Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

DualX-VSR: Dual Axial Spatial$\times$Temporal Transformer for Real-World Video Super-Resolution without Motion Compensation

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read DualX-VSR claims that attending along height-time and width-time planes lets a video super-resolution transformer skip motion compensation and still beat flow-based baselines on synthetic benchmarks.

desk verdict Solid flow-free attention idea, but the paper's real-world superiority claim is contradicted by its own VideoLQ table. read the letter →

arxiv 2506.04830 v2 pith:BKLKSWST submitted 2025-06-05 cs.CV

Shuo Cao , Yihao Liu , Xiaohui Li , Yuanting Gao , Yu Zhou , Chao Dong This is my paper
classification cs.CV MSC 68T4568T07
keywords videosuper-resolutionspatiotemporalattentionaxialtransformermotioncompensationreal-worlddegradationREDSdatasetrestoration
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

The paper tries to establish that video super-resolution can be done without optical-flow motion compensation if a transformer attends along the two space-time planes, height-time ($H\times T$) and width-time ($W\times T$), instead of treating space and time sequentially. The authors build DualX-VSR, a ViViT-style network with vertical-temporal and horizontal-temporal attention blocks, and report that this dual axial spatial$\times$temporal attention outperforms sequential spatial-temporal attention on REDS4 while running at the same speed, and that the full model beats established flow-based and window-based VSR baselines on synthetic benchmarks. A sympathetic reader would care because the result suggests that global spatiotemporal context, rather than explicit alignment, is what matters most for video restoration, and because a flow-free transformer avoids the cumulative errors of flow estimation in degraded video. The paper also shows that horizontal-temporal attention contributes more than vertical-temporal attention, consistent with measured horizontal motion dominance in common test videos.

What carries the argument

The load-bearing mechanism is Dual Axial Spatial$\times$Temporal Attention, defined as two attention blocks operating on orthogonal space-time planes: the vertical-temporal attention block (VTAB) rearranges tokens into the $H\times T$ plane and the horizontal-temporal attention block (HTAB) into the $W\times T$ plane. Each token in these planes contains both a spatial axis and time, so a single attention pass fuses texture and motion instead of stacking separate spatial and temporal modules. The blocks are connected serially (VTAB then HTAB), rotary positional encoding (RoPE) is applied to token positions, and a spatiotemporal reconstruction module adds temporal attention before per-frame 2D convolution and pixel shuffle. This construction keeps a global receptive field along the two space-time planes at lower computational cost than full 3D attention, and it is what lets the network drop optical-flow alignment and propagation entirely.

What would settle it

A controlled human-preference study or a paired real-world low-resolution/high-resolution benchmark comparing DualX-VSR against RealBasicVSR and RealViformer on genuinely degraded videos would settle the claim; if the flow-based baselines win there, the synthetic superiority does not transfer.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that space and time are better modeled jointly along orthogonal axes than sequentially. In controlled ablations on REDS4 with identical model size, only-spatial attention reaches 27.96 dB PSNR, only-temporal 28.17 dB, sequential spatial-temporal 28.22 dB, vertical-temporal 28.22 dB, horizontal-temporal 28.32 dB, and the full dual axial combination 28.38 dB with 0.8099 SSIM. The paper attributes the extra gains to each token carrying both spatial and temporal information and to horizontal-temporal attention capturing the dominant horizontal motion in the benchmarks. The resulting DualX-VSR (MSE) reports 26.03 dB PSNR on REDS4, 30.13 dB on UDM10, and 25.81 dB on SPMC30, ahead of RealBasicVSR and RealViformer, and it does so without alignment or propagation modules. The authors interpret this as evidence that a simplified, motion-compensation-free transformer can provide a cohesive spatiotemporal representation for real-world VSR.

Load-bearing premise

The load-bearing premise is that the first-order synthetic degradation pipeline closely matches real-world video degradation, so that PSNR/SSIM wins on synthetic benchmarks transfer to real-world video super-resolution.

