Pith. sign in

REVIEW 3 major objections 44 references

NanoVSR: Towards Real-Time Video Super-Resolution on Edge Devices

T0 review · 3 major / 0 minor · reviewed 2026-07-14 · grok-4.5

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

Pith's one-line read NanoVSR runs competitive video super-resolution in real time on edge hardware by collapsing a multi-branch convolutional network into plain convolutions and learning motion without optical flow.

desk verdict Clean edge-systems VSR recipe with real Jetson numbers and full reparam collapse; quality is solid for the size class but not close to flow SOTA, and the no-flow sufficiency claim is the softest part. read the letter →

arxiv 2607.10495 v1 pith:PQ3QM4U4 submitted 2026-07-11 cs.CV

Filip Pawlicki , Marcel Kańdu{l}a , Marcin Pucek , Kamil Dobies This is my paper
classification cs.CV
keywords VideoSuper-ResolutionEdgeAIStructuralReparameterizationReal-TimeInferenceBidirectionalRecurrentNetworksTensorRT
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

NanoVSR is a fully convolutional, bidirectional recurrent network built for video super-resolution on power- and memory-limited edge devices. It trains with multi-branch blocks that are later fused by structural reparameterization into ordinary 3 imes3 convolutions, so the entire inference graph stays TensorRT- and ONNX-native. Temporal context is carried by simple additive hidden-state updates rather than channel concatenation or explicit optical flow; a two-stage curriculum first teaches short-sequence spatial features and then long-sequence residual motion. On the REDS4 benchmark the 644 k-parameter model reaches 28.64 dB PSNR while sustaining 27.2 FPS on a 25 W Jetson Orin NX, and a 1.7 M-parameter variant still delivers 19.6 FPS at 29.15 dB. The result is a practical quality–speed trade-off that existing heavy transformer and flow-based models cannot match under the same edge constraints.

What carries the argument

Structural reparameterization of multi-branch (3 imes3 / 1 imes1 / identity) blocks into single dense 3 imes3 convolutions, together with additive bidirectional hidden-state propagation and a two-stage progressive training schedule.

What would settle it

On a held-out suite of sequences containing extreme non-rigid motion, measure whether NanoVSR’s PSNR falls more than roughly 1 dB behind an otherwise identical model that adds an explicit SPyNet alignment module; a larger gap would falsify the claim that implicit learning is sufficient for competitive restoration.

Watch

Extended reading notes

Core claim

A lightweight bidirectional recurrent convolutional architecture that never uses optical flow or custom operators can still produce competitive 4 imes video super-resolution on edge devices once it is trained with a progressive short-to-long curriculum and reparameterized into a pure stream of standard convolutions at inference.

Load-bearing premise

Progressive training on short then long sequences is enough for the network to learn residual spatio-temporal alignments that replace explicit optical flow without a large quality drop.

Editorial extensions

If this is right

  • Local 4 imes video upscaling becomes feasible on battery-powered cameras and mobile SoCs without cloud off-load.
  • Because the deployed graph contains only standard convolutions, TensorRT and similar accelerators can fully fuse and schedule the network.
  • Compact VSR models can keep real-time frame rates while still recovering high-frequency detail that pure single-image methods miss.
  • The same topology scales smoothly from a few hundred thousand to a few million parameters before quality saturates, giving designers a clear capacity–latency curve.

Reading between the lines

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

  • The reparameterization-plus-curriculum pattern is likely portable to other temporal edge tasks such as video denoising or frame interpolation.
  • Under severe non-rigid motion a lightweight hybrid that injects cheap flow only when residual error spikes could close the remaining quality gap without destroying the pure-convolution inference path.
  • A pure-convolution inference graph is an especially good fit for fixed-function or FPGA accelerators that cannot host deformable or attention kernels.
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 / 0 minor

Summary. NanoVSR proposes a fully convolutional bidirectional recurrent video super-resolution architecture aimed at real-time edge deployment. It discards explicit optical flow and deformable convolutions, uses element-wise additive hidden-state propagation (Eqs. 2–3) instead of channel concatenation, and builds blocks with RepVGG-style structural reparameterization so multi-branch training collapses to plain 3×3 convolutions at inference. A two-stage progressive curriculum (7-frame Vimeo-90K pre-training then 30-frame REDS fine-tuning) is used to stabilize implicit alignment learning. On REDS4 the 644k model reports 28.64 dB (RGB) at 27.2 FPS and the 1.7M model 29.15 dB at 19.58 FPS on Jetson Orin NX 16GB (25W) under TensorRT FP16 with T=15; H100 runtimes, Vid4/Vimeo90K-T metrics, a scaling curve, and ablations on fusion, single-branch training, pre-training, and SPyNet are also provided. Code is linked.

