Pith. sign in

REVIEW 3 major objections 5 minor 26 references

Image Motion Blur Removal in the Temporal Dimension with Video Diffusion Models

T0 review · 3 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read Motion blur can be undone by treating a blurry image as the temporal average of a sharp video and letting a video diffusion model recover that video.

desk verdict A genuinely novel framing of single-image deblurring as video reconstruction from a temporal average, but the empirical claims need fair baselines and real-blur evaluation before they hold. read the letter →

arxiv 2501.12604 v1 pith:4HO77SSG submitted 2025-01-22 eess.IV cs.CVcs.LG

Wang Pang , Zhihao Zhan , Xiang Zhu , Yechao Bai This is my paper · ORCID
classification eess.IVcs.CVcs.LG
keywords motiondeblurringvideodiffusionmodelposteriorsamplingtemporalaveraginglatenttransformersingle-image
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

This paper tries to establish that a single motion-blurred image can be turned back into a short sharp video by modeling blur as an average of frames rather than as a spatial convolution. If true, deblurring would no longer depend on estimating a blur kernel, and one photograph would yield a temporally coherent sequence instead of a single static image. The claim is tested in two controlled video worlds, and the recovered frames are reported to be sharper than those from conventional single-image deblurring baselines.

What carries the argument

The load-bearing mechanism is the pair formed by the temporal-average degradation operator $H(X)=\frac{1}{N}\sum_{n=0}^{N-1}x_n$ and a pre-trained transformer-based video diffusion prior. In each reverse-diffusion step, the latent video estimate is decoded and the gradient of $\|y-H(D(\hat{Z}_0(Z_t)))\|^2$ is subtracted from the sample update, so the diffusion prior determines which motions are plausible while the blurry image decides which motion actually produced the observation.

What would settle it

Record a long-exposure frame and a synchronized high-speed burst of the same real scene, then run the method and compare the recovered frames with the burst frames; if the recovered video is no closer to the burst than the sharpest single burst frame is, the temporal-average-plus-video-prior claim is falsified.

Watch

Extended reading notes

Core claim

The central claim is that a single motion-blurred image contains a short sharp video hidden by frame averaging, and that a pre-trained video diffusion model can reconstruct that video when the blur is treated as the forward model y = H(X) + e with H(X) the average over frames. The method runs Diffusion Posterior Sampling in a VQ-GAN latent space, so no blur kernel is estimated and the same pipeline covers camera shake, object motion, and non-linear trajectories. The output is a deblurred video rather than one image, with the caveat that the recovered motion can be reversed in time when the blurry image carries no directional information.

Load-bearing premise

The entire method rests on the assumption that a motion-blurred photo is the average of a short burst of sharp frames drawn from the same kind of world the video diffusion model was trained on.

Editorial extensions

If this is right

  • A single motion-blurred photograph would become a source of video, not just a static image, whenever a suitable video prior is available.
  • Deblurring would no longer require kernel estimation or explicit motion segmentation, since the temporal-average model accommodates any motion captured by the video prior.
  • Pre-trained video diffusion models could be repurposed for motion deblurring without retraining on blurry-sharp image pairs.
  • The reported robustness to frame-rate mismatch suggests that ordinary video datasets with frame averaging may suffice for training, without high-speed camera data.

Reading between the lines

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

  • Going beyond the paper, the same temporal-average inversion could be applied to other temporal degradations, such as rolling-shutter distortion or exposure bracketing, whenever the forward operator is differentiable and a video prior exists.
  • The observed time-reversal ambiguity implies that a single blurry frame cannot fix the direction of motion; adding a weak directional cue, such as a shadow or one extra frame, would be a natural testable extension.
  • One could relax the training requirement further by synthesizing blurry images from arbitrary video datasets through frame averaging, potentially scaling the method to domains outside the two toy worlds tested here.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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. This paper introduces VDM-MD, a single-image motion deblurring method that models a blurry image as the temporal average of a short video sequence and reconstructs the sharp video frames using a pre-trained video diffusion model within a Diffusion Posterior Sampling (DPS) loop. The degradation model is the linear frame-averaging operator of Eq. (4), so no explicit blur kernel is estimated. Experiments on synthetic blur generated from CLEVRER and BAIR videos report high PSNR/SSIM values, and a comparison on BAIR shows large gains over MPRNet, MTRNN, and Restormer.

