Pith. sign in

REVIEW 3 major objections 5 minor 64 references

Oneta: Multi-Style Image Enhancement Using Eigentransformation Functions

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 single vision transformer with switchable style tokens can handle six image-enhancement tasks across 30 datasets.

desk verdict A clever multi-style enhancement framework with a mislabeled oracle ceiling and an overstated headline claim; worth refereeing for the mechanism, not for the numbers. read the letter →

arxiv 2506.23547 v1 pith:WIHXVBO6 submitted 2025-06-30 cs.CV

Jiwon Kim , Soohyun Hwang , Dong-O Kim , Changsu Han , Min Kyu Park , Chang-Su Kim This is my paper
classification cs.CV
keywords Onetamulti-styleimageenhancementstyletokeneigentransformationfunctionvisiontransformerintensitylookuptablecolorcorrectionmatrixlow-light
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

Oneta is presented as the first multi-style image enhancement network: one model that can retouch, process camera raw signals, brighten low-light images, dehaze, enhance underwater photos, and white-balance by switching a learned style token. The paper's central claim is that all 30 datasets' enhancement mappings reduce to two global pixelwise operations—an intensity transformation function followed by a $3\times3$ color correction matrix—and that this two-step model has a high PSNR (peak signal-to-noise ratio) upper bound on most datasets. To make the intensity curve tractable, Oneta represents it in an 'eigentransformation function' basis, predicting only ten coefficients. On 20 of the 30 datasets the single network reports PSNR comparable to task-specific algorithms, which would matter because it replaces many separately trained systems with one checkpoint.

What carries the argument

The central object is the eigentransformation function (eigenTF): the first $M$ left singular vectors of the matrix of per-image optimal intensity transformation functions, obtained by SVD, so that any $256$-point lookup table is reconstructed as a linear combination of $M$ basis curves. Oneta's Y-Net regresses the $M$ coefficients from image patches attended by the selected style token, and C-Net regresses the six free entries of the $3\times3$ color correction matrix from the intermediate image. Quadratic programming supplies both the upper-bound ceiling and the regression targets for this two-step model, and the style tokens make the same weights switchable across datasets.

What would settle it

Recompute the QP upper bound on a held-out split for each of the 30 datasets: any dataset where that ceiling is below the best specialized baseline falsifies the claim that Oneta's global two-step model can match specialized methods there, and any dataset where Oneta sits far below its own ceiling would show a network-capacity failure instead.

Watch

Extended reading notes

Core claim

Oneta's discovery is that multi-style enhancement can be organized as a single token-conditioned prediction problem. The network uses a ViT-based Y-Net to predict coefficients of the intensity transformation function and a ViT-based C-Net to predict the color correction matrix, with a learned style token selecting which of $K$ enhancement styles is active. The paper derives the optimal transformation function and color matrix for each training pair by quadratic programming, uses those solutions both as an upper-bound analysis and as training targets, and compresses the $256$-point intensity functions into an $M$-dimensional eigenTF space by SVD, with $M=10$. The reported result is that this single network matches specialized per-task models on 20 datasets and beats the best low-light baseline on LOL-v2-real by more than 5 dB.

Load-bearing premise

The load-bearing premise is that every dataset's input-to-output mapping can be represented by one global $256$-step intensity lookup table followed by one $3\times3$ color matrix applied identically to all pixels; the paper's own upper-bound numbers show this premise fails on SID and SOTS, where the model-class ceiling is 20.40 dB versus 24.44 dB and 22.35 dB versus 42.56 dB, respectively.

Editorial extensions

If this is right

  • With one trained checkpoint, a user can retouch the same photo in eight expert styles, enhance a low-light photo in seven styles, or apply styles in series by switching or combining tokens.
  • Multi-style training is the better training strategy: on most datasets it beats training the same architecture on each dataset alone, and on the 689-image LOL-v2-real set it gains more than 5 dB over the best single-task baseline.
  • Interpolating style tokens yields output images in intermediate styles, so the token set behaves like a discrete sampling of a continuous enhancement-style space.
  • The QP upper bound is a hard ceiling for the two-point-operator model class; on SID and SOTS it is already below specialized baselines, so no extra capacity in this class can close those gaps.

Reading between the lines

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

  • The paper leaves unexplored the possibility of treating style tokens as a continuous control dial, letting users blend styles or extrapolate to enhancement styles not present in training.
  • The QP ceiling calculation is reusable as a cheap feasibility test for any global enhancement design: compute the two-point-operator ceiling on a target dataset before training to decide whether global operations suffice or local refinements are required.
  • Because LLIE and ISP inputs are preprocessed differently from the baselines, a direct comparison on identical preprocessed inputs would separate preprocessing effects from the model's own behavior, especially on SID and SOTS.
  • If the eigenTF basis is shared across all 30 styles, each style reduces to ten coefficients plus six color matrix entries, which could serve as a compact style descriptor for transferring or editing enhancement styles between images.
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 Oneta, a single ViT-based network that performs multi-style image enhancement across six tasks (retouching, ISP, LLIE, dehazing, UIE, white balancing) using K learnable style tokens. The enhancement model consists of an intensity transformation function (TF), compactly represented in an eigenTF basis, followed by a color correction matrix (CCM). The authors report PSNR comparisons against task-specific baselines on 30 datasets, and they present a quadratic-programming-based oracle analysis intended to show that the two-step model class has high performance ceilings. The paper is candid about several weak points, including failure cases and datasets where Oneta underperforms.

