Pith. sign in

REVIEW 2 major objections 4 minor 36 references

RepNet-VSR: Reparameterizable Architecture for High-Fidelity Video Super-Resolution

T0 review · 2 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 RepNet-VSR claims a higher MAI2025 composite score than previous champions with real-time mobile NPU runtime.

desk verdict The architecture is a sensible engineering mix, but the score column does not match the paper's own formula, so the headline comparison is not reproducible. read the letter →

arxiv 2504.15649 v1 pith:COBLKY4V submitted 2025-04-22 eess.IV cs.CV

Biao Wu , Diankai Zhang , Shaoli Liu , Si Gao , Chengjian Zheng , Ning Wang This is my paper · ORCID
classification eess.IVcs.CV
keywords videosuper-resolutionreparameterizationneuralarchitecturesearchmobileNPUreal-timeinferenceREDSdatasetmodelcompression
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper is trying to establish that real-time 4× video super-resolution on mobile NPUs can be accurate enough to beat the previous champion algorithms in the MAI challenge series. The authors propose RepNet-VSR, a convolutional architecture that uses reparameterizable bottleneck blocks and a 1×1 convolution in place of the usual 3×3 channel-reduction layer before depth-to-space upsampling. They report 27.79 dB PSNR on the REDS validation set for 180p→720p upscaling, with a runtime of 103 ms per 10 frames on a MediaTek Dimensity NPU. Plugging those numbers into the MAI2025 score formula gives 0.029, higher than the EVSRNet and RCBSR scores of 0.0174 and 0.0154, which is the paper's central evidence that accuracy and efficiency can be co-optimized on edge hardware.

What carries the argument

RepConv is the load-bearing module: during training it is a bottleneck of a 1×1 expansion, a 3×3 convolution, a 1×1 reduction, plus a 1×1 residual shortcut; at inference these branches are merged into one 3×3 convolution so the deployed graph is as cheap as a plain convolutional block. The second component is the 1×1 convolution placed right before the depth-to-space (pixel-shuffle) layer, which cuts the channel count from 48 to 3 and avoids the expensive 3×3 upsampling bottleneck. A neural architecture search over channel count (0–32) and number of RepConv blocks (0–8), regularized by FLOPs, selects the configuration used for the final model, and a second phase fine-tunes that model with an L2 loss.

What would settle it

Run the exported TensorFlow Lite model on a MediaTek Dimensity device under the official AI Benchmark protocol in FP16 with the MediaTek neuron delegate, measuring the time for 10 frames; if 22·(PSNR−27)/runtime falls below EVSRNet's 0.0174, the paper's central superiority claim fails.

Watch

Extended reading notes

Core claim

The central claim is that the architecture, RepNet-VSR, obtains a better accuracy-efficiency trade-off on the MAI2025 scoring metric than the previous challenge winners. The key quantitative claim is a PSNR of 27.79 dB on REDS validation for 4× super-resolution (180p→720p) and a runtime of 103 ms per 10 frames on a MediaTek Dimensity NPU, giving a composite score of 0.029 versus 0.0174 for EVSRNet and 0.0154 for RCBSR. The authors attribute the gain to replacing the 3×3 convolution that compresses 48 channels to 3 before depth-to-space with a 1×1 convolution (which they say speeds up upsampling by a factor of 4 and improves quality), and to reparameterized RepConv blocks that train with a multi-branch structure and fold into a single 3×3 convolution at inference.

Load-bearing premise

The 103 ms per 10 frames runtime on a MediaTek Dimensity NPU was produced by the official MAI2025 benchmarking procedure, even though the paper reports that the specified hardware was unavailable during the evaluation and does not describe how the final NPU measurement was made.

Editorial extensions

If this is right

  • If the reported runtime holds under the official protocol, 4× video super-resolution at 720p output becomes practical on current phone NPUs without cloud offloading.
  • The 1×1-before-depth-to-space substitution is a drop-in change for other super-resolution backbones and should cut upsampling cost whenever the pre-upsampling channel count is large.
  • The NAS search over channels and RepConv blocks yields a configurable trade-off curve, so the same architecture can be tuned for faster or higher-fidelity deployment.
  • On the MAI2025 score formula, the reported configuration beats both prior champions, making it a reference point for future mobile SR challenge entries.

