Pith. sign in

REVIEW 3 major objections 4 minor 46 references

Personalized photo color grading can be learned from a handful of private ratings, without ever uploading the photos or ratings, by training a global scorer federatively and letting a frozen per-user version guide a small on-device enhancer

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 03:49 UTC pith:W3I6K73F

load-bearing objection A solid federated engineering contribution with a real evaluation gap: the personalization evidence is a self-referential proxy, so it deserves peer review but needs a human preference check before the headline claim is trusted. the 3 major comments →

arxiv 2607.27659 v1 pith:W3I6K73F submitted 2026-07-30 cs.CV cs.AIcs.DC

Learning Color Grading, No Photo Sharing: Federated Aesthetic Preference Learning for Personalized Image Enhancement

Chuanzhi Xu , Ziyuan Tao , Jean Julien KNell , Yanrong Chen , Haolan Guo , Xuanhua Yin , Adnan Mahmood , Weidong Cai This is my paper
classification cs.CV cs.AIcs.DC
keywords federated learningpersonalized image enhancementaesthetic preferencecolor grading3D LUTfew-shot calibrationproxy-score regularizationprivacy-preserving vision
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper tries to show that a user's aesthetic taste for photo color adjustments can be learned from a handful of private ratings, without uploading photos or ratings to a server. It proposes FedPAIE, which trains a lightweight global aesthetic scorer through federated averaging, calibrates that scorer to each new user on a 10- to 100-rating local support set, and then freezes it to guide adaptation of a compressed color-lookup-table enhancer on unpaired local photos. The claimed payoff is a personalized enhancer with about a third of a million parameters that adjusts warmth, exposure, and saturation to individual preference while staying close to the original image, with an excess-gap penalty plus pixel and perceptual losses preventing the enhancer from gaming the imperfect scorer. If true, it makes privacy-preserving personalized photo editing feasible on resource-constrained devices.

Core claim

The central claim is that a three-stage decomposition—federated population-level scorer learning, per-user local scorer calibration, and frozen-scorer-guided enhancer adaptation—makes personalized color grading learnable from sparse, unpaired, private data. Concretely, FedPAIE claims to be the first federated method for personalized aesthetic image enhancement, and its experiments on MIT-Adobe FiveK and Flickr-AES report that the full objective outperforms the generic enhancement prior for every evaluated user under the frozen-scorer proxy, with paired tests giving p≤3.6×10⁻⁶, while ablations show that the excess-gap penalty and fidelity terms are what protect against proxy-score over-optimi

What carries the argument

The engine is the frozen personalized aesthetic scorer: a lightweight dual-cue model (24-dimensional HSV/Lab color statistics plus a frozen MobileNetV3 semantic embedding, fused by an MLP with a learnable temperature) that is first trained via square-root-weighted federated averaging on decentralized ratings, then calibrated locally with regression, pairwise-ranking, and variance-preservation losses under a support-dependent parameter mask. Once frozen, it supplies a differentiable preference gradient that guides adaptation of a CLUT-Net enhancer—a compressed 3D color lookup-table—by updating only the coefficient predictor's CNN backbone and head while the LUT bases stay fixed. The regulariz

Load-bearing premise

The frozen personalized scorer's predicted rating is a faithful proxy for the user's true aesthetic judgment of the enhanced image; if the scorer can be gamed or is miscalibrated, the measured preference gain does not establish real personalization.

