Pith. sign in

REVIEW 3 major objections 4 minor 71 references

FMA-Net++: Motion- and Exposure-Aware Joint Video Super-Resolution and Deblurring

T0 review · 3 major / 4 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read FMA-Net++ claims that joint video super-resolution and deblurring under dynamically varying exposure is best solved by conditioning degradation-kernel prediction on per-frame exposure and processing the sequence in parallel through hierarch

desk verdict Solid extension of the authors' FMA-Net line with a credible architecture and thorough ablations, but the headline real-world generalization claims rest on a self-made synthetic simulator, so the SOTA numbers should be read as in-distribution wins until code and real paired data appear. read the letter →

arxiv 2512.04390 v2 pith:EZVME6WT submitted 2025-12-04 cs.CV cs.AI

Geunhyuk Youk , Jihyong Oh , Munchurl Kim This is my paper
classification cs.CVcs.AI
keywords videosuper-resolutiondeblurringdynamicexposuremotionblurdegradationkernelestimationtemporalmodelingflow-guidedfilteringexposure-awaremodulation
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

FMA-Net++ addresses joint video super-resolution and deblurring when exposure time changes frame to frame, a realistic condition caused by auto-exposure and low-light capture. The paper claims that explicitly conditioning restoration features on per-frame exposure, and processing the whole sequence in parallel through hierarchically expanding bidirectional blocks, is both more accurate and faster than sliding-window or recurrent alternatives. On new REDS-ME and REDS-RE benchmarks, and on unseen GoPro and real smartphone footage, the method reports the best PSNR/SSIM/temporal consistency with lower runtime than prior state-of-the-art. A reader should care because it offers a practical route to restoring real videos where blur severity varies continuously.

What carries the argument

The load-bearing components are: (1) HRBP blocks, which stack hierarchical bidirectional propagation to expand temporal receptive fields while keeping all frames processed in parallel; (2) ETM, a lightweight spatial-feature-transform layer that modulates features with per-frame exposure embeddings from a frozen ETE; (3) an exposure-aware FGDF module, which samples filtering weights along estimated flow trajectories under exposure conditioning to predict degradation kernels; and (4) decoupled degradation-learning (Net_D) and restoration (Net_R) networks, where Net_R uses the predicted kernels as queries in degradation-aware attention. The new REDS-ME and REDS-RE benchmarks, built by averaging

What would settle it

Evaluate the trained model on real exposure-varying footage recorded with known shutter-speed metadata, comparing it to a fixed-exposure baseline trained without the ETE. If FMA-Net++ does not beat that baseline on such real data, or if its PSNR gap shrinks drastically relative to the synthetic benchmarks, the exposure-aware claim fails.

Watch

Extended reading notes

Core claim

The central claim is that the coupled effect of motion and dynamically varying exposure can be modeled by a non-recurrent, sequence-level network in which every refinement block is conditioned on per-frame exposure. A frozen Exposure Time-aware Feature Extractor supplies exposure embeddings; an Exposure Time-aware Modulation layer injects them as affine feature transforms; and an exposure-aware Flow-Guided Dynamic Filtering module predicts physically grounded, position-, motion-, and exposure-dependent degradation kernels. Degradation learning is decoupled from restoration: Net_D estimates the kernels and motion priors, while Net_R consumes them through degradation-aware attention to produce

Load-bearing premise

The central assumption is that blurry low-resolution video is faithfully produced by linearly averaging high-frame-rate frames and then bicubic downsampling; if real sensors' nonlinear response, noise, or continuous auto-exposure transitions invalidate that linear model, the reported SOTA numbers mostly measure the simulator, not real exposure-varying degradation.

Editorial extensions

If this is right

  • If correct, sequence-level hierarchical processing is a better default than recurrent propagation for VSRDB, because it avoids the sequential bottleneck while capturing long-range context.
  • Explicit exposure conditioning matters most when exposure changes within a clip: on REDS-RE the gap over fixed-exposure baselines widens, and removing the ETE degrades out-of-distribution results the most.
  • The decoupled design means the expensive exposure-aware FGDF is used only for prior estimation in Net_D, so restoration stays cheap, which is why the full model runs at 0.074 s/frame.
  • A frozen ETE trained on only five discrete exposure labels still produces a feature space that appears to interpolate smoothly enough for real-world auto-exposure videos, suggesting exposure conditioning need not be end-to-end optimized.
  • The REDS-ME and REDS-RE benchmarks give the community a standard way to test VSRDB under controlled exposure variation.