Editorial extensions

If this is right

  • A VSR network can dispense with optical-flow alignment and propagation modules and still reach higher PSNR/SSIM than RealBasicVSR, RealViformer, and Upscale-A-Video on REDS4, UDM10, and SPMC30 in the paper's experiments.
  • Because horizontal-temporal attention alone outperforms vertical-temporal attention and the dual combination outperforms both, the design exploits the dataset bias toward horizontal motion, which the paper measures with FlowFormer motion amplitudes.
  • The method's runtime in the attention ablation (45 ms per frame) is lower than full 3D omnidirectional attention while retaining a global receptive field, so efficiency is part of the claimed advantage.
  • The pretrain-on-bicubic then fine-tune-on-degraded-data schedule is load-bearing: without pre-training, REDS4 PSNR drops from 26.03 to 25.71 dB in the paper's Table 9, and larger patch sizes and more input frames add further gains.

Reading between the lines

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

  • Inference: because the architecture itself is not tied to the degradation pipeline, the same dual axial space-time attention could be trained on paired real-world data or with a learned degradation model once such data is available.
  • Inference: the motion-amplitude result predicts a measurable asymmetry: on video with dominant vertical motion, the vertical-temporal block should contribute more than the horizontal-temporal block, a direction the paper does not test.
  • Inference: the mechanism likely transfers to other video restoration tasks where flow alignment is fragile, such as video deblurring, denoising, and compression artifact removal.
  • Inference: since the authors cite limited compute as preventing full scaling, the claim that the model benefits from larger scale is testable by scaling depth, width, and sequence length.
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 DualX-VSR, a transformer-based video super-resolution model that replaces the commonly used sequential spatial-temporal attention with a dual axial spatial×temporal attention, applying vertical-temporal and horizontal-temporal attention in sequence. The model dispenses with optical-flow-based motion compensation entirely, using a ViViT-style architecture with input video embedding, a dual axial transformer module, and a spatiotemporal reconstruction module, trained with a pretrain-finetune strategy against first-order synthetic degradations. The authors evaluate on synthetic benchmarks REDS4, UDM10, and SPMC30 using PSNR/SSIM/LPIPS, and on the real-world VideoLQ dataset using the non-reference metrics LIQE and NIQE, also reporting runtime, parameter counts, and temporal consistency. The central claims are that the dual axial mechanism outperforms sequential spatial-temporal attention and that the resulting model achieves high fidelity and superior performance in real-world VSR without motion compensation.

Significance. If the claims are supported, the paper makes a useful architectural contribution: it shows that a flow-free transformer with factorized spatial-temporal attention along height-time and width-time planes can compete with flow-based recurrent methods on synthetic VSR benchmarks, while substantially reducing MACs compared to scaled-up baselines (Table 4). The controlled attention comparison in Table 1 provides direct evidence that the dual axial design improves over sequential spatial-temporal attention under the same parameter budget, and the appendix promises code and pretrained models, which would facilitate reproduction. The main significance is therefore in the architecture and its efficiency. However, the headline claim of superior performance in real-world VSR is not supported by the paper's own VideoLQ results, and the synthetic evaluation is conducted under the same degradation pipeline used for training, so the real-world generalizability of the reported gains remains unestablished.

