Pith. sign in

REVIEW 2 major objections 4 minor 66 references

Learning Pixel-adaptive Multi-layer Perceptrons for Real-time Image Enhancement

T0 review · 2 major / 4 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 The paper claims that storing a distinct tiny 3-8-3 MLP per pixel inside a bilateral grid yields state-of-the-art real-time image enhancement, outperforming 3D LUT and prior bilateral-grid methods on FiveK, PPR10K, and LCDP while…

desk verdict Legitimate new architecture with supporting ablations, but the state-of-the-art claim rests on thin margins and the grid depth and loss weights were tuned on the FiveK test set. read the letter →

arxiv 2507.12135 v1 pith:QBB6XHCH submitted 2025-07-16 cs.CV

Junyu Lou , Xiaorui Zhao , Kexuan Shi , Shuhang Gu This is my paper · ORCID
classification cs.CV
keywords imageenhancementbilateralgridpixel-adaptiveMLPreal-timeinferencetonemappingphotoretouchingexposurecorrection3Dlookuptables
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 the two dominant fast-enhancement paradigms, 3D lookup tables and bilateral grids with affine transformations, can be fused and outperformed by making a small multilayer perceptron pixel-adaptive. Its claim is that a bilateral grid can store, not affine coefficients, but the weights and biases of a tiny 3-8-3 MLP, so that every pixel retrieves its own unique nonlinear color mapping through the usual slicing operation. To make this work with RGB images, the paper groups MLP parameters into categories stored in separate subgrids, each steered by its own guidance-map channel. On three public benchmarks the method reports the best PSNR and SSIM among real-time methods, up to 0.55 dB over the strongest LUT-with-bilateral-grid rival on full-resolution FiveK tone mapping, while running 4K images in 27.8 ms. If the claim holds, real-time enhancement need not choose between nonlinear expressiveness and spatial adaptivity.

What carries the argument

The load-bearing object is the bilateral grid re-purposed as a parameter store: instead of affine matrices, each grid cell holds the 32 numbers (24 weights and 8 biases) for the first layer and the 27 numbers (24 weights and 3 biases) for the output layer of a 3-8-3 MLP with ReLU activation. Two slicing operations retrieve these per-pixel parameters through trilinear interpolation over the grid's eight neighboring cells, with the first grid's hidden vector feeding a second guidance network that drives the second slicing pass. The grid-decomposition strategy partitions the first grid into four subgrids (one per input color channel, sharing one bias group) and the second into nine subgrids (one per hidden neuron, sharing one bias group), so each guidance-map channel steers its own subgrid. A three-layer U-Net-style NAFNet backbone plus pixel unshuffle produces the grids, and the whole chain is differentiable so the grids learn what the per-pixel MLPs should be.

What would settle it

Train BPAM on the FiveK full-resolution tone-mapping task with grid depth raised from 8 to 64 and spatial downsampling removed, holding everything else fixed; if PSNR does not improve over the reported 25.12 dB, the claim that an 8-level, heavily downsampled grid faithfully represents per-pixel MLP parameters loses its load-bearing status.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that the bilateral grid, long used to store per-pixel affine coefficients for fast enhancement, can instead store the parameters of a small multilayer perceptron so that every pixel effectively runs its own 3-8-3 network with weights drawn from the grid at its spatial and intensity coordinates. The paper argues that affine mappings, not the grid itself, were the limiting factor in prior bilateral-grid methods, and that globally shared MLP weights, not MLPs themselves, limited earlier learning-based color transforms. Its grid-decomposition strategy splits the MLP's weights into per-channel groups held in separate subgrids, with a multi-channel guidance map assigning each group its own slicing coordinate, which the paper says uses the RGB channels fully during retrieval and sharpens parameter estimation. Measured this way, the method reports the best numbers among real-time approaches on FiveK, PPR10K, and LCDP, including the largest margin of 0.55 dB PSNR over the leading LUT-with-bilateral-grid method on full-resolution FiveK tone mapping and a 4K runtime of 27.8 ms.

Load-bearing premise

The method's gains rest on the premise that a bilateral grid with only 8 intensity levels and spatial downsampling factors of 4 to 32 can store per-pixel MLP parameters such that trilinear interpolation of those parameters yields accurate, artifact-free color mappings for every pixel.

Editorial extensions

If this is right

  • Real-time image enhancement can combine spatial adaptivity and nonlinear color mapping in a single pass, without the resolution-dependent quality loss the paper attributes to fixed-size 3D LUTs.
  • The same per-pixel MLP retrieval recipe generalizes across three distinct tasks, tone mapping, photo retouching, and exposure correction, including the extreme illumination regime of LCDP where LUT-based methods degrade.
  • Because the slicing operation is differentiable, the parameter-generation pipeline trains end-to-end, so the grids learn task-specific parameter layouts rather than relying on hand-designed coefficients.
  • On 4K inputs the method sustains 36 FPS, which the paper argues places it in the real-time regime competitive with 3D LUT approaches while delivering higher quality on high-resolution images.

