in related work and are not load-bearing; no uniqueness theorem or ansatz is imported from the authors' prior work. No circular step satisfies the quoted-reduction standard.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central contributions rest on a blur-formation model, a midpoint-time alignment assumption, the quadratic displacement interpolation, and a reparameterization equivalence that is stated imprecisely for depthwise convolutions. The method introduces no entities beyond network modules, and the only tuned scalars are loss weights, the mask threshold, and the number of exposure instants.

free parameters (4)
  • alpha
    Weight on the total-variation term in the offset prediction loss, Eq. 10. Its value is not reported anywhere in the manuscript.
  • mask threshold epsilon = 0.5
    Inference-time threshold in Eq. 3 converts predicted probabilities to a binary pruning mask. Hand-chosen.
  • number of exposure instants N2 = 9
    Sets both the number of warped frames in the reblur loss and the number of deformable-convolution offsets. Chosen from the ablation in Table 6.
  • loss weights lambda and gamma = lambda=0.01, gamma=0.1
    Balance mask, offset, and reconstruction terms in Eqs. 7 and 8. Taken from prior work [13,14], not independently justified.
assumptions (5)
  • domain assumption A blurred frame is approximated by the sum of sharp frames warped forward to discrete time instants during exposure, as in Eq. 11.
    Used to supervise predicted displacements; ignores occlusion, lighting changes, and non-rigid appearance changes.
  • domain assumption The sharp ground-truth frame corresponds to the midpoint of the exposure interval.
    Needed to define o(t0) and o(t_{N-1}) in Section 3.2 and Eq. 11. The paper only says this typically holds.
  • domain assumption Pixel displacements between the start and end of exposure follow a quadratic-in-time curve.
    Used in Eq. 4 to interpolate N displacement fields. Cited to [37], not derived for motion blur.
  • standard math A 3x3 convolution can be exactly replaced by an unfold operation plus a 1x1 convolution.
    The inference-time reparameterization in Section 3.3.1 relies on this equivalence. For depthwise convolutions the equivalent 1x1 weight must be block-structured, which the paper does not state.
  • domain assumption Dataset blur masks from ReLoBlur, with the LMD-ViT annotation scheme, and the all-one masks for GoPro correctly supervise the mask predictor.
    Mask loss in Eq. 9 treats these annotations as ground truth; errors propagate to the pruning and deblurring.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Motion-Aware Adaptive Pixel Pruning for Efficient Local Motion Deblurring." pith.science (2026). https://pith.science/paper/SJ5XDFV6

@misc{pith2026250707708,
  author       = {Pith},
  title        = {Pith review of: Motion-Aware Adaptive Pixel Pruning for Efficient Local Motion Deblurring},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SJ5XDFV6}},
  note         = {Machine review of arXiv:2507.07708}
}
abstract

Local motion blur in digital images originates from the relative motion between dynamic objects and static imaging systems during exposure. Existing deblurring methods face significant challenges in addressing this problem due to their inefficient allocation of computational resources and inadequate handling of spatially varying blur patterns. To overcome these limitations, we first propose a trainable mask predictor that identifies blurred regions in the image. During training, we employ blur masks to exclude sharp regions. For inference optimization, we implement structural reparameterization by converting $3\times 3$ convolutions to computationally efficient $1\times 1$ convolutions, enabling pixel-level pruning of sharp areas to reduce computation. Second, we develop an intra-frame motion analyzer that translates relative pixel displacements into motion trajectories, establishing adaptive guidance for region-specific blur restoration. Our method is trained end-to-end using a combination of reconstruction loss, reblur loss, and mask loss guided by annotated blur masks. Extensive experiments demonstrate superior performance over state-of-the-art methods on both local and global blur datasets while reducing FLOPs by 49\% compared to SOTA models (e.g., LMD-ViT). The source code is available at https://github.com/shangwei5/M2AENet.

Figures

Figures reproduced from arXiv: 2507.07708 by the authors.

Figure 1
Figure 1. Comparison with recent deblurring models on [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of key differences with the original NAFNet Block. (a) The original NAFNet block [ [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. The pipeline of our motion-aware deformable con [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (10 more)
Figure 4
Figure 4. Figure 4: Visual comparison with the state-of-the-art deblurring methods on ReLoBlur [ [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Visual comparison with the state-of-the-art deblur [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Visual comparison with the state-of-the-art deblur [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: The position where mask-aware convolution can [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: (a) Locally Blurred Image. (b) Motion trajectory w/o [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]
Figure 9
Figure 9. Figure 9: Visual comparison with the state-of-the-art deblurring methods on ReLoBlur [ [PITH_FULL_IMAGE:figures/full_fig_p012_9.png]
Figure 10
Figure 10. Figure 10: Visual comparison with the state-of-the-art deblurring methods on GoPro [ [PITH_FULL_IMAGE:figures/full_fig_p013_10.png]
Figure 11
Figure 11. Figure 11: Visual comparison with the state-of-the-art deblurring methods on real-world locally blurred images. [PITH_FULL_IMAGE:figures/full_fig_p014_11.png]
Figure 12
Figure 12. Figure 12: Visualization of deblurring results and motion trajectories on ReLoBlur [ [PITH_FULL_IMAGE:figures/full_fig_p014_12.png]
Figure 13
Figure 13. Figure 13: Visualization of deblurring results and motion trajectories on GoPro [ [PITH_FULL_IMAGE:figures/full_fig_p015_13.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

