Pith. sign in

REVIEW 3 major objections 4 minor 50 references

Small Clips, Big Gains: Learning Long-Range Refocused Temporal Information for Video Super-Resolution

T0 review · 3 major / 4 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read Recurrent video super-resolution can learn long-range temporal detail by backpropagating through short clips while a single forward pass supplies full-video context.

desk verdict A useful TBPTT-for-VSR training idea and a clean RITB block, but the reported training-time numbers don't match Algorithm 1; the efficiency claim needs a fix before I'd trust it. read the letter →

arxiv 2505.02159 v1 pith:BELGBNKB submitted 2025-05-04 cs.CV

Xingyu Zhou , Wei Long , Jingbo Lu , Shiyin Jiang , Weiyi You , Haifeng Wu , Shuhang Gu This is my paper · ORCID
classification cs.CV
keywords videosuper-resolutionrecurrentnetworkstruncatedbackpropagationthroughtimelong-rangetemporalinformationtransformerrefocusedattentionReLU2activationREDSdataset
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

Recurrent video super-resolution models get sharper results when they see long-range temporal context, but training them on long clips is memory-hungry and slow. This paper argues that a model can learn those long-range patterns while training on short clips: run one forward pass over the whole long video to obtain hidden states, then sample short clips and backpropagate only through them, using the precomputed context from just outside each clip. The paper also introduces a transformer block whose attention uses $\mathrm{ReLU}^2$ instead of SoftMax and whose feed-forward network is gated by the aligned previous-frame hidden state, so the model suppresses misaligned temporal features and concentrates on useful ones. The paper reports that the strategy is generic: it improves BasicVSR, BasicVSR++, and PSRT by 0.10–0.14 dB on REDS4, and the complete LRTI-VSR model reaches 33.06 dB on REDS4 and 34.81 dB on ToS3 at $4\times$ super-resolution.

What carries the argument

The load-bearing mechanism is the truncated-backpropagation loop in Algorithm 1: the bidirectional recurrent VSR model first runs a forward-only pass over the entire long video and stores hidden states for every frame; then it samples a short clip at a random start position and backpropagates through that clip alone, feeding the stored hidden states from two frames before and after the clip as context. The companion component is the refocused intra&inter-frame transformer block (RITB), which draws Query tokens from the current frame, Key/Value tokens from the current frame plus the aligned hidden states of the previous two frames, computes attention with $\mathrm{ReLU}^2$ in place of SoftMax, and multiplies the aligned previous-frame hidden state into the feed-forward network through a refocused gated unit. This splits the cost of long-range context (one forward pass per outer iteration) from the cost of learning (many short backpropagations), which is what makes the efficiency claim work.

What would settle it

Train the same recurrent VSR architecture on REDS4 under two protocols matched in model, clip length, sampling schedule, and iterations: the paper's one-forward-pass-per-outer-loop protocol, and a version that recomputes the long-video hidden states after every short-clip gradient update. If the recomputed-context version outperforms the paper's by a clear margin, the fixed-context assumption is the bottleneck; if the two are close, a single forward pass is sufficient context.

Watch

Extended reading notes

Core claim

The paper's central claim is that long-range temporal learning and short-clip training need not conflict. A recurrent VSR model absorbs long-range information during a forward-only pass over the whole video, whose hidden states are stored; backpropagation then runs only on a randomly sampled short clip, with the stored states just outside the clip serving as accurate context. Because the clip boundaries carry the same propagation states the model would use at inference, the gradient signal reflects true long-range dependencies rather than a short-window approximation. The paper further claims that this context is best used selectively: replacing SoftMax with the sparse activation $\mathrm{ReLU}^2$ in intra- and inter-frame attention and gating the feed-forward network with the aligned previous hidden state improves the masked intra&inter-frame attention baseline by 0.24 dB on REDS4 with no added parameters, and the combined system reports 33.06 dB on REDS4 and 34.81 dB on ToS3.