Reading between the lines

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

  • The paper's real-world evidence is qualitative only; a direct test on exposure-varying footage with recorded shutter times would settle whether the exposure anchors generalize to continuous, unseen exposure states.
  • Because the synthetic pipeline is linear frame averaging, adding realistic sensor noise, a camera response function, or spatially varying illumination to the training data would likely improve real-world robustness and would test the model's limits.
  • The same HRBP plus ETM recipe could transfer to neighbouring tasks such as video frame interpolation or low-light video enhancement, where exposure-dependent motion blur also couples with restoration.
  • The finding that deliberately wrong exposure guidance degrades but does not catastrophically break restoration suggests the network uses exposure cues as an auxiliary prior rather than a hard gate; quantifying this with adversarial exposure-label errors would clarify its robustness.
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 / 4 minor

Summary. The paper proposes FMA-Net++, a non-recurrent, sequence-level architecture for joint video super-resolution and deblurring (VSRDB) that explicitly conditions on per-frame exposure information. The core building block is the Hierarchical Refinement with Bidirectional Propagation (HRBP) module, which processes all frames in parallel while expanding the temporal receptive field across stacked blocks. Exposure conditioning is introduced through a pretrained Exposure Time-aware Feature Extractor (ETE) whose embeddings drive ETM layers inside each HRBP block, and an exposure-aware Flow-Guided Dynamic Filtering module predicts degradation kernels. The authors also introduce REDS-ME and REDS-RE synthetic benchmarks derived from REDS by high-frame-rate interpolation, linear frame averaging at five exposure ratios, and bicubic downsampling. Experiments report state-of-the-art PSNR/SSIM/tOF on REDS4-ME, REDS-RE, and GoPro, with a runtime of 0.074s/frame, together with qualitative results on real smartphone videos.

Significance. If the results hold, FMA-Net++ makes a meaningful advance: it offers a parallelizable alternative to recurrent and sliding-window temporal modeling for VSRDB, and it is one of the few methods that explicitly address frame-wise exposure variation. The empirical program is thorough in several respects: ablations cover the number of HRBP blocks (Table 5), the number of multi-flow pairs (Table 7), loss coefficients (Table 9), and temporal modeling strategy (Table 3); the sensitivity study in Table 6 directly varies the ETE guidance and shows a monotonic effect, providing some evidence of a causal role; and the GoPro evaluation is an external check whose blur is real (only the LR downsampling is synthetic), which is a genuine strength. However, the central quantitative evidence for the exposure-aware contribution currently rests on benchmarks generated with the same linear averaging pipeline that produced the ETE's supervised exposure labels, and the ETE ablation is confounded by parameter count. These issues block full confidence in the claimed real-world dynamic-exposure generalization.

major comments (3)
  1. [Sec. 8.2; Eq. (15); Table 4] The ETE is pretrained on the same five discrete exposure levels (5:1-5:5) that define the REDS-ME and REDS-RE test sets. Because the ETE is frozen and then used for ETM conditioning, the reported improvements (+0.41 dB on REDS-RE, +0.71 dB on GoPro) may partly reflect the model learning the simulator's discrete exposure labels rather than a continuous physical quantity. The GoPro dataset is real blur but does not contain dynamic exposure variation, and the real-world smartphone results use only no-reference metrics. To support the central 'real-world dynamic exposure' claim, please add evaluation on a more realistic degradation model (e.g., including sensor noise, nonlinear response, and continuous exposure values) or on real auto-exposure sequences with reference ground truth.
  2. [Sec. 5.2; Table 4] The ETE ablation in Table 4 varies not only the presence of ETE guidance but also the parameter count: 'w/o ETE' has 9.8M parameters while 'w/ ETE' has 12.8M (+3M parameters). The larger gains on REDS-RE and GoPro could therefore be attributable to extra capacity rather than exposure awareness. Please provide a matched-capacity control (e.g., a learned per-frame embedding of the same dimension and parameter cost but without exposure supervision) or otherwise decorrelate capacity from the exposure-conditioning signal. Without this, the causal role of the ETE is not established.
  3. [Sec. 3.4; Table 6; Fig. 9] The paper claims the predicted degradation kernels are 'physically grounded' in motion and exposure. The supporting evidence is qualitative (Fig. 9) and a small sensitivity gap: in Table 6, correct 5:5 guidance gives 29.24 dB vs 29.07 dB for the most incorrect 5:1 guidance, a 0.17 dB range with the no-ETE baseline at 29.12 dB. This is consistent with ETE having an effect, but it is not a quantitative test of physical grounding. Please provide a metric that links kernel statistics (e.g., spatial spread or variance) to the actual exposure time, ideally on data outside the training label set or on the real GoPro frames.
minor comments (4)
  1. [Table 2] The row label 'EV-DeblurVSR*' uses inconsistent capitalization; elsewhere in the paper it is 'Ev-DeblurVSR'.
  2. [Eq. (3)] The notation x_{t→r} and the offsets p_k should be defined in the main text. The same operation is expressed as Y_{t→i} in Eq. (4); using a single notation for warped features would improve readability.
  3. [Sec. 4.2; Table 1] Please report the hardware and measurement conditions for the runtime numbers (e.g., whether model loading, warping, and I/O are included). This is relevant because the claimed speedup is a headline result.
  4. [Suppl. Sec. 9.1; Fig. 8] The t-SNE visualization should state how many frames per exposure level were used and whether the embeddings come from the training or test split. The acknowledged overlap between 5:4 and 5:5 merits a brief discussion of its implications for the most challenging exposure levels.

