PaperlayerView exact revision on arXiv

Source-linked AI summary

PrefLUT: Reusable and Refinable Personalized Color Editing from Pairwise Preferences

Chuanzhi Xu, Langyi Chen, Chengkun Yue, Xuanhua Yin, Boyu Wei, Qingwen Zeng, Zihan Deng, Weidong Cai

arXiv:2609.34133v1cs.CVcs.AI

TL;DR

Most LUT and reference-guided methods target a specified appearance rather than persistent preferences from repeated user choices. PrefLUT builds a reusable, refinable profile from ordered preference pairs and predicts query-conditioned 3D LUTs; on PPSD it leads the reported fidelity comparisons and passes all five PCVP tests.

  • Problem

    Most LUT and reference-guided methods target a specified appearance rather than persistent preferences from repeated user choices, which may be stable only along selected color dimensions and must be separated from unrelated image content.

    Source: §1 · p. 1

  • Method

    PrefLUT encodes ordered preference pairs into a reusable profile, combines it with each query image to predict a LUT and edit strength, and refines the profile with additional pairs without per-user optimization.

    Source: §1 · p. 2

  • Results

    On PPSD, PrefLUT leads all four direct-fidelity metrics and CQS under ∆E00, LPIPS, and SSIM, passes all five PCVP tests, and edits at 1.365 ms/image.

    Source: §6 · pp. 9–10

  • Takeaways & Limitations

    Each quantized profile requires only 260 bytes and supports repeated editing without re-encoding reference images.

    Source: §6 · p. 9

    PPSD's official user split and episode generator are unavailable, and validation estimates may be optimistic because validation users inform checkpoint selection.

    Source: G LIMITATIONS · p. 28

Abstract

from arXiv · show

Photographic color editing is inherently personal: the same image can appear too warm, too muted, or already satisfactory to different users. Most lookup table (LUT) and reference-guided methods target a specified appearance rather than model persistent preferences from repeated user choices. To address this gap, we introduce PrefLUT, a reusable and refinable user-preference modeling framework for deployable 3D LUTs, encoding ordered preferred/non-preferred image pairs into a lightweight Reusable User Profile that is reused across queries and refined using additional user preference pairs, without per-user optimization. A Query-Conditioned LUT Predictor combines this profile with each image to predict a LUT latent vector and edit strength. An Identity-Residual LUT Decoder and Edit-Strength Controller then produce an exportable 3D LUT. Experiments on three datasets demonstrate effective personalized editing and general-purpose enhancement. Each quantized profile requires only 260 bytes, and editing takes 1.365 ms/image on an RTX 5090 GPU. We also introduce the Preference-Conditioning Verification Protocol (PCVP), an evaluation protocol to verify whether personalized image edits depend on user preferences and the query image through controlled changes to user profiles, preference orders, pair correspondences, and query images.

1 INTRODUCTION

Most color-editing methods target a shared appearance or a single reference, while persistent personal preferences are expressed through repeated choices and may vary by color dimension. PrefLUT turns ordered preference pairs into reusable profiles and predicts query-conditioned LUTs; experiments report personalized editing gains and preference-conditioning verification.

1 INTRODUCTION

  • Most learned-LUT methods optimize a shared target, while reference-guided methods use one style image to specify an appearance.These approaches do not directly model preferences expressed through repeated user choices.Source: §1 · p. 1
  • Personal preferences can be stable only along selected color dimensions, must be separated from unrelated image content, and can be refined as feedback accumulates.PPSD evaluates whether outputs remain faithful while moving toward a user's preferred appearance, but its baselines do not explicitly output LUTs.Source: §1 · p. 1
  • PCVP tests whether personalized edits reflect the intended user's preferences and the query image rather than tendencies shared across users.The question arises because preference-alignment scores alone cannot distinguish user-specific preferences from shared editing tendencies.Source: §1 · p. 1
  • PrefLUT encodes ordered preference pairs into a reusable profile, combines it with each query image to predict a LUT and edit strength, and refines the profile with new pairs while network weights remain frozen.The framework uses an Ordered Preference Encoder and Preference Set Aggregator for profile construction, then an Identity-Residual LUT Decoder and Edit-Strength Controller for the edit.Source: §1 · p. 2
  • PrefLUT leads all four direct fidelity metrics and achieves the highest CQS under ∆E00, LPIPS, and SSIM among compared PPSD methods, while passing all five PCVP tests.Additional preference feedback improves editing on unseen queries through profile refinement with frozen network weights.Source: §1 · p. 2