Reading between the lines

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

  • The paper's Section 4.3 says the required MediaTek hardware was unavailable and runtime was measured on a Qualcomm Snapdragon 870 GPU, while Section 4.5 states MediaTek NPU results without detailing the measurement; if the 103 ms figure was obtained differently, the composite score advantage could shrink or disappear.
  • The CPU runtime is 2.44× slower than RCBSR (273 ms vs 112 ms), so the reported advantage is specific to NPU execution; on systems where parts of the graph run on the CPU, RCBSR may be the better choice.
  • A natural test of generality would be to insert the same 1×1 upsampling bottleneck into a transformer-based VSR model and check whether the 4× upsampling speedup and quality retention replicate outside convolutional backbones.
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

2 major / 4 minor

Summary. This paper proposes RepNet-VSR, a reparameterizable convolutional architecture for 4x video super-resolution on mobile NPUs. The architecture uses NAS to select the number of channels and RepBlocks, 1x1 convolutions for channel compression before depth-to-space, and multi-level feature fusion via channel concatenation. On the REDS validation set the authors report 27.79 dB PSNR and 103 ms per 10 frames on a MediaTek Dimensity NPU, and a MAI2025 composite score of 0.029 that is claimed to be 66.7% higher than EVSRNet and 88.3% higher than RCBSR. The paper also presents an ablation over channel/block configurations and a qualitative visual comparison.

Significance. If substantiated, the method would be a useful contribution to efficient edge VSR: the use of reparameterization, NAS-guided configuration, and 1x1 channel compression before upsampling is a sensible design direction, and the reported PSNR/runtime combination would be competitive. The paper is commendable for including multiple model configurations in Table 2 and for making the architectural recipe in Figures 1 and 2 concrete. However, the central comparative claim depends entirely on score and runtime values that cannot be reproduced from the paper's own formulas and measurement statements; until those are corrected and verified on the official hardware, the significance of the claimed superiority is not established.

major comments (2)
  1. [Section 3.1, Eq. (1), Tables 2-3] The composite scores in Tables 2 and 3 do not follow the scoring formula stated in Eq. (1). For example, substituting EVSRNet's reported PSNR 27.42 and NPU runtime 103 ms into Eq. (1) gives 22*0.42/103 ≈ 0.0897, not the reported 0.0174; RCBSR gives 22*0.28/95.8 ≈ 0.0643, not 0.0154; and the proposed model gives 22*0.79/103 ≈ 0.169, not 0.029. Table 2 has the same discrepancy, e.g. row 2 gives 22*0.79/89.6 ≈ 0.194 versus the reported 0.0334. The discrepancies are not a constant rescaling, so no simple unit conversion repairs them. Consequently, the claimed 66.7% and 88.3% improvements over the previous champions are not reproducible from the paper's own data; either Eq. (1) is not the scoring formula actually used, or the Score columns are incorrect, and in both cases the central comparative claim is not supported.
  2. [Section 4.3 vs. Section 4.5] The runtime measurement that feeds the score is not supported. Section 4.3 explicitly states that the MediaTek hardware was unavailable and that efficiency assessments were conducted on a Qualcomm Snapdragon 870 GPU, yet the abstract and Section 4.5 report 103 ms per 10 frames on a MediaTek Dimensity NPU without describing any measurement procedure. Section 4.4 labels Table 2 as Snapdragon 870 GPU results, while Table 3 is labeled MediaTek NPU, and no protocol for the NPU benchmark (delegate, quantization, number of runs, warm-up) is given. Moreover, the architecture selection in Table 2 uses runtimes from a different hardware platform than the final NPU comparison in Table 3, so the selected 'optimal' configuration is not validated on the target hardware. This undermines the runtime input to Eq. (1) and the comparability of the claimed NPU results.
minor comments (4)
  1. [Table 1] The search space lists 'The number of channels 0,1,..,32'; a channel count of 0 is degenerate and presumably a typo for a positive range.
  2. [Equation (2)] Equation (2) uses symbols θ and ψ without defining them, and the value of the regularizer weight λ is not reported; please clarify the notation and provide the loss configuration used in the NAS phase.
  3. [Section 4.6] The heading contains a typo ('Comparision' should be 'Comparison'), and Figure 3 would benefit from explicit crop locations or zoomed insets to support the claimed qualitative superiority.
  4. [General] No model weights or code are released, and no statistical variation (e.g., multiple runs) is reported for the PSNR and runtime values; adding these would improve reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reported scores are empirical benchmark evaluations, not predictions derived from the paper's own equations or from load-bearing self-citations.