major comments (3)
  1. [Abstract; Sec. 5.2; Table 2] The claim of 'superior performance in real-world VSR task' is contradicted by the paper's own real-world numbers. On VideoLQ, DualX-VSR (MSE) achieves LIQE 1.8052 and NIQE 6.7903, and DualX-VSR (GAN) achieves LIQE 2.5515 and NIQE 4.0272, whereas RealBasicVSR achieves LIQE 3.0239 and NIQE 3.6921 and RealViformer achieves LIQE 2.4022 and NIQE 4.0581. On both metrics at least one baseline is better than the proposed MSE variant, and on LIQE both proposed variants are worse than RealBasicVSR. The abstract and Section 5.2 should either be revised to claim only competitive real-world performance, or the authors should add stronger real-world evidence such as a reference-based real-world benchmark or a human study.
  2. [Sec. 5.1; Appendix C.2; Fig. 7; Table 2] The synthetic evaluation does not provide independent evidence for real-world superiority because the model is trained and tested under the same first-order degradation pipeline, and the choice of first-order degradation over second-order is justified only by a visual comparison in Fig. 7. Consequently, the PSNR/SSIM/LPIPS gains on REDS4, UDM10, and SPMC30 in Table 2 may reflect the model's fit to the training degradation rather than general real-world robustness. The authors should provide quantitative validation of the degradation proxy, cross-degradation evaluation, or real-world reference-based results, and should report variance or error bars given that the degradation pipeline is stochastic.
  3. [Sec. 5.3; Table 5; Table 9; Table 2] The design choices of serial connection, 2D pre-feature convolution, training patch size, and frame count are selected by evaluating on the REDS4 test set (Tables 5 and 9), and REDS4 is then reported as one of the headline synthetic benchmarks in Table 2. This creates a mild test-set-driven selection bias for the REDS4 row. The authors should either validate the chosen hyperparameters on a held-out dataset or clearly disclose this selection process in the main text; ideally, the headline comparison should also include a validation split that was not used for architecture decisions.
minor comments (4)
  1. [Sec. 5.1; Table 2 caption] There are several typos in the manuscript, including 'appoarches' in the Table 2 caption, 'empoly' in Section 5.1, and 'Traning' in Table 9; these should be corrected.
  2. [Sec. 4.2; Eq. (1)-(2)] The notation in Eq. (1) is unconventional: the subscript 'u' is used for both the key index and the position in f_q(q_u, u), and the definitions of q_u and k_v are not precisely specified. Please clarify the indexing of queries and keys under RoPE.
  3. [Sec. 5.2; Table 3] Table 3 reports that DualX-VSR has a runtime of 0.40 s/frame while RealBasicVSR has 0.07 s/frame, yet the text says DualX-VSR 'still perform efficiency'; the meaning of 'efficiency' is unclear given the slower runtime, and the text should reconcile these numbers or report conditions more explicitly.
  4. [Appendix B.2] The discussion of Stripformer is informative, but the claim that DualX-VSR is 'the first to introduce 3D dual axial spatial×temporal attention in the real-world video super-resolution task' should be phrased more cautiously, since axial attention variants already exist in other video and image restoration contexts; a citation to the earliest axial attention works would help.

Circularity Check

1 steps flagged · score 4.0 of 10

Headline REDS4 gains are partly self-confirming because the same test set was used to select architecture hyperparameters; the core dual-axial attention comparison is still an independent controlled ablation.

  1. fitted input called prediction [Sec. 3.2 and Sec. 5.3 (Tables 1, 5, 9) vs. Sec. 5.2 (Table 2)]
    "We conducted experiments by training models on REDS dataset [31] and testing them on REDS4 dataset [31], varying only the type of attention mechanism. ... However, as shown in Tab. 5 (a) and (d), our experiments indicate that serial connection outperforms interleaved connection, whereas DualX-VSR utilizes serial connection."

    The attention type (Table 1), the block connection and pre-feature extraction (Table 5), and the training patch size, frame count, and pretraining strategy (Table 9) are all selected by comparing PSNR/SSIM on the REDS4 test set. That same REDS4 set is then reported as the headline synthetic benchmark in Table 2. The REDS4 advantage of the final DualX-VSR configuration over baselines is therefore not an out-of-sample prediction; it is partly an artifact of having chosen the configuration using those exact test numbers. The controlled within-model attention ablation remains informative, but the headline 'superior performance' claim on synthetic data is statistically forced by the selection protocol.

full rationale

The paper's central architectural claim—that dual axial spatial×temporal attention improves over sequential spatial-temporal attention and over window-based/flow-based alternatives—is supported by a controlled comparison in Table 1, where only the attention mechanism is varied. That core comparison is not circular: the vertical-temporal and horizontal-temporal attention variants are evaluated under the same training setup as the sequential spatial-temporal baseline, and the result is an empirical, falsifiable improvement. The main circularity concern is evaluation-protocol level rather than derivation-level: the final DualX-VSR configuration is selected using REDS4 test-set scores (Tables 5 and 9), and the same REDS4 scores are then presented as evidence of superiority in Table 2. This makes the headline synthetic benchmark partly self-confirming. The real-world claim is additionally weakened by the paper's own VideoLQ numbers, where both DualX-VSR variants trail RealBasicVSR on LIQE and NIQE, and by the choice of a first-order degradation pipeline validated only visually against VideoLQ; however, these are correctness/evidence weaknesses rather than circular derivations. No load-bearing self-citation chain, uniqueness theorem, or ansatz-smuggling pattern is present, and the core attention mechanism retains independent content, so a moderate score of 4 is appropriate.