Significance. If the reported Jetson throughputs and quality numbers hold under independent reproduction, the work supplies a practical, ONNX/TensorRT-native VSR design that substantially improves the speed–quality Pareto front for sub-2M-parameter models on constrained hardware. Strengths that should be credited include: (i) end-to-end collapse of the bidirectional recurrent path into a plain convolution stream via reparameterization, (ii) re-measured H100 baselines plus dual-TDP Jetson TensorRT profiles rather than proxy FLOPs, (iii) a clear scaling study (Fig. 5) and targeted ablations (Tab. 3), and (iv) public code. The contribution is primarily systems/engineering rather than a new alignment theory, but that is valuable for edge CV.

major comments (3)
  1. Abstract and §4.2 claim “competitive restoration quality” and an “exceptional balance” for compact architectures. Table 1 shows NanoVSR-644k at 28.64 dB and NanoVSR-1.7M at 29.15 dB on REDS4 (RGB), i.e., 2.3–3.1 dB behind BasicVSR/IconVSR and ~1.4 dB behind EDVR-M (3.3M). The speed advantage is clear and well measured; the quality language overstates the result relative to the same table. Please reframe claims around the demonstrated efficiency–quality operating point (and, if possible, add other lightweight/real-time VSR baselines) so that “competitive/exceptional” is defined against a stated compact Pareto set rather than against full SOTA.
  2. The load-bearing design claim that progressive training plus additive bidirectional propagation (Sec. 3.1–3.3, Eqs. 2–3) lets the network “implicitly learn spatio-temporal alignments” sufficiently to replace explicit motion compensation is only weakly supported. Tab. 3: NanoVSR-226k-NOPRET matches the reference at 28.23 dB on REDS4 (only a small Vid4 drop); adding SPyNet yields +1.21 dB (28.23→29.44) at >2× runtime. Limitations §5 already note failures under extreme non-rigid motion, yet no motion-magnitude or failure-case stratification on REDS4 is given. Either strengthen the evidence (e.g., hard-motion subset, qualitative failure analysis, or a unidirectional streaming ablation) or narrow the abstract/intro claim to “acceptable quality under standard benchmarks without explicit flow,” consistent with the ablation gap.
  3. Edge evaluation (Tab. 2) uses a fixed T=15 bidirectional chunk. Limitations §5 correctly flag look-ahead buffering, but the paper never quantifies the quality/latency cost of a unidirectional or causal mode that would be required for strict frame-by-frame streaming. Because the central selling point is real-time edge VSR, a short unidirectional ablation (PSNR and FPS on the same Jetson configs) is needed to bound how much of the reported trade-off depends on non-causal context.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: empirical architecture + held-out metrics, not algebraic rearrangement of inputs.

full rationale

NanoVSR is a standard empirical systems paper. Its central claims (competitive PSNR/SSIM without explicit flow via progressive training + additive bidirectional propagation, plus TensorRT-native speed via structural reparameterization) are design choices whose outcomes are measured on held-out benchmarks (REDS4 clips 000/011/015/020 strictly excluded from training; also Vid4 and Vimeo90K-T). Equations 1–5 define a feed-forward architecture; they do not algebraically force the reported PSNR numbers. Ablations (Table 3) and scaling curves (Fig. 5) are independent empirical checks, not tautologies. Citations (RepVGG, BasicVSR, SPyNet, etc.) are external prior art; there is no self-citation chain, uniqueness theorem imported from the authors, fitted free parameter renamed as a prediction, or self-definitional loop. The paper is therefore self-contained against external benchmarks with score 0.