What would settle it

Show users a blinded comparison of FedPAIE's output versus the generic enhancement prior (and the original) for the same photo and ask which they prefer. If users do not prefer the personalized output at a rate meaningfully above chance—or if adversarial inputs that raise the scorer's Δ are judged visually worse—the personalization claim fails. The paper contains no human perceptual validation, so this experiment is the direct test.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If the central claim holds, personalized image enhancement no longer requires collecting private photos or paired user retouches; a small set of scalar ratings per user suffices.
  • The frozen-scorer-guidance design, with its fidelity and excess-gap regularization, offers a template for safely using imperfect proxy objectives in on-device adaptation.
  • The 0.293M-parameter single-pass enhancer is light enough for mobile inference, and the pipeline's per-stage update counts (≤0.787M scorer, ≤0.265M enhancer) fit resource-constrained devices.
  • Open-world personalization is reported to work at both 10 and 100 support ratings, with the federated initialization competitive with centralized training in scorer calibration metrics.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A natural decisive test is blinded human evaluation; the paper itself defers such validation, so a user study comparing personalized outputs against the generic prior and the original would directly test whether the proxy gain reflects real taste.
  • The same architecture—global preference model, local calibration, frozen-scorer-guided transformation—should transfer beyond color grading to other operations such as tone curves, local retouching, denoising, or even non-photographic content preferences, since the frozen scorer does not depend on the LUT representation.
  • The privacy boundary is protocol-level (only parameters and counts are exchanged); a deployment would need to layer on secure aggregation or differential privacy and measure the resulting utility loss, which the paper states as compatible but does not experiment with.
  • The image-conditioned variation analysis suggests users diverge most on bright, low-saturation inputs; a testable extension is to use that signal to recommend which photos a user should rate first to accelerate personalization.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper proposes FedPAIE, a federated framework for personalized aesthetic image enhancement and color grading. The pipeline first trains a lightweight dual-cue aesthetic scorer through Federated Aesthetic Preference Learning, calibrates it to a new user from a small local rated support set, freezes the calibrated scorer, and then uses it to guide adaptation of a CLUT-based enhancer on unpaired local photographs. Fidelity terms and an excess-gap penalty are included to limit proxy over-optimization. The authors report evaluations on MIT-Adobe FiveK and Flickr-AES, including scorer calibration accuracy, scorer-predicted preference gains, fidelity metrics relative to input and Expert C, ablations of the enhancement objective, and resource accounting. The central claim is that FedPAIE is the first federated method for personalized aesthetic image enhancement and that it achieves effective open-world personalization while keeping raw photos and ratings local.

Significance. If the personalization claim were independently validated, the paper would make a meaningful contribution: it connects federated preference learning to a lightweight, on-device image transformation without paired user retouches, and its staged design (global scorer, local calibration, frozen-scorer-guided enhancer adaptation) is technically coherent. The manuscript is unusually careful about resource accounting, cohort splits, matched ablations, and disclosure of evaluation scope. The explicit statement in Appendix H that blinded user evaluation is deferred is honest but exposes the main weakness: the reported preference gains are produced by the same scorer that is optimized and used for checkpoint selection, so the evidence for genuine personalization is currently indirect.

major comments (3)
  1. [§3.4, Eqs. (13)–(18); Tables 2 and 4] The core evidence of personalization is the scorer-predicted preference gain Δ = S_{θu}(E_{ϕu}(I)) − S_{θu}(I). This same frozen personalized scorer is (a) the training signal in Eq. (17), (b) the checkpoint-selection criterion in Eq. (18), and (c) the reported success metric in Tables 2 and 4. Positive Δ is therefore encouraged by construction; the paired user-level tests in Table 22 quantify consistency of this proxy, not of human preference. Appendix H explicitly defers blinded user evaluation. Without an independent perceptual ground truth, the reported gains could reflect generic brightness/saturation shifts that inflate the scorer rather than user-specific taste. This is load-bearing for the central claim of "effective open-world personalization." I recommend adding a human A/B preference study on held-out users, or substantially softening the personalization claims to "scorer-pred
  2. [§4.1 and Appendix E.2] The experimental protocol filters users by requiring validation SRCC ≥ 0.10 for the personalized scorer before enhancer adaptation. This means the reported 36/37-user cohorts consist of users whose scorers already correlate with their ratings. The open-world claim is thus conditional on this eligibility rule, and the evidence may not generalize to weaker or noisier preference signals. The paper should report results without this filtering (or a sensitivity analysis with different thresholds), and should discuss how the eligibility rule affects the interpretation of the average gains.
  3. [§4.3, Table 4; Appendix G.1] The ablation "without scorer guidance" converges to an identity-like transformation (Δ≈0.0002/0.0001), which is expected because only L1 and LPIPS remain. This confirms that the enhancer is indeed moving along scorer gradients, but it does not establish that the direction of movement is human-preferred. Similarly, the excess-gap penalty is shown to protect fidelity, but the preference axis in the trade-off is still the same proxy. The ablations are internally consistent, but they do not break the circularity identified above; they should be presented as validating the optimization behavior rather than as evidence of perceptual personalization.
minor comments (4)
  1. [Eq. (7)] The notation eE_ϕ(I) is awkward; consider using a tilde or a separate symbol for the unclipped output to improve readability.
  2. [Table 4] The abbreviations PC, SC, LC are used in the table but defined only in the text. Define them in the caption for self-containedness.
  3. [Appendix H] The limitation that blinded user evaluation is deferred appears only in the appendix. Given that it concerns the central claim, it should be acknowledged in the main text, e.g., in the experimental scope or conclusion.
  4. [Table 3] The AdaInt comparison uses a shared protocol but relies on Expert C fidelity and scorer-proxy differences. The text should reiterate that Expert C similarity is not a measure of user preference, to avoid overinterpretation.