full rationale

RepNet-VSR makes no derivation of its central claims; it reports benchmark measurements (PSNR, runtime, composite score) on the REDS validation set. The NAS objective in Eq. (2) is an optimization criterion, not an input that definitionally fixes the reported Score. The model configurations in Table 2 are empirical search results, and preferring the configuration with the highest Score is model selection, not a prediction. The comparison to EVSRNet and RCBSR is an external empirical benchmark; the citation to [32] (EVSRNet, co-authored by one of the present authors) is used only as a baseline name and carries no load-bearing evidence. No equation in the paper defines a predicted quantity in terms of the target claim, and no fitted parameter is renamed as a prediction. The apparent arithmetic mismatch between Eq. (1) and the Score columns is a correctness or consistency concern, not circularity. Therefore no circular step is identifiable under the stated standards.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim depends on the following free parameters and assumptions. The number of channels and Repblocks were fitted via NAS to the REDS validation set, which is also used for the final score report. The scoring formula and hardware measurement protocol are assumed to be consistent, but the paper's own data contradict these assumptions.

free parameters (3)
  • Number of channels (nc) = 16
    Selected by NAS on the REDS validation set to maximize the challenge score; Table 2 shows the search results.
  • Number of Repblocks (nb) = 4
    Selected by NAS alongside channels; Table 2 shows the search results.
  • Loss regularizer weight lambda
    Equation (2) includes lambda times R(psi) to penalize FLOPs, but the paper never specifies the value of lambda.
assumptions (4)
  • domain assumption The challenge score formula in Eq. (1) is used consistently to rank models.
    The reported scores in Table 2 and Table 3 do not match this formula, so the ranking depends on this assumption being true.
  • standard math Reparameterized multi-branch training networks produce outputs identical to the merged single-branch inference network.
    This is a standard property of linear operations like convolutions, relied on when reporting NPU runtime for the merged model.
  • domain assumption The REDS validation set serves as an unbiased evaluation set.
    The paper uses NAS on the same validation set to choose the architecture, so the final PSNR is not an unbiased estimate.
  • domain assumption MediaTek NPU runtime measurements were made under a protocol comparable to that used for the previous champions.
    The paper does not describe the NPU measurement procedure after stating the hardware was unavailable.

how reviews work

0 comments
Cite this review

Pith. "Pith review of RepNet-VSR: Reparameterizable Architecture for High-Fidelity Video Super-Resolution." pith.science (2026). https://pith.science/paper/COBLKY4V

@misc{pith2026250415649,
  author       = {Pith},
  title        = {Pith review of: RepNet-VSR: Reparameterizable Architecture for High-Fidelity Video Super-Resolution},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/COBLKY4V}},
  note         = {Machine review of arXiv:2504.15649}
}
read the original abstract

As a fundamental challenge in visual computing, video super-resolution (VSR) focuses on reconstructing highdefinition video sequences from their degraded lowresolution counterparts. While deep convolutional neural networks have demonstrated state-of-the-art performance in spatial-temporal super-resolution tasks, their computationally intensive nature poses significant deployment challenges for resource-constrained edge devices, particularly in real-time mobile video processing scenarios where power efficiency and latency constraints coexist. In this work, we propose a Reparameterizable Architecture for High Fidelity Video Super Resolution method, named RepNet-VSR, for real-time 4x video super-resolution. On the REDS validation set, the proposed model achieves 27.79 dB PSNR when processing 180p to 720p frames in 103 ms per 10 frames on a MediaTek Dimensity NPU. The competition results demonstrate an excellent balance between restoration quality and deployment efficiency. The proposed method scores higher than the previous champion algorithm of MAI video super-resolution challenge.

Figures

Figures reproduced from arXiv: 2504.15649 by the authors.