Assumptions & free parameters 4 free parameters · 5 assumptions · 1 invented entities

This is an empirical systems paper. Load-bearing content is architectural design choices and hyperparameters, not new physical entities. Free parameters are capacity and training knobs that define the reported operating points. Axioms are standard CV/ML assumptions plus the paper’s key modeling bet that progressive training can substitute for explicit flow. The main invented construct is the NanoVSR additive-propagation reparameterizable stack itself.

free parameters (4)
  • Model capacity (channels C and number N of temporal RepVGG blocks) = e.g. NanoVSR-644k, NanoVSR-1.7M
    Defines the 226k/644k/1.7M/5.4M variants and the reported PSNR–FPS Pareto points; chosen by authors, not derived.
  • Progressive training schedule (50k iters on 7-frame Vimeo-90K then 100k on 30-frame REDS) = 50k + 100k iterations
    Curriculum lengths and switch point are hand-chosen; ablation shows they affect generalization (Vid4).
  • Learning-rate cosine schedule bounds = 3e-4 to 1e-7
    3e-4 → 1e-7 with Adam and Charbonnier loss; standard but free knobs that affect final PSNR.
  • Inference temporal window T=15 and spatial patch/batch settings = T=15; 256×256; batch 12
    T=15 chunking for Jetson memory/power and 256×256 patches / batch 12 for training are deployment and training choices that condition reported FPS and quality.
assumptions (5)
  • domain assumption Multi-branch (3×3 + 1×1 + identity) blocks with BN can be fused into a single dense 3×3 convolution with identical inference function (structural reparameterization).
    Invoked in Sec. 3.2 and Fig. 2; taken from RepVGG literature and treated as exact for deployment.
  • ad hoc to paper Element-wise additive bidirectional hidden-state propagation can carry residual motion and long-range context without channel concatenation or explicit warping.
    Core design in Sec. 3.1 Eqs. (2)–(3); not proven, supported only by end metrics and ablations.
  • ad hoc to paper A short-then-long progressive curriculum stabilizes training of implicit-alignment VSR and induces useful spatio-temporal alignment.
    Sec. 3.3; ablation NOPRET shows modest Vid4 drop; still an unproved modeling bet for the central no-flow claim.
  • domain assumption Bilinear 4× residual skip plus learned high-frequency residual is an adequate reconstruction path for edge VSR.
    Eq. (5); common SR practice, chosen for hardware efficiency over bicubic.
  • domain assumption REDS4, Vid4, and Vimeo90K-T under bicubic 4× degradation are valid proxies for the targeted edge upscaling use cases.
    Standard VSR evaluation protocol used throughout Sec. 4; real camera degradations may differ.
invented entities (1)
  • NanoVSR additive bidirectional reparameterizable architecture independent evidence
    purpose: Provide a TensorRT-native VSR stack that trains multi-branch and deploys as plain convolutions without optical flow or custom CUDA ops.
    The specific combination of additive temporal fusion, RepVGG blocks throughout propagation, and progressive no-flow training is introduced here; independent evidence is the public code and benchmark numbers, not external theory.

how reviews work

0 comments
Cite this review

Pith. "Pith review of NanoVSR: Towards Real-Time Video Super-Resolution on Edge Devices." pith.science (2026). https://pith.science/paper/PQ3QM4U4

@misc{pith2026260710495,
  author       = {Pith},
  title        = {Pith review of: NanoVSR: Towards Real-Time Video Super-Resolution on Edge Devices},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PQ3QM4U4}},
  note         = {Machine review of arXiv:2607.10495}
}
read the original abstract

Recent Video Super-Resolution (VSR) methods rely heavily on transformers and explicit optical flow, creating computational overhead and custom operations that hinder deployment on hardware accelerators like TensorRT. To address this, we introduce NanoVSR, a scalable, fully convolutional architecture designed for resource-constrained edge devices. Using structural reparameterization, NanoVSR collapses into standard convolutions during inference, ensuring seamless hardware compatibility and negligible runtime overhead. Furthermore, despite lacking explicit motion compensation, it maintains competitive restoration quality by implicitly learning spatio-temporal alignments through progressive training. Evaluated on the REDS4 benchmark, NanoVSR demonstrates an exceptional balance between accuracy and computational efficiency, significantly improving the trade-off for compact architectures. Our NanoVSR-644k baseline yields 28.64 dB PSNR while delivering 27.2 FPS on the NVIDIA Jetson Orin NX 16GB (25W), offering massive speed gains over heavier models. The scaled NanoVSR-1.7M variant reaches 29.15 dB with a throughput of 19.58 FPS, providing superior, edge-optimized upscaling. Code is available at https://github.com/filippawlicki/nanovsr.