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

The central claim depends on standard transformer machinery plus a set of training and evaluation choices (degradation model, hyperparameters, comparison protocol) rather than new physical entities. None of these choices are derived from external first principles; they are validated empirically, and some are tuned on the same test sets used for headline results.

free parameters (4)
  • First-order degradation constraint = first-order
    Chosen because it closely resembles real-world VideoLQ (Appendix C.2). This choice determines all synthetic training and test data for the reported comparisons.
  • Train patch size and frame count = 112x112, 16 frames
    Selected via ablations on REDS4 (Table 9); increasing patch size and frames improves PSNR and is tuned on the test set.
  • Loss weights for perceptual and adversarial losses = lambda_pix=1e-2, lambda_per=1, lambda_adv=5e-3
    Set following BasicVSR (Appendix C.1); these weights affect the GAN-stage metrics.
  • Attention connection and pre-feature convolution type = serial VTAB-HTAB; 2D conv
    Selected on REDS4 (Table 5), giving about 0.1 dB over alternatives.
assumptions (4)
  • domain assumption First-order degradation is a sufficient proxy for real-world low-quality videos.
    Central to all synthetic evaluations; validated only by visual comparison with VideoLQ (Appendix C.2, Fig. 7).
  • domain assumption Horizontal motion amplitude exceeds vertical motion amplitude in natural video datasets.
    Motivates horizontal-temporal attention; measured on four datasets with FlowFormer (Sec. 3.2, Fig. 3).
  • domain assumption A ViViT-style tokenization can preserve enough detail for pixel-level VSR if attention is redesigned.
    The plain ViViT-VSR baseline is poor; the whole approach assumes a redesigned attention fixes this (Sec. 1).
  • domain assumption All compared baselines were trained with the same degradation pipeline as DualX-VSR.
    Stated in Appendix C.2 but no training protocol or evidence is provided; if false, the comparison is unfair.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DualX-VSR: Dual Axial Spatial$\times$Temporal Transformer for Real-World Video Super-Resolution without Motion Compensation." pith.science (2026). https://pith.science/paper/BKLKSWST

@misc{pith2026250604830,
  author       = {Pith},
  title        = {Pith review of: DualX-VSR: Dual Axial Spatial$\times$Temporal Transformer for Real-World Video Super-Resolution without Motion Compensation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BKLKSWST}},
  note         = {Machine review of arXiv:2506.04830}
}
abstract

Transformer-based models like ViViT and TimeSformer have advanced video understanding by effectively modeling spatiotemporal dependencies. Recent video generation models, such as Sora and Vidu, further highlight the power of transformers in long-range feature extraction and holistic spatiotemporal modeling. However, directly applying these models to real-world video super-resolution (VSR) is challenging, as VSR demands pixel-level precision, which can be compromised by tokenization and sequential attention mechanisms. While recent transformer-based VSR models attempt to address these issues using smaller patches and local attention, they still face limitations such as restricted receptive fields and dependence on optical flow-based alignment, which can introduce inaccuracies in real-world settings. To overcome these issues, we propose Dual Axial Spatial$\times$Temporal Transformer for Real-World Video Super-Resolution (DualX-VSR), which introduces a novel dual axial spatial$\times$temporal attention mechanism that integrates spatial and temporal information along orthogonal directions. DualX-VSR eliminates the need for motion compensation, offering a simplified structure that provides a cohesive representation of spatiotemporal information. As a result, DualX-VSR achieves high fidelity and superior performance in real-world VSR task.

Figures

Figures reproduced from arXiv: 2506.04830 by the authors.

Figure 1
Figure 1. Comparison of different VSR models. ViViT-VSR fol [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Diagram of different attention. Time (T), Width (W), [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Average motion amplitude across different video [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Overview of the proposed DualX-VSR. The top section shows the pipeline, which converts low-quality video to high-quality [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Qualitative comparisons on both synthetic datasets and real-world datasets. The top two comparisons are from the REDS [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Different structures of attention blocks. (a) illustrates the [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Comparison of different degradations. (a) and (b) are [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]
Figure 8
Figure 8. Figure 8: More qualitative comparisons on both synthetic datasets and real-world datasets. The top two comparisons are from the UDM10 [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]
Figure 9
Figure 9. Figure 9: Qualitative comparison with diffusion-based real-world [PITH_FULL_IMAGE:figures/full_fig_p015_9.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. ArtiMuse: Fine-Grained Image Aesthetics Assessment with Joint Scoring and Expert-Level Understanding

    cs.CV 2025-07 conditional novelty 6.0 of 10

    ArtiMuse is an MLLM that jointly scores image aesthetics and writes expert-style 8-attribute critiques, trained on a new 10,000-image expert-annotated dataset with a token-based continuous scoring method.