Circularity Check

0 steps flagged · score 2.0 of 10

No construction-level circularity; exposure-aware gains are empirically tested on external GoPro data, though REDS-ME/RE are author-created synthetic benchmarks.

full rationale

Stepping through the paper's derivation chain, I find no circular reduction. The degradation model of Eq. (9) is explicitly presented as a 'practical, learnable approximation' of the continuous physical integral (Eq. 8); Eq. (4) is a reconstruction objective for the intermediate kernels, not a fitted formula renamed as a prediction of the final HR output. The ETE is pretrained on the same five discrete exposure pseudo-labels used to build REDS-ME, so REDS-ME/RE results are partly in-distribution with respect to those labels, and the benchmarks are author-created; however, the final restoration is still measured against sharp HR ground truth, all baselines are retrained on the same synthetic set, and the evaluation includes the external GoPro dataset (real blurry HR frames, bicubic-downsampled only for LR) plus no-reference real-world clips. That makes the exposure-aware gains empirically testable rather than forced by construction. The self-citations of FMA-Net are heavy—FGDF and the three-frame kernel parameterization are inherited from [54]—but this is architectural ancestry and a retrained baseline, not a load-bearing citation that substitutes for evidence; no uniqueness theorem is imported. Suppl. 11.1 candidly concedes that linear frame averaging may not capture nonlinear sensor response and noise; this threatens the real-world-generalization interpretation, but it is a dataset-validity caveat, not a circularity. Overall, the strongest caveats (author-created synthetic benchmarks, discrete exposure anchors, no GT on real clips) are correctness/robustness risks, not construction-level circularity.

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

The paper makes no analytic derivation: its physics-based degradation model is a stated approximation (Eqs. 9-10), and the claimed SOTA rests on hand-chosen hyperparameters and synthetic-data assumptions. The main free parameters are the discrete exposure bins, loss weights, and architectural sizes, all tuned on the evaluation benchmarks the authors designed. There are no invented physical entities; the only new objects are learned features and kernels whose independent validity is not established outside the paper.

free parameters (4)
  • Exposure level ratios (5:1..5:5) = 5 discrete levels; 5:1 to 5:5
    Hand-chosen synthetic exposure bins define both the ETE's contrastive labels and the REDS-ME/RE evaluation, so SOTA claims are conditioned on this discretization.
  • Loss coefficients lambda_1, lambda_2, lambda_3 = 1e-4, 1e-4, 0.1
    Ablated on REDS4-ME-5:5 (Table 9); final values chosen for best PSNR/tOF, so performance is partly tuned to the test split.
  • Architecture sizes M, n, k_d = M=4 HRBP blocks, n=9 multi-flow pairs, k_d=20
    Selected by ablations (Tables 5, 7 and Supplement), trading accuracy and runtime; the reported SOTA holds for this configuration.
  • ETE contrastive temperature alpha = 0.5
    Chosen for ETE pretraining; affects exposure embedding quality and hence all downstream conditioning.
assumptions (4)
  • domain assumption Linear averaging of interpolated frames models exposure-dependent motion blur
    Used to synthesize REDS-ME/RE and to justify Eqs. (8)-(10); Suppl. 11.1 concedes this omits nonlinear sensor response and noise.
  • domain assumption ETE exposure embeddings trained on five discrete synthetic levels transfer to continuous real-world exposures
    Central to real-world generalization claims (Suppl. Sec. 10); no real exposure-estimation accuracy is measured.
  • domain assumption 2D optical flow plus occlusion masks approximate the continuous motion field
    Used in HRBP and exposure-aware FGDF (Sec. 3.3-3.4); Suppl. 11.2 concedes breakdown under out-of-plane and non-rigid motion.
  • domain assumption RAFT pseudo-GT flow is reliable supervision
    Used in the Net_D loss (Eq. 16, Suppl. 7.3); errors in RAFT propagate into the learned motion priors.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FMA-Net++: Motion- and Exposure-Aware Joint Video Super-Resolution and Deblurring." pith.science (2026). https://pith.science/paper/EZVME6WT

@misc{pith2026251204390,
  author       = {Pith},
  title        = {Pith review of: FMA-Net++: Motion- and Exposure-Aware Joint Video Super-Resolution and Deblurring},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EZVME6WT}},
  note         = {Machine review of arXiv:2512.04390}
}
read the original abstract

