Pith. sign in

REVIEW 3 major objections 5 minor 41 references

LCS: An AI-based Low-Complexity Scaler for Power-Efficient Super-Resolution of Game Content

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A 0.21-million-parameter neural upscaler outperforms FSR1 and EASF on perceptual quality metrics for game content.

desk verdict Solid architecture-and-training paper whose 'better perceptual quality' claim is not established because the two winning metrics align with the training objective and the one human-calibrated metric shows parity. read the letter →

arxiv 2507.22873 v1 pith:X35RXI6S submitted 2025-07-30 cs.CV cs.LG

Simon Pochinda , Momen K. Tageldeen , Mark Thompson , Tony Rinaldi , Troy Giorshev , Keith Lee , Jie Zhou , Frederick Walls This is my paper · ORCID
classification cs.CVcs.LG
keywords low-complexityscalerefficientsuper-resolutionadversarialtrainingreparameterizationquantization-awaregamecontentupscalingperceptualqualitymetricsNPUoffload
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 proposes a low-complexity neural scaler (LCS) that upscales game frames from low to high resolution using only about 0.21 million parameters after reparameterization and 8-bit quantization. It is trained adversarially on native low-resolution and high-resolution image pairs from the GameIR dataset rather than on bicubic-downsampled images, and the paper claims the resulting model achieves better perceptual quality than AMD's FSR1 and EASF upscalers on the NIQE and LPIPS metrics. If this holds, it would show that a small, quantized neural network can be offloaded to a low-power device such as an NPU while still producing perceptually competitive game upscaling. The paper also frames the result as evidence that maximizing traditional metrics like PSNR and SSIM does not automatically produce perceptually satisfying images.

What carries the argument

The load-bearing mechanism is the residual reparameterization feature block (RRFB), a residual feature block built from residual-in-residual reparameterization blocks (RRRBs) plus an enhanced spatial attention (ESA) module. During training the RRRBs use multi-branch convolutions for higher learning capacity, and at inference they are folded into a single 3x3 convolution, cutting the generator from about 0.74 million to 0.21 million parameters and from 672 to 175 GMACs. The model is trained adversarially with a relativistic VGG-style discriminator and a generator loss combining adversarial, L1, and perceptual terms, then quantized to INT8 using quantization-aware training so the efficiency optimizations preserve the perceptual scores.

What would settle it

Run a human pairwise preference study on the GameIR validation images and other game content, asking viewers to choose between LCS INT8, FSR1, and EASF upscales; if viewers do not prefer LCS at rates consistent with its NIQE and LPIPS ordering, the central claim fails.

Watch

Extended reading notes

Core claim

The paper's central claim is that the LCS generator and its reparameterized and reparameterized-plus-quantized variants beat both EASF and FSR1 on NIQE and LPIPS over the GameIR validation set, with non-overlapping 68% confidence intervals on those two metrics. The reported point estimates are NIQE 3.43 for all LCS variants versus 5.55 for FSR1 and 5.20 for EASF, and LPIPS 0.152, 0.152, and 0.150 for the LCS variants versus 0.199 for FSR1 and 0.210 for EASF. At the same time, FSR1 and EASF score higher on PSNR and SSIM, though the paper notes those differences may not be statistically significant, and JOD scores show essentially no difference across all methods. The paper concludes that the LCS demonstrates the potential of efficient super-resolution models for upscaling on resource-constrained devices.

Load-bearing premise

The central claim rests on NIQE and LPIPS being trustworthy proxies for perceived image quality, so if those metrics do not track what viewers actually prefer, the advantage over FSR1 and EASF could vanish.

Editorial extensions

If this is right

  • Game upscaling could be moved off the GPU to an NPU or other low-power device using a model with roughly 0.21 million parameters and 175 GMACs.
  • Training on natively rendered game low/high-resolution pairs avoids the domain mismatch that bicubic-downsampled training data creates for game content.
  • Adversarial training is what separates the LCS from L1-only training, which the paper shows produces visibly blurrier upscaled images.
  • Reparameterization and INT8 quantization keep NIQE and LPIPS essentially unchanged, so the efficiency gains do not appear to cost perceptual quality on these metrics.
  • The paper identifies real-time video upscaling, temporal stability, and more diverse game content as necessary future evaluation steps.