Figures

Figures reproduced from arXiv: 2607.10495 by the authors.

Figure 1
Figure 1. Detailed schematic of the NanoVSR architecture. LR input frames are embed￾ded into a latent space and passed through a bidirectional recurrent structure built with reparameterizable blocks (see [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Diagram of the structural reparameterization process [7]. During training, the block utilizes a multi-branch topology (3 × 3, 1 × 1, and identity pathways) to ensure robust feature extraction. For deployment, Batch Normalization statistics are first ab￾sorbed into their respective kernels. Finally, the parallel branches are mathematically fused into a single, dense 3 × 3 convolution, eliminating memory fragmentation… view at source ↗
Figure 3
Figure 3. Qualitative comparison on REDS4 [23] dataset. Left: full frame, Right: magni￾fied patches. 4.2 Quantitative Results We evaluate NanoVSR against several state-of-the-art VSR baselines on the Vid4, REDS4, and Vimeo90K-T benchmarks. Our analysis considers both restora￾tion quality (PSNR, SSIM) and computational efficiency, including parameter count and runtime per frame for a 180 × 320 input. As detailed in Tab. 1, Nan… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison on Vid4 [19] dataset. Left: full frame, Right: magnified patches. that NanoVSR can deliver solid restoration quality at a fraction of the com￾putational cost of traditional architectures. Qualitative results for the REDS4 dataset are illustrated …
Figure 5
Figure 5. Figure 5: Model scaling analysis on the REDS4 [23] dataset. The plot illustrates the trade-off between the number of parameters and restoration quality (PSNR). While scaling up the architecture consistently improves performance, the curve demonstrates diminishing returns beyond …

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

44 extracted references · 6 canonical work pages

  1. [1]

    In: 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    Caballero, J., Ledig, C., Aitken, A., Acosta, A., Totz, J., Wang, Z., Shi, W.: Real- time video super-resolution with spatio-temporal networks and motion compen- sation. In: 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2848–2857 (2017).https://doi.org/10.1109/CVPR.2017.304

    work page doi:10.1109/cvpr.2017.304 2017

  2. [2]

    Cao, J., Li, Y., Zhang, K., Van Gool, L.: Video super-resolution transformer (2023), arXiv preprint arXiv:2106.06847

    arXiv 2023

  3. [3]

    In: 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Chan, K.C., Wang, X., Yu, K., Dong, C., Loy, C.C.: BasicVSR: The search for essential components in video super-resolution and beyond. In: 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4945–4954 (2021).https://doi.org/10.1109/CVPR46437.2021.00491

    doi:10.1109/cvpr46437.2021.00491 2021

  4. [4]

    In: 2022 IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR)

    Chan, K.C., Zhou, S., Xu, X., Loy, C.C.: BasicVSR++: Improving video super- resolution with enhanced propagation and alignment. In: 2022 IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR). pp. 5962–5971 (2022). https://doi.org/10.1109/CVPR52688.2022.00588

    doi:10.1109/cvpr52688.2022.00588 2022

  5. [5]

    In: Proceedings of 1st International Conference on Image Processing

    Charbonnier, P., Blanc-Feraud, L., Aubert, G., Barlaud, M.: Two deterministic half-quadratic regularization algorithms for computed imaging. In: Proceedings of 1st International Conference on Image Processing. vol. 2, pp. 168–172 (1994). https://doi.org/10.1109/ICIP.1994.413553

    work page doi:10.1109/icip.1994.413553 1994

  6. [6]

    In: 2017 IEEE International Conference on Computer Vision (ICCV)

    Dai, J., Qi, H., Xiong, Y., Li, Y., Zhang, G., Hu, H., Wei, Y.: Deformable con- volutional networks. In: 2017 IEEE International Conference on Computer Vision (ICCV). pp. 764–773 (2017).https://doi.org/10.1109/ICCV.2017.89

    doi:10.1109/iccv.2017.89 2017

  7. [7]