Joint video super-resolution and deblurring (VSRDB) requires both efficient long-range temporal modeling and robustness to frame-wise exposure-duration variation, which changes the extent of motion blur across video frames. We propose FMA-Net++, a non-recurrent, sequence-level framework built from Hierarchical Refinement with Bidirectional Aggregation (HRBA) blocks. By stacking HRBA blocks, FMA-Net++ processes video frames in parallel while hierarchically expanding the temporal receptive field, avoiding the limited temporal receptive field of sliding-window designs and the sequential bottleneck of recurrent ones. To handle exposure-duration-dependent blur, we introduce an Exposure Time-aware Modulation (ETM) layer that conditions HRBA features on exposure embeddings from an Exposure Time-aware Feature Extractor (ETE). The conditioned features guide an exposure-aware flow-guided dynamic filtering module to predict motion- and exposure-aware degradation kernels. FMA-Net++ decouples degradation learning from restoration: the former predicts degradation priors and the latter exploits them for efficient high-resolution restoration. To evaluate VSRDB under controlled exposure-duration variation, we introduce the REDS-ME (multi-exposure) and REDS-RE (random-exposure) benchmarks. Trained solely on synthetic data, FMA-Net++ achieves state-of-the-art accuracy and temporal consistency on these benchmarks. It further shows strong out-of-distribution performance on GoPro and challenging real-world videos, while outperforming recent methods in both restoration quality and inference speed.

Figures

Figures reproduced from arXiv: 2512.04390 by the authors.

Figure 1
Figure 1. FMA-Net++ outperforms state-of-the-art methods in real-world qualitative results and quantitative benchmarks for VSRDB. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Conceptual illustration and overview of the FMA-Net++ framework. [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Architecture of FMA-Net++ for joint video super-resolution and deblurring (VSRDB). [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (12 more)
Figure 4
Figure 4. Figure 4: Details of an HRBP block. (a) Structure of the HRBP [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Qualitative comparisons of ×4 VSRDB on REDS4-ME￾5: 5 and GoPro [33]. Each scene contains severe motion blur with different characteristics. Best viewed in zoom. HAT [7]), single-image deblurring (Restormer [55], FFT￾former [23]), VSR (BasicVSR++ [5], IART [53]), video …
Figure 6
Figure 6. Figure 6: An overview of our three-stage training strategy. [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]
Figure 7
Figure 7. Figure 7: Example frames from our REDS-ME dataset across five exposure levels and two different scenes. Each column corresponds [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]
Figure 9
Figure 9. Figure 9: Effect of ETE guidance on the exposure-aware degrada [PITH_FULL_IMAGE:figures/full_fig_p015_9.png]
Figure 8
Figure 8. Figure 8: t-SNE [44] visualization of exposure time-aware fea￾tures ui extracted by ETE, showing their distinguishability across different exposure levels. 9. Further Ablation Studies In this section, we provide further ablation studies and de￾tailed visual analyses that were om…
Figure 10
Figure 10. Figure 10: Effect of the number of multi-flow-mask pairs ( [PITH_FULL_IMAGE:figures/full_fig_p016_10.png]
Figure 11
Figure 11. Figure 11: Visualization of the progressive feature refinement [PITH_FULL_IMAGE:figures/full_fig_p016_11.png]
Figure 12
Figure 12. Figure 12: Visualization of synthesized exposure trajectories in the [PITH_FULL_IMAGE:figures/full_fig_p017_12.png]
Figure 13
Figure 13. Figure 13: Qualitative comparison between FMA-Net++ (Ours) [PITH_FULL_IMAGE:figures/full_fig_p017_13.png]
Figure 14
Figure 14. Figure 14: Additional qualitative comparisons on the REDS4-ME- [PITH_FULL_IMAGE:figures/full_fig_p019_14.png]
Figure 15
Figure 15. Figure 15: Additional qualitative comparisons on challenging real-world videos captured with smartphones. These videos contain con [PITH_FULL_IMAGE:figures/full_fig_p020_15.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

71 extracted references · 5 linked inside Pith

  1. [1]

    Self-supervised deep blind video super-resolution.IEEE Transactions on Pattern Anal- ysis and Machine Intelligence, 46(7):4641–4653, 2024

    Haoran Bai and Jinshan Pan. Self-supervised deep blind video super-resolution.IEEE Transactions on Pattern Anal- ysis and Machine Intelligence, 46(7):4641–4653, 2024. 2

    2024

  2. [2]

    Video super-resolution transformer.arXiv preprint arXiv:2106.06847, 2021

    Jiezhang Cao, Yawei Li, Kai Zhang, and Luc Van Gool. Video super-resolution transformer.arXiv preprint arXiv:2106.06847, 2021. 3

    arXiv 2021

  3. [3]

    Basicvsr: The search for essential com- ponents in video super-resolution and beyond

    Kelvin CK Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Basicvsr: The search for essential com- ponents in video super-resolution and beyond. InProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4947–4956, 2021. 2, 3, 5, 7, 1

    2021

  4. [4]