Figure 1
Figure 1. RepNet-VSR architecture overview [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. RepConv overview [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Qualitative comparison on the REDS val datasets. Zoom in for better visualization. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

36 extracted references · 36 canonical work pages

  1. [1]

    Ai benchmark: All about deep learning on smartphones in 2019

    Andrei Kulik Seungsoo Yang Ke Wang Felix Baum Max Wu Lirong Xu Andrey Ignatov, Radu Timofte and Luc Van Gool. Ai benchmark: All about deep learning on smartphones in 2019

    work page 2019

  2. [2]

    Ai benchmark: Running deep neural networks on android smartphones

    William Chou Ke Wang Max Wu Tim Hartley An- drey Ignatov, Radu Timofte and Luc Van Gool. Ai benchmark: Running deep neural networks on android smartphones

    work page

  3. [3]

    Fast, accurate, and lightweight super-resolution with cascading residual network

    Byungkon Kang Namhyuk Ahn and Kyung-Ah Sohn. Fast, accurate, and lightweight super-resolution with cascading residual network. In Proceedings of the European Conference on Computer Vision, pages 252–268, 2018

    work page 2018

  4. [4]

    Fast and memory-efficient network to- wards efficient image super-resolution

    Jie Liu Jie Tang Gangshan Wu Zongcai Du, Ding Liu and Lean Fu. Fast and memory-efficient network to- wards efficient image super-resolution. page 853–862, 2022

    work page 2022

  5. [5]

    Feature distillation interac- tion weighting network for lightweight image super- resolution

    Juncheng Li Fei Wu Huimin Lu Guangwei Gao, Wenjie Li and Yi Yu. Feature distillation interac- tion weighting network for lightweight image super- resolution. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 661–669 , 2022

    work page 2022

  6. [6]

    Lightweight image super-resolution with in- formation multi-distillation network

    Yunchu Yang Zheng Hui, Xinbo Gao and Xiumei Wang. Lightweight image super-resolution with in- formation multi-distillation network. In Proceedings of the 27th ACM International Conference on Multi- media, pages 2024–2032, 2019

    work page 2024

  7. [7]

    C.; He K.; Dong, C.; Loy and X

    C. C.; He K.; Dong, C.; Loy and X. 2015. Tang. Image super-resolution using deep convolutional networks. page 38(2): 295–307

    work page 2015

  8. [8]

    Ac- celerating the super-resolution convolutional neural network

    Chen Change Loy Chao Dong and Xiaoou Tang. Ac- celerating the super-resolution convolutional neural network. 2016

    work page 2016

Show all 36 references
  1. [9]

    Deeply-recursive convolutional network for image super-resolution

    Jung Kwon Lee Jiwon Kim and Kyoung Mu Lee. Deeply-recursive convolutional network for image super-resolution. 2016

    2016

  2. [10]

    Image super- resolution via deep recursive residual network

    Jian Yang Ying Tai and Xiaoming Liu. Image super- resolution via deep recursive residual network. 2017

    2017

  3. [11]

    Fast and accurate single image super-resolution via information distillation network

    Xiumei Wang Zheng Hui and Xinbo Gao. Fast and accurate single image super-resolution via information distillation network

  4. [12]

    Lightweight image super-resolution with in- formation multi-distillation network

    Yunchu Yang Zheng Hui, Xinbo Gao and Xiumei Wang. Lightweight image super-resolution with in- formation multi-distillation network

  5. [13]

    Residual fea- ture distillation network for lightweight image super- resolution

    Jie Tang Jie Liu and Gangshan Wu. Residual fea- ture distillation network for lightweight image super- resolution

  6. [14]

    Ef- ficient long-range attention network for image super- resolution

    Shi Guo Xindong Zhang, Hui Zeng and Lei Zhang. Ef- ficient long-range attention network for image super- resolution

  7. [15]

    Swinfir:revisiting the swinir with fast fourier convolution and improved training for image super-resolution

    Shizhuo Liu Xiaobing Wang Dafeng Zhang, Feiyu Huang and Zhezhu Jin. Swinfir:revisiting the swinir with fast fourier convolution and improved training for image super-resolution

  8. [16]

    Lu, Zhisheng

    et al. Lu, Zhisheng. Transformer for single image super-resolution

  9. [17]

    Recurrent back-projection network for video superresolution

    Greg Shakhnarovich Muhammad Haris and Norimichi Ukita. Recurrent back-projection network for video superresolution

  10. [18]

    Tdan: Temporally deformable alignment network for video super-resolution

    Yun Fu Yapeng Tian, Yulun Zhang and Chenliang Xu. Tdan: Temporally deformable alignment network for video super-resolution