Load-bearing premise

The method assumes that hidden states from a single forward pass over a long video, computed while the model's weights are still changing, are accurate enough to serve as fixed context for backpropagation on short clips; if those bootstrapped states are too noisy early in training, the gradients are misdirected and the reported efficiency gain would not transfer.

Editorial extensions

If this is right

  • Applied unchanged to BasicVSR, BasicVSR++, and PSRT, the training strategy lifts REDS4 PSNR by 0.14, 0.10, and 0.12 dB respectively, so any recurrent VSR model can take the long-context benefit without architectural changes.
  • Training with 8-frame clips plus long-video context matches training with 24-frame clips, at roughly 2.5 times less training time and 2.9 times less GPU memory.
  • The complete LRTI-VSR model reaches 33.06 dB on REDS4 and 34.81 dB on ToS3 with 12.9 million parameters and 1.54 T FLOPs, the latter about 40 percent lower than the IART baseline it surpasses.
  • Replacing SoftMax with $\mathrm{ReLU}^2$ attention and adding the refocused gated unit contribute separately (0.19 dB and 0.14 dB over the masked intra&inter-frame attention baseline), so the gains do not hinge on one monolithic design.

Reading between the lines

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

  • A natural next test is to apply the forward-long/backward-short protocol to other recurrent low-level vision tasks, such as video deblurring or video denoising, where the paper only provides a qualitative real-world VSR example.
  • The paper fixes the long-context forward pass until the next outer-loop iteration; periodically refreshing it (say every $K$ short-clip updates) would map the trade-off between context staleness and compute that the paper leaves unexplored.
  • Because $\mathrm{ReLU}^2$ is a sparsity prior on attention scores, the same replacement could be tested in single-image transformer super-resolution models, where suppressing low-scoring tokens may behave differently from distributing probability mass with SoftMax.
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 / 4 minor

Summary. The paper proposes LRTI-VSR, a training and architecture framework for recurrent video super-resolution. The main contributions are (i) a truncated-backpropagation-style training strategy that first runs a forward pass over a long video to obtain hidden states and then performs backpropagation on short clips conditioned on those hidden states, and (ii) a Refocused Intra&Inter-frame Transformer Block (RITB) that replaces softmax attention with a ReLU2 activation and injects the previous-frame aligned hidden state into the FFN through a Refocused Gated Unit (RGU). The method is evaluated on REDS4 and ToS3 for 4x VSR, reporting 33.06 dB PSNR on REDS4 and 34.81 dB on ToS3, and the training strategy is shown to improve BasicVSR, BasicVSR++, and PSRT. The paper also reports training time and memory comparisons, claiming a nearly 2.5x training acceleration relative to training on 24-frame clips.

Significance. If the efficiency claims are accurate, the proposed training strategy would be a broadly useful and generic technique for recurrent VSR, since it is validated on multiple external architectures (BasicVSR, BasicVSR++, PSRT) and the RITB block improves over the authors' own prior IIAB baseline. The paper releases code, uses standard benchmarks, and provides component ablations that are generally clean. However, the central efficiency claim is internally inconsistent with the algorithmic description, and the training strategy relies on an unanalyzed stale-hidden-state mechanism. These issues must be resolved before the contribution can be fully assessed.