Reading between the lines

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

  • The grid-decomposition idea is separable from MLPs: the same subgrid-per-parameter-category trick could sharpen other interpolated operators such as 3D LUT weights or polynomial coefficients by giving each coefficient type its own guidance channel, a transfer the paper does not discuss.
  • The second guidance network consumes the hidden-layer vector, coupling the two slicing stages, which suggests an iterative or recurrent extension where later stages refine parameters using features of the intermediate transform, something the current single-pass design does not explore.
  • A testable consequence of the representation premise is that if the 8-level intensity grid is genuinely faithful, images with very steep local tonal transitions such as hard shadows and specular highlights should show no banding or blocking at grid boundaries; a stress dataset of such images would probe the limit of that assumption.
  • The reported 0.55 dB gain over the strongest LUT rival on full-resolution FiveK suggests the advantage may grow with input resolution, since LUT methods use fixed-size transforms while BPAM's grid scales with the image; plotting PSNR against resolution would test whether the advantage is resolution-driven.
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

2 major / 4 minor

Summary. This paper proposes BPAM, a real-time image enhancement architecture that stores per-pixel MLP parameters in bilateral grids. A low-resolution NAFNet backbone generates two bilateral grids containing the weights and biases of a 3-8-3 MLP; a slicing operation with learned multi-channel guidance maps retrieves per-pixel parameters, and a grid-decomposition strategy places different parameter categories in separate subgrids. The method is evaluated on MIT-Adobe FiveK (tone mapping and photo retouching), PPR10K, and LCDP, reporting state-of-the-art PSNR/SSIM values while processing 4K images at about 36 FPS. Ablations in Table 6 and the supplementary material support the two proposed components.

Significance. If the empirical claims hold, this is a useful architectural contribution: it extends bilateral-grid processing from affine transformations to nonlinear per-pixel MLP mappings while retaining real-time performance, and the grid-decomposition scheme offers a clean way to exploit multi-channel guidance information. The method is clearly specified, ablations support both proposed components, and the authors provide a code repository. The main weakness is the strength of the evidence for the headline comparison: margins over the strongest baseline are small on several tasks, and the manuscript does not demonstrate that these margins are robust to hyperparameter selection or training stochasticity.

major comments (2)
  1. [Sections 4.1 and Supp. C; Eq. (9)] The hyperparameter-selection protocol is unclear. Section 4.1 describes only a 4500/500 train/test split for FiveK, with no validation split, and Supp. Section C states that the grid depth was chosen by running experiments "on the FiveK dataset to determine the optimal settings." The loss weights in Eq. (9) are likewise described as "empirically set." If these choices were made using the 500 test images, the reported test metrics are optimistically biased. This is load-bearing because the PSNR advantage over LUTBGrid is 0.07 dB in Table 2, 0.09-0.13 dB in Table 3, and 0.24 dB in Table 1 at 480p; only the full-resolution tone-mapping gain of 0.55 dB is large. Please specify the validation protocol used for all hyperparameter choices, or, if none was used, re-select hyperparameters on a held-out split and report test metrics with frozen hyperparameters.
  2. [Tables 1-4, 6; Supp. Tables 7-8] The paper reports only single-run point estimates, with no error bars, standard deviations, or paired significance tests. Given margins as small as 0.07 dB on FiveK photo retouching and 0.09-0.13 dB on PPR10K, the state-of-the-art claim rests on differences that could easily lie within run-to-run noise. Please report mean and standard deviation over at least three seeds for the main comparisons, or paired-difference statistics against the strongest baseline, and do the same for the ablation results in Table 6 and Supp. Table 8.
minor comments (4)
  1. [Section 4.2, Table 3] The PPR10K model configuration is not fully reported. Section 4.2 says the model's parameter count was increased to a level comparable to the 11.7M ResNet-18 used by LUT baselines, but Table 3 has no parameter column and the other tables report 624K parameters. Please report the exact architecture and parameter count used for each dataset.
  2. [Table 6] The "Affine Trans" ablation baseline is under-specified. Please state whether it uses the same two-grid design, the same guidance-map architecture, and the same parameter count as the MLP setting, so that the gain in Table 6 can be attributed to the nonlinear transformation rather than to other architectural differences.
  3. [Supp. Section C, Table 8] The discussion of MLP depth is internally inconsistent: the text first says increasing the number of intermediate layers degrades performance, then says increasing depth leads to improvement, while Table 8 shows 3-8-8-3 improving PSNR by 0.04 dB over 3-8-3. Please clarify whether "depth" and "number of intermediate layers" refer to the same quantity and revise the explanation accordingly.
  4. [Eq. (9)] Please define L_ssim precisely (presumably 1 - SSIM) and specify which VGG19 layers are used to compute L_per, since these choices affect reproducibility and the reported ablation results.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the architecture's output is not defined in terms of its reported metrics, and no load-bearing self-citation or construction-level reduction is present.