Significance. The conceptual contribution is attractive: video diffusion models encode temporal dynamics that could resolve the N-to-one ambiguity inherent in temporal averaging, and the DPS formulation is standard. If the empirical claims were fully supported, this would open a promising new direction for deblurring. However, the evaluation currently conflates method quality with domain match: the VDM is trained on the same video distribution used to synthesize the blurry inputs, while the baseline deblurring networks are not. The claimed superiority over existing techniques is therefore not yet established, and the 'real-world dataset' description overstates what is tested.

major comments (3)
  1. [Section 4.2, Table 2] The comparison with MPRNet, MTRNN, and Restormer is not a fair test of the claimed superiority. VDM-MD is trained on 260K BAIR clips, whereas the baselines are pre-trained on other datasets (e.g., GoPro) and are not fine-tuned or retrained on the BAIR distribution. The large PSNR gap (24.24 vs 16.58 dB on BAIR Main) likely reflects the baselines' out-of-distribution inputs rather than the intrinsic advantage of the temporal-averaging formulation. The paper itself notes at the end of Section 4.2 that the baselines 'were never designed to handle this kind of motion-blur scenario.' To support the abstract's claim that the method outperforms existing techniques, the authors should at least fine-tune the baselines on the same synthetic blur distribution, or evaluate on a common benchmark with a matched video prior for the proposed method.
  2. [Abstract and Section 4] The phrase 'real-world datasets' is misleading. While BAIR contains real camera recordings, all blurry inputs are synthesized by averaging consecutive frames, so the degradation model is exactly the assumed H(·). No experiment uses genuine motion-blurred images captured with finite exposure, nor does any test use a blur formation model that deviates from frame averaging (e.g., with sensor noise, non-linear response, or rolling shutter). This matters because the paper's motivation is that real blur is non-convolutional and complex; without tests on actual blurred images, the method's practical advantage over existing deblurring methods is not demonstrated.
  3. [Section 4.1, Table 1] The 'temporal down-sampling' robustness experiment does not test mismatch with a real blur formation model. Both 'smoothly' and 'less smoothly blurred' images are averages of frames from the same video, just at different frame rates; they still obey the temporal-averaging assumption in Eq. (4). The conclusion that 'H(·) may not perfectly match real-motion conditions' is therefore unsupported. A meaningful mismatch test would require a physical motion-blur synthesis, such as integrating scene radiance along a camera or object trajectory over an exposure interval, or a real blurred image benchmark.
minor comments (5)
  1. [References, [18]] Reference [18] is incorrect for the CLEVRER dataset; it cites Srivastava et al. (2015) on unsupervised video representation learning. The CLEVRER dataset should be cited as Yi et al., 'CLEVRER: Collision Events for Video Representation and Reasoning,' ICLR 2020.
  2. [Section 4.2, Table 2] The terms 'BAIR Main' and 'BAIR Aux1' are used in Table 2 but never defined. Please specify what these subsets are.
  3. [Algorithm 1] Algorithm 1 uses the update Z_{t-1} = Z'_{t-1} - η_t ∇_{Z_t} ||y - \hat y_{t-1}||^2 without the 1/σ^2 factor that appears in Eq. (15). Clarify whether η_t absorbs σ^2 or whether the algorithm is missing the noise-variance normalization.
  4. [Equation (15)] Equation (15) contains a typo: the term '- β (s_{θ*}(Z_t,t) - ...)' should read '- β(t) ( ... )' to match the notation used elsewhere.
  5. [Section 4.1] CLEVRER is described as having 'minimal motion between consecutive frames,' which suggests the blur being removed may be relatively small. Reporting the average amount of motion or the effective blur kernel width would help assess task difficulty.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the forward model, learned prior, and DPS update are explicit and do not reduce to the target output by construction.