Reference graph

Works this paper leans on

51 extracted references · 45 canonical work pages · cited by 1 Pith paper

  1. [1]

    Vivit: A video vi- sion transformer, 2021

    Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lu ˇci´c, and Cordelia Schmid. Vivit: A video vi- sion transformer, 2021. 1, 2, 3, 4, 5, 7, 12

    work page 2021

  2. [2]

    Vidu: a highly consistent, dynamic and skilled text-to-video generator with diffusion models, 2024

    Fan Bao, Chendong Xiang, Gang Yue, Guande He, Hongzhou Zhu, Kaiwen Zheng, Min Zhao, Shilong Liu, Yaole Wang, and Jun Zhu. Vidu: a highly consistent, dynamic and skilled text-to-video generator with diffusion models, 2024. 4

    work page 2024

  3. [3]

    Is space-time attention all you need for video understanding?,

    Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding?,

  4. [4]

    Align your latents: High-resolution video synthesis with la- tent diffusion models, 2023

    Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dock- horn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with la- tent diffusion models, 2023. 3

    2023

  5. [5]

    Video generation models as world simulators

    Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luh- man, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators

  6. [6]

    Video super-resolution transformer, 2023

    Jiezhang Cao, Yawei Li, Kai Zhang, and Luc Van Gool. Video super-resolution transformer, 2023. 2

    work page 2023

  7. [7]

    Kelvin C. K. Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Basicvsr: The search for essential com- ponents in video super-resolution and beyond, 2021. 2, 13

    work page 2021

  8. [8]

    Kelvin C. K. Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Basicvsr++: Improving video super- resolution with enhanced propagation and alignment, 2021. 2

    work page 2021

Show all 51 references
  1. [9]

    Kelvin C. K. Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Investigating tradeoffs in real-world video super-resolution, 2021. 2, 6, 7, 8, 9, 13, 14, 15

    2021

  2. [10]

    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. InProceedings of 1st international conference on image processing, pages 168–172. IEEE, 1994. 12

    1994

  3. [11]

    Pre-trained image processing transformer, 2021

    Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer, 2021. 2

    2021

  4. [12]

    Hat: Hybrid attention transformer for image restoration, 2023

    Xiangyu Chen, Xintao Wang, Wenlong Zhang, Xiangtao Kong, Yu Qiao, Jiantao Zhou, and Chao Dong. Hat: Hybrid attention transformer for image restoration, 2023. 6

    2023

  5. [13]

    An image is worth 16x16 words: Transformers for image recognition at scale, 2021

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition ...

    2021

  6. [14]

    Venhancer: Generative space-time enhancement for video generation, 2024

    Jingwen He, Tianfan Xue, Dongyang Liu, Xinqi Lin, Peng Gao, Dahua Lin, Yu Qiao, Wanli Ouyang, and Ziwei Liu. Venhancer: Generative space-time enhancement for video generation, 2024. 7, 14

    2024

  7. [15]

    Flowformer: A transformer architecture for optical flow,

    Zhaoyang Huang, Xiaoyu Shi, Chao Zhang, Qiang Wang, Ka Chun Cheung, Hongwei Qin, Jifeng Dai, and Hongsheng Li. Flowformer: A transformer architecture for optical flow,

  8. [16]

    VBench: Com- prehensive benchmark suite for video generative models

    Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, Yaohui Wang, Xinyuan Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. VBench: Com- prehensive benchmark suite for video generative models....

    2024

  9. [17]

    Vbench++: Comprehensive and versatile bench- mark suite for video generative models.arXiv preprint arXiv:2411.13503, 2024

    Ziqi Huang, Fan Zhang, Xiaojie Xu, Yinan He, Jiashuo Yu, Ziyue Dong, Qianli Ma, Nattapol Chanpaisit, Chenyang Si, Yuming Jiang, Yaohui Wang, Xinyuan Chen, Ying- Cong Chen, Limin Wang, Dahua Lin, Yu Qiao, and Zi- wei Liu. Vbench++: Comprehensive and versatile bench- mark suite ...