major comments (3)
  1. [Section 4.2 (Table 2) and Algorithm 1] The reported GPU days are inconsistent with the per-iteration workload implied by Algorithm 1. With T=40, L=8, N=5, each iteration requires a full forward pass over 40 frames (plus a backward-direction pass if the bidirectional model is used) and then five short-clip forward/backward passes over 8 frames, i.e., at least 120 frame-steps per iteration, versus 16 for the L=8 baseline and 48 for the L=24 baseline. The reported times (3.75 vs 3.33 and 9.12 GPU days for 300K iterations with batch size 1) imply that the TB training is only 13% slower than L=8 and 2.4x faster than L=24, which is arithmetically impossible under the stated protocol unless the long forward pass is amortized, performed once per epoch, or excluded from the timing. The abstract and Section 4.2 explicitly advertise 'nearly 2.5x acceleration' and 'training and computational efficiency,' so this discrepancy directly affects a central claim; please clarify the timing protocol or correct the efficiency numbers.
  2. [Algorithm 1 (lines 3-10)] The long forward pass computes hidden states H with the current model parameters, after which N short-clip backward passes update the weights while H remains fixed. Therefore, after the first clip update, all subsequent clips in the same outer iteration are trained against stale hidden states from the pre-update model. The paper does not discuss or analyze this bootstrapping mismatch, nor does it provide a sensitivity study with respect to N. Because the entire efficiency argument rests on this strategy, please provide a justification (e.g., why staleness is benign for N up to 6 or 7) or an empirical analysis showing the effect of N on convergence and final accuracy.
  3. [Section 4.2 and Table 1] The ablations and hyperparameter choices are evaluated on REDS4, the same test set used for the final comparison in Table 3, and no error bars or multi-seed runs are reported. The SSIM gains in Table 1 are very large relative to the PSNR gains (e.g., BasicVSR: +0.14 dB PSNR but +0.0151 SSIM; BasicVSR++: +0.10 dB but +0.0107 SSIM), which is unusual and suggests either a systematic bias in the evaluation protocol or a typo. Please clarify the evaluation procedure and, ideally, validate the training strategy on a held-out set (e.g., ToS3) to guard against overfitting to REDS4.
minor comments (4)
  1. [Table 4] The rows after 'RITB' are not labeled with method names; the checkmarks are placed ambiguously, making it difficult to read which configuration corresponds to each row. Please reformat the table with explicit labels (e.g., 'RITB w/o RGU', 'RITB w/o ReLU2') and align the checkmarks correctly.
  2. [Throughout] There are several typos and grammatical issues: 'paopagation' in Figure 1's caption, 'dateset' in Section 4.1, 'to to build' in Section 4.2, 'an reduction' in Section 4.2, and 'Transformaer-based' in the supplementary material. Please proofread the manuscript.
  3. [Section 3.2, Eq. (3)] The equation defines \hat{I}^{HR}_{clip} as the output of the model, but the text describing it says 'the output of the forward propagation process during the VSR model's training.' Consider clarifying that the short-clip forward pass also uses the hidden states H_clip as inputs, as shown in the algorithm.
  4. [Section 4.1 and Supplementary C.2] The main text states that all models are trained from scratch for 600K iterations, but the supplementary material specifies a batch size of 24 for the final model. The main text should state the batch size as well, especially since Table 2 uses batch size 1 and the difference in batch size may affect the reported GPU times.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the training strategy and RITB block are validated empirically against external architectures and benchmarks, and no derivation reduces to its own inputs.

full rationale

The paper's central claims are (i) a generic truncated-backpropagation training strategy that improves BasicVSR, BasicVSR++, and PSRT, and (ii) a refocused intra&inter-frame transformer block (RITB) that improves over prior attention blocks. Neither claim is obtained by definitional construction. The training strategy is tested on existing external architectures (BasicVSR, BasicVSR++, PSRT) and produces PSNR gains of 0.10–0.14 dB that are not baked into the loss or the sampling procedure. The RITB ablation compares against MFSAB from PSRT and IIAB from the authors' own MIA-VSR paper, but this is an experimental baseline comparison, not a load-bearing self-citation: the improvements come from replacing SoftMax with ReLU2 and adding a gated FFN unit, and the final comparison in Table 3 is against multiple outside methods (RVRT, IART, PSRT, MIA-VSR, etc.). There is no fitted parameter that is later renamed as a prediction, no uniqueness theorem imported from the authors' prior work, and no equation that reduces to its own input by construction. The reported training-efficiency numbers in Table 2 may warrant scrutiny as an internal-consistency or correctness issue, but that is not a circularity concern. The derivation chain is self-contained and empirically grounded, so the circularity score is 0.