full rationale

The derivation chain is self-contained. The paper assumes the temporal-averaging degradation y = H(X) + e (Eqs. 4 and 13), uses a pre-trained video diffusion model as the prior over the latent video Z, and applies the DPS likelihood update (Eqs. 14 and 15). None of these steps fits a parameter to the target sharp frames and then relabels that fit as a prediction; the recovered output depends on the learned prior and the data-consistency gradient, so it is not forced by construction. The synthetic blur is generated by averaging held-out test frames, which makes the evaluation in-distribution and the forward model exact, but that is a benchmark-design limitation, not a logical circularity. The authors explicitly acknowledge the scope limitation in the conclusions, noting the setup 'cannot yet serve as a fully general-purpose solution.' There are no load-bearing self-citations: DPS, VQ-GAN, OpenSora, and the baselines are all external prior works. The abstract's general superiority claim over real complex motion blur is stronger than the synthetic/BAIR evidence supports, but that is a correctness and external-validity concern, not a circularity in the derivation.

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

The method introduces no new physical entities. It depends on the temporal-averaging degradation model, the existence of a sufficiently expressive video diffusion prior, and the requirement that the blurred image comes from the prior's training distribution. The main free parameters are the DPS hyperparameters and latent compression settings, which are not fully reported.

free parameters (5)
  • DPS step size eta_t = not reported
    Gradient update step size in Algorithm 1, chosen by hand and not specified in experiments; central to convergence of the reverse diffusion.
  • Likelihood noise variance sigma^2 = not reported
    Assumed noise level in Eq. (14); used in DPS gradient weighting but not given for CLEVRER or BAIR tests.
  • Number of averaged frames N = 10 (CLEVRER); not stated for BAIR
    Used in the degradation model H; averaging 10 frames creates the blurry inputs.
  • Latent compression factor p and channels c = p=8, c=12
    VQ-GAN latent settings chosen to reduce video dimensionality.
  • Diffusion steps T = not reported
    Number of reverse diffusion iterations is not specified.
assumptions (4)
  • domain assumption Motion blur is approximated by uniform averaging over N sharp frames (Eq. 4).
    The entire degradation model H is this average; real camera blur may include noise, saturation, and non-uniform exposure, which are not modeled.
  • domain assumption A video diffusion model trained on a dataset provides an accurate prior over sharp videos for that world.
    The method relies on the VDM's score to resolve the ill-posed inverse problem; if the prior is weak for the test scene, reconstruction fails.
  • ad hoc to paper The blurred image belongs to the same distribution as the VDM training videos.
    Section 3: 'as long as the image is about the given world'; this is stated as a requirement, not proven for real images.
  • standard math DPS gradient approximation (Eq. 10) is valid in latent space.
    Borrowed from Chung et al. [11]; the paper applies it without modification to the latent variable Z.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Image Motion Blur Removal in the Temporal Dimension with Video Diffusion Models." pith.science (2026). https://pith.science/paper/4HO77SSG

@misc{pith2026250112604,
  author       = {Pith},
  title        = {Pith review of: Image Motion Blur Removal in the Temporal Dimension with Video Diffusion Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4HO77SSG}},
  note         = {Machine review of arXiv:2501.12604}
}
read the original abstract

Most motion deblurring algorithms rely on spatial-domain convolution models, which struggle with the complex, non-linear blur arising from camera shake and object motion. In contrast, we propose a novel single-image deblurring approach that treats motion blur as a temporal averaging phenomenon. Our core innovation lies in leveraging a pre-trained video diffusion transformer model to capture diverse motion dynamics within a latent space. It sidesteps explicit kernel estimation and effectively accommodates diverse motion patterns. We implement the algorithm within a diffusion-based inverse problem framework. Empirical results on synthetic and real-world datasets demonstrate that our method outperforms existing techniques in deblurring complex motion blur scenarios. This work paves the way for utilizing powerful video diffusion models to address single-image deblurring challenges.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

26 extracted references · 19 canonical work pages

  1. [1]