Circularity Check

1 steps flagged

Reported preference gains are the same frozen-scorer objective used for enhancer training and checkpoint selection; no independent perceptual validation is provided.

specific steps
  1. fitted input called prediction [Section 3.4, Eqs. (13), (17), (18); Section 4.1; Appendix H]
    "... let ˆI = Eϕ(I) and define: s+u = Sθu(ˆI), s0u = Sθu(I), ∆u = s+u − s0u. (13) ... LEu = λpref Lpref + λaes Laes + λ1L1 + λpercLperc + λgapLgap. (17) ... The fixed-hyperparameter (fixed-HP) configuration sets γ1 = γp = 0 and therefore selects the checkpoint with the largest validation preference gain. ... Enhancement uses the scorer-predicted preference gain ∆u = Sθu(ˆI) − Sθu(I) as an optimization-aligned personalization proxy."

    The enhancer is trained to maximize Sθu on enhanced images: Eq. (17) includes Lpref = −E log σ(∆u) and Laes = −E s+u. The fixed-HP checkpoint selector (Eq. 18 with γ1 = γp = 0) chooses the checkpoint with the largest validation ∆. The same ∆ is then reported in Tables 2 and 4 as 'preference gain' and as evidence of personalization. Since Sθu is a fitted proxy rather than an independent judgment of enhanced-image preference, positive ∆ is the training objective itself, not an external prediction. The paper concedes this: 'The frozen personalized scorer supplies both the training signal and the reported scorer-predicted preference gain,' and defers 'Blinded user evaluation.' PSNR/SSIM/LPIPS are independent but measure fidelity or reference similarity, not taste. Thus the central personalizat

full rationale

The circularity is localized to the enhancement-preference claim and is acknowledged in the paper itself. The scorer's rating regression is independently tested on held-out ratings (Table 1), so the scorer is not circular as a rating predictor. However, the enhancer's success metric ∆ = Sθu(Eϕu(I)) − Sθu(I) is exactly what Eq. (17) maximizes and what the fixed-HP checkpoint rule in Eq. (18) selects on; reporting this same quantity as 'preference gain' in Tables 2 and 4 means the reported gains reduce to the fitted objective. The paper's own Appendix H states that the frozen personalized scorer 'supplies both the training signal and the reported scorer-predicted preference gain,' and it defers blinded user evaluation, which would be the missing independent ground truth. There is no load-bearing self-citation chain or imported uniqueness theorem; the novelty of the federated pipeline and the independent fidelity metrics are genuine. But because the headline evidence for 'personalized aesthetic preference' is the optimization-aligned proxy, the central claim is only partially independently supported. Score 6 reflects one prediction metric that reduces by construction, while other components retain independent content.

Axiom & Free-Parameter Ledger

10 free parameters · 8 axioms · 0 invented entities

The central claim rests mostly on domain assumptions: that Flickr-AES ratings represent usable preference signals, that a 10/100-shot calibrated scorer is a valid proxy for a user's taste, and that a FiveK-pretrained CLUT prior transfers to arbitrary users. The many tuned weights are documented in the paper and are not hidden, but they are fitted to the same validation ecosystem.