Assumptions & free parameters 5 free parameters · 4 assumptions · 2 invented entities

The central claim is empirical rather than derived, so there are no fitted physical constants. The listed hyperparameters were chosen on REDS4, the same test set as the final comparisons, which is the main selection-bias concern. The two novel architectural components are validated only within this paper.

free parameters (5)
  • truncation length L = 16 (final), 8 (ablation)
    Chosen to balance training efficiency and performance; the paper states the product N times L should approximate the full video length.
  • sampling times N = 6 (final), 5 (ablation)
    Chosen by the heuristic that N times L approximates the total sequence length; this controls how many short-clip updates reuse one long-sequence forward pass.
  • ReLU2 exponent = 2
    The squared-ReLU activation is selected as the sparsifying attention activation; no derivation from theory is provided.
  • number of RITB blocks per module = 18 (final), 6 (ablation)
    Architecture capacity chosen consistent with PSRT/IART; performance depends on this capacity.
  • attention window size, heads, channels = 8x8, 6, 120
    Follows the settings of prior transformer VSR models.
assumptions (4)
  • domain assumption TBPTT-style training with fixed hidden states reused across short-clip gradient updates is valid for recurrent video super-resolution
    The entire training strategy rests on this premise; the paper validates it empirically but provides no theoretical justification. It enters at Algorithm 1.
  • domain assumption REDS4 and ToS3 benchmarks are representative and comparable across methods
    The state-of-the-art comparison assumes that reported numbers from different papers, trained under different protocols, are directly comparable.
  • ad hoc to paper ReLU2 attention focuses on useful tokens better than softmax for this task
    This is a design hypothesis supported only by ablation evidence and the qualitative attention map in Figure 3.
  • standard math Standard backpropagation and gradient-based optimization (Adam) converge to a good solution
    Assumed as in all deep learning papers.
invented entities (2)
  • RITB block (refocused intra&inter-frame transformer block) independent evidence
    purpose: Selectively prioritize useful temporal information in attention and FFN for current frame restoration
    Its contribution is quantified in ablations versus MFSAB and IIAB, and it is reproducible as an architectural component.
  • RGU (refocused gated unit) independent evidence
    purpose: Injects aligned hidden state from the previous frame into the FFN via a gate
    Ablation shows a 0.14 dB gain over baseline; it is a testable component.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Small Clips, Big Gains: Learning Long-Range Refocused Temporal Information for Video Super-Resolution." pith.science (2026). https://pith.science/paper/BELGBNKB

@misc{pith2026250502159,
  author       = {Pith},
  title        = {Pith review of: Small Clips, Big Gains: Learning Long-Range Refocused Temporal Information for Video Super-Resolution},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BELGBNKB}},
  note         = {Machine review of arXiv:2505.02159}
}
read the original abstract

Video super-resolution (VSR) can achieve better performance compared to single image super-resolution by additionally leveraging temporal information. In particular, the recurrent-based VSR model exploits long-range temporal information during inference and achieves superior detail restoration. However, effectively learning these long-term dependencies within long videos remains a key challenge. To address this, we propose LRTI-VSR, a novel training framework for recurrent VSR that efficiently leverages Long-Range Refocused Temporal Information. Our framework includes a generic training strategy that utilizes temporal propagation features from long video clips while training on shorter video clips. Additionally, we introduce a refocused intra&inter-frame transformer block which allows the VSR model to selectively prioritize useful temporal information through its attention module while further improving inter-frame information utilization in the FFN module. We evaluate LRTI-VSR on both CNN and transformer-based VSR architectures, conducting extensive ablation studies to validate the contribution of each component. Experiments on long-video test sets demonstrate that LRTI-VSR achieves state-of-the-art performance while maintaining training and computational efficiency.

Figures

Figures reproduced from arXiv: 2505.02159 by the authors.