2 RELATED WORK

Prior LUT, preset, reference-guided, and personalized enhancement methods offer reusable or flexible color transforms, but the cited approaches do not combine persistent preferences from repeated ordered choices with refinable, exportable LUT editing. PrefLUT is presented as that combination.

2 RELATED WORK

  • Learned LUT and preset methods provide lightweight or reusable color transforms, but typically target a shared appearance, fixed styles, or a single reference image.The supplied passage names Image-Adaptive 3D-LUT, Spatial-Aware 3D LUT, 4D LUT, BGrid, AdaInt, SepLUT, CLUT-Net, NILUT, neural LUT encoding, Deep Preset, AdaCM, and Neural Preset.Source: §2 · pp. 2–3
  • Reference-guided and generative LUT methods condition each edit independently, whereas PrefLUT aggregates repeated ordered choices into a persistent profile and reuses it for query-conditioned LUT prediction.PrefLUT updates the profile when feedback arrives and does not reprocess preference references at query time.Source: §2 · p. 3
  • Personalized enhancement methods represent preferences or use user feedback, but the cited methods do not jointly support pairwise-feedback profile refinement and exportable LUT-based editing.PrefLUT uses a compact reusable profile refined from preference pairs to predict exportable, query-conditioned 3D LUTs.Source: §2 · p. 3

3 PREFLUT

PrefLUT constructs a reusable profile from ordered preference pairs, then combines it with each query image to produce a personalized, exportable 3D LUT. Its training uses separate query pairs and a wrong-user contrast to encourage preference-consistent, user-specific edits.

3 PREFLUT

  • PrefLUT separates profile construction from query-time editing, predicting a LUT latent vector and edit strength from a user profile and a new image.The resulting LUT is applied by trilinear mapping and can be saved for images at different resolutions.Source: §3 · pp. 3–4, §3.2 · p. 4

3.1 PROFILE CONSTRUCTION

  • Ordered pairs become direction-aware tokens, which an order-invariant aggregator summarizes into a reusable profile that can be updated from new feedback with network weights frozen.The retained feedback may be cumulative or drawn from a recent window.Source: §3.1 · p. 4

3.2 QUERY-TIME EDITING

  • A low-dimensional decoder predicts a clipped residual around the identity LUT, and the query-conditioned edit strength scales that residual before application.The decoder is pretrained on fitted target LUTs and frozen during personalized training; preference training therefore does not require a target LUT for every pair.Source: §3.2 · pp. 4–5

3.3 TRAINING WITH SEPARATE QUERY PAIRS

  • Training uses separate query pairs and color-statistics distances to rank edits toward preferred appearances, while a wrong-user contrast compares outputs for the same query under correct and incorrect profiles.Auxiliary losses address aligned-pair reconstruction, preservation, edit-strength supervision, and limited deviation from the identity LUT.Source: §3.3 · p. 5

4 PREFERENCE-CONDITIONING VERIFICATION PROTOCOL

PCVP tests whether personalized edits respond to preference evidence and the current query by changing one condition at a time while holding evaluation procedures fixed. It uses five controls and requires positive paired gains across every metric and control for a full pass.