Reading between the lines

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

  • Because the generator is trained with a VGG-based perceptual loss and LPIPS is also a VGG-based learned metric, part of the measured LPIPS advantage may reflect the model matching the metric's own feature space, so a human preference study would be the decisive test.
  • The reported advantage is measured on a 240-image validation subset from a single driving simulator, so the margin may shrink on other game genres and rendering styles.
  • An adversarial generator that sharpens static frames could introduce temporal flicker in video, a failure mode the static-image metrics in this paper would not expose.
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 / 5 minor

Summary. The paper proposes a low-complexity super-resolution model (LCS) based on DIPNet/RLFN, with reparameterized residual feature blocks, adversarial training, and INT8 quantization, trained on GameIR native LR-HR pairs. The authors compare LCS (including reparameterized and quantized variants) against AMD's FSR1 and EASF using PSNR, SSIM, NIQE, JOD, and LPIPS, and report 68% confidence intervals. They conclude that LCS achieves better perceptual quality on NIQE and LPIPS, and that this demonstrates the potential of ESR models for upscaling on resource-constrained devices.

Significance. If the central claim were fully supported, a 0.21M-parameter model with 175 GMACs that produces perceptually preferable game upscales compared to FSR1 and EASF would be a valuable contribution to power-efficient super-resolution. The paper provides concrete architectural details, training hyperparameters, and confidence intervals for the main metrics, which is commendable. However, the perceptual superiority claim currently rests on NIQE and LPIPS, both of which are aligned with the VGG-based training objective, while the human-calibrated JOD metric shows parity across all methods. The power-efficiency claim also lacks direct power or NPU measurements. The paper is a solid proof-of-concept for a narrow dataset, but its headline conclusions need strengthening or reframing.

major comments (3)
  1. [Section II.C / Section III.B (Table I)] The central claim that LCS 'achieves better perceptual quality' (abstract and Section IV) is not established by the reported evidence. The only metrics with non-overlapping 68% CIs are NIQE and LPIPS, but LPIPS is a VGG-based learned metric and the generator is trained with a VGG-based perceptual loss and a relativistic VGG-style discriminator; GAN-trained images also tend to lower the no-reference NIQE score. The JOD metric, which is derived from human discrimination data, shows no statistically significant difference across all methods (8.50–8.51, overlapping intervals). In Section III.B the authors state that NIQE and LPIPS are 'more aligned with our subjective evaluation,' but no subjective evaluation data are reported. A human preference study or a detailed justification of why JOD should be discounted is required before claiming perceptual superiority.
  2. [Abstract / Section II.D / Table I] The title and abstract emphasize 'power-efficient' super-resolution, but no power or energy measurements are reported. Runtime is measured only on an AMD Instinct MI210 GPU for the FP32 LCS and the reparameterized LCS, and the quantized model's runtime is not given. Since the stated motivation is offloading to a low-power NPU, the paper should either measure power/energy on a representative low-power device or reframe the claim as 'low-complexity' rather than 'power-efficient.'
  3. [Section II.A / Section III.B] All quantitative conclusions are drawn from a 240-image validation subset of a single driving simulator (CARLA rendered with Unreal Engine 4, via the GameIR dataset). The authors should explicitly state that the results are limited to this dataset and avoid the broader generalization implied by the abstract's 'demonstrating the potential of ESR models for upscaling.' A single simulator, even with 240 images, is too narrow to support claims about general game content without further evidence or discussion of representativeness.
minor comments (5)
  1. [Section III.B] The statement that 'the JOD scores shows little variance across the board, making it a less useful metric for our purposes' is not a sound interpretation: a metric that fails to differentiate methods may indicate that the methods are perceptually similar, not that the metric is uninformative. Please rephrase or provide evidence that JOD is insensitive in this application.
  2. [Table I] The 'Param.' row is implicitly in millions of parameters, but this is not stated; please label the row as 'Params [M]' for clarity.
  3. [Section II.D] The claim that reparameterization yields 'similar performance to the original model' should point to the quantitative comparison in Table I, where the LCS and LCS Rep. rows are identical across all metrics.
  4. [Section III.A / Figure 2] The qualitative description refers to a 'dark line below the taillights' in the LR image, but the LR image is not shown in Figure 2; please add a pointer to Figure 3 in Appendix A or include the LR panel.
  5. [Section II.C] The learning-rate schedule 'halved at iteration [50k, 100k, 200k, 300k]' would read more cleanly as 'halved at iterations 50k, 100k, 200k, and 300k.'