Significance. If the claims were fully supported, the paper would demonstrate that a single, parameter-light network with switchable tokens can cover a broad range of enhancement tasks, and that a simple global two-step model has substantial representational capacity. The oracle analysis, style-token mechanism, and extensive dataset coverage are valuable. However, the validity of the oracle computation and the fairness of the preprocessing protocol are load-bearing; until these are resolved, the central claims about model-class sufficiency and cross-task effectiveness are not established.

major comments (3)
  1. [Sec. 3.2, Eqs. (5)-(6)] The numbers labeled 'Upper bound' in Table 1 are not valid upper bounds on the two-step model class. The optimization of the TF in Eq. (5) and the CCM in Eq. (6) is performed sequentially, not jointly. The full objective is bilinear in the TF x and the CCM kappa because the intermediate image, and hence the matrix B in Eq. (6), depends on x. The reported procedure therefore yields a feasible point, and the resulting PSNR is a lower bound on the true oracle optimum, not an upper bound. For instance, the SOTS-in row reports 22.35 dB, but a jointly optimized TF+CCM could in principle achieve a higher PSNR, potentially even exceeding the 42.56 dB of C2PNet. This weakens the conclusion in Sec. 4.3 that the two-step global model is insufficient for dehazing and SID. The authors should either compute a proper upper bound (e.g., by alternating optimization or a discretized search over TFs) or relabel the row as 'QP feasible PSNR' and rephrase the model-sufficiency conclusions accordingly.
  2. [Sec. 4.1 and Supplement A] The input preprocessing protocol may confound the comparison with baselines. LLIE images are pre-processed with BT.709 gamma correction and a bilateral filter, and ISP images are gamma-corrected, but the paper does not state that the baselines (Retinexformer, SNR-Net, InvISP, ParamISP, etc.) received the same preprocessing. If they did not, the 5.1 dB margin over Retinexformer on LOL-v2-real (27.91 dB vs. 22.80 dB) could be entirely due to the input transformation rather than to Oneta's enhancement capability. The authors should rerun all baselines under identical preprocessing, or provide a clear justification (e.g., showing that gamma correction alone changes baseline PSNR negligibly).
  3. [Abstract and Sec. 4.3] The abstract claims that Oneta 'can effectively undertake six enhancement tasks across 30 datasets,' but the paper's own Table 1 shows substantial failures: On SOTS-in and SOTS-out, Oneta is about 20-22 dB behind the best dehazing baselines, and on SID it is more than 6 dB behind Retinexformer. While Sec. 4.3 and footnote 1 narrow this claim, the abstract and introduction do not. The authors should qualify the claim in the abstract to say that Oneta achieves comparable performance on 20 of the 30 datasets and underperforms on the remaining ones, matching the more cautious language in Sec. 4.3.
minor comments (5)
  1. [Fig. 31] The caption reads 'Dehzing results' and should be corrected to 'Dehazing results.'
  2. [Eq. (7)] The row sum constraint is written as 'P3 j=1 κij = 1for each i' with a missing space; it should be 'Σ_{j=1}^3 κ_ij = 1 for each i.'
  3. [Sec. 3.2] The definition of the matrix A would be clearer if the paper explicitly stated that each row of A is a one-hot vector with a single 1 at the column corresponding to the input intensity, and that the number of rows N equals the number of pixels.
  4. [Table 1 caption] The caption lists 'Upper bound' without explaining the methodology; a short note pointing to Sec. 3.2 and Supplement B would help readers interpret this row.
  5. [Supplement D.4] The inference-only monotonicity post-processing (setting x_k = x_{k-1} when x_k < x_{k-1}) is an ad-hoc fix; the authors should state whether any monotonicity regularization is included in the training loss, because otherwise the train/test setup is inconsistent.

Circularity Check

0 steps flagged · score 2.0 of 10

No material circularity; oracle ceilings are external model-class estimates, and self-citations are background only.

full rationale

The paper's central derivation—a two-step global TF+CCM model with eigenTF compression and token-switched ViT predictors—is evaluated against external benchmarks. The 'Upper bound' row in Table 1 is obtained by fitting the model's parameters with quadratic programming to ground-truth test images (Eqs. 5-6), which is a legitimate oracle/capacity analysis, not a network prediction; the network is then trained on training pairs (Eq. 12) and tested on held-out data, so the reported PSNRs are not fitted inputs renamed as predictions. The eigenTF basis is a standard SVD low-rank approximation of oracle TFs, not a self-justifying construction. The paper's few self-citations (e.g., [21], [24]) appear only as related-work background and do not carry the argument. Two issues noted by the reader—the sequential QP solution of Eqs. (5)-(6) means the reported 'upper bound' may not be a true joint upper bound, and LLIE/ISP inputs are preprocessed differently from baselines—are correctness/fairness concerns rather than circularity. The paper itself narrows the 'all' claim in footnote 1 and acknowledges SID/dehazing shortcomings in Sec. 4.3, further reducing any concern that the headline overpromises by construction.