full rationale

This is an empirical architecture paper rather than a formal derivation. The pipeline in Eqs. (7)-(9) computes each output color from the input color through per-pixel MLP weights obtained by trilinear slicing of predicted bilateral grids; no reported quantity (PSNR, SSIM, DeltaE, LPIPS) is defined in terms of those weights, and no fitted constant is renamed as a prediction. The only self-citation is Ref. [39] in Related Work, which descriptively mentions a Fourier reparameterization method and is not load-bearing for any claim. Comparisons are made against external baselines on three public benchmarks. One evaluation-hygiene concern is that Supp. Sec. C selects the bilateral-grid depth by running experiments on the FiveK dataset, while no separate validation split is described in Sec. 4.1, and the loss weights in Eq. (9) are described as 'empirically set'; if those choices were made on the FiveK test split, the reported margins could be optimistically biased. However, this is a test-set overfitting and reproducibility concern, not a self-referential derivation: the model output is not constructed to match the benchmark metric, and the architecture does not reduce to its inputs by construction. Therefore no circular step is established.

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

The central claim rests on several standard assumptions inherited from bilateral grid and MLP literature, plus a set of hyperparameters (grid depth, downsampling, loss weights, MLP size) that are tuned on the evaluation datasets. There are no newly posited physical entities. The grid decomposition is an invented representational device with support only from the paper's own ablations.

free parameters (4)
  • Per-pixel MLP width and depth = 3-8-3
    Chosen manually; the ablation in Supp. Table 8 shows 3-16-3 and 3-8-8-3 give similar or slightly worse performance, so this specific choice is not critical to the central claim.
  • Grid depth (number of intensity levels) = 8
    Selected by ablation on the FiveK tone mapping task (Supp. Figure 8). This hyperparameter balances capacity and trainability.
  • Grid spatial downsampling factors = 1/4 (PPR10K), 1/8 (default), 1/32 (full-res FiveK)
    Set per dataset in Sec. 4.2 and Supp. B; a design choice not derived from first principles.
  • Loss weighting coefficients = 1.0 for L2, 0.5 for SSIM, 0.005 for VGG perceptual
    Empirically set in Eq. (9); typical values for image enhancement, but not independently justified.
assumptions (4)
  • domain assumption Trilinear interpolation of grid coefficients is differentiable and preserves the edge-aware behavior of the bilateral grid.
    Assumed from prior bilateral grid works [3, 9] and used throughout Sec. 3.1; no proof is given in this paper.
  • domain assumption MLP parameters can be meaningfully interpolated in the spatial-plus-intensity grid without significant loss of mapping quality.
    Core modeling premise of Sec. 3.2: per-pixel MLP weights are read out by trilinear interpolation; if this interpolation were lossy, the whole framework would degrade.
  • domain assumption The combined MSE, SSIM and VGG perceptual loss with fixed weights is a valid proxy for perceptual quality in enhancement.
    Used in Eq. (9); the weights are not derived from a perceptual model but are standard practice.
  • domain assumption The public benchmark splits and the reported baselines are representative and not cherry-picked.
    The paper follows standard splits for FiveK, PPR10K and LCDP; however, independent verification of the comparisons is not possible from the text alone.
invented entities (1)
  • Category-specific subgrids (grid decomposition)
    purpose: Store weights and biases of different types in separate grids so that multi-channel guidance maps can retrieve each parameter class separately.
    This is an architectural construct validated only by internal ablations (Tables 6 and 7). It does not make a falsifiable prediction outside the paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning Pixel-adaptive Multi-layer Perceptrons for Real-time Image Enhancement." pith.science (2026). https://pith.science/paper/QBB6XHCH

@misc{pith2026250712135,
  author       = {Pith},
  title        = {Pith review of: Learning Pixel-adaptive Multi-layer Perceptrons for Real-time Image Enhancement},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QBB6XHCH}},
  note         = {Machine review of arXiv:2507.12135}
}
read the original abstract

Deep learning-based bilateral grid processing has emerged as a promising solution for image enhancement, inherently encoding spatial and intensity information while enabling efficient full-resolution processing through slicing operations. However, existing approaches are limited to linear affine transformations, hindering their ability to model complex color relationships. Meanwhile, while multi-layer perceptrons (MLPs) excel at non-linear mappings, traditional MLP-based methods employ globally shared parameters, which is hard to deal with localized variations. To overcome these dual challenges, we propose a Bilateral Grid-based Pixel-Adaptive Multi-layer Perceptron (BPAM) framework. Our approach synergizes the spatial modeling of bilateral grids with the non-linear capabilities of MLPs. Specifically, we generate bilateral grids containing MLP parameters, where each pixel dynamically retrieves its unique transformation parameters and obtain a distinct MLP for color mapping based on spatial coordinates and intensity values. In addition, we propose a novel grid decomposition strategy that categorizes MLP parameters into distinct types stored in separate subgrids. Multi-channel guidance maps are used to extract category-specific parameters from corresponding subgrids, ensuring effective utilization of color information during slicing while guiding precise parameter generation. Extensive experiments on public datasets demonstrate that our method outperforms state-of-the-art methods in performance while maintaining real-time processing capabilities.