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the claimed perceptual advantage is evaluated with external metrics (LPIPS, NIQE), and no fitted parameter or self-citation chain is shown to force the result.

full rationale

The paper's comparison is an empirical evaluation rather than a formal derivation, and no step reduces by construction to its own inputs. Section II.C states that the generator is trained with an adversarial loss, an L1 loss, and a perceptual loss adopted from ESRGAN; Section III.B evaluates the model on PSNR, SSIM, NIQE, JOD, and LPIPS. LPIPS is an external learned perceptual metric and NIQE is a no-reference quality estimator; neither is identical to the training loss, and the validation images are separate from the training crops. The concern that a VGG-based perceptual loss and adversarial training may inflate LPIPS and NIQE is a legitimate external-validity threat, not a circularity, because no equation in the paper makes the evaluation metric equal to the optimized objective. The paper's own statement that 'JOD scores shows little variance across the board, making it a less useful metric for our purposes' (Section III.B) is an evidentiary weakness, especially since JOD is human-calibrated and shows parity, but it is a correctness-risk issue rather than a definitional identity. The architecture and dataset citations (RLFN, DIPNet, ESRGAN, GameIR) are external prior work, and no fitted parameter is renamed as a prediction. Therefore no circular step can be exhibited from the paper's equations or cited inputs.

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

The central claim is empirical, not derived. The hand-chosen training losses and architecture dimensions are free parameters that shape the result; the domain assumptions about the dataset subset, the validity of perceptual metrics, and the comparability of LR-HR pairs are load-bearing. No invented entities are introduced.

free parameters (4)
  • Adversarial loss weight lambda = 5e-3
    Hand-chosen in Section II.C; balances sharpness against reconstruction fidelity.
  • L1 loss weight eta = 1e-2
    Hand-chosen in Section II.C; controls pixel-level fidelity.
  • Network width and depth = 4 RRFBs, 38 channels, expansion factor 2
    Architecture hyperparameters selected to fit the low-complexity target; not derived.
  • Training schedule = 500k iterations, batch 4, crop 128, LR 1e-4 halved at 50k/100k/200k/300k
    Chosen training recipe in Section II.C; affects the final model.
assumptions (4)
  • domain assumption The 1,434-pair subset of GameIR is representative of the full dataset and of game content broadly.
    Section II.A selects a smaller subset without describing the selection mechanism; all training and validation use this subset.
  • domain assumption NIQE and LPIPS are valid arbiters of perceptual quality for this comparison.
    Section III.B uses these metrics to conclude LCS is better; JOD shows no difference, and no human study is reported.
  • domain assumption The natively rendered LR and HR pairs are sufficiently aligned for image-quality metrics.
    Section III.A notes LR and HR differ in fine details and color, which can distort pixel-level metrics like PSNR and SSIM.
  • domain assumption FSR1 and EASF are appropriate and correctly configured baselines.
    The paper compares against these two AMD upscalers without reporting their configuration or tuning; hardware EASF behavior may vary.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LCS: An AI-based Low-Complexity Scaler for Power-Efficient Super-Resolution of Game Content." pith.science (2026). https://pith.science/paper/X35RXI6S

@misc{pith2026250722873,
  author       = {Pith},
  title        = {Pith review of: LCS: An AI-based Low-Complexity Scaler for Power-Efficient Super-Resolution of Game Content},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/X35RXI6S}},
  note         = {Machine review of arXiv:2507.22873}
}
read the original abstract