    Understanding deformable alignment in video super-resolution

    Kelvin CK Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Understanding deformable alignment in video super-resolution. InProceedings of the AAAI Confer- ence on Artificial Intelligence, pages 973–981, 2021. 4

    2021

  5. [5]

    Basicvsr++: Improving video super- resolution with enhanced propagation and alignment

    Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Basicvsr++: Improving video super- resolution with enhanced propagation and alignment. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5972–5981, 2022. 2, 3, 4, 5, 6, 7, 1

    2022

  6. [6]

    Learning to see in the dark

    Chen Chen, Qifeng Chen, Jia Xu, and Vladlen Koltun. Learning to see in the dark. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3291–3300, 2018. 1

    2018

  7. [7]

    Activating more pixels in image super- resolution transformer

    Xiangyu Chen, Xintao Wang, Jiantao Zhou, Yu Qiao, and Chao Dong. Activating more pixels in image super- resolution transformer. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22367–22377, 2023. 6, 7

    2023

  8. [8]

    Stable long-term recurrent video super-resolution

    Benjamin Naoto Chiche, Arnaud Woiselle, Joana Frontera- Pons, and Jean-Luc Starck. Stable long-term recurrent video super-resolution. InProceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition, pages 837–846, 2022. 3

    2022

Show all 71 references
  1. [9]

    Learning temporal coherence via self- supervision for gan-based video generation.ACM Transac- tions on Graphics (TOG), 39(4):75–1, 2020

    Mengyu Chu, You Xie, Jonas Mayer, Laura Leal-Taix ´e, and Nils Thuerey. Learning temporal coherence via self- supervision for gan-based video generation.ACM Transac- tions on Graphics (TOG), 39(4):75–1, 2020. 6

    2020

  2. [10]

    An image is worth 16x16 words: Trans- formers for image recognition at scale.arXiv preprint arXiv:2010.11929, 2020

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, et al. An image is worth 16x16 words: Trans- formers for image recognition at scale.arXiv preprint ar...

    2010 arXiv

  3. [11]

    High-resolution optical flow and frame-recurrent network for video super-resolution and deblurring.Neurocomputing, 489:128–138, 2022

    Ning Fang and Zongqian Zhan. High-resolution optical flow and frame-recurrent network for video super-resolution and deblurring.Neurocomputing, 489:128–138, 2022. 1, 3

    2022

  4. [12]

    Recurrent back-projection network for video super- resolution

    Muhammad Haris, Gregory Shakhnarovich, and Norimichi Ukita. Recurrent back-projection network for video super- resolution. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3897– 3906, 2019. 2

    2019

  5. [13]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770–778, 2016. 2

    2016

  6. [14]

    Many-to-many splatting for efficient video frame interpola- tion

    Ping Hu, Simon Niklaus, Stan Sclaroff, and Kate Saenko. Many-to-many splatting for efficient video frame interpola- tion. InProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, pages 3553–3562,

  7. [15]

    Spatial transformer networks.Advances in Neural Informa- tion Processing Systems, 28, 2015

    Max Jaderberg, Karen Simonyan, Andrew Zisserman, et al. Spatial transformer networks.Advances in Neural Informa- tion Processing Systems, 28, 2015. 4, 1

    2015

  8. [16]

    Dynamic filter networks.Advances in Neural Infor- mation Processing Systems, 29, 2016

    Xu Jia, Bert De Brabandere, Tinne Tuytelaars, and Luc V Gool. Dynamic filter networks.Advances in Neural Infor- mation Processing Systems, 29, 2016. 5, 6

    2016

  9. [17]

    Deep video super-resolution network us- ing dynamic upsampling filters without explicit motion com- pensation

    Younghyun Jo, Seoung Wug Oh, Jaeyeon Kang, and Seon Joo Kim. Deep video super-resolution network us- ing dynamic upsampling filters without explicit motion com- pensation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3224– 3232, 20...

    2018

  10. [18]

    Event-enhanced blurry video super-resolution

    Dachun Kai, Yueyi Zhang, Jin Wang, Zeyu Xiao, Zhiwei Xiong, and Xiaoyan Sun. Event-enhanced blurry video super-resolution. InProceedings of the AAAI Conference on Artificial Intelligence, pages 4175–4183, 2025. 1, 2, 3, 6, 7

    2025

  11. [19]

    Musiq: Multi-scale image quality transformer

    Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 5148–5157, 2021. 6

    2021

  12. [20]

    Supervised contrastive learning.Advances in Neural Information Processing Systems, 33:18661–18673,

    Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning.Advances in Neural Information Processing Systems, 33:18661–18673,

  13. [21]

    Event-guided deblurring of unknown exposure time videos

    Taewoo Kim, Jeongmin Lee, Lin Wang, and Kuk-Jin Yoon. Event-guided deblurring of unknown exposure time videos. InEuropean Conference on Computer Vision, pages 519–

  14. [22]