    Removing such blur is never trivial

    INTRODUCTION Motion of the camera or objects during the exposure time leads to motion blur, which is very common in imaging pro- cesses [1]. Removing such blur is never trivial. In the past two decades, numerous algorithms have been proposed for motion deblurring (MD), and they are generally categorized into two types: those with explicit kernel estimatio...

    arXiv 2025

  2. [2]

    world models

    RELA TED WORK 2.1. Diffusion Models Diffusion models have recently shown remarkable success in generating multi-dimensional signals such as images, videos and audios. The core idea is to learn the prior distribution of data x by gradually adding Gaussian noise to a clean sample until it becomes pure noise, then training a network to reverse this noising p...

    work page

  3. [3]

    PROPOSED APPROACH We present VDM-MD, a VDM based method that formu- lates motion deblurring as an inverse problem within the DPS framework. Our key premise is that once the VDM has learned the underlying dynamics of a world represented by a training video dataset, it can naturally resolve single image motion blur as long as the image is about the given wo...

    work page

  4. [4]

    toy world

    EXPERIMENTS 4.1. Synthetic Dataset To analyze our algorithm’s performance without requiring an extensive, large-scale transformer, we used the CLEVRER dataset [18] as a “toy world.” CLEVRER features relatively simple objects obeying basic physics, with minimal motion Blurry input 17326 GT Output Blurry input 17032 GT Output Fig. 2. Motion deblurring examp...

    work page

  5. [5]

    CONCLUSIONS We introduced a single image motion deblurring approach that reinterprets the task as a video diffusion problem, re- covering multiple sharp frames instead of a single deblurred image. Central to our method is the ability to learn not Algorithm 1: VDM-MD Input: y, T Initialize ZT ∼ N(0, I); for t ← T − 1 to 0 do ˆs = sθ∗ (Zt, t); ˆZ0 = 1√¯αt (...

    work page

  6. [6]

    Motion from blur,

    Shengyang Dai and Ying Wu, “Motion from blur,” in 2008 IEEE Conference on Computer Vision and Pattern Recognition. IEEE, 2008, pp. 1–8

    work page 2008

  7. [7]

    High-quality motion deblurring from a single image,

    Qi Shan, Jiaya Jia, and Aseem Agarwala, “High-quality motion deblurring from a single image,” Acm transac- tions on graphics (tog), vol. 27, no. 3, pp. 1–10, 2008

    work page 2008

  8. [8]

    Efficient blind deblurring under high noise lev- els,

    J ´er´emy Anger, Mauricio Delbracio, and Gabriele Fac- ciolo, “Efficient blind deblurring under high noise lev- els,” in 2019 11th International Symposium on Image and Signal Processing and Analysis (ISPA). IEEE, 2019, pp. 123–128

    work page 2019

Show all 26 references
  1. [9]

    Deblurring using analysis-synthesis networks pair,

    Adam Kaufman and Raanan Fattal, “Deblurring using analysis-synthesis networks pair,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 5811–5820

    2020

  2. [10]

    Event-guided multi-patch network with self-supervision for non-uniform motion deblurring,

    Hongguang Zhang, Limeng Zhang, Yuchao Dai, Hong- dong Li, and Piotr Koniusz, “Event-guided multi-patch network with self-supervision for non-uniform motion deblurring,” International Journal of Computer Vision , vol. 131, no. 2, pp. 453–470, 2023

    2023

  3. [11]

    Multi-temporal recurrent neural net- works for progressive non-uniform single image deblur- ring with incremental temporal training,

    Dongwon Park, Dong Un Kang, Jisoo Kim, and Se Young Chun, “Multi-temporal recurrent neural net- works for progressive non-uniform single image deblur- ring with incremental temporal training,” in European Conference on Computer Vision . Springer, 2020, pp. 327–343

    2020

  4. [12]

    Image deblurring by exploring in-depth properties of transformer,

    Pengwei Liang, Junjun Jiang, Xianming Liu, and Jiayi Ma, “Image deblurring by exploring in-depth properties of transformer,” IEEE Transactions on Neural Networks and Learning Systems, 2024

    2024

  5. [13]