Evaluation Principle

  • PCVP holds the method, users, queries, image processing, and metric computation fixed, changing one condition at a time so each quality difference measures response to that condition.It is intended for methods conditioned on preference examples or a user representation derived from them.Source: §4 · p. 6

PCVP Controls

  • The five controls alter the user profile, preference order, pair correspondence, population profile, or conditioning query to test distinct sources of personalization.The Wrong Query control uses a different query from the same user while retaining the original query on any separate image-application path.Source: §4 · p. 6

PCVP pass criterion

  • A control passes only when every metric’s paired 95% bootstrap interval has a lower endpoint above zero; a full PCVP pass requires all five controls to pass.Repeated preregistered episodes must show the same direction for the repeated estimate.Source: §4 · p. 6

5 EXPERIMENTS AND RESULTS

PrefLUT is evaluated for personalized editing on PPSD, preference-conditioning dependence, profile refinement, and general-purpose enhancement on FiveK and PPR10K. The reported results show strong fidelity and preference alignment, sensitivity to intended user and query conditions, and improvement from accumulating feedback.

PPSD Personalized Editing

  • PrefLUT leads the compared PPSD methods in CQS under ∆E00, LPIPS, and SSIM and in direct fidelity to Preferred under ∆E00, LPIPS, PSNR, and SSIM.Qualitative examples show it often follows users’ differing brightness and color preferences more closely than PIE-MSM and DiffRetouch.Source: §5 · p. 8

PCVP Evaluation

  • PrefLUT passes all five PCVP controls across four CQS measures, while none of eight baseline adaptations passes all five.All five controls significantly reduce CQS relative to the correct condition.Source: §5 · p. 8, Table 2 · p. 8

Profile Refinement

  • With network weights and unseen queries fixed, adding feedback improves editing through profile updates alone.Correct new feedback also outperforms reversed new feedback on identical images.Source: §5 · p. 9

Ablation Studies

  • Removing explicit feature differences or replacing Set Transformer aggregation with masked mean pooling reduces CQS and weakens the correct-versus-wrong-user advantage.Removing the query feature weakens user-specific editing despite higher PSNR CQS; removing Wrong-User Contrast Loss further improves wrong-user PCVP gain.Source: §5 · p. 9, Table 3 · p. 9

Extended Experiments

  • On FiveK Expert C, PrefLUT achieves 24.572 dB PSNR, 0.917 SSIM, 7.833 ∆E76, and 0.067 LPIPS, averaged over eight training seeds.The experiments also evaluate general-purpose enhancement on PPR10K Experts A/B/C.Source: §5 · p. 9
  • On PPR10K, one shared model across Experts A/B/C achieves 25.105 dB PSNR, 7.822 ∆E76, 28.368 dB PSNR-HC, and 5.069 ∆E76-HC.Results average three experts and two training seeds on 2,286 validation images per expert.Source: §5 · p. 9

6 CONCLUSION

The conclusion presents PrefLUT as a reusable, refinable preference-profile framework for query-conditioned 3D LUT editing, alongside PCVP for testing preference and query dependence. The reported results include all five PCVP tests passed, leading PPSD comparisons on specified measures, fast editing, and general-purpose enhancement on FiveK and PPR10K.

6 CONCLUSION

  • PrefLUT encodes ordered preferences in a reusable, refinable profile and predicts query-conditioned 3D LUTs without per-user fine-tuning.PCVP tests dependence on preferences and the query image through five controlled tests.Source: §6 · p. 9
  • Each quantized profile requires 260 bytes and supports repeated editing without re-encoding reference images.Source: §6 · p. 9
  • PrefLUT passes all five PCVP tests and leads compared personalized PPSD methods on four direct-fidelity metrics and CQS under ∆E00, LPIPS, and SSIM, with editing at 1.365 ms/image.FiveK and PPR10K experiments also demonstrate general-purpose enhancement across expert targets.Source: §6 · pp. 9–10
  • The paper uses released preference and retouched-image datasets, conducts no new human-participant interaction, and does not redistribute raw user data.It states that dataset licenses are followed and evaluation procedures are documented in the appendix.Source: ETHICS STATEMENT · p. 10

A METHOD DETAILS

The appendix details profile construction, query-time editing, batching, LUT preparation, training losses, and benchmark settings. Together, these passages specify how preference pairs become reusable profiles and how the resulting LUT editing system is trained and evaluated.