The increasing complexity of content rendering in modern games has led to a problematic growth in the workload of the GPU. In this paper, we propose an AI-based low-complexity scaler (LCS) inspired by state-of-the-art efficient super-resolution (ESR) models which could offload the workload on the GPU to a low-power device such as a neural processing unit (NPU). The LCS is trained on GameIR image pairs natively rendered at low and high resolution. We utilize adversarial training to encourage reconstruction of perceptually important details, and apply reparameterization and quantization techniques to reduce model complexity and size. In our comparative analysis we evaluate the LCS alongside the publicly available AMD hardware-based Edge Adaptive Scaling Function (EASF) and AMD FidelityFX Super Resolution 1 (FSR1) on five different metrics, and find that the LCS achieves better perceptual quality, demonstrating the potential of ESR models for upscaling on resource-constrained devices.

Figures

Figures reproduced from arXiv: 2507.22873 by the authors.

Figure 1
Figure 1. Architecture of the low-complexity scaler and overview of reparameterization residual feature blocks (RRFB), residual in residual reparameterization [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Comparison of the LCS, reparameterized LCS, reparameterized and quantized LCS, FSR1, and EASF on a GameIR dynamic validation image. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Comparison of the HR ground truth, LR upscaled with a bicubic filter, LCS, and LCS trained with L1 loss only. The LCS trained with L1 loss [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

41 extracted references · 38 canonical work pages

  1. [1]

    Agustsson and R

    E. Agustsson and R. Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, July 2017

    work page 2017

  2. [2]

    L. Chen, X. Chu, X. Zhang, and J. Sun. Simple baselines for image restoration. In Computer Vision - ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 2- 27, 2022, Proceedings, Part VII, pages 17–33, Berlin, Heidelberg, 2022. Springer-Verlag

    work page 2022

  3. [3]

    Colbert, A

    I. Colbert, A. Pappalardo, and J. Petri-Koenig. A2q: Accumulator-aware quantization with guaranteed over- flow avoidance. In Proceedings of the IEEE/CVF Inter- national Conference on Computer Vision (ICCV), pages 16989–16998, October 2023

    work page 2023

  4. [4]

    X. Ding, X. Zhang, N. Ma, J. Han, G. Ding, and J. Sun. Repvgg: Making vgg-style convnets great again. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13728–13737, 2021

    work page 2021

  5. [5]

    Dosovitskiy, G

    A. Dosovitskiy, G. Ros, F. Codevilla, A. Lopez, and V . Koltun. CARLA: An open urban driving simulator. In S. Levine, V . Vanhoucke, and K. Goldberg, editors, Proceedings of the 1st Annual Conference on Robot Learning, volume 78 of Proceedings of Machine Learn- ing Research, pages 1–16. PMLR, 13–15 Nov 2017

    work page 2017

  6. [6]

    Z. Du, D. Liu, J. Liu, J. Tang, G. Wu, and L. Fu. Fast and memory-efficient network towards efficient image super- resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 853–862, June 2022

    work page 2022

  7. [7]

    Elsken, J

    T. Elsken, J. H. Metzen, and F. Hutter. Neural ar- chitecture search: a survey. J. Mach. Learn. Res., 20(1):1997–2017, Jan. 2019

    work page 1997

  8. [8]

    Franco, A

    G. Franco, A. Pappalardo, and N. J. Fraser. Xil- inx/brevitas, 2025

    work page 2025

Show all 41 references
  1. [9]

    Gholami, S

    A. Gholami, S. Kim, Z. Dong, Z. Yao, M. W. Mahoney, and K. Keutzer. A Survey of Quantization Methods for Efficient Neural Network Inference. arXiv e-prints, page arXiv:2103.13630, Mar. 2021

    2021 arXiv

  2. [10]

    I. J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y . Ben- gio. Generative adversarial nets. In Z. Ghahramani, M. Welling, C. Cortes, N. Lawrence, and K. Weinberger, editors, Advances in Neural Information Processing Sys- tems, v...