Figures

Figures reproduced from arXiv: 2507.12135 by the authors.

Figure 1
Figure 1. Pixel-adaptive MLP. S represents the intensity value lo￾cated at (x, y) in corresponding guidance map. The red lines in the MLP represent the contributions of the input layer’s R channel to the hidden layer, while the purple lines indicate the contributions of the first neuron in the hidden layer to the output layer. Each pixel on the image can obtain its unique MLP parameters based on its coordinates and pixel valu… view at source ↗
Figure 2
Figure 2. Illustration of proposed bilateral grid-based pixel-adaptive multi-layer perceptron (BPAM) framework. It comprises three main [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Illustration of grid decomposition for the first grid, which [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (13 more)
Figure 4
Figure 4. Figure 4: Visual comparison with state-of-the-art methods on the [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Visual comparison with state-of-the-art methods on the [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Visual comparison with state-of-the-art methods on the [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: NAF-block Our proposed method utilize a three-layer U-net-style NAFNet [5] as our backbone to gener￾ate the bilateral grids, with each layer in both the encoder and decoder comprising two NAF￾blocks (see [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: Ablation study on different numbers of bilateral grid [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]
Figure 9
Figure 9. Figure 9: Visual comparison with state-of-the-art methods on the LCDP dataset for exposure correction. [PITH_FULL_IMAGE:figures/full_fig_p014_9.png]
Figure 10
Figure 10. Figure 10: Visual comparison with state-of-the-art methods on the LCDP dataset for exposure correction. [PITH_FULL_IMAGE:figures/full_fig_p014_10.png]
Figure 11
Figure 11. Figure 11: Visual comparison with state-of-the-art methods on the LCDP dataset for exposure correction. [PITH_FULL_IMAGE:figures/full_fig_p015_11.png]
Figure 12
Figure 12. Figure 12: Visual comparison with state-of-the-art methods on the LCDP dataset for exposure correction. [PITH_FULL_IMAGE:figures/full_fig_p015_12.png]
Figure 13
Figure 13. Figure 13: Visual comparison with state-of-the-art methods on the FiveK dataset (4K) for tone mapping. [PITH_FULL_IMAGE:figures/full_fig_p016_13.png]
Figure 14
Figure 14. Figure 14: Visual comparison with state-of-the-art methods on the FiveK dataset (4K) for tone mapping. [PITH_FULL_IMAGE:figures/full_fig_p016_14.png]
Figure 15
Figure 15. Figure 15: Visual comparison with state-of-the-art methods on the FiveK dataset (4K) for tone mapping. [PITH_FULL_IMAGE:figures/full_fig_p017_15.png]
Figure 16
Figure 16. Figure 16: Visual comparison with state-of-the-art methods on the FiveK dataset (4K) for tone mapping. [PITH_FULL_IMAGE:figures/full_fig_p017_16.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

66 extracted references · 58 canonical work pages

  1. [1]

    Learning photographic global tonal adjustment with a database of input/output image pairs

    Vladimir Bychkovsky, Sylvain Paris, Eric Chan, and Fr ´edo Durand. Learning photographic global tonal adjustment with a database of input/output image pairs. InCVPR 2011, pages 97–104. IEEE, 2011. 5, 1

    work page 2011

  2. [2]

    Learning to see in the dark

    Chen Chen, Qifeng Chen, Jia Xu, and Vladlen Koltun. Learning to see in the dark. In Proceedings of the IEEE con- ference on computer vision and pattern recognition , pages 3291–3300, 2018. 1, 2, 7

    work page 2018

  3. [3]

    Real-time edge-aware image processing with the bilateral grid

    Jiawen Chen, Sylvain Paris, and Fr ´edo Durand. Real-time edge-aware image processing with the bilateral grid. ACM Transactions on Graphics (TOG), 26(3):103–es, 2007. 2, 3

    work page 2007

  4. [4]

    Bilateral guided upsampling

    Jiawen Chen, Andrew Adams, Neal Wadhwa, and Samuel W Hasinoff. Bilateral guided upsampling. ACM Transactions on Graphics (TOG), 35(6):1–8, 2016. 2, 3, 5

    work page 2016

  5. [5]

    Simple baselines for image restoration

    Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In European confer- ence on computer vision , pages 17–33. Springer, 2022. 6, 1

    work page 2022

  6. [6]

    Cyclemlp: A mlp-like architecture for dense visual predictions

    Shoufa Chen, Enze Xie, Chongjian Ge, Runjian Chen, Ding Liang, and Ping Luo. Cyclemlp: A mlp-like architecture for dense visual predictions. IEEE Transactions on Pattern Analysis and Machine Intelligence , 45(12):14284–14300,

    work page

  7. [7]