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

The central claim rests on a dominant modeling assumption: that a spatially uniform two-step point-operator model suffices for all six tasks. The paper's own QP ceiling refutes this for dehazing (SOTS-in ceiling 22.35 dB vs C2PNet 42.56 dB) and hard low-light (SID ceiling 20.40 dB vs 24.44 dB), so the assumption is false by the paper's numbers on those tasks. The two learned components, the eigentransformation basis and the 30 style tokens, are both fit to the training data and are not externally verified. Hyperparameters M=10 and lambda=0.1 are set by a partial ablation and by hand. One ad hoc inference-only rule (monotonicity post-processing) introduces a train/test mismatch. Net: the honest contribution is a demonstrated multi-style conditioning mechanism plus an oracle-ceiling methodology, with the model class itself being the principal limitation.

free parameters (4)
  • M, number of eigentransformation basis vectors = 10
    Set by the ablation in Figure 9 on a 9-dataset subset of 500 train and 50 test images; PSNR saturates around M=10. The choice affects the capacity of every predicted transformation function.
  • eigentransformation basis U_10 (256 by 10 matrix) = top-10 left singular vectors of the oracle TF matrix (Eq. 8)
    The basis is fit to the training set's QP-optimal transformations, so the 10 coefficients predicted by Y-Net are expressed in data-fitted coordinates rather than a fixed analytic family.
  • lambda, loss weight on coefficient regression = 0.1
    Chosen by hand in Eq. 12; no ablation or sensitivity analysis is reported.
  • LLIE preprocessing parameters (BT.709 gamma and bilateral filter settings) = not reported
    The bilateral filter parameters (sigma space and range) are unspecified, yet this preprocessing is load-bearing for low-light inputs because the model cannot remove noise itself (Section 4.1).
assumptions (5)
  • domain assumption A single global intensity transformation function and a single 3x3 color correction matrix, applied identically to every pixel, can represent the input-to-output mapping of each of the 30 datasets.
    Entered in Section 3.1 (two-step point operators, Eqs. 1 to 3). The paper attempts to justify the model class with the QP oracle ceiling in Section 4.3, but the same ceiling is below existing algorithms for SOTS-in (22.35 vs 42.56), SOTS-out (29.92 vs 36.68), SID (20.40 vs 24.44), and ISP 20D, D7000, and D90, so the assumption fails for those tasks by the paper's own numbers.
  • domain assumption Noise in low-light inputs can be removed by preprocessing (BT.709 gamma plus bilateral filtering) without changing the task being compared.
    Section 4.1: LLIE images are preprocessed to reduce noise, which is not removable by point operators, and ISP images are gamma-corrected. This moves Oneta's input distribution toward the point-operator model and may differ from what baselines such as SNR-Net or Retinexformer received.
  • domain assumption The top-10 eigentransformation basis, computed once from the oracle transformations of the training set, is a sufficient coordinate system for all 30 styles.
    Section 3.3 (Eqs. 8 to 11): a single basis serves all styles. The M ablation (Figure 9) covers only 9 of the 30 datasets, so the adequacy of the basis for the other 21 styles is untested.
  • standard math Best rank-M approximation via SVD and convexity of the quadratic programs (Eqs. 5 and 6).
    Invoked with citations [6] (rank-M approximation) and [7] (Boyd and Vandenberghe) in Sections 3.2 and 3.3; these are standard results applied correctly as far as can be checked from the text.
  • ad hoc to paper The inference-only monotonicity enforcement (setting xk = xk-1 when xk < xk-1) preserves the learned enhancement behavior.
    Supplement D.4: applied only at inference, creating a train/test mismatch; Table 4 shows small PSNR effects (up to +0.58 dB), but the rule is hand-designed and not derived from the QP optimum of Eq. 5.
invented entities (1)
  • Style tokens (30 learnable vectors, one per dataset or expert)
    purpose: Condition the two ViT encoders so a single network emits the enhancement style of the selected dataset; token interpolation provides intermediate styles.
    All evidence for token behavior is internal to the paper: t-SNE of the tokens (Figures 10 to 12) and Oneta's own outputs (Figures 5 to 7). The closest external handle, transferring the LSUI token to UIEB, yields 19.87 dB versus 22.15 dB with the UIEB token, so token specificity is not independently confirmed. The vectors are learned parameters, not externally verified entities.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Oneta: Multi-Style Image Enhancement Using Eigentransformation Functions." pith.science (2026). https://pith.science/paper/WIHXVBO6