    2024 arXiv

  10. [18]

    Percep- tual losses for real-time style transfer and super-resolution

    Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Percep- tual losses for real-time style transfer and super-resolution. InComputer Vision–ECCV 2016: 14th European Confer- ence, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 694–711. Springer, 2016. 12

    2016

  11. [19]

    Unsupervised real-world super resolution with cycle genera- tive adversarial network and domain discriminator

    Gwantae Kim, Jaihyun Park, Kanghyu Lee, Junyeop Lee, Jeongki Min, Bokyeung Lee, David K Han, and Hanseok Ko. Unsupervised real-world super resolution with cycle genera- tive adversarial network and domain discriminator. InPro- ceedings of the IEEE/CVF Conference on Computer Vi...

    2020

  12. [20]

    Spatio-temporal transformer network for video restoration

    Tae Hyun Kim, Mehdi SM Sajjadi, Michael Hirsch, and Bernhard Scholkopf. Spatio-temporal transformer network for video restoration. InProceedings of the European con- ference on computer vision (ECCV), pages 106–122, 2018. 2

    2018

  13. [21]

    Adam: A method for stochastic opti- mization.arXiv preprint arXiv:1412.6980, 2014

    Diederik P Kingma. Adam: A method for stochastic opti- mization.arXiv preprint arXiv:1412.6980, 2014. 6

    2014 arXiv

  14. [22]

    Towards ef- fective multiple-in-one image restoration: A sequential and prompt learning strategy, 2024

    Xiangtao Kong, Chao Dong, and Lei Zhang. Towards ef- fective multiple-in-one image restoration: A sequential and prompt learning strategy, 2024. 6

    2024

  15. [23]

    Learning blind video temporal consistency

    Wei-Sheng Lai, Jia-Bin Huang, Oliver Wang, Eli Shechtman, Ersin Yumer, and Ming-Hsuan Yang. Learning blind video temporal consistency. InEuropean Conference on Computer Vision, 2018. 7, 8, 14

    2018

  16. [24]

    Vrt: A video restoration transformer, 2022

    Jingyun Liang, Jiezhang Cao, Yuchen Fan, Kai Zhang, Rakesh Ranjan, Yawei Li, Radu Timofte, and Luc Van Gool. Vrt: A video restoration transformer, 2022. 2, 3, 5, 13

    2022

  17. [25]

    Recurrent video restoration transformer with guided deformable attention, 2022

    Jingyun Liang, Yuchen Fan, Xiaoyu Xiang, Rakesh Ranjan, Eddy Ilg, Simon Green, Jiezhang Cao, Kai Zhang, Radu Timofte, and Luc Van Gool. Recurrent video restoration transformer with guided deformable attention, 2022. 2, 3, 12, 13

    2022

  18. [26]

    Diff- bir: Towards blind image restoration with generative diffu- sion prior, 2024

    Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Bo Dai, Fanghua Yu, Wanli Ouyang, Yu Qiao, and Chao Dong. Diff- bir: Towards blind image restoration with generative diffu- sion prior, 2024. 7

    2024

  19. [27]

    Swin trans- former: Hierarchical vision transformer using shifted win- dows, 2021

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin trans- former: Hierarchical vision transformer using shifted win- dows, 2021. 2, 3

    2021

  20. [28]

    Vdt: General-purpose video diffusion transformers via mask modeling, 2023

    Haoyu Lu, Guoxing Yang, Nanyi Fei, Yuqi Huo, Zhiwu Lu, Ping Luo, and Mingyu Ding. Vdt: General-purpose video diffusion transformers via mask modeling, 2023. 1, 2, 3

    2023

  21. [29]

    Latte: La- tent diffusion transformer for video generation, 2024

    Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Ziwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: La- tent diffusion transformer for video generation, 2024. 1, 3, 4, 7

    2024

  22. [30]

    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. 7, 15

    2012

  23. [31]

    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. InCVPR Workshops, 2019. 4, 6, 7, 13, 14

    2019

  24. [32]

    Deep blind video super-resolution, 2020

    Jinshan Pan, Songsheng Cheng, Jiawei Zhang, and Jinhui Tang. Deep blind video super-resolution, 2020. 7

    2020

  25. [33]

    Scalable diffusion models with transformers, 2023

    William Peebles and Saining Xie. Scalable diffusion models with transformers, 2023. 1

    2023

  26. [34]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684–10695, 2022. 7

    2022

  27. [35]

    Rethinking alignment in video super- resolution transformers, 2022

    Shuwei Shi, Jinjin Gu, Liangbin Xie, Xintao Wang, Yujiu Yang, and Chao Dong. Rethinking alignment in video super- resolution transformers, 2022. 2, 3, 5, 12, 13