A.1 PROFILE CONSTRUCTION AND QUERY-TIME EDITING

  • Profile construction encodes ordered reference pairs, aggregates valid preference tokens into a stored profile, and query-time editing combines that profile with a query feature to predict a LUT latent vector and edit strength.The decoded LUT residual is scaled from identity, applied by trilinear interpolation, and refinement rebuilds the profile from added pairs with frozen weights for reuse on later queries.Source: §A.1 · p. 13

A.2 BATCHING AND VALID-PAIR MASKING

  • Batching pads reference sets and masks padded pairs; valid tokens use a fixed feature-based order, while permutation invariance comes from the set architecture without positional embeddings.Source: §A.2 · p. 13

A.3 TARGET LUT FITTING; A.4 IDENTITY-RESIDUAL LUT DECODER PRETRAINING

  • For pixel-aligned preference pairs, target LUTs map non-preferred images toward preferred images and are fitted on 128-pixel crops with reconstruction and regularization objectives.The fitted LUTs supervise decoder pretraining; the Identity-Residual LUT Decoder is trained as an autoencoder, then frozen during personalized training, which uses image-based supervision, edit-strength supervision, and LUT regularization without fitted target LUTs for preference pairs.Source: §A.3 · pp. 13–14, §A.4 · p. 14

A.5 GLOBAL COLOR-STATISTICS DESCRIPTOR

  • The global color-statistics descriptor uses means and standard deviations for RGB, luminance, and saturation, with equal-weight absolute differences across ten components.These global statistics reduce sensitivity to spatial misalignment but do not impose pixelwise correspondence; aligned pairs also receive reconstruction supervision.Source: §A.5 · p. 14

A.6 AUXILIARY TRAINING LOSSES

  • Auxiliary training losses include aligned-pair reconstruction, preservation of preferred queries, binary edit-strength supervision, and penalization of deviations from the identity LUT.The reconstruction loss is zero when a minibatch has no aligned query pairs.Source: §A.6 · pp. 14–15

A.7 WRONG-USER PROFILE SELECTION

  • Wrong-User Contrast compares outputs for the same query and preferred target while replacing the current user’s profile with another user’s profile.The other profile is selected by a fixed nonzero cyclic shift among distinct users in the minibatch.Source: §A.7 · p. 15

B.1 PPSD

  • PPSD evaluation uses pairwise choices rather than absolute aesthetic scores, with a deterministic split of 471 training users and 50 validation users and disjoint reference and query pairs, scenes, and images.The main quality and PCVP episodes use 16 reference and 16 query pairs per user; CQS combines base fidelity with comparative margin ratio.Source: §B.1 · pp. 15–16

B.2 MIT-ADOBE FIVEK; B.3 PPR10K

  • FiveK Expert C uses 4,500 training images and 500 test images, with one profile built from 16 fixed training pairs and results averaged over eight predefined seeds.PPR10K evaluates one shared model across Experts A, B, and C on the released 2,286-image validation split; results average two training seeds.Source: §B.2 · p. 16, §B.3 · p. 16

B.4 METRICS AND STATISTICAL TESTING

This section defines the quality metrics and statistical procedures used across PPSD, FiveK, and PPR10K, then details PrefLUT and baseline configurations. It also reports a validation-selected PrefLUT checkpoint and describes the evaluation settings.

B.4 METRICS AND STATISTICAL TESTING

  • CQS combines BFS and CMR separately for ΔE00, LPIPS, PSNR, and SSIM; higher values are better, and positive CMR indicates closer agreement with the preferred target.For aggregate results, raw scores are averaged across preferred and non-preferred inputs and query pairs before the nonlinear CQS calculation.Source: §B.4 · pp. 16–17
  • PPSD metrics include ΔE00, LPIPS, PSNR, and SSIM, while FiveK and PPR10K use their specified PSNR, SSIM, ΔE76, Alex-LPIPS, and human-centered variants.FiveK reports means and sample standard deviations across eight seeds; PPR10K averages results over two seeds and then equally across three experts.Source: §B.4 · p. 17
  • PPSD uses paired user-bootstrap intervals, resampling 50 users with replacement and including all queries for each selected user under both conditions.A paired comparison is statistically positive when the lower endpoint of its 95% user-bootstrap interval is above zero.Source: §B.4 · p. 17