@misc{pith2026250623547,
  author       = {Pith},
  title        = {Pith review of: Oneta: Multi-Style Image Enhancement Using Eigentransformation Functions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WIHXVBO6}},
  note         = {Machine review of arXiv:2506.23547}
}
abstract

The first algorithm, called Oneta, for a novel task of multi-style image enhancement is proposed in this work. Oneta uses two point operators sequentially: intensity enhancement with a transformation function (TF) and color correction with a color correction matrix (CCM). This two-step enhancement model, though simple, achieves a high performance upper bound. Also, we introduce eigentransformation function (eigenTF) to represent TF compactly. The Oneta network comprises Y-Net and C-Net to predict eigenTF and CCM parameters, respectively. To support $K$ styles, Oneta employs $K$ learnable tokens. During training, each style token is learned using image pairs from the corresponding dataset. In testing, Oneta selects one of the $K$ style tokens to enhance an image accordingly. Extensive experiments show that the single Oneta network can effectively undertake six enhancement tasks -- retouching, image signal processing, low-light image enhancement, dehazing, underwater image enhancement, and white balancing -- across 30 datasets.

Figures

Figures reproduced from arXiv: 2506.23547 by the authors.

Figure 1
Figure 1. Given an input image Iin and a style token, the proposed Oneta algorithm generates an output image Iout in the desired style. Oneta can retouch an identical image in different styles and perform a variety of tasks, including LLIE and WB, simply by changing the style token. respectively. All these algorithms were designed to undertake each specific task only. Meanwhile, at￾tempts have been made to perform multiple im… view at source ↗
Figure 2
Figure 2. Oneta can cope with retouching, ISP, LLIE, de [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Overview of the proposed Oneta algorithm: In Oneta, a user selects one of [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (32 more)
Figure 4
Figure 4. Figure 4: Illustration of the eigenTF representation. As [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Oneta enhancement results for the retouching task: the left five images are for experts A, B, C, D, and E in MIT [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: Oneta enhancement results for the LLIE, dehazing, UIE, ISP, and WB tasks. [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: (d) illustrates that a user can perform the enhancement by employing multiple style tokens in series. These use cases show that users can utilize the single Oneta network to enhance images in vari￾ous manners according to their preferences. The supplemental document pr…
Figure 9
Figure 9. Figure 9: Ablation study on the number M of eigenTFs. (b) Retouching ISP LLIE Dehazing UIE WB (a) Epoch 0 Epoch 15 Epoch 40 [PITH_FULL_IMAGE:figures/full_fig_p008_9.png]
Figure 10
Figure 10. Figure 10: t-SNE visualization: (a) clustering of style tokens [PITH_FULL_IMAGE:figures/full_fig_p008_10.png]
Figure 11
Figure 11. Figure 11: visualizes how style tokens si , s ′ i , and s ′′ i are aligned through training. Also, [PITH_FULL_IMAGE:figures/full_fig_p013_11.png]
Figure 12
Figure 12. Figure 12: t-SNE visualization of style tokens si, s ′ i , and s ′′ i for each task [PITH_FULL_IMAGE:figures/full_fig_p014_12.png]
Figure 13
Figure 13. Figure 13: Retouching examples on the FiveK dataset. For each image, the retouching results of expert styles A, B, C, D, E, a, [PITH_FULL_IMAGE:figures/full_fig_p016_13.png]
Figure 14
Figure 14. Figure 14: Retouching examples on the PPR10K dataset. For each image, the retouching results of expert styles A, B, C, D, E, [PITH_FULL_IMAGE:figures/full_fig_p017_14.png]
Figure 15
Figure 15. Figure 15: ISP results on the FiveK-Canon EOS 20D dataset. [PITH_FULL_IMAGE:figures/full_fig_p018_15.png]
Figure 16
Figure 16. Figure 16: ISP results on the FiveK-Canon EOS 40D dataset. [PITH_FULL_IMAGE:figures/full_fig_p019_16.png]
Figure 17
Figure 17. Figure 17: ISP results on the FiveK-Canon EOS 5D dataset. [PITH_FULL_IMAGE:figures/full_fig_p020_17.png]
Figure 18
Figure 18. Figure 18: ISP results on the FiveK-Canon EOS XSi dataset. [PITH_FULL_IMAGE:figures/full_fig_p021_18.png]
Figure 19
Figure 19. Figure 19: ISP results on the FiveK-Nikon D700 dataset. [PITH_FULL_IMAGE:figures/full_fig_p022_19.png]
Figure 20
Figure 20. Figure 20: ISP results on the FiveK-Nikon D70s dataset. [PITH_FULL_IMAGE:figures/full_fig_p023_20.png]
Figure 21
Figure 21. Figure 21: ISP results on the RAISE-Nikon D7000 dataset. [PITH_FULL_IMAGE:figures/full_fig_p024_21.png]
Figure 22
Figure 22. Figure 22: ISP results on the Samsung Galaxy D90 dataset. [PITH_FULL_IMAGE:figures/full_fig_p025_22.png]
Figure 23
Figure 23. Figure 23: ISP results on the RAISE-Nikon S7 dataset. [PITH_FULL_IMAGE:figures/full_fig_p026_23.png]
Figure 24
Figure 24. Figure 24: LLIE results on the LOL-v2-real dataset. [PITH_FULL_IMAGE:figures/full_fig_p027_24.png]
Figure 25
Figure 25. Figure 25: LLIE results on the LOL-v2-syn dataset [PITH_FULL_IMAGE:figures/full_fig_p028_25.png]
Figure 26
Figure 26. Figure 26: LLIE results on the LOL-v1 dataset. Input Output Ground-truth Input Output Ground-truth [PITH_FULL_IMAGE:figures/full_fig_p029_26.png]
Figure 27
Figure 27. Figure 27: LLIE results on the SID dataset [PITH_FULL_IMAGE:figures/full_fig_p029_27.png]
Figure 28
Figure 28. Figure 28: LLIE results on the SMID dataset [PITH_FULL_IMAGE:figures/full_fig_p030_28.png]
Figure 29
Figure 29. Figure 29: LLIE results on the SDSD-indoor dataset. [PITH_FULL_IMAGE:figures/full_fig_p031_29.png]
Figure 30
Figure 30. Figure 30: LLIE results on the SDSD-outdoor dataset. [PITH_FULL_IMAGE:figures/full_fig_p031_30.png]
Figure 31
Figure 31. Figure 31: Dehzing results on the SOTS-in dataset [PITH_FULL_IMAGE:figures/full_fig_p032_31.png]
Figure 32
Figure 32. Figure 32: Dehazing results on the SOTS-out dataset. [PITH_FULL_IMAGE:figures/full_fig_p033_32.png]
Figure 33
Figure 33. Figure 33: UIE results on the LSUI dataset [PITH_FULL_IMAGE:figures/full_fig_p034_33.png]
Figure 34
Figure 34. Figure 34: Underwater image enhancement results on the UIEB dataset. [PITH_FULL_IMAGE:figures/full_fig_p035_34.png]
Figure 35
Figure 35. Figure 35: WB results on the WB dataset [PITH_FULL_IMAGE:figures/full_fig_p036_35.png]
Figure 36
Figure 36. Figure 36: WB results on the Cube+ dataset [PITH_FULL_IMAGE:figures/full_fig_p037_36.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