    Adam: A method for stochastic optimization.arXiv preprint arXiv:1412.6980,

    Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization.arXiv preprint arXiv:1412.6980,

    arXiv

  15. [23]

    Efficient frequency domain-based trans- formers for high-quality image deblurring

    Lingshun Kong, Jiangxin Dong, Jianjun Ge, Mingqiang Li, and Jinshan Pan. Efficient frequency domain-based trans- formers for high-quality image deblurring. InProceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, pages 5886–5895, 2023. 6, 7 9

    2023

  16. [24]

    Dy- navsr: Dynamic adaptive blind video super-resolution

    Suyoung Lee, Myungsub Choi, and Kyoung Mu Lee. Dy- navsr: Dynamic adaptive blind video super-resolution. In Proceedings of the IEEE/CVF Winter Conference on Appli- cations of Computer Vision, pages 2093–2102, 2021. 2

  17. [25]

    A simple baseline for video restoration with grouped spatial- temporal shift

    Dasong Li, Xiaoyu Shi, Yi Zhang, Ka Chun Cheung, Simon See, Xiaogang Wang, Hongwei Qin, and Hongsheng Li. A simple baseline for video restoration with grouped spatial- temporal shift. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 98...

    2023

  18. [26]

    Mucan: Multi-correspondence aggregation network for video super-resolution

    Wenbo Li, Xin Tao, Taian Guo, Lu Qi, Jiangbo Lu, and Jiaya Jia. Mucan: Multi-correspondence aggregation network for video super-resolution. InEuropean Conference on Com- puter Vision, pages 335–351. Springer, 2020. 2, 3, 5

    2020

  19. [27]

    Swinir: Image restoration us- ing swin transformer

    Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration us- ing swin transformer. InProceedings of the IEEE/CVF Inter- national Conference on Computer Vision, pages 1833–1844,

  20. [28]

    Vrt: A video restoration transformer.arXiv preprint arXiv:2201.12288, 2022

    Jingyun Liang, Jiezhang Cao, Yuchen Fan, Kai Zhang, Rakesh Ranjan, Yawei Li, Radu Timofte, and Luc Van Gool. Vrt: A video restoration transformer.arXiv preprint arXiv:2201.12288, 2022. 3

    2022 arXiv

  21. [29]

    Recurrent video restoration trans- former with guided deformable attention.Advances in Neu- ral Information Processing Systems, 35:378–393, 2022

    Jingyun Liang, Yuchen Fan, Xiaoyu Xiang, Rakesh Ranjan, Eddy Ilg, Simon Green, Jiezhang Cao, Kai Zhang, Radu Timofte, and Luc V Gool. Recurrent video restoration trans- former with guided deformable attention.Advances in Neu- ral Information Processing Systems, 35:378–393, 202...

    2022

  22. [30]

    Fdan: Flow-guided deformable alignment network for video super-resolution

    Jiayi Lin, Yan Huang, and Liang Wang. Fdan: Flow-guided deformable alignment network for video super-resolution. arXiv preprint arXiv:2105.05640, 2021. 2

    2021 arXiv

  23. [31]

    Learning trajectory-aware transformer for video super- resolution

    Chengxu Liu, Huan Yang, Jianlong Fu, and Xueming Qian. Learning trajectory-aware transformer for video super- resolution. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5687– 5696, 2022. 2, 3, 5

    2022

  24. [32]

    completely blind

    Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Mak- ing a “completely blind” image quality analyzer.IEEE Sig- nal Processing Letters, 20(3):209–212, 2012. 6

    2012

  25. [33]

    Deep multi-scale convolutional neural network for dynamic scene deblurring

    Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3883– 3891, 2017. 1, 2, 4, 6, 7, 3, 8

    2017

  26. [34]

    Ntire 2019 challenge on video deblurring and super- resolution: Dataset and study

    Seungjun Nah, Sungyong Baik, Seokil Hong, Gyeongsik Moon, Sanghyun Son, Radu Timofte, and Kyoung Mu Lee. Ntire 2019 challenge on video deblurring and super- resolution: Dataset and study. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Works...

    2019

  27. [35]

    Motion deblurring in the wild

    Mehdi Noroozi, Paramanand Chandramouli, and Paolo Favaro. Motion deblurring in the wild. InGerman Confer- ence on Pattern Recognition, pages 65–77. Springer, 2017. 3, 6

    2017

  28. [36]

    Demfi: deep joint deblur- ring and multi-frame interpolation with flow-guided attentive correlation and recursive boosting

    Jihyong Oh and Munchurl Kim. Demfi: deep joint deblur- ring and multi-frame interpolation with flow-guided attentive correlation and recursive boosting. InEuropean Conference on Computer Vision, pages 198–215. Springer, 2022. 1, 3, 4

    2022

  29. [37]