    Deblurgan: Blind motion deblurring using conditional adversarial net- works,

    Orest Kupyn, V olodymyr Budzan, Mykola Mykhailych, Dmytro Mishkin, and Ji ˇr´ı Matas, “Deblurgan: Blind motion deblurring using conditional adversarial net- works,” in Proceedings of the IEEE conference on com- puter vision and pattern recognition , 2018, pp. 8183– 8192

    2018

  6. [14]

    Deep multi-scale convolutional neural network for dy- namic scene deblurring,

    Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee, “Deep multi-scale convolutional neural network for dy- namic scene deblurring,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2017, pp. 3883–3891

    2017

  7. [15]

    Id-blau: Im- age deblurring by implicit diffusion-based reblurring augmentation,

    Jia-Hao Wu, Fu-Jen Tsai, Yan-Tsung Peng, Chung-Chi Tsai, Chia-Wen Lin, and Yen-Yu Lin, “Id-blau: Im- age deblurring by implicit diffusion-based reblurring augmentation,” in Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition , 2024, pp. 25847–25856

    2024

  8. [16]

    Diffusion poste- rior sampling for general noisy inverse problems,

    Hyungjin Chung, Jeongsol Kim, Michael T Mccann, Marc L Klasky, and Jong Chul Ye, “Diffusion poste- rior sampling for general noisy inverse problems,”arXiv preprint arXiv:2209.14687, 2022

    2022 arXiv

  9. [17]

    Parallel diffusion models of operator and image for blind inverse problems,

    Hyungjin Chung, Jeongsol Kim, Sehui Kim, and Jong Chul Ye, “Parallel diffusion models of operator and image for blind inverse problems,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 6059–6069

    2023

  10. [18]

    Score-based generative modeling through stochastic differential equations,

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole, “Score-based generative modeling through stochastic differential equations,” arXiv preprint arXiv:2011.13456, 2020

    2011 arXiv

  11. [19]

    Photorealistic video generation with diffusion models,

    Agrim Gupta, Lijun Yu, Kihyuk Sohn, Xiuye Gu, Meera Hahn, Li Fei-Fei, Irfan Essa, Lu Jiang, and Jos´e Lezama, “Photorealistic video generation with diffusion models,” arXiv preprint arXiv:2312.06662, 2023

    2023 arXiv

  12. [20]

    Sora: Creating video from text,

    OpenAI, “Sora: Creating video from text,” https: //openai.com/sora, 2024

    2024

  13. [21]

    Open-sora: Democratiz- ing efficient video production for all,

    Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chen- hui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You, “Open-sora: Democratiz- ing efficient video production for all,” arXiv preprint arXiv:2412.20404, 2024

    2024 arXiv

  14. [22]

    Taming transformers for high-resolution image syn- thesis,

    Patrick Esser, Robin Rombach, and Bjorn Ommer, “Taming transformers for high-resolution image syn- thesis,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2021, pp. 12873–12883

    2021

  15. [23]

    Unsupervised learning of video repre- sentations using lstms,

    Nitish Srivastava, Elman Mansimov, and Ruslan Salakhudinov, “Unsupervised learning of video repre- sentations using lstms,” in International conference on machine learning. PMLR, 2015, pp. 843–852

    2015

  16. [24]

    Multi-stage progressive image restora- tion,

    Syed Waqas Zamir, Aditya Arora, Salman Khan, Mu- nawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao, “Multi-stage progressive image restora- tion,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2021, pp. 14821–14831

    2021

  17. [25]

    Restormer: Efficient transformer for high- resolution image restoration,

    Syed Waqas Zamir, Aditya Arora, Salman Khan, Mu- nawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang, “Restormer: Efficient transformer for high- resolution image restoration,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 5728–5739

    2022

  18. [26]

    Self-supervised visual planning with temporal skip connections.,

    Frederik Ebert, Chelsea Finn, Alex X Lee, and Sergey Levine, “Self-supervised visual planning with temporal skip connections.,” CoRL, vol. 12, no. 16, pp. 23, 2017

    2017

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.