free parameters (10)
  • Global scorer parameters θ=(θ_c, θ_s, θ_f, τ) = 0.787M params; final model round 13
    Learned from federated Flickr-AES ratings; defines the initial preference model for all users.
  • Personalized scorer calibration blocks = 10-shot: 526,850 (fusion+τ); 100-shot: 787,202
    Fitted per user on private support ratings via regression, pairwise, and variance terms.
  • Generic CLUT enhancer prior (ψ_g, β_g) = 292,541 params, pretrained on FiveK paired retouches
    Supplies the transformation dictionary; not learned from user data but all personalization starts from it.
  • Personalized enhancer predictor ψ_u = 0.265M updated params; bases frozen
    Fitted on unpaired local photos by maximizing frozen scorer output under fidelity/gap constraints.
  • Calibration hyperparameters δ, ρ, κ, ε_c, N0 = 0.1, 0.7, 0.5, 0.01, 20
    Hand-chosen implementation constants controlling pair construction, variance preservation, collapse protection, and support mask.
  • Enhancer objective weights λ_pref, λ_aes, λ_1, λ_perc, λ_gap, μ = Tab. 7: e.g., 10-shot 0.0411/0.5996/0.1007/0.0543/0.5107/0.1048; 100-shot differs
    Tuned by shared HPO on validation partitions of eight identities; set the preference-fidelity trade-off.
  • Checkpoint selection penalties γ_1, γ_p = 0 in fixed-HP; positive in shared HPO
    Control validation-based selection; configuration-dependent.
  • Eligibility threshold and support sizes = validation SRCC ≥0.10; 10/100 support ratings
    Quality filter for scorer guidance; affects cohort and excludes weak personalizers.
  • Global model selection round = round 13
    Selected by peak validation SRCC; a model-selection choice.
  • Federated aggregation weights α_k = sqrt(m_k)/Σ sqrt(m_j)
    Chosen design to reduce domination by high-volume users; not data-fitted but affects global scorer.
axioms (8)
  • domain assumption Flickr-AES ratings capture meaningful user aesthetic preferences for color grading.
    The entire preference signal is derived from these crowdsourced ratings; if ratings reflect photo quality more than personal taste, the personalization target is misspecified.
  • domain assumption A scorer calibrated on 10-100 ratings generalizes to the user's preference on unseen photos.
    Few-shot personalization assumes the support set is representative.
  • domain assumption The frozen scorer's output is a valid proxy for user satisfaction with enhanced images.
    Central to the enhancer objective and evaluation; untested against human judgment.
  • domain assumption CLUT-Net pretrained on FiveK expert retouches provides a useful generic color-transform prior for arbitrary users.
    All personalization builds on this pretrained dictionary.
  • domain assumption LPIPS and L1 fidelity terms capture perceptually relevant preservation.
    Used to prevent proxy exploitation; LPIPS is a learned proxy, not ground truth.
  • domain assumption Federated optimization (square-root FedAvg) converges to a useful shared scorer under non-IID aesthetic data.
    Assumed by the framework; no convergence guarantees are given.
  • domain assumption Unpaired local photos are sufficient for enhancer adaptation despite no per-user retouch target.
    The method relies on scorer gradients on ordinary photos to transfer to retouching behavior.
  • domain assumption Rating normalization to [0,1] and cross-user regression are valid.
    Used to train a single global scorer across heterogeneous users.

pith-pipeline@v1.3.0-daily-deepseek · 34018 in / 16413 out tokens · 146928 ms · 2026-08-01T03:49:31.579033+00:00 · methodology

0 comments
read the original abstract

Personalized image enhancement should reflect individual aesthetic taste, yet learning such preferences commonly depends on private photos and ratings that are unsuitable for centralized collection. The task must infer preference from sparse, heterogeneous feedback and translate it into natural-looking color transformations on resource-constrained user devices. We introduce FedPAIE, a federated personalized aesthetic image enhancement framework for user-adaptive color grading without centralizing raw photos or ratings. FedPAIE trains a lightweight dual-cue aesthetic scorer, calibrates it into a personalized scorer on a small local support set, and freezes it to guide regularized adaptation of a lightweight CLUT enhancer from unpaired local photographs. Fidelity constraints and an excess-gap penalty regularize scorer-guided adaptation to limit proxy-score over-optimization while preserving content and natural appearance. Training remains lightweight throughout the pipeline: scorer learning updates at most 0.787M parameters, enhancer adaptation updates 0.265M, and inference retains only a 0.293M-parameter personalized enhancer. Experiments on MIT-Adobe FiveK and Flickr-AES demonstrate effective open-world personalization and a favorable balance between user preference and image fidelity. FedPAIE thus connects decentralized preference learning with efficient personalized image transformation without requiring paired user retouches.