    Deep blind video super-resolution

    Jinshan Pan, Haoran Bai, Jiangxin Dong, Jiawei Zhang, and Jinhui Tang. Deep blind video super-resolution. InProceed- ings of the IEEE/CVF International Conference on Com- puter Vision, pages 4811–4820, 2021. 2, 6, 7

    2021

  30. [38]

    Joint video multi-frame interpola- tion and deblurring under unknown exposure time

    Wei Shang, Dongwei Ren, Yi Yang, Hongzhi Zhang, Kede Ma, and Wangmeng Zuo. Joint video multi-frame interpola- tion and deblurring under unknown exposure time. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13935–13944, 2023. 3, 4, 6

    2023

  31. [39]

    Xvfi: extreme video frame interpolation

    Hyeonjun Sim, Jihyong Oh, and Munchurl Kim. Xvfi: extreme video frame interpolation. InProceedings of the IEEE/CVF International Conference on Computer Vision,

  32. [40]

    Deep video deblurring for hand-held cameras

    Shuochen Su, Mauricio Delbracio, Jue Wang, Guillermo Sapiro, Wolfgang Heidrich, and Oliver Wang. Deep video deblurring for hand-held cameras. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1279–1288, 2017. 3, 6

    2017

  33. [41]

    Detail-revealing deep video super-resolution

    Xin Tao, Hongyun Gao, Renjie Liao, Jue Wang, and Ji- aya Jia. Detail-revealing deep video super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4472–4480, 2017. 3, 5

    2017

  34. [42]

    Raft: Recurrent all-pairs field transforms for optical flow

    Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. InEuropean Conference on Computer Vision, pages 402–419. Springer, 2020. 2

    2020

  35. [43]

    Tdan: Temporally-deformable alignment network for video super-resolution

    Yapeng Tian, Yulun Zhang, Yun Fu, and Chenliang Xu. Tdan: Temporally-deformable alignment network for video super-resolution. InProceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition, pages 3360–3369, 2020. 2, 3

    2020

  36. [44]

    Visualizing data using t-sne.Journal of Machine Learning Research, 9 (11), 2008

    Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne.Journal of Machine Learning Research, 9 (11), 2008. 4

    2008

  37. [45]

    Attention is all you need.Advances in neural information processing systems, 30, 2017

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszko- reit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need.Advances in neural information processing systems, 30, 2017. 4, 5, 2

    2017

  38. [46]

    Deep video super-resolution using hr optical flow estimation.IEEE Transactions on Image Pro- cessing, 29:4323–4336, 2020

    Longguang Wang, Yulan Guo, Li Liu, Zaiping Lin, Xinpu Deng, and Wei An. Deep video super-resolution using hr optical flow estimation.IEEE Transactions on Image Pro- cessing, 29:4323–4336, 2020. 2

    2020

  39. [47]

    Recovering realistic texture in image super-resolution by deep spatial feature transform

    Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Recovering realistic texture in image super-resolution by deep spatial feature transform. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 606–615, 2018. 5, 2

    2018

  40. [48]

    Recovering realistic texture in image super-resolution by deep spatial feature transform

    Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Recovering realistic texture in image super-resolution by deep spatial feature transform. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018. 8

    2018

  41. [49]

    Esrgan: En- hanced super-resolution generative adversarial networks

    Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: En- hanced super-resolution generative adversarial networks. In European Conference on Computer Vision Workshops, 2018. 6 10

    2018

  42. [50]

    Edvr: Video restoration with enhanced deformable convolutional networks

    Xintao Wang, Kelvin CK Chan, Ke Yu, Chao Dong, and Chen Change Loy. Edvr: Video restoration with enhanced deformable convolutional networks. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0–0, 2019. 2, 3, 4, 5

    2019

  43. [51]

    Image quality assessment: from error visibility to structural similarity.IEEE Transactions on Image Process- ing, 13(4):600–612, 2004

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Si- moncelli. Image quality assessment: from error visibility to structural similarity.IEEE Transactions on Image Process- ing, 13(4):600–612, 2004. 6

    2004

  44. [52]

    Event- based blurry frame interpolation under blind exposure

    Wenming Weng, Yueyi Zhang, and Zhiwei Xiong. Event- based blurry frame interpolation under blind exposure. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 1588–1598, 2023. 1, 3

    2023

  45. [53]

    Enhancing video super-resolution via im- plicit resampling-based alignment

    Kai Xu, Ziwei Yu, Xin Wang, Michael Bi Mi, and Angela Yao. Enhancing video super-resolution via im- plicit resampling-based alignment. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2546–2555, 2024. 2, 6, 7

    2024

  46. [54]

    Fma- net: Flow-guided dynamic filtering and iterative feature re- finement with multi-attention for joint video super-resolution and deblurring

    Geunhyuk Youk, Jihyong Oh, and Munchurl Kim. Fma- net: Flow-guided dynamic filtering and iterative feature re- finement with multi-attention for joint video super-resolution and deblurring. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pa...