64 extracted references · 61 canonical work pages

  1. [1]

    Semantic White Balance: Semantic Color Constancy Using Convolutional Neural Network

    Mahmoud Afifi. Semantic white balance: Semantic color constancy using convolutional neural network. arXiv preprint arXiv:1802.00153, 2018. 1, 2

    work page Pith review arXiv 2018

  2. [2]

    Deep white- balance editing

    Mahmoud Afifi and Michael S Brown. Deep white- balance editing. In CVPR, 2020. 1, 2, 7, 12

    work page 2020

  3. [3]

    When color constancy goes wrong: Correcting improperly white-balanced images

    Mahmoud Afifi, Brian Price, Scott Cohen, and Michael S Brown. When color constancy goes wrong: Correcting improperly white-balanced images. In CVPR, 2019. 5, 11, 12

    work page 2019

  4. [4]

    Un- supervised learning for color constancy.arXiv preprint arXiv:1712.00436, 2017

    Nikola Bani ´c, Karlo Koˇsˇcevi´c, and Sven Lonˇcari´c. Un- supervised learning for color constancy.arXiv preprint arXiv:1712.00436, 2017. 5, 11, 12

    arXiv 2017

  5. [5]

    Bruna, Filippo Naccari, and Raimondo Schettini

    Simone Bianco, Arcangelo R. Bruna, Filippo Naccari, and Raimondo Schettini. Color correction pipeline op- timization for digital cameras. Journal of Electronic Imaging, 22(2):023014–023014, 2013. 3

    work page 2013

  6. [6]

    Foundations of Data Science

    Avrim Blum, John Hopcroft, and Ravindran Kannan. Foundations of Data Science. 2015. 2, 4

    work page 2015

  7. [7]

    Boyd and Lieven Vandenberghe

    Stephen P. Boyd and Lieven Vandenberghe. Convex Optimization. Cambridge University Press, 2004. 2, 4

    work page 2004

  8. [8]

    Parameter values for the HDTV standards for production and international pro- gramme exchange

    ITU-R Recommendation BT. Parameter values for the HDTV standards for production and international pro- gramme exchange. 2002. 5, 12

    work page 2002

Show all 64 references
  1. [9]

    Learning photographic global tonal ad- justment with a database of input/output image pairs

    Vladimir Bychkovsky, Sylvain Paris, Eric Chan, and Fr´edo Durand. Learning photographic global tonal ad- justment with a database of input/output image pairs. In CVPR, 2011. 5, 11, 12

    2011

  2. [10]

    Retinexformer: One- stage retinex-based transformer for low-light image enhancement