Figure 1
Figure 1. The pipline of our proposed training strategy. Our proposed training strategy can effectively utilize accurate long￾range temporal dependencies in long video sequences to assist training while using short video clips for high training efficiency. In this figure, −→hf means forward propagation hidden state and ←−hb means backward paopagation hidden state in the bidirec￾tional recurrent-based VSR model. of modern VSR … view at source ↗
Figure 2
Figure 2. Structure of feature propagation module. The com￾monly used feature propagation module contains bidirectional fea￾ture propagation structure (red solid lines), and second-order con￾nection structure (green solid lines) that leverages the computed hidden states of the previous two frames to recover the current frame. These two components are widely used in existing state-of￾the-art VSR models [6, 25, 35, 44, 46]. are… view at source ↗
Figure 3
Figure 3. Attention map comparison. We compare the visualiza￾tion of the intra&inter attention map under ReLU2 and the original SoftMax. The Top 50 values of all tokens (64×192) of an atten￾tion map under different network depths using these two activation functions are also counted. It is obvious that ReLU2 is significantly more focused on fewer tokens than SoftMax. projection matrices; m and n denote the m-th feature prop￾a… view at source ↗
Figures from the paper (7 more)
Figure 5
Figure 5. Figure 5: Effect of truncated back-propagation (TB) training strategy. The contribution of our proposed training strategy is more obvious in regions with fine details. The long-term propagation information from long video clips leads to marked improvements. are shown in the Supp…
Figure 6
Figure 6. Figure 6: Visual comparison for 4× VSR on REDS4 [29] dataset and ToS3 [10] dataset. length 24 when using video clips truncated to a length of 8, while attaining an acceleration of nearly 2.5× in training time and an reduction of nearly 2.9× in GPU memory con￾sumption. Additional…
Figure 7
Figure 7. Figure 7: Qualitative comparison on VideoLQ [7] dataset. Our proposed TB training strategy recovers the aircraft textures and reduces the curtain artifacts, which RealBasicVSR does not recover. room, bridge and face) and the length of each video clip is 150, 166 and 233 (534×128…
Figure 8
Figure 8. Figure 8: The overall architecture of LRTI-VSR. We develop a truncated backpropagation training strategy for the VSR model (A) learning long-term propagation patterns within long video sequences from short video clips and further proposed a refocused intra&inter Transformer bloc…
Figure 9
Figure 9. Figure 9: The structural comparison of MFSAB, IIAB and RITB. Our proposed a refocused intra&inter Transformer block (RITB) improves the previous MFSAB[35] block and IIAB[46] block in two aspects: (1) We replace SoftMax with the sparse refocus activation function ReLU2 to suppres…
Figure 10
Figure 10. Figure 10: Visual comparison for 4× VSR on REDS4 dataset. 4 [PITH_FULL_IMAGE:figures/full_fig_p014_10.png]
Figure 11
Figure 11. Figure 11: Visual comparison for 4× VSR on ToS3 dataset. 5 [PITH_FULL_IMAGE:figures/full_fig_p015_11.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

50 extracted references · 33 canonical work pages

  1. [1]

    Long- former: The long-document transformer

    Iz Beltagy, Matthew E Peters, and Arman Cohan. Long- former: The long-document transformer. arXiv preprint arXiv:2004.05150, 2020. 3

    arXiv 2004

  2. [2]

    Real- time video super-resolution with spatio-temporal networks and motion compensation

    Jose Caballero, Christian Ledig, Andrew Aitken, Alejandro Acosta, Johannes Totz, Zehan Wang, and Wenzhe Shi. Real- time video super-resolution with spatio-temporal networks and motion compensation. In Proceedings of the IEEE Con- ference on Computer Vision and Pattern Recognition (CVPR), pages 4778–4787, 2017. 2

    work page 2017

  3. [3]

    Video super-resolution transformer

    Jiezhang Cao, Yawei Li, Kai Zhang, Jingyun Liang, and Luc Van Gool. Video super-resolution transformer. arXiv preprint arXiv:2106.06847, 2021. 1, 2, 7, 8

    arXiv 2021

  4. [4]

    BasicVSR: The Search for Essential Components in Video Super-Resolution and Beyond

    Kelvin CK Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Basicvsr: The search for essential compo- nents in video super-resolution and beyond. arXiv preprint arXiv:2012.02181, 2020. 3, 7, 8, 1

    work page Pith review arXiv 2012

  5. [5]

    Basicvsr: The search for essential compo- nents in video super-resolution and beyond