    In: 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Ding, X., Zhang, X., Ma, N., Han, J., Ding, G., Sun, J.: RepVGG: Making VGG- style convnets great again. In: 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13728–13737 (2021).https://doi.org/10. 1109/CVPR46437.2021.01352

    arXiv 2021

  8. [8]

    IEEE Transactions on Pattern Analysis and Machine Intelligence 38(2), 295–307 (2016).https://doi.org/10.1109/TPAMI.2015.2439281

    Dong, C., Loy, C.C., He, K., Tang, X.: Image super-resolution using deep convolu- tional networks. IEEE Transactions on Pattern Analysis and Machine Intelligence 38(2), 295–307 (2016).https://doi.org/10.1109/TPAMI.2015.2439281

    doi:10.1109/tpami.2015.2439281 2016

Show all 44 references
  1. [9]

    In: International Conference on Learning Representations (2021) 16 F

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on L...

    2021

  2. [10]

    In: 2019 IEEE/CVF International Conference on Com- puter Vision Workshop (ICCVW)

    Fuoli, D., Gu, S., Timofte, R.: Efficient video super-resolution through recurrent latent space propagation. In: 2019 IEEE/CVF International Conference on Com- puter Vision Workshop (ICCVW). pp. 3476–3485 (2019).https://doi.org/10. 1109/ICCVW.2019.00431

    2019

  3. [11]

    In: 2015 IEEE International Confer- ence on Computer Vision (ICCV)

    He, K., Zhang, X., Ren, S., Sun, J.: Delving deep into rectifiers: Surpassing human- level performance on imagenet classification. In: 2015 IEEE International Confer- ence on Computer Vision (ICCV). pp. 1026–1034 (2015).https://doi.org/10. 1109/ICCV.2015.123

    2015

  4. [12]

    In: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 770–778 (2016).https://doi.org/10.1109/CVPR.2016.90

    2016 doi

  5. [13]

    In: 2019 IEEE/CVF International Conference on Computer Vision (ICCV)

    Howard, A., Sandler, M., Chen, B., Wang, W., Chen, L.C., Tan, M., Chu, G., Vasudevan, V., Zhu, Y., Pang, R., Adam, H., Le, Q.: Searching for MobileNetV3. In: 2019 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 1314–1324 (2019).https://doi.org/10.1109/ICCV.2019.00140

    2019 doi

  6. [14]

    In: Bach, F., Blei, D

    Ioffe, S., Szegedy, C.: Batch normalization: Accelerating deep network training by reducing internal covariate shift. In: Bach, F., Blei, D. (eds.) Proceedings of the 32nd International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 37, pp. 448–...

    2015

  7. [15]

    In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J.M

    Isobe, T., Jia, X., Gu, S., Li, S., Wang, S., Tian, Q.: Video super-resolution with recurrent structure-detail network. In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J.M. (eds.) Computer Vision – ECCV 2020. pp. 645–660. Springer International Publishing, Cham (2020)

    2020

  8. [16]

    In: Interna- tional Conference on Learning Representations (ICLR) (2015)

    Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. In: Interna- tional Conference on Learning Representations (ICLR) (2015)

    2015

  9. [17]

    IEEE Transactions on Image Processing 33, 2171–2182 (2024).https://doi.org/10.1109/TIP.2024.3372454

    Liang, J., Cao, J., Fan, Y., Zhang, K., Ranjan, R., Li, Y., Timofte, R., Van Gool, L.: VRT: A video restoration transformer. IEEE Transactions on Image Processing 33, 2171–2182 (2024).https://doi.org/10.1109/TIP.2024.3372454

    2024 doi

  10. [18]

    In: Proceedings of the 36th International Conference on Neu- ral Information Processing Systems

    Liang, J., Fan, Y., Xiang, X., Ranjan, R., Ilg, E., Green, S., Cao, J., Zhang, K., Timofte, R., Van Gool, L.: Recurrent video restoration transformer with guided deformable attention. In: Proceedings of the 36th International Conference on Neu- ral Information Processing Syste...