    Deep photo enhancer: Unpaired learning for im- age enhancement from photographs with gans

    Yu-Sheng Chen, Yu-Ching Wang, Man-Hsin Kao, and Yung- Yu Chuang. Deep photo enhancer: Unpaired learning for im- age enhancement from photographs with gans. In Proceed- ings of the IEEE conference on computer vision and pattern recognition, pages 6306–6314, 2018. 1, 2

    work page 2018

  8. [8]

    Repmlp: Re- parameterizing convolutions into fully-connected layers for image recognition

    Xiaohan Ding, Chunlong Xia, Xiangyu Zhang, Xiaojie Chu, Jungong Han, and Guiguang Ding. Repmlp: Re- parameterizing convolutions into fully-connected layers for image recognition. arXiv preprint arXiv:2105.01883, 2021. 3

    arXiv 2021

Show all 66 references
  1. [9]

    Deep bilateral learning for real- time image enhancement

    Micha ¨el Gharbi, Jiawen Chen, Jonathan T Barron, Samuel W Hasinoff, and Fr´edo Durand. Deep bilateral learning for real- time image enhancement. ACM Transactions on Graphics (TOG), 36(4):1–12, 2017. 2, 3, 4, 5, 6, 8, 1

    2017

  2. [10]

    Zero-reference deep curve estimation for low-light image enhancement

    Chunle Guo, Chongyi Li, Jichang Guo, Chen Change Loy, Junhui Hou, Sam Kwong, and Runmin Cong. Zero-reference deep curve estimation for low-light image enhancement. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 1780–1789, 2020. 3, 7

    2020

  3. [11]

    Hire-mlp: Vi- sion mlp via hierarchical rearrangement

    Jianyuan Guo, Yehui Tang, Kai Han, Xinghao Chen, Han Wu, Chao Xu, Chang Xu, and Yunhe Wang. Hire-mlp: Vi- sion mlp via hierarchical rearrangement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 826–836, 2022. 3

    2022

  4. [12]

    An fpga-based fully pipelined bilateral grid for real-time image denoising

    Nobuho Hashimoto and Shinya Takamaeda-Yamazaki. An fpga-based fully pipelined bilateral grid for real-time image denoising. In 2021 31st International Conference on Field- Programmable Logic and Applications (FPL) , pages 167–

    2021

  5. [13]

    Con- ditional sequential modulation for efficient global image re- touching

    Jingwen He, Yihao Liu, Yu Qiao, and Chao Dong. Con- ditional sequential modulation for efficient global image re- touching. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceed- ings, Part XIII 16 , pages 679–695. Springer, 2020. 2, 3, 6, 8

    2020

  6. [14]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceed- ings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016. 6, 1

    2016

  7. [15]

    Vision permutator: A per- mutable mlp-like architecture for visual recognition

    Qibin Hou, Zihang Jiang, Li Yuan, Ming-Ming Cheng, Shuicheng Yan, and Jiashi Feng. Vision permutator: A per- mutable mlp-like architecture for visual recognition. IEEE transactions on pattern analysis and machine intelligence , 45(1):1328–1334, 2022. 3

    2022

  8. [16]

    Exposure normalization and compensation for multiple-exposure correction

    Jie Huang, Yajing Liu, Xueyang Fu, Man Zhou, Yang Wang, Feng Zhao, and Zhiwei Xiong. Exposure normalization and compensation for multiple-exposure correction. In Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6043–6052, 2022. 7

    2022

  9. [17]

    Deep fourier-based exposure correction network with spatial- frequency interaction

    Jie Huang, Yajing Liu, Feng Zhao, Keyu Yan, Jinghao Zhang, Yukun Huang, Man Zhou, and Zhiwei Xiong. Deep fourier-based exposure correction network with spatial- frequency interaction. In European Conference on Computer Vision, pages 163–180. Springer, 2022. 7

    2022

  10. [18]

    Context-based automatic local image enhancement

    Sung Ju Hwang, Ashish Kapoor, and Sing Bing Kang. Context-based automatic local image enhancement. In Com- puter Vision–ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Pro- ceedings, Part I 12, pages 569–582. Springer, 2012. 1

    2012

  11. [19]

    Dslr-quality photos on mobile devices with deep convolutional networks

    Andrey Ignatov, Nikolay Kobyshev, Radu Timofte, Kenneth Vanhoey, and Luc Van Gool. Dslr-quality photos on mobile devices with deep convolutional networks. InProceedings of the IEEE international conference on computer vision, pages 3277–3285, 2017. 1, 2

    2017

  12. [20]

    Enlightengan: Deep light enhancement without paired supervision

    Yifan Jiang, Xinyu Gong, Ding Liu, Yu Cheng, Chen Fang, Xiaohui Shen, Jianchao Yang, Pan Zhou, and Zhangyang Wang. Enlightengan: Deep light enhancement without paired supervision. IEEE transactions on image processing, 30:2340–2349, 2021. 1, 2