Figures

Figures reproduced from arXiv: 2607.27659 by Adnan Mahmood, Chuanzhi Xu, Haolan Guo, Jean Julien KNell, Weidong Cai, Xuanhua Yin, Yanrong Chen, Ziyuan Tao.

Figure 1
Figure 1. Figure 1: FedPAIE learns user-specific color grading without [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of FedPAIE. Global initialization learns the Lightweight Dual-Cue Aesthetic Scorer and CLUT enhancer [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Five shared-input FiveK comparisons. The CLUT enhancer obtained through Generic Enhancement Prior Learning [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative objective ablation for user 204 [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Scorer-predicted preference-gain–fidelity trade-off in the objective ablation for Frozen-Scorer-Guided Enhancer [PITH_FULL_IMAGE:figures/full_fig_p018_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Cross-distribution analysis for the objective ablation of Frozen-Scorer-Guided Enhancer Adaptation. Each point [PITH_FULL_IMAGE:figures/full_fig_p020_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Full-image and detail-crop comparison for Examples A–C. Yellow boxes mark the enlarged regions. The crops make [PITH_FULL_IMAGE:figures/full_fig_p021_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Additional preference context and Full-objective [PITH_FULL_IMAGE:figures/full_fig_p022_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Pairwise output diversity over the full 100-shot evaluation cohort. Each cell is the mean absolute RGB difference [PITH_FULL_IMAGE:figures/full_fig_p023_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Additional 100-shot objective-ablation cases for user 199 (left) and user 26 (right). Within each case, panels show [PITH_FULL_IMAGE:figures/full_fig_p024_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Randomly selected shared-input comparison for 100-shot Clients 194, 199, 200, 204, and 210. Each row contains a [PITH_FULL_IMAGE:figures/full_fig_p026_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Attribute changes from the input for the Generic Enhancement Prior and five main-paper clients. Bars and error bars [PITH_FULL_IMAGE:figures/full_fig_p026_13.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

46 extracted references · 3 linked inside Pith

  1. [1]

    Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , pages =

    Learning Photographic Global Tonal Adjustment with a Database of Input/Output Image Pairs , author =. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , pages =. 2011 , doi =

    2011

  2. [2]

    ACM Transactions on Graphics , volume =

    Deep Bilateral Learning for Real-Time Image Enhancement , author =. ACM Transactions on Graphics , volume =. 2017 , doi =

    2017

  3. [3]

    Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , pages =

    Deep Photo Enhancer: Unpaired Learning for Image Enhancement from Photographs with GANs , author =. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , pages =. 2018 , doi =

    2018

  4. [4]

    Proceedings of the 26th ACM International Conference on Multimedia , pages =

    Aesthetic-Driven Image Enhancement by Adversarial Learning , author =. Proceedings of the 26th ACM International Conference on Multimedia , pages =. 2018 , doi =

    2018

  5. [5]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , volume =

    Learning Image-Adaptive 3D Lookup Tables for High Performance Photo Enhancement in Real-Time , author =. IEEE Transactions on Pattern Analysis and Machine Intelligence , volume =. 2022 , doi =

    2022

  6. [6]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    AdaInt: Learning Adaptive Intervals for 3D Lookup Tables on Real-Time Image Enhancement , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =. 2022 , doi =

    2022

  7. [7]

    Proceedings of the 30th ACM International Conference on Multimedia , pages =

    CLUT-Net: Learning Adaptively Compressed Representations of 3DLUTs for Lightweight Image Enhancement , author =. Proceedings of the 30th ACM International Conference on Multimedia , pages =. 2022 , publisher =

    2022

  8. [8]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages =

    Lightweight and Fast Real-time Image Enhancement via Decomposition of the Spatial-aware Lookup Tables , author =. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages =. 2025 , doi =

    2025

  9. [9]

    IEEE Transactions on Image Processing , volume =

    NIMA: Neural Image Assessment , author =. IEEE Transactions on Image Processing , volume =. 2018 , doi =

    2018

  10. [10]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    VILA: Learning Image Aesthetics from User Comments with Vision-Language Pretraining , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =. 2023 , doi =

    2023

  11. [11]