    Kelvin CK Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Basicvsr: The search for essential compo- nents in video super-resolution and beyond. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4947–4956, 2021. 1, 2, 3, 6

    work page 2021

  6. [6]

    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. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 5972–5981, 2022. 1, 3, 4, 5, 6, 7, 8, 2

    work page 2022

  7. [7]

    Investigating tradeoffs in real-world video super-resolution

    Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Investigating tradeoffs in real-world video super-resolution. In Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition , pages 5962–5971, 2022. 2, 3, 6, 1

    work page 2022

  8. [8]

    Two deterministic half-quadratic regular- ization algorithms for computed imaging

    Pierre Charbonnier, Laure Blanc-Feraud, Gilles Aubert, and Michel Barlaud. Two deterministic half-quadratic regular- ization algorithms for computed imaging. In Proceedings of 1st international conference on image processing, pages 168–172. IEEE, 1994. 5

    work page 1994

Show all 50 references
  1. [9]

    Generating long sequences with sparse transformers

    Rewon Child, Scott Gray, Alec Radford, and Ilya Sutskever. Generating long sequences with sparse transformers. arXiv preprint arXiv:1904.10509, 2019. 3

    1904 arXiv

  2. [10]

    Learning temporal coherence via self- supervision for gan-based video generation

    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, 7, 1, 3

    2020

  3. [11]

    Deformable convolutional net- works

    Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional net- works. In Proceedings of the IEEE international conference on computer vision (ICCV), pages 764–773, 2017. 2

    2017

  4. [12]

    Flashattention: Fast and memory-efficient exact attention with io-awareness

    Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christo- pher R ´e. Flashattention: Fast and memory-efficient exact attention with io-awareness. Advances in Neural Information Processing Systems, 35:16344–16359, 2022. 3

    2022

  5. [13]

    Longnet: Scaling transformers to 1,000,000,000 tokens.arXiv preprint arXiv:2307.02486, 2023

    Jiayu Ding, Shuming Ma, Li Dong, Xingxing Zhang, Shao- han Huang, Wenhui Wang, Nanning Zheng, and Furu Wei. Longnet: Scaling transformers to 1,000,000,000 tokens.arXiv preprint arXiv:2307.02486, 2023. 3

    2023 arXiv

  6. [14]

    Efficient video super-resolution through recurrent latent space propagation

    Dario Fuoli, Shuhang Gu, and Radu Timofte. Efficient video super-resolution through recurrent latent space propagation. In 2019 IEEE/CVF International Conference on Computer Vision Workshop (ICCVW), pages 3476–3485. IEEE, 2019. 1, 2, 3

    2019

  7. [15]

    Efficiently mod- eling long sequences with structured state spaces

    Albert Gu, Karan Goel, and Christopher R´e. Efficiently mod- eling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021. 3

    2021 arXiv

  8. [16]

    Video super-resolution with recurrent structure-detail network

    Takashi Isobe, Xu Jia, Shuhang Gu, Songjiang Li, Shengjin Wang, and Qi Tian. Video super-resolution with recurrent structure-detail network. arXiv preprint arXiv:2008.00455,

    2008 arXiv

  9. [17]

    Video super-resolution with temporal group attention

    Takashi Isobe, Songjiang Li, Xu Jia, Shanxin Yuan, Gregory Slabaugh, Chunjing Xu, Ya-Li Li, Shengjin Wang, and Qi Tian. Video super-resolution with temporal group attention. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2

    2020

  10. [18]

    Look back and forth: Video super-resolution with explicit temporal difference modeling

    Takashi Isobe, Xu Jia, Xin Tao, Changlin Li, Ruihuang Li, Yongjie Shi, Jing Mu, Huchuan Lu, and Yu-Wing Tai. Look back and forth: Video super-resolution with explicit temporal difference modeling. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recog...