C.1 PREFLUT SETTINGS

  • PrefLUT’s standard configuration encodes 16 ordered reference pairs into a 256-wide profile and combines that profile with query features to predict a LUT latent vector and edit strength.The decoder maps the 256-dimensional latent vector to a 173 × 4 × 4 × 4 feature volume and adds a scaled residual to the identity LUT.Source: §C.1 · p. 17
  • The LUT autoencoder is pretrained for 50 epochs on 5,711 fitted training LUTs, followed by a 30-epoch pair-to-LUT pretraining stage using the same aligned pairs and fitted LUTs.Personalized training optimizes the aggregator and prediction heads for 100 epochs; the ordered preference encoder, query encoder, and decoder remain frozen.Source: §C.1 · pp. 17–18
  • Epoch 91 gives the highest CQS on all four measures among PrefLUT’s five lowest-validation-loss checkpoints, and is used for PPSD quality, PCVP, and profile refinement.PPSD quality results are validation estimates after checkpoint selection on the same users and evaluation episode.Source: §C.1 · p. 18

C.2 BASELINE REPRODUCTION AND ADAPTATION

  • Eight baseline adaptations use ten PPSD training epochs without validation-based checkpoint selection, while PieNet additionally receives ten epochs of triplet pretraining.Baseline evaluation uses 50 validation users and 800 query pairs; PPSD training uses 471 users with disjoint reference and query pairs.Source: §C.2 · pp. 18–19

C.2.1 PERSONALIZED ENHANCEMENT METHODS NOT COMPARED ON PPSD

This block explains exclusions from PPSD comparisons, extended expert-target configurations, efficiency measurements, and the controlled PCVP tests. The PCVP changes profile or query conditions while holding other evaluation elements fixed.

C.2.1 PERSONALIZED ENHANCEMENT METHODS NOT COMPARED ON PPSD

  • RefRetouch is excluded from quantitative PPSD comparisons because its original/retouched-pair interface and training objective require adaptation to ordered preferences that need not be pixel-aligned.FedPAIE is also outside the comparison because PPSD lacks the scalar rating targets its scorer supervision requires, and adaptation would require redefining its supervision and client-training protocol.Source: §C.2.1 · pp. 19–20

C.3.1 FIVEK EXPERT-TARGET ENHANCEMENT

  • The FiveK PrefLUT configuration uses 6.127M parameters and is evaluated on all 500 test images after architecture and training-duration selection on the training-image validation subset.The PPR10K configuration uses 17.156M parameters, trains jointly for three experts, and is evaluated on 2,286 released validation images per expert.Source: §C.3.1 · p. 20, §C.3.2 · pp. 20–21

C.3.2 PPR10K EXPERT-TARGET ENHANCEMENT

  • The 17.156M-parameter PrefLUT configuration jointly trains monotone color curves, a rank-64 173 LUT, an ImageNet-pretrained ResNet-18 query encoder, and spatial residual; each expert profile uses the same 16 training images, with the retouched target preferred over the source.Evaluation uses the released 360p validation set of 2,286 source images, averaging equally the epoch-8 checkpoints from seeds 2034 and 2036.Source: §C.3.2 · pp. 20–21

C.4 EFFICIENCY MEASUREMENT

  • PrefLUT’s compact deployment stores an int8 profile in 260 bytes per user, including its four-byte FP32 scale; the quality tables instead report an unquantized 1,024-byte profile.Quantization and dequantization are included in timed stages.Source: §C.4 · p. 21
  • The separate profile-reuse benchmark takes 77.204236 ms when reference sets are encoded for all 64 queries and 8.437795 ms when two reusable profiles are reused, a 9.150× ratio.Including initial reference encoding yields 10.569326 ms for the first batch and a 7.305× speedup; this benchmark excludes LUT decoding and image application.Source: §C.4 · p. 21

D PREFERENCE-CONDITIONING VERIFICATION PROTOCOL (PCVP)