    2021

  13. [21]

    Per- sonalization of image enhancement

    Sing Bing Kang, Ashish Kapoor, and Dani Lischinski. Per- sonalization of image enhancement. In2010 IEEE Computer Society Conference on Computer Vision and Pattern Recog- nition, pages 1799–1806. IEEE, 2010. 1

    2010

  14. [22]

    Global and local enhancement networks for paired and unpaired im- age enhancement

    Han-Ul Kim, Young Jun Koh, and Chang-Su Kim. Global and local enhancement networks for paired and unpaired im- age enhancement. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXV 16, pages 339–354. Springer, 2020. 3

    2020

  15. [23]

    Jsi-gan: Gan- based joint super-resolution and inverse tone-mapping with pixel-wise task-specific filters for uhd hdr video

    Soo Ye Kim, Jihyong Oh, and Munchurl Kim. Jsi-gan: Gan- based joint super-resolution and inverse tone-mapping with pixel-wise task-specific filters for uhd hdr video. InProceed- ings of the AAAI Conference on Artificial Intelligence, pages 11287–11295, 2020. 1, 2

    2020

  16. [24]

    Image-adaptive 3d lookup tables for real-time image enhancement with bilateral grids

    Wontae Kim and Nam Ik Cho. Image-adaptive 3d lookup tables for real-time image enhancement with bilateral grids. In European Conference on Computer Vision, pages 91–108. Springer, 2024. 2, 3, 6, 7, 8, 1

    2024

  17. [25]

    Adam: A method for stochastic optimization

    Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 ,

    arXiv

  18. [26]

    Automatic content-aware color and tone stylization

    Joon-Young Lee, Kalyan Sunkavalli, Zhe Lin, Xiaohui Shen, and In So Kweon. Automatic content-aware color and tone stylization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2470–2478,

  19. [27]

    Flexible piecewise curves estimation for photo enhancement

    Chongyi Li, Chunle Guo, Qiming Ai, Shangchen Zhou, and Chen Change Loy. Flexible piecewise curves estimation for photo enhancement. arXiv preprint arXiv:2010.13412, 2020. 3

    2010 arXiv

  20. [28]

    Real-time expo- sure correction via collaborative transformations and adap- tive sampling

    Ziwen Li, Feng Zhang, Meng Cao, Jinpu Zhang, Yuanjie Shao, Yuehuan Wang, and Nong Sang. Real-time expo- sure correction via collaborative transformations and adap- tive sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 2984–...

    2024

  21. [29]

    Ppr10k: A large-scale portrait photo retouch- ing dataset with human-region mask and group-level consis- tency

    Jie Liang, Hui Zeng, Miaomiao Cui, Xuansong Xie, and Lei Zhang. Ppr10k: A large-scale portrait photo retouch- ing dataset with human-region mask and group-level consis- tency. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, pages 653–661...

    2021

  22. [30]

    Iterative prompt learning for unsupervised backlit image enhancement

    Zhexin Liang, Chongyi Li, Shangchen Zhou, Ruicheng Feng, and Chen Change Loy. Iterative prompt learning for unsupervised backlit image enhancement. In Proceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 8094–8103, 2023. 7

    2023

  23. [31]

    4d lut: learnable context-aware 4d lookup table for image enhancement

    Chengxu Liu, Huan Yang, Jianlong Fu, and Xueming Qian. 4d lut: learnable context-aware 4d lookup table for image enhancement. IEEE Transactions on Image Processing, 32: 4742–4756, 2023. 2, 3

    2023

  24. [32]

    Pay attention to mlps

    Hanxiao Liu, Zihang Dai, David So, and Quoc V Le. Pay attention to mlps. Advances in neural information processing systems, 34:9204–9215, 2021. 3

    2021

  25. [33]

    Pixel-learnable 3dlut with saturation-aware compensation for image enhancement

    Jing Liu, Qingying Li, Xiongkuo Min, Yuting Su, Guang- tao Zhai, and Xiaokang Yang. Pixel-learnable 3dlut with saturation-aware compensation for image enhancement. IEEE Transactions on Multimedia, 2024. 2, 3, 6

    2024

  26. [34]

    Autostyle: Automatic style transfer from im- age collections to users’ images

    Yiming Liu, Michael Cohen, Matt Uyttendaele, and Szymon Rusinkiewicz. Autostyle: Automatic style transfer from im- age collections to users’ images. In Computer Graphics Fo- rum, pages 21–31. Wiley Online Library, 2014. 1

    2014

  27. [35]

    Deeplpf: Deep local para- metric filters for image enhancement

    Sean Moran, Pierre Marza, Steven McDonagh, Sarah Parisot, and Gregory Slabaugh. Deeplpf: Deep local para- metric filters for image enhancement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12826–12835, 2020. 1, 2, 6

    2020

  28. [36]