41 extracted references · 41 canonical work pages

  1. [1]

    Taiebeh Askari Javaran, Hamid Hassanpour, and Vahid Abolghasemi. 2017. Local motion deblurring using an effective image prior based on both the first-and second-order gradients.Machine Vision and Applications28 (2017), 431–444

    work page 2017

  2. [2]

    Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. 2022. Simple baselines for image restoration. InEur. Conf. Comput. Vis.17–33

    work page 2022

  3. [3]

    Liangyu Chen, Xin Lu, Jie Zhang, Xiaojie Chu, and Chengpeng Chen. 2021. HINet: Half instance normalization network for image restoration. InIEEE Conf. Comput. Vis. Pattern Recog.182–192

    work page 2021

  4. [4]

    Sung-Jin Cho, Seo-Won Ji, Jun-Pyo Hong, Seung-Won Jung, and Sung-Jea Ko

    work page

  5. [5]

    Rob Fergus, Barun Singh, Aaron Hertzmann, Sam T Roweis, and William T Freeman. 2006. Removing camera shake from a single photograph. InACM SIGGRAPH. 787–794

    work page 2006

  6. [6]

    Eric Jang, Shixiang Gu, and Ben Poole. 2017. Categorical reparameterization with gumbel-softmax. InInt. Conf. Learn. Represent

    work page 2017

  7. [7]

    Jinho Jeong, Jinwoo Kim, Younghyun Jo, and Seon Joo Kim. 2024. Accelerating Image Super-Resolution Networks with Pixel-Level Classification. InEur. Conf. Comput. Vis.236–251

    work page 2024

  8. [8]

    Diederik P Kingma and Jimmy Ba. 2014. Adam: A method for stochastic opti- mization. InInt. Conf. Learn. Represent

    work page 2014

Show all 41 references
  1. [9]

    Lingshun Kong, Jiangxin Dong, Jianjun Ge, Mingqiang Li, and Jinshan Pan. 2023. Efficient frequency domain-based transformers for high-quality image deblurring. InIEEE Conf. Comput. Vis. Pattern Recog.5886–5895

    2023

  2. [10]

    Xiangtao Kong, Hengyuan Zhao, Yu Qiao, and Chao Dong. 2021. ClassSR: Agen- eral framework to accelerate super-resolution networks by data characteristic. InIEEE Conf. Comput. Vis. Pattern Recog.12016–12025

    2021

  3. [11]

    Orest Kupyn, Volodymyr Budzan, Mykola Mykhailych, Dmytro Mishkin, and Jiří Matas. 2018. Deblurgan: Blind motion deblurring using conditional adversarial networks. InIEEE Conf. Comput. Vis. Pattern Recog.8183–8192

    2018

  4. [12]

    Anat Levin. 2006. Blind motion deblurring using image statistics. InAdv. Neural Inform. Process. Syst., Vol. 19

    2006

  5. [13]

    Haoying Li, Ziran Zhang, Tingting Jiang, Peng Luo, Huajun Feng, and Zhihai Xu. 2023. Real-world deep local motion deblurring. InAss. Advance. Arti. Intell., Vol. 37. 1314–1322

    2023

  6. [14]

    Haoying Li, Jixin Zhao, Shangchen Zhou, Huajun Feng, Chongyi Li, and Chen Change Loy. 2024. Adaptive Window Pruning for Efficient Local Motion Deblurring. InInt. Conf. Learn. Represent

    2024

  7. [15]

    Ilya Loshchilov and Frank Hutter. 2017. SGDR: Stochastic Gradient Descent with Warm Restarts. InInt. Conf. Learn. Represent

    2017

  8. [16]

    Xintian Mao, Qingli Li, and Yan Wang. 2024. AdaRevD: Adaptive Patch Exiting Reversible Decoder Pushes the Limit of Image Deblurring. InIEEE Conf. Comput. Vis. Pattern Recog.25681–25690

    2024

  9. [17]

    Xintian Mao, Yiming Liu, Fengze Liu, Qingli Li, Wei Shen, and Yan Wang. 2023. Intriguing findings of frequency selection for image deblurring. InAss. Advance. Arti. Intell., Vol. 37. 1905–1913

    2023

  10. [18]

    Xintian Mao, Jiansheng Wang, Xingran Xie, Qingli Li, and Yan Wang. 2024. LoFormer: Local frequency Transformer for image deblurring. InACM Int. Conf. Multimedia. 10382–10391

    2024

  11. [19]

    Seungjun Nah, Sungyong Baik, Seokil Hong, Gyeongsik Moon, Sanghyun Son, Radu Timofte, and Kyoung Mu Lee. 2019. NTIRE 2019 challenge on video deblur- ring and super-resolution: Dataset and study. InIEEE Conf. Comput. Vis. Pattern Recog. Worksh.0–0