    2022

  28. [36]

    Roformer: Enhanced transformer with rotary position embedding, 2023

    Jianlin Su, Yu Lu, Shengfeng Pan, Ahmed Murtadha, Bo Wen, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding, 2023. 5, 12

    2023

  29. [37]

    Detail-revealing deep video super-resolution

    Xin Tao, Hongyun Gao, Renjie Liao, Jue Wang, and Jiaya Jia. Detail-revealing deep video super-resolution. InThe IEEE International Conference on Computer Vision (ICCV),

  30. [38]

    Detail-revealing deep video super-resolution, 2017

    Xin Tao, Hongyun Gao, Renjie Liao, Jue Wang, and Jiaya Jia. Detail-revealing deep video super-resolution, 2017. 6, 7, 14, 15

    2017

  31. [39]

    Stripformer: Strip transformer for fast image deblurring, 2022

    Fu-Jen Tsai, Yan-Tsung Peng, Yen-Yu Lin, Chung-Chi Tsai, and Chia-Wen Lin. Stripformer: Strip transformer for fast image deblurring, 2022. 12

    2022

  32. [40]

    Jeya Maria Jose Valanarasu, Poojan Oza, Ilker Hacihaliloglu, and Vishal M. Patel. Medical transformer: Gated axial- attention for medical image segmentation, 2021. 3

    2021

  33. [41]

    Xintao Wang, Kelvin C. K. Chan, Ke Yu, Chao Dong, and Chen Change Loy. Edvr: Video restoration with enhanced deformable convolutional networks, 2019. 2

    2019

  34. [42]

    Real-esrgan: Training real-world blind super-resolution with pure synthetic data, 2021

    Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data, 2021. 6, 7, 13

    2021

  35. [43]

    Lavie: High-quality video gen- eration with cascaded latent diffusion models, 2023

    Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, Yuwei Guo, Tianxing Wu, Chenyang Si, Yum- ing Jiang, Cunjian Chen, Chen Change Loy, Bo Dai, Dahua Lin, Yu Qiao, and Ziwei Liu. Lavie: High-quality video gen...

    2023

  36. [44]

    Enhancing video super-resolution via implicit resampling-based alignment, 2024

    Kai Xu, Ziwei Yu, Xin Wang, Michael Bi Mi, and An- gela Yao. Enhancing video super-resolution via implicit resampling-based alignment, 2024. 2, 3, 5, 12, 13

    2024

  37. [45]

    Videogigagan: Towards detail-rich video super-resolution,

    Yiran Xu, Taesung Park, Richard Zhang, Yang Zhou, Eli Shechtman, Feng Liu, Jia-Bin Huang, and Difan Liu. Videogigagan: Towards detail-rich video super-resolution,

  38. [46]

    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. 2

    2019

  39. [47]

    Real- world video super-resolution: A benchmark dataset and a decomposition based learning scheme

    Xi Yang, Wangmeng Xiang, Hui Zeng, and Lei Zhang. Real- world video super-resolution: A benchmark dataset and a decomposition based learning scheme. In2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 4761–4770, 2021. 7, 8, 14

    2021

  40. [48]

    The unreasonable effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shecht- man, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. InProceedings of the IEEE conference on computer vision and pattern recogni- tion, pages 586–595, 2018. 7, 15

    2018

  41. [49]

    Blind image quality assessment via vision- language correspondence: A multitask learning perspective,

    Weixia Zhang, Guangtao Zhai, Ying Wei, Xiaokang Yang, and Kede Ma. Blind image quality assessment via vision- language correspondence: A multitask learning perspective,

  42. [50]

    Realviformer: Investigating attention for real-world video super-resolution, 2024

    Yuehan Zhang and Angela Yao. Realviformer: Investigating attention for real-world video super-resolution, 2024. 2, 5, 7, 8, 9, 13, 14

    2024

  43. [51]

    Upscale-a-video: Temporal- consistent diffusion model for real-world video super- resolution, 2023

    Shangchen Zhou, Peiqing Yang, Jianyi Wang, Yihang Luo, and Chen Change Loy. Upscale-a-video: Temporal- consistent diffusion model for real-world video super- resolution, 2023. 2, 7, 8, 14 Appendix A. Overview In the Appendix, we provide additional details and experi- ments to ...

    2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.