    2014

  3. [11]

    S. Han, J. Pool, J. Tran, and W. Dally. Learning both weights and connections for efficient neural network. In C. Cortes, N. Lawrence, D. Lee, M. Sugiyama, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 28. Curran Associates, Inc., 2015

    2015

  4. [12]

    Hinton, O

    G. Hinton, O. Vinyals, and J. Dean. Distilling the knowledge in a neural network. In NIPS Deep Learning and Representation Learning Workshop, 2015

    2015

  5. [13]

    Isola, J.-Y

    P. Isola, J.-Y . Zhu, T. Zhou, and A. A. Efros. Image-to- image translation with conditional adversarial networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017

    2017

  6. [14]

    Johnson, A

    J. Johnson, A. Alahi, and L. Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In B. Leibe, J. Matas, N. Sebe, and M. Welling, editors, Computer Vision – ECCV 2016, pages 694–711, Cham, 2016. Springer International Publishing

    2016

  7. [15]

    Jolicoeur-Martineau

    A. Jolicoeur-Martineau. The relativistic discriminator: a key element missing from standard GAN. arXiv e-prints, page arXiv:1807.00734, July 2018

    2018 arXiv

  8. [16]

    D. P. Kingma and J. Ba. Adam: A method for stochastic optimization. In Y . Bengio and Y . LeCun, editors, 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015

    2015

  9. [17]

    F. Kong, M. Li, S. Liu, D. Liu, J. He, Y . Bai, F. Chen, and L. Fu. Residual local feature network for efficient super- resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 766–776, June 2022

    2022

  10. [18]

    LeCun, J

    Y . LeCun, J. Denker, and S. Solla. Optimal brain damage. In D. Touretzky, editor, Advances in Neural Information Processing Systems, volume 2. Morgan-Kaufmann, 1989

    1989

  11. [19]

    H. Li, A. Kadav, I. Durdanovic, H. Samet, and H. P. Graf. Pruning filters for efficient convnets. In International Conference on Learning Representations, 2017

    2017

  12. [20]

    Y . Li, K. Zhang, J. Liang, J. Cao, C. Liu, R. Gong, Y . Zhang, H. Tang, Y . Liu, D. Demandolx, R. Ranjan, R. Timofte, and L. Van Gool. Lsdir: A large scale dataset for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVP...

    2023

  13. [21]

    Y . Li, K. Zhang, R. Timofte, L. Van Gool, F. Kong, M. Li, S. Liu, Z. Du, D. Liu, C. Zhou, J. Chen, Q. Han, Z. Li, Y . Liu, X. Chen, H. Cai, Y . Qiao, C. Dong, L. Sun, J. Pan, Y . Zhu, Z. Zong, X. Liu, Z. Hui, T. Yang, P. Ren, X. Xie, X.-S. Hua, Y . Wang, X. Ji, C. Lin, D. Luo...

    2022

  14. [22]

    Y . Li, Y . Zhang, R. Timofte, L. Van Gool, L. Yu, Y . Li, X. Li, T. Jiang, Q. Wu, M. Han, W. Lin, C. Jiang, J. Luo, H. Fan, S. Liu, Y . Wang, M. Cai, M. Li, Y . Zhang, X.-j. Fan, Y . Sheng, Y . Mao, N. Zhang, Q. Wang, M. Zheng, L. Sun, J. Pan, J. Dong, J. Tang, Z. Yang, Y . W...