    Yuanhao Cai, Hao Bian, Jing Lin, Haoqian Wang, Radu Timofte, and Yulun Zhang. Retinexformer: One- stage retinex-based transformer for low-light image enhancement. In ICCV, 2023. 1, 2, 3, 7

    2023

  3. [11]

    Learning to see in the dark

    Chen Chen, Qifeng Chen, Jia Xu, and Vladlen Koltun. Learning to see in the dark. In CVPR, 2018. 5, 11, 12

    2018

  4. [12]

    Seeing motion in the dark

    Chen Chen, Qifeng Chen, Minh N Do, and Vladlen Koltun. Seeing motion in the dark. In ICCV, 2019. 5, 11, 12

    2019

  5. [13]

    Pre-trained image processing transformer

    Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In CVPR, 2021. 1, 3

    2021

  6. [14]

    Raise: A raw images dataset for digital image forensics

    Duc-Tien Dang-Nguyen, Cecilia Pasquini, Valentina Conotter, and Giulia Boato. Raise: A raw images dataset for digital image forensics. InACM MM, pages 219–224, 2015. 5, 11, 12

    2015

  7. [15]

    ImageNet: A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In CVPR, 2009. 4

    2009

  8. [16]

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

    Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR,

  9. [17]

    R. C. Gonzalez and R. E. Woods. Digital Image Pro- cessing. Prentice Hall, 3rd edition, 2007. 1, 2

    2007

  10. [18]

    Image dehazing transformer with transmission-aware 3D position em- bedding

    Chun-Le Guo, Qixin Yan, Saeed Anwar, Runmin Cong, Wenqi Ren, and Chongyi Li. Image dehazing transformer with transmission-aware 3D position em- bedding. In CVPR, 2022. 1, 3

    2022

  11. [19]

    Anil K. Jain. Fundamentals of Digital Image Process- ing. Prentice Hall, 1989. 1, 2, 6

    1989

  12. [20]

    When fast Fourier transform meets transformer for image restoration

    Xingyu Jiang, Xiuhui Zhang, Ning Gao, and Yue Deng. When fast Fourier transform meets transformer for image restoration. In ECCV, 2024. 1, 2

    2024

  13. [21]

    PieNet: Personalized image enhancement network

    Han-Ul Kim, Young Jun Koh, and Chang-Su Kim. PieNet: Personalized image enhancement network. In ECCV, 2020. 1, 2

    2020

  14. [22]

    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 ECCV, 2024. 1, 2, 7

    2024

  15. [23]

    ParamISP: Learned forward and inverse ISPs using camera parameters

    Woohyeok Kim, Geonu Kim, Junyong Lee, Seungy- ong Lee, Seung-Hwan Baek, and Sunghyun Cho. ParamISP: Learned forward and inverse ISPs using camera parameters. In CVPR, 2024. 1, 2, 7

    2024

  16. [24]

    Contrast enhancement based on layered difference representa- tion of 2D histograms

    Chulwoo Lee, Chul Lee, and Chang-Su Kim. Contrast enhancement based on layered difference representa- tion of 2D histograms. IEEE TIP, 22(12):5372–5384,

  17. [25]

    Feature mod- ulation transformer: Cross-refinement of global rep- resentation via high-frequency prior for image super- resolution

    Ao Li, Le Zhang, Yun Liu, and Ce Zhu. Feature mod- ulation transformer: Cross-refinement of global rep- resentation via high-frequency prior for image super- resolution. In ICCV, 2023. 3

    2023

  18. [26]

    Bench- marking single-image dehazing and beyond

    Boyi Li, Wenqi Ren, Dengpan Fu, Dacheng Tao, Dan Feng, Wenjun Zeng, and Zhangyang Wang. Bench- marking single-image dehazing and beyond. IEEE TIP, 28(1):492–505, 2018. 5, 11, 12

    2018

  19. [27]

    An un- derwater image enhancement benchmark dataset and beyond

    Chongyi Li, Chunle Guo, Wenqi Ren, Runmin Cong, Junhui Hou, Sam Kwong, and Dacheng Tao. An un- derwater image enhancement benchmark dataset and beyond. IEEE TIP, 29:4376–4389, 2019. 5, 11, 12

    2019

  20. [28]

    Underwater image enhancement via medium transmission-guided multi- color space embedding

    Chongyi Li, Saeed Anwar, Junhui Hou, Runmin Cong, Chunle Guo, and Wenqi Ren. Underwater image enhancement via medium transmission-guided multi- color space embedding. IEEE TIP , 30:4985–5000,

  21. [29]

    Efficient and explicit modelling of image hierarchies for image restoration

    Yawei Li, Yuchen Fan, Xiaoyu Xiang, Denis De- mandolx, Rakesh Ranjan, Radu Timofte, and Luc Van Gool. Efficient and explicit modelling of image hierarchies for image restoration. In CVPR, 2023. 1, 3

    2023

  22. [30]

    SwinIR: Image restoration using swin transformer

    Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. SwinIR: Image restoration using swin transformer. In ICCV, 2021. 1, 3

    2021

  23. [31]