    2022

  11. [19]

    Deep video super-resolution network using dynamic upsampling filters without explicit motion compensation

    Younghyun Jo, Seoung Wug Oh, Jaeyeon Kang, and Seon Joo Kim. Deep video super-resolution network using dynamic upsampling filters without explicit motion compensation. In Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR), pages 3224–3232, 2018. 2

    2018

  12. [20]

    Transformers are rnns: Fast autoregressive transformers with linear attention

    Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, and Franc ¸ois Fleuret. Transformers are rnns: Fast autoregressive transformers with linear attention. InInternational conference on machine learning, pages 5156–5165. PMLR, 2020. 3

    2020

  13. [21]

    Efficient memory management for large lan- guage model serving with pagedattention

    Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph Gonzalez, Hao Zhang, and Ion Stoica. Efficient memory management for large lan- guage model serving with pagedattention. In Proceedings of the 29th Symposium on Operating Systems Principles,...

    2023

  14. [22]

    Sequence parallelism: Long sequence training from system perspective

    Shenggui Li, Fuzhao Xue, Chaitanya Baranwal, Yongbin Li, and Yang You. Sequence parallelism: Long sequence training from system perspective. arXiv preprint arXiv:2105.13120,

    arXiv

  15. [23]

    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. arXiv preprint arXiv:2007.11803 ,

    2007 arXiv

  16. [24]

    Vrt: A video restoration transformer

    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. 1, 2, 3, 7, 8

    2022 arXiv

  17. [25]

    Recurrent video restoration transformer with guided deformable attention

    Jingyun Liang, Yuchen Fan, Xiaoyu Xiang, Rakesh Ranjan, Eddy Ilg, Simon Green, Jiezhang Cao, Kai Zhang, Radu Tim- ofte, and Luc V Gool. Recurrent video restoration transformer with guided deformable attention. Advances in Neural Infor- mation Processing Systems, 35:378–393, 20...

    2022

  18. [26]

    Learning trajectory-aware transformer for video super- 9 resolution

    Chengxu Liu, Huan Yang, Jianlong Fu, and Xueming Qian. Learning trajectory-aware transformer for video super- 9 resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5687–5696,

  19. [27]

    Robust video super-resolution with learned temporal dynamics

    Ding Liu, Zhaowen Wang, Yuchen Fan, Xianming Liu, Zhangyang Wang, Shiyu Chang, and Thomas Huang. Robust video super-resolution with learned temporal dynamics. In Proceedings of the IEEE International Conference on Com- puter Vision (ICCV), pages 2507–2515, 2017. 2

    2017

  20. [28]

    Generating wikipedia by summarizing long sequences

    Peter J Liu, Mohammad Saleh, Etienne Pot, Ben Goodrich, Ryan Sepassi, Lukasz Kaiser, and Noam Shazeer. Generating wikipedia by summarizing long sequences. arXiv preprint arXiv:1801.10198, 2018. 3

    2018 arXiv

  21. [29]

    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. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Work...

    2019

  22. [30]

    Train- ing recurrent networks online without backtracking

    Yann Ollivier, Corentin Tallec, and Guillaume Charpiat. Train- ing recurrent networks online without backtracking. arXiv preprint arXiv:1507.07680, 2015. 3

    2015 arXiv

  23. [31]

    Pgt: A pro- gressive method for training models on long videos

    Bo Pang, Gao Peng, Yizhuo Li, and Cewu Lu. Pgt: A pro- gressive method for training models on long videos. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11379–11389, 2021. 3

    2021

  24. [32]

    Learning spatiotemporal frequency-transformer for com- pressed video super-resolution

    Zhongwei Qiu, Huan Yang, Jianlong Fu, and Dongmei Fu. Learning spatiotemporal frequency-transformer for com- pressed video super-resolution. In European Conference on Computer Vision, pages 257–273. Springer, 2022. 1

    2022

  25. [33]

    Frame-recurrent video super-resolution

    Mehdi SM Sajjadi, Raviteja Vemulapalli, and Matthew Brown. Frame-recurrent video super-resolution. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recog- nition (CVPR), pages 6626–6634, 2018. 2

    2018

  26. [34]