  11. [19]

    Chan and Chen Change Loy

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

  12. [20]

    Efficient video super-resolution through recurrent latent space propagation

    Shuhang Gu Dario Fuoli and Radu Timofte. Efficient video super-resolution through recurrent latent space propagation

  13. [21]

    Bidirectional recurrent convolutional networks for multi-frame su- perresolution

    Wei Wang Yan Huang and Liang Wang. Bidirectional recurrent convolutional networks for multi-frame su- perresolution

  14. [22]

    Video super- resolution with recurrent structure-detail network

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

  15. [23]

    Chan, Xintao Wang and Chen Change Loy

    Ke Yu Chao Dong Kelvin C.K. Chan, Xintao Wang and Chen Change Loy. Basicvsr: The search for es- sential components in video super-resolution and be- yond

  16. [24]

    Takashi Isobe, Fang Zhu

    Xu Jia ShengjinWang. Takashi Isobe, Fang Zhu. Revisiting temporal modeling for video super- resolution

  17. [25]

    Neural architecture search for lightweight non-local networks

    Zhang et al. Neural architecture search for lightweight non-local networks

  18. [26]

    Fast, accurate, and lightweight super- resolution with neural architecture search

    Chu et al. Fast, accurate, and lightweight super- resolution with neural architecture search

  19. [27]

    Multi-objective neural architecture search for fast and accurate image super-resolution

    Mei et al. Multi-objective neural architecture search for fast and accurate image super-resolution

  20. [28]

    Repvgg: Making vgg-style convnets great again

    Xiaohan Ding, Xiangyu Zhang, Ningning Ma, Jun- gong Han, Guiguang Ding, and Jian Sun. Repvgg: Making vgg-style convnets great again. In IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion (CVPR), pages 13733–13742, 2021

    2021

  21. [29]

    Ecbsr: Edge-oriented convolution block for real-time super- resolution

    Jie Liu, Jie Tang, and Gangshan Wu. Ecbsr: Edge-oriented convolution block for real-time super- resolution. IEEE Transactions on Image Processing , 31:4529–4543, 2022

    2022

  22. [30]

    Repsr: Training efficient vgg- style super-resolution networks with structural re- parameterization

    Xiangyu Chen, Xintao Wang, Jiantao Zhou, Yu Qiao, and Chao Dong. Repsr: Training efficient vgg- style super-resolution networks with structural re- parameterization. In IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR) , pages 5720–5729, 2022

    2022

  23. [31]

    Swinrep: Swin-transformer with structural re- parameterization for efficient image restoration

    Yi Zhang, Kun Li, Wenqi Li, and Yunhui Liu. Swinrep: Swin-transformer with structural re- parameterization for efficient image restoration. arXiv preprint arXiv:2303.15348, 2023

    2023 arXiv

  24. [32]

    Evsrnet: Efficient video super- resolution with neural architecture search

    Liu, Shaoli et al. Evsrnet: Efficient video super- resolution with neural architecture search. IEEE Con- ference on Computer Vision and Pattern Recognition (CVPR), 2021

    2021

  25. [33]

    M., and Meng Z

    Romero A., Kim H., Timofte R., Ho C. M., and Meng Z. et al. Real-time video super-resolution on smart- phones with deep learning, mobile ai 2021 challenge: report. IEEE Conference on Computer Vision and Pat- tern Recognition (CVPR), 2021

    2021

  26. [34]

    Power efficient video super-resolution on mo- bile npus with deep learning, mobile ai & aim 2022 challenge: Report

    Andrey Ignatov, Radu Timofte, Cheng-Ming Chiang et al. Power efficient video super-resolution on mo- bile npus with deep learning, mobile ai & aim 2022 challenge: Report. IEEE Conference on Computer Vi- sion and Pattern Recognition (CVPR), 2022

    2022

  27. [35]

    Kim H., Hong S., and Han B. et al. Fine-grained neu- ral architecture search. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2019

    2019

  28. [36]

    Ntire 2019 challenge on video deblur- ring and superresolution: Dataset and study

    Seokil Hong Gyeongsik Moon Sanghyun Son Radu Timofte Seungjun Nah, Sungyong Baik and Ky- oung Mu Lee. Ntire 2019 challenge on video deblur- ring and superresolution: Dataset and study. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Works...

    2019

Pith tools

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