    PPR10k: A large-scale portrait photo retouching dataset with human-region mask and group-level consistency

    Jie Liang, Hui Zeng, Miaomiao Cui, Xuansong Xie, and Lei Zhang. PPR10k: A large-scale portrait photo retouching dataset with human-region mask and group-level consistency. In CVPR, 2021. 5, 11

    2021

  24. [32]

    Recurrent video restoration transformer with guided deformable attention

    Jingyun Liang, Yuchen Fan, Xiaoyu Xiang, Rakesh Ranjan, Eddy Ilg, Simon Green, Jiezhang Cao, Kai Zhang, Radu Timofte, and Luc V Gool. Recurrent video restoration transformer with guided deformable attention. In NeurIPS, 2022. 1, 3

    2022

  25. [33]

    Decoupled weight decay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 5

    2019

  26. [34]

    U-shape transformer for underwater image enhancement

    Lintao Peng, Chunli Zhu, and Liheng Bian. U-shape transformer for underwater image enhancement. IEEE TIP, 32:3066–3079, 2023. 1, 3, 5, 7, 8, 11, 12

    2023

  27. [35]

    MB-TaylorFormer: Multi-branch efficient transformer expanded by Tay- lor formula for image dehazing

    Yuwei Qiu, Kaihao Zhang, Chenxi Wang, Wenhan Luo, Hongdong Li, and Zhi Jin. MB-TaylorFormer: Multi-branch efficient transformer expanded by Tay- lor formula for image dehazing. In ICCV, 2023. 1, 3

    2023

  28. [36]

    DeepISP: Toward learning an end-to-end image pro- cessing pipeline

    Eli Schwartz, Raja Giryes, and Alex M Bronstein. DeepISP: Toward learning an end-to-end image pro- cessing pipeline. IEEE TIP, 28(2):912–923, 2018. 5, 11, 12

    2018

  29. [37]

    Tempformer: Temporally consistent transformer for video denoising

    Mingyang Song, Yang Zhang, and Tunc ¸ O Aydın. Tempformer: Temporally consistent transformer for video denoising. In ECCV, 2022. 3

    2022

  30. [38]

    Underwater image enhancement by transformer-based diffusion model with non-uniform sampling for skip strategy

    Yi Tang, Hiroshi Kawasaki, and Takafumi Iwaguchi. Underwater image enhancement by transformer-based diffusion model with non-uniform sampling for skip strategy. In ACM MM, 2023. 1, 2

    2023

  31. [39]

    Bilateral filter- ing for gray and color images

    Carlo Tomasi and Roberto Manduchi. Bilateral filter- ing for gray and color images. In ICCV, 1998. 5, 12

    1998

  32. [40]

    Training data-efficient image transformers & distillation through attention

    Hugo Touvron, Matthieu Cord, Matthijs Douze, Fran- cisco Massa, Alexandre Sablayrolles, and Herv ´e J´egou. Training data-efficient image transformers & distillation through attention. In ICML, 2021. 3

    2021

  33. [41]

    Stripformer: Strip trans- former for fast image deblurring

    Fu-Jen Tsai, Yan-Tsung Peng, Yen-Yu Lin, Chung- Chi Tsai, and Chia-Wen Lin. Stripformer: Strip trans- former for fast image deblurring. In ECCV, 2022. 3

    2022

  34. [42]

    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 CVPR, 2022. 1, 2

    2022

  35. [43]

    Visu- alizing data using t-SNE

    Laurens Van der Maaten and Geoffrey Hinton. Visu- alizing data using t-SNE. Journal of machine learning research, 9(11), 2008. 8

    2008

  36. [44]

    Underexposed photo enhancement using deep illumination estima- tion

    Ruixing Wang, Qing Zhang, Chi-Wing Fu, Xiaoyong Shen, Wei-Shi Zheng, and Jiaya Jia. Underexposed photo enhancement using deep illumination estima- tion. In CVPR, 2019. 1, 2

    2019

  37. [45]

    Seeing dynamic scene in the dark: A high-quality video dataset with mechatronic alignment

    Ruixing Wang, Xiaogang Xu, Chi-Wing Fu, Jiangbo Lu, Bei Yu, and Jiaya Jia. Seeing dynamic scene in the dark: A high-quality video dataset with mechatronic alignment. In ICCV, 2021. 5, 11, 12

    2021

  38. [46]

    Low-light image enhance- ment with illumination-aware gamma correction and complete image modelling network

    Yinglong Wang, Zhen Liu, Jianzhuang Liu, Songcen Xu, and Shuaicheng Liu. Low-light image enhance- ment with illumination-aware gamma correction and complete image modelling network. In ICCV, 2023. 1, 3

    2023

  39. [47]

    Uformer: A general U-shaped transformer for image restoration

    Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wen- gang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general U-shaped transformer for image restoration. In CVPR, 2022. 1, 3

    2022

  40. [48]

    Atten- tion is all you need

    A Waswani, N Shazeer, N Parmar, J Uszkoreit, L Jones, A Gomez, L Kaiser, and I Polosukhin. Atten- tion is all you need. In NeurIPS, 2017. 4