    2022

  11. [19]

    IEEE Transactions on Pattern Analysis and Machine Intelligence36(2), 346–360 (2014).https:// doi.org/10.1109/TPAMI.2013.127

    Liu, C., Sun, D.: On bayesian adaptive video super resolution. IEEE Transactions on Pattern Analysis and Machine Intelligence36(2), 346–360 (2014).https:// doi.org/10.1109/TPAMI.2013.127

    2014 doi

  12. [20]

    In: International Conference on Learning Representations (2017)

    Loshchilov, I., Hutter, F.: SGDR: Stochastic gradient descent with warm restarts. In: International Conference on Learning Representations (2017)

    2017

  13. [21]

    In: Computer Vision – ECCV 2018

    Ma, N., Zhang, X., Zheng, H.T., Sun, J.: ShuffleNet V2: Practical guidelines for efficient CNN architecture design. In: Computer Vision – ECCV 2018. pp. 122–138. Springer International Publishing, Cham (2018)

    2018

  14. [22]

    In: International Conference on Learning Representations (2018)

    Micikevicius, P., Narang, S., Alben, J., Diamos, G., Elsen, E., Garcia, D., Ginsburg, B., Houston, M., Kuchaiev, O., Venkatesh, G., Wu, H.: Mixed precision training. In: International Conference on Learning Representations (2018)

    2018

  15. [23]

    In: 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW)

    Nah, S., Baik, S., Hong, S., Moon, G., Son, S., Timofte, R., Lee, K.M.: NTIRE 2019 challenge on video deblurring and super-resolution: Dataset and study. In: 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). pp. 1996–2005 (2019).https://doi...

    2019 doi

  16. [24]

    com/tensorrt, last accessed 2026-06-25 NanoVSR: Towards Real-Time VSR on Edge Devices 17

    NVIDIA Corporation: NVIDIA TensorRT (2017),https://developer.nvidia. com/tensorrt, last accessed 2026-06-25 NanoVSR: Towards Real-Time VSR on Edge Devices 17

    2017

  17. [25]

    ONNX Contributors: ONNX: Open neural network exchange (2017),https:// github.com/onnx/onnx, last accessed 2026-06-25

    2017

  18. [26]

    In: 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    Ranjan, A., Black, M.J.: Optical flow estimation using a spatial pyramid network. In: 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2720–2729 (2017).https://doi.org/10.1109/CVPR.2017.291

    2017 doi

  19. [27]

    In: 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Sajjadi, M.S.M., Vemulapalli, R., Brown, M.: Frame-recurrent video super- resolution. In: 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6626–6634 (2018).https://doi.org/10.1109/CVPR.2018.00693

    2018 doi

  20. [28]

    In: 2018 IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR)

    Sandler, M., Howard, A., Zhu, M., Zhmoginov, A., Chen, L.C.: MobileNetV2: In- verted residuals and linear bottlenecks. In: 2018 IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR). pp. 4510–4520. IEEE Computer Society, Los Alamitos, CA, USA (Jun 2018).https...

    2018 doi

  21. [29]

    Schwartz, R., Dodge, J., Smith, N.A., Etzioni, O.: Green AI. Commun. ACM 63(12), 54–63 (Nov 2020).https://doi.org/10.1145/3381831

    2020 doi

  22. [30]

    In: Proceedings of the 36th International Con- ference on Neural Information Processing Systems

    Shi, S., Gu, J., Xie, L., Wang, X., Yang, Y., Dong, C.: Rethinking alignment in video super-resolution transformers. In: Proceedings of the 36th International Con- ference on Neural Information Processing Systems. NIPS ’22, Curran Associates Inc., Red Hook, NY, USA (2022)

    2022

  23. [31]

    In: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    Shi, W., Caballero, J., Huszár, F., Totz, J., Aitken, A.P., Bishop, R., Rueckert, D., Wang, Z.: Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVP...

    2016 doi

  24. [32]

    In: Korhonen, A., Traum, D., Màrquez, L

    Strubell, E., Ganesh, A., McCallum, A.: Energy and policy considerations for deep learning in NLP. In: Korhonen, A., Traum, D., Màrquez, L. (eds.) Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics. pp. 3645–3650. Association for Computatio...