    Curl: Neural curve layers for global image enhancement

    Sean Moran, Steven McDonagh, and Gregory Slabaugh. Curl: Neural curve layers for global image enhancement. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 9796–9803. IEEE, 2021. 3

    2020

  29. [37]

    Distort-and-recover: Color enhancement using deep reinforcement learning

    Jongchan Park, Joon-Young Lee, Donggeun Yoo, and In So Kweon. Distort-and-recover: Color enhancement using deep reinforcement learning. In Proceedings of the IEEE con- ference on computer vision and pattern recognition , pages 5928–5936, 2018. 3

    2018

  30. [38]

    Color transfer between images

    Erik Reinhard, Michael Adhikhmin, Bruce Gooch, and Peter Shirley. Color transfer between images. IEEE Computer graphics and applications, 21(5):34–41, 2001. 1

    2001

  31. [39]

    Improved implicit neural representation with fourier reparameterized training

    Kexuan Shi, Xingyu Zhou, and Shuhang Gu. Improved implicit neural representation with fourier reparameterized training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 25985– 25994, 2024. 3

    2024

  32. [40]

    Very deep convo- lutional networks for large-scale image recognition

    Karen Simonyan and Andrew Zisserman. Very deep convo- lutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 5

    2014 arXiv

  33. [41]

    Starenhancer: Learning real-time and style-aware image enhancement

    Yuda Song, Hui Qian, and Xin Du. Starenhancer: Learning real-time and style-aware image enhancement. In Proceed- ings of the IEEE/CVF International Conference on Com- puter Vision, pages 4126–4135, 2021. 3

    2021

  34. [42]

    Mlp-mixer: An all-mlp architecture for vision

    Ilya O Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lu- cas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, et al. Mlp-mixer: An all-mlp architecture for vision. Advances in neural information processing systems, 34:2...

  35. [43]

    Resmlp: Feedforward networks for image classification with data-efficient training

    Hugo Touvron, Piotr Bojanowski, Mathilde Caron, Matthieu Cord, Alaaeldin El-Nouby, Edouard Grave, Gautier Izac- ard, Armand Joulin, Gabriel Synnaeve, Jakob Verbeek, et al. Resmlp: Feedforward networks for image classification with data-efficient training. IEEE transactions on ...

    2022

  36. [44]

    Maxim: Multi-axis mlp for image processing

    Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxim: Multi-axis mlp for image processing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5769–5780, 2022. 3

    2022

  37. [45]

    Lgabl: Uhd multi-exposure image fusion via local and global aware bilateral learning

    Di Wang, Zhuoran Zheng, Weiping Ding, and Xiuyi Jia. Lgabl: Uhd multi-exposure image fusion via local and global aware bilateral learning. IEEE Transactions on Emerg- ing Topics in Computational Intelligence , 8(2):1362–1375,

  38. [46]

    Local color distributions prior for image enhancement

    Haoyuan Wang, Ke Xu, and Rynson WH Lau. Local color distributions prior for image enhancement. InEuropean con- ference on computer vision, pages 343–359. Springer, 2022. 5, 7, 8, 1

    2022

  39. [47]

    Underexposed photo enhance- ment using deep illumination estimation

    Ruixing Wang, Qing Zhang, Chi-Wing Fu, Xiaoyong Shen, Wei-Shi Zheng, and Jiaya Jia. Underexposed photo enhance- ment using deep illumination estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6849–6857, 2019. 2, 3, 6

    2019

  40. [48]

    Real-time image en- hancer via learnable spatial-aware 3d lookup tables

    Tao Wang, Yong Li, Jingyang Peng, Yipeng Ma, Xian Wang, Fenglong Song, and Youliang Yan. Real-time image en- hancer via learnable spatial-aware 3d lookup tables. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2471–2480, 2021. 2, 3, 6

    2021

  41. [49]

    Image quality assessment: from error visibility to structural similarity

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Si- moncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600–612, 2004. 7

    2004

  42. [50]

    Deep retinex decomposition for low-light enhancement

    Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep retinex decomposition for low-light enhancement. arXiv preprint arXiv:1808.04560, 2018. 1, 2, 7

    2018 arXiv

  43. [51]

    Uretinex-net: Retinex-based deep unfolding network for low-light image enhancement

    Wenhui Wu, Jian Weng, Pingping Zhang, Xu Wang, Wen- han Yang, and Jianmin Jiang. Uretinex-net: Retinex-based deep unfolding network for low-light image enhancement. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 5901–5910, 2022. 1, 2, 7

    2022

  44. [52]

    Semi-supervised learning for low-light image enhancement by pseudo low-light image

    Shuo Xie, Yingjie Ma, Wei Xu, Song Qiu, and Yifei Sun. Semi-supervised learning for low-light image enhancement by pseudo low-light image. In 2023 16th International Congress on Image and Signal Processing, BioMedical En- gineering and Informatics (CISP-BMEI) , pages 1–6. IEEE,

    2023

  45. [53]