    2603.12575 , archivePrefix =

    Yin, Xuanhua and Xu, Chuanzhi and Zhou, Haoxian and Wei, Boyu and Cai, Weidong , year =. 2603.12575 , archivePrefix =

    Pith/arXiv arXiv

  12. [12]

    2605.05155 , archivePrefix =

    Xu, Chuanzhi and Wei, Boyu and Zhou, Haoxian and Yin, Xuanhua and Deng, Zihan and Chen, Haodong and Qu, Qiang and Cai, Weidong , year =. 2605.05155 , archivePrefix =

    Pith/arXiv arXiv

  13. [13]

    Proceedings of the IEEE International Conference on Computer Vision , pages =

    Personalized Image Aesthetics , author =. Proceedings of the IEEE International Conference on Computer Vision , pages =. 2017 , doi =

    2017

  14. [14]

    Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence , pages =

    Collaborative and Attentive Learning for Personalized Image Aesthetic Assessment , author =. Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence , pages =. 2018 , doi =

    2018

  15. [15]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    Personalized Image Aesthetics Assessment With Rich Attributes , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =. 2022 , doi =

    2022

  16. [16]

    IEEE Transactions on Cybernetics , volume =

    Personalized Image Aesthetics Assessment via Meta-Learning With Bilevel Gradient Optimization , author =. IEEE Transactions on Cybernetics , volume =. 2022 , doi =

    2022

  17. [17]

    Knowledge-Based Systems , volume =

    Personalized Image Aesthetics Assessment based on Graph Neural Network and Collaborative Filtering , author =. Knowledge-Based Systems , volume =. 2024 , doi =

    2024

  18. [18]

    IEEE Transactions on Multimedia , volume =

    Multi-Level Transitional Contrast Learning for Personalized Image Aesthetics Assessment , author =. IEEE Transactions on Multimedia , volume =. 2024 , doi =

    2024

  19. [19]

    Computer Vision -- ECCV 2024: 18th European Conference, Milan, Italy, September 29--October 4, 2024, Proceedings, Part XL , series =

    Scaling Up Personalized Image Aesthetic Assessment via Task Vector Customization , author =. Computer Vision -- ECCV 2024: 18th European Conference, Milan, Italy, September 29--October 4, 2024, Proceedings, Part XL , series =. 2024 , publisher =

    2024

  20. [20]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    Rethinking Personalized Aesthetics Assessment: Employing Physique Aesthetics Assessment as An Exemplification , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =. 2025 , doi =

    2025

  21. [21]

    2025 , doi =

    Maerten, Anne-Sofie and Chen, Li-Wei and De Winter, Stefanie and Bossens, Christophe and Wagemans, Johan , booktitle =. 2025 , doi =

    2025

  22. [22]

    Advances in Neural Information Processing Systems , volume =

    Deep Reinforcement Learning from Human Preferences , author =. Advances in Neural Information Processing Systems , volume =

  23. [23]

    IEEE Transactions on Multimedia , volume =

    User-Guided Personalized Image Aesthetic Assessment Based on Deep Reinforcement Learning , author =. IEEE Transactions on Multimedia , volume =. 2023 , doi =

    2023

  24. [24]

    Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , pages =

    Personalization of Image Enhancement , author =. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , pages =. 2010 , doi =

    2010

  25. [25]

    Computer Vision -- ECCV 2020 , series =

    PieNet: Personalized Image Enhancement Network , author =. Computer Vision -- ECCV 2020 , series =. 2020 , publisher =

    2020

  26. [26]

    IEEE Transactions on Image Processing , volume =

    Personalized Image Enhancement Using Neural Spline Color Transforms , author =. IEEE Transactions on Image Processing , volume =. 2020 , doi =

    2020

  27. [27]

    IEEE Transactions on Circuits and Systems for Video Technology , volume =

    Personalized Image Enhancement Featuring Masked Style Modeling , author =. IEEE Transactions on Circuits and Systems for Video Technology , volume =. 2024 , doi =

    2024

  28. [28]

    2025 International Technical Conference on Circuits/Systems, Computers, and Communications (

    Personalized Image Enhancement Using Global and Local Style Information , author =. 2025 International Technical Conference on Circuits/Systems, Computers, and Communications (. 2025 , publisher =

    2025

  29. [29]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    Learning Personalized Photographic Style from Pairwise User Preferences , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