    2019 doi

  25. [33]

    In: 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., Erhan, D., Vanhoucke, V., Rabinovich, A.: Going deeper with convolutions. In: 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1–9 (2015). https://doi.org/10.1109/CVPR.2015.7298594

    2015 doi

  26. [34]

    In: Chaudhuri, K., Salakhutdinov, R

    Tan, M., Le, Q.: EfficientNet: Rethinking model scaling for convolutional neural networks. In: Chaudhuri, K., Salakhutdinov, R. (eds.) Proceedings of the 36th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 97, pp. 6105–6114. PMLR (...

    2019

  27. [35]

    In: 2017 IEEE International Conference on Computer Vision (ICCV)

    Tao, X., Gao, H., Liao, R., Wang, J., Jia, J.: Detail-revealing deep video super- resolution. In: 2017 IEEE International Conference on Computer Vision (ICCV). pp. 4482–4490 (2017).https://doi.org/10.1109/ICCV.2017.479

    2017 doi

  28. [36]

    In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Tian, Y., Zhang, Y., Fu, Y., Xu, C.: TDAN: Temporally-deformable alignment network for video super-resolution. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3357–3366 (2020).https://doi. org/10.1109/CVPR42600.2020.00342

    2020 doi

  29. [37]

    IEEE Transactions on Pattern Analysis and Machine Intelligence44(9), 4555–4576 (2022).https: //doi.org/10.1109/TPAMI.2021.3069908

    Wang, X., Chen, Y., Zhu, W.: A survey on curriculum learning. IEEE Transactions on Pattern Analysis and Machine Intelligence44(9), 4555–4576 (2022).https: //doi.org/10.1109/TPAMI.2021.3069908

    2022 doi

  30. [38]

    In: 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW)

    Wang, X., Chan, K.C., Yu, K., Dong, C., Loy, C.C.: EDVR: Video restoration with enhanced deformable convolutional networks. In: 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). pp. 1954–1963 (2019).https://doi.org/10.1109/CVPRW.2019.00247 ...

    2019 doi

  31. [39]

    ACM Comput

    Wang, X., Tang, Z., Guo, J., Meng, T., Wang, C., Wang, T., Jia, W.: Empowering edge intelligence: A comprehensive survey on on-device ai models. ACM Comput. Surv.57(9) (Apr 2025).https://doi.org/10.1145/3724420

    2025 doi

  32. [40]

    IEEE Transactions on Image Processing 13(4), 600–612 (2004).https://doi.org/10.1109/TIP.2003.819861

    Wang, Z., Bovik, A., Sheikh, H., Simoncelli, E.: Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing 13(4), 600–612 (2004).https://doi.org/10.1109/TIP.2003.819861

    2004 doi

  33. [41]

    International Journal of Computer Vision127(8), 1106–1125 (2019)

    Xue, T., Chen, B., Wu, J., Wei, D., Freeman, W.T.: Video enhancement with task- oriented flow. International Journal of Computer Vision127(8), 1106–1125 (2019). https://doi.org/10.1007/s11263-018-01144-2

    2019 doi

  34. [42]

    In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Yoo, J., Ahn, N., Sohn, K.A.: Rethinking data augmentation for image super- resolution: A comprehensive analysis and a new strategy. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8372–8381 (2020).https://doi.org/10.1109/CVPR42600.2020.00840

    2020 doi

  35. [43]

    In: Ferrari, V., Hebert, M., Sminchisescu, C., Weiss, Y

    Zhang, Y., Li, K., Li, K., Wang, L., Zhong, B., Fu, Y.: Image super-resolution using very deep residual channel attention networks. In: Ferrari, V., Hebert, M., Sminchisescu, C., Weiss, Y. (eds.) Computer Vision – ECCV 2018. pp. 294–310. Springer International Publishing, Cham (2018)

    2018

  36. [44]

    In: 2019 IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion(CVPR).pp.9300–9308(2019).https://doi.org/10.1109/CVPR.2019.00953

    Zhu, X., Hu, H., Lin, S., Dai, J.: Deformable convnets v2: More deformable, better results. In: 2019 IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion(CVPR).pp.9300–9308(2019).https://doi.org/10.1109/CVPR.2019.00953

    2019 doi

Pith tools

Reviewed July 14, 2026 · model on record in the stance chip above.