    2023

  15. [23]

    J. Liu, J. Tang, and G. Wu. Residual feature distilla- tion network for lightweight image super-resolution. In A. Bartoli and A. Fusiello, editors, Computer Vision – ECCV 2020 Workshops, pages 41–55, Cham, 2020. Springer International Publishing

    2020

  16. [24]

    J. Liu, W. Zhang, Y . Tang, J. Tang, and G. Wu. Residual feature aggregation network for image super-resolution. In IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (CVPR), June 2020

    2020

  17. [25]

    Z. Liu, M. Sun, T. Zhou, G. Huang, and T. Darrell. Rethinking the value of network pruning. In International Conference on Learning Representations, 2019

    2019

  18. [26]

    R. K. Mantiuk, P. Hanji, M. Ashraf, Y . Asano, and A. Chapiro. Colorvideovdp: A visual difference predictor for image, video and display distortions. ACM Trans. Graph., 43(4), July 2024

    2024

  19. [27]

    completely blind

    A. Mittal, R. Soundararajan, and A. C. Bovik. Making a “completely blind” image quality analyzer. IEEE Signal Processing Letters, 20(3):209–212, 2013

    2013

  20. [28]

    Pathak, P

    D. Pathak, P. Krahenbuhl, J. Donahue, T. Darrell, and A. A. Efros. Context encoders: Feature learning by inpainting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2016

    2016

  21. [29]

    B. Ren, H. Guo, L. Sun, Z. Wu, R. Timofte, and Y . Li. The tenth ntire 2025 efficient super-resolution challenge report. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR) Workshops, pages 917–966, June 2025

    2025

  22. [30]

    B. Ren, Y . Li, N. Mehta, R. Timofte, H. Yu, C. Wan, Y . Hong, B. Han, Z. Wu, Y . Zou, Y . Liu, J. Li, K. He, C. Fan, H. Zhang, X. Zhang, X. Yin, K. Zuo, B. Liao, P. Xia, L. Peng, Z. Du, X. Di, W. Li, Y . Wang, W. Zhai, R. Pei, J. Guo, S. Xu, Y . Cao, Z. Zha, Y . Wang, Y . Liu...

    2024

  23. [31]

    Romero, N

    A. Romero, N. Ballas, S. E. Kahou, A. Chassang, C. Gatta, and Y . Bengio. Fitnets: Hints for thin deep nets. In 3rd International Conference on Learning Rep- resentations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015

    2015

  24. [32]

    W. Shi, J. Caballero, F. Husz ´ar, J. Totz, A. P. Aitken, R. Bishop, D. Rueckert, and Z. Wang. 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 (CVPR...

    2016

  25. [33]

    Simonyan and A

    K. Simonyan and A. Zisserman. Very deep convolutional networks for large-scale image recognition. In Y . Bengio and Y . LeCun, editors, 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015

    2015

  26. [34]

    Timofte, E

    R. Timofte, E. Agustsson, L. Van Gool, M.-H. Yang, L. Zhang, B. Lim, et al. Ntire 2017 challenge on single image super-resolution: Methods and results. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, July 2017

    2017

  27. [35]

    X. Wang, K. Yu, S. Wu, J. Gu, Y . Liu, C. Dong, Y . Qiao, and C. Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In Proceedings of the European Conference on Computer Vision (ECCV) Workshops, September 2018

    2018

  28. [36]

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

    2004

  29. [37]

    White, M

    C. White, M. Safari, R. Sukthanker, B. Ru, T. Elsken, A. Zela, D. Dey, and F. Hutter. Neural Architecture Search: Insights from 1000 Papers. arXiv e-prints, page arXiv:2301.08727, Jan. 2023

    2023 arXiv

  30. [38]

    L. Yu, X. Li, Y . Li, T. Jiang, Q. Wu, H. Fan, and S. Liu. Dipnet: Efficiency distillation and iterative pruning for image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 1692–1701, June 2023

    2023

  31. [39]

    Zamfir, M

    E. Zamfir, M. V . Conde, and R. Timofte. Towards real- time 4k image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 1522–1532, June 2023

    2023

  32. [40]

    Zhang, P

    R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang. The unreasonable effectiveness of deep fea- tures as a perceptual metric. In CVPR, 2018

    2018

  33. [41]

    L. Zhou, K. Han, N. Ling, W. Wang, and W. Jiang. Gameir: A large-scale synthesized ground-truth dataset for image restoration over gaming content. In M. Cho, I. Laptev, D. Tran, A. Yao, and H.-B. Zha, editors, Computer Vision – ACCV 2024 Workshops, pages 179– 196, Singapore, 2...

    2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.