    2017

  41. [49]

    Deep retinex decomposition for low-light en- hancement

    Chen Wei, Wenjing Wang, Wenhan Yang, and Jiay- ing Liu. Deep retinex decomposition for low-light en- hancement. arXiv preprint arXiv:1808.04560 , 2018. 5, 11, 12

    2018 arXiv

  42. [50]

    Invertible image signal processing

    Yazhou Xing, Zian Qian, and Qifeng Chen. Invertible image signal processing. In CVPR, 2021. 1, 2, 7, 8

    2021

  43. [51]

    Multi-class token trans- former for weakly supervised semantic segmentation

    Lian Xu, Wanli Ouyang, Mohammed Bennamoun, Farid Boussaid, and Dan Xu. Multi-class token trans- former for weakly supervised semantic segmentation. In CVPR, 2022. 3

    2022

  44. [52]

    SNR-aware low-light image enhancement

    Xiaogang Xu, Ruixing Wang, Chi-Wing Fu, and Ji- aya Jia. SNR-aware low-light image enhancement. In CVPR, 2022. 1, 2, 3, 7

    2022

  45. [53]

    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 CVPR, 2022. 1, 2

    2022

  46. [54]

    SepLUT: separable image- adaptive lookup tables for real-time image enhance- ment

    Canqian Yang, Meiguang Jin, Yi Xu, Rui Zhang, Ying Chen, and Huaida Liu. SepLUT: separable image- adaptive lookup tables for real-time image enhance- ment. In ECCV, 2022. 1, 2, 7

    2022

  47. [55]

    Sparse gradient regularized deep retinex network for robust low-light image en- hancement

    Wenhan Yang, Wenjing Wang, Haofeng Huang, Shiqi Wang, and Jiaying Liu. Sparse gradient regularized deep retinex network for robust low-light image en- hancement. IEEE TIP, 30:2072–2086, 2021. 5, 11, 12

    2021

  48. [56]

    Per- ceiving and modeling density for image dehazing

    Tian Ye, Yunchen Zhang, Mingchao Jiang, Liang Chen, Yun Liu, Sixiang Chen, and Erkang Chen. Per- ceiving and modeling density for image dehazing. In ECCV, 2022. 1, 2, 7

    2022

  49. [57]

    Restormer: Efficient transformer for high- resolution image restoration

    Syed Waqas Zamir, Aditya Arora, Salman Khan, Mu- nawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high- resolution image restoration. In CVPR, 2022. 1, 3

    2022

  50. [58]

    Learning image-adaptive 3D lookup tables for high performance photo enhancement in real-time

    Hui Zeng, Jianrui Cai, Lida Li, Zisheng Cao, and Lei Zhang. Learning image-adaptive 3D lookup tables for high performance photo enhancement in real-time. IEEE TPAMI, 44(4):2058–2073, 2022. 1, 2

    2022

  51. [59]

    Accurate image restora- tion with attention retractable transformer

    Jiale Zhang, Yulun Zhang, Jinjin Gu, Yongbing Zhang, Linghe Kong, and Xin Yuan. Accurate image restora- tion with attention retractable transformer. In ICLR,

  52. [60]

    RealViformer: Investigating attention for real-world video super- resolution

    Yuehan Zhang and Angela Yao. RealViformer: Investigating attention for real-world video super- resolution. In ECCV, 2024. 3

    2024

  53. [61]

    STAR: A structure- aware lightweight transformer for real-time image en- hancement

    Zhaoyang Zhang, Yitong Jiang, Jun Jiang, Xiaogang Wang, Ping Luo, and Jinwei Gu. STAR: A structure- aware lightweight transformer for real-time image en- hancement. In ICCV, 2021. 1, 3, 7, 12

    2021

  54. [62]

    Wavelet-based Fourier information interac- tion with frequency diffusion adjustment for underwa- ter image restoration

    Chen Zhao, Weiling Cai, Chenyu Dong, and Cheng- wei Hu. Wavelet-based Fourier information interac- tion with frequency diffusion adjustment for underwa- ter image restoration. In CVPR, 2024. 1, 2, 3

    2024

  55. [63]

    Curricular contrastive regularization for physics-aware single image dehazing

    Yu Zheng, Jiahui Zhan, Shengfeng He, Junyu Dong, and Yong Du. Curricular contrastive regularization for physics-aware single image dehazing. In CVPR, 2023. 1, 2, 7 Oneta: Multi-Style Image Enhancement Using Eigentransformation Functions Supplementary Material A. Datasets We us...

    2023

  56. [64]

    We reduce the number of training images by a factor of 1/10 to prevent overfitting on the RESIDE dataset

    and RESIDE-OTS [26] for training and SOTS-indoor [26] and SOTS-outdoor [26] for testing. We reduce the number of training images by a factor of 1/10 to prevent overfitting on the RESIDE dataset. RESIDE images consist of multiple data samples of the same scene with varying degr...

Pith tools

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