    Fast transformer decoding: One write-head is all you need

    Noam Shazeer. Fast transformer decoding: One write-head is all you need. arXiv preprint arXiv:1911.02150, 2019. 3

    1911 arXiv

  27. [35]

    Rethinking alignment in video super- resolution transformers

    Shuwei Shi, Jinjin Gu, Liangbin Xie, Xintao Wang, Yujiu Yang, and Chao Dong. Rethinking alignment in video super- resolution transformers. arXiv preprint arXiv:2207.08494,

    arXiv

  28. [36]

    Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network

    Wenzhe Shi, Jose Caballero, Ferenc Husz´ar, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In Proceedings of the IEEE conference on comput...

    2016

  29. [37]

    Unbiased online recurrent optimization

    Corentin Tallec and Yann Ollivier. Unbiased online recurrent optimization. arXiv preprint arXiv:1702.05043, 2017. 3

    2017 arXiv

  30. [38]

    Unbiasing truncated back- propagation through time

    Corentin Tallec and Yann Ollivier. Unbiasing truncated back- propagation through time. arXiv preprint arXiv:1705.08209,

    arXiv

  31. [39]

    Detail-revealing deep video super-resolution

    Xin Tao, Hongyun Gao, Renjie Liao, Jue Wang, and Jiaya Jia. Detail-revealing deep video super-resolution. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 4472–4480, 2017. 2

    2017

  32. [40]

    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. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages 3360–3369, 2020. 1, 2

    2020

  33. [41]

    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. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recogni- tion Workshops (CVPRW), pages 0–0, 2019. 1, 2, 7, 8

    2019

  34. [42]

    Backpropagation through time: what it does and how to do it.Proceedings of the IEEE, 78(10):1550–1560,

    Paul J Werbos. Backpropagation through time: what it does and how to do it.Proceedings of the IEEE, 78(10):1550–1560,

  35. [43]

    Gradient-based learning algorithms for recurrent networks and their computational complexity

    Ronald J Williams and David Zipser. Gradient-based learning algorithms for recurrent networks and their computational complexity. In Backpropagation, pages 433–486. Psychology Press, 2013. 3, 4

    2013

  36. [44]

    An implicit alignment for video super-resolution

    Kai Xu, Ziwei Yu, Xin Wang, Michael Bi Mi, and Angela Yao. An implicit alignment for video super-resolution. arXiv preprint arXiv:2305.00163, 2023. 1, 3, 4, 5, 7, 8

    2023 arXiv

  37. [45]

    Video enhancement with task-oriented flow

    Tianfan Xue, Baian Chen, Jiajun Wu, Donglai Wei, and William T Freeman. Video enhancement with task-oriented flow. International Journal of Computer Vision, 127(8):1106– 1125, 2019. 7, 8

    2019

  38. [46]

    Video super-resolution trans- former with masked inter&intra-frame attention

    Xingyu Zhou, Leheng Zhang, Xiaorui Zhao, Keze Wang, Leida Li, and Shuhang Gu. Video super-resolution trans- former with masked inter&intra-frame attention. In Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 25399–25408, 2024. 1...

    2024

  39. [48]

    and the attention structure of MIA-VSR [46]. Besides the commonly used shallow feature extraction, recurrent feature refinement and feature reconstruction parts, we utilize the proposed truncated backpropagation training method and refocused intra&inter Transformer block (RITB...

  40. [49]

    It is obvious in the Fig

    model. It is obvious in the Fig. 7 that our training method can help produce more realistic and fine-grained results. C. Dataset and Implementation Details C.1. Datasets REDS [29] REDS is a widely-used video dataset for evalu- ating video restoration tasks. It has 270 clips wi...

  41. [50]

    We calculate PSNR and SSIM on the RGB channel for these datatsets

    In the testing phase, we evaluate LRTI-VSR model’s per- formance on the REDS4 [29] dataset and ToS3 [10] dataset. We calculate PSNR and SSIM on the RGB channel for these datatsets. For the calculation of the FLOPs in Table 4, we compare all VSR models with an input low-resolut...

  42. [2022]

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

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.