    2019

  12. [20]

    Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. 2017. Deep multi-scale con- volutional neural network for dynamic scene deblurring. InIEEE Conf. Comput. Vis. Pattern Recog.3883–3891

    2017

  13. [21]

    Simon Niklaus and Feng Liu. 2020. Softmax splatting for video frame interpolation. InIEEE Conf. Comput. Vis. Pattern Recog.5437–5446

    2020

  14. [22]

    Jinshan Pan, Deqing Sun, Hanspeter Pfister, and Ming-Hsuan Yang. 2016. Blind image deblurring using dark channel prior. InIEEE Conf. Comput. Vis. Pattern Recog.1628–1636

    2016

  15. [23]

    Bang-Dang Pham, Phong Tran, Anh Tran, Cuong Pham, Rang Nguyen, and Minh Hoai. 2024. Blur2Blur: Blur Conversion for Unsupervised Image Deblurring on Unknown Domains. InIEEE Conf. Comput. Vis. Pattern Recog.2804–2813

    2024

  16. [24]

    Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. 2021. DynamicViT: Efficient vision Transformers with dynamic token sparsification.Adv. Neural Inform. Process. Syst.34 (2021), 13937–13949

    2021

  17. [25]

    Dongwei Ren, Kai Zhang, Qilong Wang, Qinghua Hu, and Wangmeng Zuo. 2020. Neural blind deconvolution using deep priors. InIEEE Conf. Comput. Vis. Pattern Recog.3341–3350

    2020

  18. [26]

    Wenqi Ren, Jiawei Zhang, Jinshan Pan, Sifei Liu, Jimmy S Ren, Junping Du, Xiaochun Cao, and Ming-Hsuan Yang. 2021. Deblurring dynamic scenes via spatially varying recurrent neural networks.IEEE Trans. Pattern Anal. Mach. Intell.44, 8 (2021), 3974–3987

    2021

  19. [27]

    Qi Shan, Jiaya Jia, and Aseem Agarwala. 2008. High-quality motion deblurring from a single image.ACM Trans. Graph.27, 3 (2008), 1–10

    2008

  20. [28]

    Wei Shang, Dongwei Ren, Yi Yang, and Wangmeng Zuo. 2025. Aggregating nearest sharp features via hybrid transformers for video deblurring.Information Sciences694 (2025), 121689

    2025

  21. [29]

    Jian Sun, Wenfei Cao, Zongben Xu, and Jean Ponce. 2015. Learning a convo- lutional neural network for non-uniform motion blur removal. InIEEE Conf. Comput. Vis. Pattern Recog.769–777

    2015

  22. [30]

    Xin Tao, Hongyun Gao, Xiaoyong Shen, Jue Wang, and Jiaya Jia. 2018. Scale- recurrent network for deep image deblurring. InIEEE Conf. Comput. Vis. Pattern Recog.8174–8182

    2018

  23. [31]

    Fu-Jen Tsai, Yan-Tsung Peng, Yen-Yu Lin, Chung-Chi Tsai, and Chia-Wen Lin

  24. [32]

    John YA Wang and Edward H Adelson. 1993. Layered representation for motion analysis. InIEEE Conf. Comput. Vis. Pattern Recog.361–366

    1993

  25. [33]

    Shizun Wang, Jiaming Liu, Kaixin Chen, Xiaoqi Li, Ming Lu, and Yandong Guo

  26. [34]

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. 2004. Image quality assessment: from error visibility to structural similarity.IEEE Trans. Image Process.13, 4 (2004), 600–612

    2004

  27. [35]

    Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. 2022. Uformer: A general U-shaped transformer for image restoration. InIEEE Conf. Comput. Vis. Pattern Recog.17683–17693

    2022

  28. [36]

    Adaptive patch exiting for scalable single image super-resolution. InEur. Conf. Comput. Vis.292–307

  29. [37]

    Xiangyu Xu, Siyao Li, Wenxiu Sun, Qian Yin, and Ming-Hsuan Yang. 2019. Qua- dratic video interpolation. InAdv. Neural Inform. Process. Syst.1–10

    2019

  30. [38]

    Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. 2022. Restormer: Efficient transformer for high- resolution image restoration. InIEEE Conf. Comput. Vis. Pattern Recog.5728–5739. Conference’25, Oct 2025, Dublin, Ireland Wei S...

    2022

  31. [39]

    Zeyu Xiao, Zhihe Lu, Michael Bi Mi, Zhiwei Xiong, and Xinchao Wang. 2024. Unraveling Motion Uncertainty for Local Motion Deblurring. InACM Int. Conf. Multimedia. 3065–3074

    2024

  32. [2021]

    Rethinking coarse-to-fine approach in single image deblurring. InInt. Conf. Comput. Vis.4641–4650

  33. [2022]

    Stripformer: Strip Transformer for fast image deblurring. InEur. Conf. Comput. Vis.Springer, 146–162

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.