  47. [55]

    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. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5728– 5739,...

    2022

  48. [56]

    Extracting motion and ap- pearance via inter-frame attention for efficient video frame interpolation

    Guozhen Zhang, Yuhan Zhu, Haonan Wang, Youxin Chen, Gangshan Wu, and Limin Wang. Extracting motion and ap- pearance via inter-frame attention for efficient video frame interpolation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 568...

    2023

  49. [57]

    Spatio- temporal deformable attention network for video deblurring

    Huicong Zhang, Haozhe Xie, and Hongxun Yao. Spatio- temporal deformable attention network for video deblurring. InEuropean Conference on Computer Vision, pages 581–

  50. [58]

    Blur-aware spatio-temporal sparse transformer for video deblurring

    Huicong Zhang, Haozhe Xie, and Hongxun Yao. Blur-aware spatio-temporal sparse transformer for video deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 2673–2681, 2024. 2, 3, 6, 7

    2024

  51. [59]

    Adversarial spatio-temporal learning for video deblurring.IEEE Transactions on Image Processing, 28(1):291–301, 2018

    Kaihao Zhang, Wenhan Luo, Yiran Zhong, Lin Ma, Wei Liu, and Hongdong Li. Adversarial spatio-temporal learning for video deblurring.IEEE Transactions on Image Processing, 28(1):291–301, 2018. 2, 3

    2018

  52. [60]

    Video frame interpolation without temporal priors.Advances in Neural Information Processing Systems, 33:13308–13318,

    Youjian Zhang, Chaoyue Wang, and Dacheng Tao. Video frame interpolation without temporal priors.Advances in Neural Information Processing Systems, 33:13308–13318,

  53. [61]

    Efficient spatio-temporal recurrent neural network for video deblurring

    Zhihang Zhong, Ye Gao, Yinqiang Zheng, and Bo Zheng. Efficient spatio-temporal recurrent neural network for video deblurring. InEuropean Conference on Computer Vision, pages 191–207. Springer, 2020. 2

    2020

  54. [62]

    Davanet: Stereo deblur- ring with view aggregation

    Shangchen Zhou, Jiawei Zhang, Wangmeng Zuo, Haozhe Xie, Jinshan Pan, and Jimmy S Ren. Davanet: Stereo deblur- ring with view aggregation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10996–11005, 2019. 3, 6

    2019

  55. [63]

    Led- net: Joint low-light enhancement and deblurring in the dark

    Shangchen Zhou, Chongyi Li, and Chen Change Loy. Led- net: Joint low-light enhancement and deblurring in the dark. InEuropean conference on computer vision, pages 573–589. Springer, 2022. 1

    2022

  56. [64]

    Deep recurrent neural net- work with multi-scale bi-directional propagation for video deblurring

    Chao Zhu, Hang Dong, Jinshan Pan, Boyang Liang, Yuhao Huang, Lean Fu, and Fei Wang. Deep recurrent neural net- work with multi-scale bi-directional propagation for video deblurring. InProceedings of the AAAI Conference on Arti- ficial Intelligence, pages 3598–3607, 2022. 3 11 ...

    2022

  57. [67]

    Detailed Problem Formulation As mentioned in Sec

    Detailed Method 7.1. Detailed Problem Formulation As mentioned in Sec. 3.1 of the main paper, we address VSRDB under dynamically varying exposure. While most existing video restoration methods [3, 5, 29] have advanced temporal modeling, they typically assume a fixed exposure t...

  58. [68]

    4 of the main paper

    Detailed Experimental Setup We provide the implementation details omitted from Sec. 4 of the main paper. 2 5 ∶ 1 5 ∶ 2 5 ∶ 3 5 ∶ 4 5 ∶ 5 Figure 7. Example frames from our REDS-ME dataset across five exposure levels and two different scenes. Each column corresponds to an exposu...

    1920

  59. [69]

    5 of the main paper

    Further Ablation Studies In this section, we provide further ablation studies and de- tailed visual analyses that were omitted from Sec. 5 of the main paper. We first present detailed quantitative and quali- tative analyses of our core contributions, namely exposure- aware mod...

  60. [70]

    1(a) and Fig

    Additional Qualitative Results We provide additional qualitative comparisons comple- menting the results shown in the main paper (Fig. 1(a) and Fig. 5). Further results on challenging scenes from the REDS4-ME-5 : 5and GoPro test sets are shown in Fig. 14. Finally, examples on ...

  61. [71]

    Limitations 11.1. Limitations of Synthetic Datasets Our proposed benchmarks, REDS-ME and REDS-RE, are constructed by averaging high-framerate frames to simulate motion blur under varying exposure conditions. While this approach follows standard protocols in video deblurring an...

  62. [538]

    Springer, 2022. 1, 3

    2022

  63. [2024]

    1, 2, 3, 4, 5, 6, 7, 8

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.