    Bilateral grid learning for stereo matching networks

    Bin Xu, Yuhua Xu, Xiaoli Yang, Wei Jia, and Yulan Guo. Bilateral grid learning for stereo matching networks. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12497–12506, 2021. 2, 3

    2021

  46. [54]

    Deep bilateral learning for stereo image super-resolution

    Qingyu Xu, Longguang Wang, Yingqian Wang, Weidong Sheng, and Xinpu Deng. Deep bilateral learning for stereo image super-resolution. IEEE Signal Processing Letters, 28: 613–617, 2021. 2, 3

    2021

  47. [55]

    Automatic photo adjustment using deep neu- ral networks

    Zhicheng Yan, Hao Zhang, Baoyuan Wang, Sylvain Paris, and Yizhou Yu. Automatic photo adjustment using deep neu- ral networks. ACM Transactions on Graphics (TOG), 35(2): 1–15, 2016. 1, 2

    2016

  48. [56]

    Adaint: Learning adaptive intervals for 3d lookup tables on real-time image enhancement

    Canqian Yang, Meiguang Jin, Xu Jia, Yi Xu, and Ying Chen. Adaint: Learning adaptive intervals for 3d lookup tables on real-time image enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17522–17531, 2022. 2, 3, 6

    2022

  49. [57]

    Seplut: Separable image-adaptive lookup tables for real-time image enhancement

    Canqian Yang, Meiguang Jin, Yi Xu, Rui Zhang, Ying Chen, and Huaida Liu. Seplut: Separable image-adaptive lookup tables for real-time image enhancement. In European Con- ference on Computer Vision, pages 201–217. Springer, 2022. 3, 6

    2022

  50. [58]

    Learning image-adaptive 3d lookup tables for high perfor- mance photo enhancement in real-time

    Hui Zeng, Jianrui Cai, Lida Li, Zisheng Cao, and Lei Zhang. Learning image-adaptive 3d lookup tables for high perfor- mance photo enhancement in real-time. IEEE Transactions on Pattern Analysis and Machine Intelligence , 44(4):2058– 2073, 2020. 3, 6, 8, 1

    2020

  51. [59]

    Clut-net: Learning adaptively compressed representations of 3dluts for lightweight image enhancement

    Fengyi Zhang, Hui Zeng, Tianjun Zhang, and Lin Zhang. Clut-net: Learning adaptively compressed representations of 3dluts for lightweight image enhancement. In Proceedings of the 30th ACM International Conference on Multimedia , pages 6493–6501, 2022. 3

    2022

  52. [60]

    Lookup table meets lo- cal laplacian filter: pyramid reconstruction network for tone mapping

    Feng Zhang, Ming Tian, Zhiqiang Li, Bin Xu, Qingbo Lu, Changxin Gao, and Nong Sang. Lookup table meets lo- cal laplacian filter: pyramid reconstruction network for tone mapping. Advances in Neural Information Processing Sys- tems, 36:57558–57569, 2023. 2, 3, 6, 8, 1

    2023

  53. [61]

    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. In Proceedings of the IEEE conference on computer vision and pattern recogni- tion, pages 586–595, 2018. 7

    2018

  54. [62]

    Kindling the darkness: A practical low-light image enhancer

    Yonghua Zhang, Jiawan Zhang, and Xiaojie Guo. Kindling the darkness: A practical low-light image enhancer. In Pro- ceedings of the 27th ACM international conference on mul- timedia, pages 1632–1640, 2019. 1, 2

    2019

  55. [63]

    Ultra-high-definition image hdr reconstruc- tion via collaborative bilateral learning

    Zhuoran Zheng, Wenqi Ren, Xiaochun Cao, Tao Wang, and Xiuyi Jia. Ultra-high-definition image hdr reconstruc- tion via collaborative bilateral learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4449–4458, 2021. 2, 3

    2021

  56. [64]

    4k-resolution photo exposure correction at 125 fps with˜ 8k parameters

    Yijie Zhou, Chao Li, Jin Liang, Tianyi Xu, Xin Liu, and Jun Xu. 4k-resolution photo exposure correction at 125 fps with˜ 8k parameters. In Proceedings of the IEEE/CVF Win- ter Conference on Applications of Computer Vision , pages 1587–1597, 2024. 2, 3 Learning Pixel-adaptive M...

    2024

  57. [65]

    The width in NAF-block is set as

    as our backbone to gener- ate the bilateral grids, with each layer in both the encoder and decoder comprising two NAF- blocks (see Figure 7), and the number of blocks at the bot- tom level is the same. The width in NAF-block is set as

  58. [66]

    We then apply the pixel unshuffle operation to reduce its resolution by a factor of 4 while expanding the channel count by 16 times

    The feature map output from NAFNet maintains a chan- nel numble equal to the model’s width. We then apply the pixel unshuffle operation to reduce its resolution by a factor of 4 while expanding the channel count by 16 times. Next, two 1×1 convo- lutions are used to further adj...

Pith tools

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