  30. [30]

    2026 , doi =

    Chang, Zewei and Duan, Zheng-Peng and Zhang, Jianxing and Guo, Chun-Le and Liu, Siyu and Chun, Hyungju and Park, Hyunhee and Liu, Zikun and Li, Chongyi , journal =. 2026 , doi =

    2026

  31. [31]

    Proceedings of the 20th International Conference on Artificial Intelligence and Statistics , volume =

    Communication-Efficient Learning of Deep Networks from Decentralized Data , author =. Proceedings of the 20th International Conference on Artificial Intelligence and Statistics , volume =. 2017 , publisher =

    2017

  32. [32]

    Proceedings of Machine Learning and Systems , volume =

    Federated Optimization in Heterogeneous Networks , author =. Proceedings of Machine Learning and Systems , volume =

  33. [33]

    Proceedings of the 36th International Conference on Machine Learning , pages =

    Agnostic Federated Learning , author =. Proceedings of the 36th International Conference on Machine Learning , pages =. 2019 , volume =

    2019

  34. [34]

    2019 , eprint =

    Federated Collaborative Filtering for Privacy-Preserving Personalized Recommendation System , author =. 2019 , eprint =

    2019

  35. [35]

    IEEE Intelligent Systems , volume =

    Secure Federated Matrix Factorization , author =. IEEE Intelligent Systems , volume =. 2021 , doi =

    2021

  36. [36]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume =

    FedRec++: Lossless Federated Recommendation with Explicit Feedback , author =. Proceedings of the AAAI Conference on Artificial Intelligence , volume =. 2021 , doi =

    2021

  37. [37]

    Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing , pages =

    Efficient-FedRec: Efficient Federated Learning Framework for Privacy-Preserving News Recommendation , author =. Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing , pages =. 2021 , doi =

    2021

  38. [38]

    Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence , pages =

    Federated Probabilistic Preference Distribution Modelling with Compactness Co-Clustering for Privacy-Preserving Multi-Domain Recommendation , author =. Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence , pages =. 2023 , doi =

    2023

  39. [39]

    Proceedings of the IEEE International Conference on Multimedia and Expo , pages =

    Federated Learning for Personalized Image Aesthetics Assessment , author =. Proceedings of the IEEE International Conference on Multimedia and Expo , pages =. 2023 , doi =

    2023

  40. [40]

    2020 , doi =

    Liu, Yang and Huang, Anbu and Luo, Yun and Huang, He and Liu, Youzhi and Chen, Yuanyuan and Feng, Lican and Chen, Tianjian and Yu, Han and Yang, Qiang , journal =. 2020 , doi =

    2020

  41. [41]

    2111.11066 , archivePrefix =

    He, Chaoyang and Shah, Alay Dilipbhai and Tang, Zhenheng and Fan, Di and Sivashunmugam, Adarshan Naiynar and Bhogaraju, Keerti and Shimpi, Mita and Shen, Li and Chu, Xiaowen and Soltanolkotabi, Mahdi and Avestimehr, Salman , year =. 2111.11066 , archivePrefix =

    Pith/arXiv arXiv

  42. [42]

    Foundations and Trends in Machine Learning , volume =

    Advances and Open Problems in Federated Learning , author =. Foundations and Trends in Machine Learning , volume =. 2021 , doi =

    2021

  43. [43]

    2016 , eprint =

    Concrete Problems in AI Safety , author =. 2016 , eprint =

    2016

  44. [44]

    2025 , eprint =

    FedVideoMAE: Efficient Privacy-Preserving Federated Video Moderation , author =. 2025 , eprint =

    2025

  45. [45]

    and Adam, Hartwig , title =

    Howard, Andrew and Sandler, Mark and Chu, Grace and Chen, Liang-Chieh and Chen, Bo and Tan, Mingxing and Wang, Weijun and Zhu, Yukun and Pang, Ruoming and Vasudevan, Vijay and Le, Quoc V. and Adam, Hartwig , title =. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages =. 2019 , doi =

    2019

  46. [46]

    and Shechtman, Eli and Wang, Oliver , title =

    Zhang, Richard and Isola, Phillip and Efros, Alexei A. and Shechtman, Eli and Wang, Oliver , title =. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , pages =. 2018 , doi =

    2018