  • PCVP tests whether editing quality benefits from the intended user’s profile, preference order, original pair correspondence, individual rather than mean preferences, and the current query image.Each test changes one condition while keeping model parameters, users, query-target pairs, image processing, and metric computation fixed.Source: D PREFERENCE-CONDITIONING VERIFICATION… · p. 21, §D.1 · p. 22

D.1 DETAILED PCVP CONTROLS

  • The four profile controls respectively substitute another user’s profile, reverse preference order, mismatch pair correspondence, or replace individual profiles with a training-user mean.The mean profile is built from training users only, with equal user weights.Source: §D.1.1 · p. 22, §D.1.2 · p. 22, §D.1.3 · p. 22

D.1.5 CYCLIC WRONG-QUERY CONTROL

  • The Cyclic Wrong-Query Control substitutes a different query’s conditioning image while retaining the original query on any separate image-application path.For PrefLUT, the substituted query determines the predicted LUT and edit strength, and both transforms are applied to the same original image.Source: §D.1.5 · p. 23

D.2 EVALUATION CRITERION AND PPSD SETTINGS

This section defines PCVP gains and pass criteria, then specifies the PPSD evaluation and repeated-episode settings.

D.2 EVALUATION CRITERION AND PPSD SETTINGS

  • A PCVP pass for a control requires the lower endpoint of the paired 95% bootstrap interval to be above zero for every evaluated metric.On PPSD, the metrics are CQS values for ΔE00, LPIPS, PSNR, and SSIM; positive gains favor the correct condition.Source: §D.2 · p. 23
  • A full PCVP pass requires 4/4 metrics to pass for each of all five controls; the pass count measures conditioning evidence, not absolute image quality.PCVP controls change profile or query conditions with model parameters fixed, unlike architecture ablations.Source: §D.2 · p. 23
  • Baseline PPSD evaluations use 50 validation users, 16 reference pairs and 16 query pairs per user, and 250,000 paired user-bootstrap resamples.The same checkpoint and sampled reference and query sets are used for each baseline control and its correct condition.Source: §D.2 · p. 23
  • PrefLUT’s reversed-order and mismatched-pair tests use 256-pixel inputs and 100,000 resamples, while its other specified controls use 250,000 resamples.Three additional episodes with seeds 2029–2031 repeat the fixed wrong-user, reversed-order, and wrong-query tests with the trained model and validation users fixed.Source: §D.2 · pp. 23–24

E.1 PPSD PCVP RESULTS

The results report PCVP outcomes and their stability, distinguish protocols for FiveK and PPR10K comparisons, and describe the profile-refinement evaluation. The supplied passages specify refinement comparisons but do not provide their numerical outcomes.

E.1 PPSD PCVP RESULTS

  • StarEnhancer’s mismatched-pair control is not applicable because it preserves both separately averaged style centers; its 1/4 count covers applicable controls, while full PCVP requires all five.It passes one of its four applicable controls.Source: §E.1 · p. 24
  • StarEnhancer, DiffRetouch, and PerTouch pass all four metrics for the wrong-query control, but none passes all four for either wrong-user or training-user-mean profiles.PerTouch’s query intervention also changes its VAE image condition, so its query gains include image-content dependence.Source: §E.1 · p. 24
  • PIE-MSM passes the reversed-order, mismatched-pair, and wrong-query tests, but its correct profile does not improve all four metrics over another user or the training-user mean.The reported conclusions apply to the evaluated PPSD adaptations and conditioning interfaces, not necessarily the original methods.Source: §E.1 · p. 24

E.1.1 PREFLUT STABILITY ACROSS EVALUATION EPISODES

  • Across three additional evaluation episodes, PrefLUT passes all four metrics in the reversed-order and wrong-query tests and passes three, three, and four metrics in the wrong-user test.The model and validation users remain fixed, so this measures evaluation-sampling sensitivity rather than variation across independently trained models.Source: §E.1.1 · p. 24

E.2 ENHANCEMENT ON FIVEK AND PPR10K

  • FiveK and PPR10K comparisons separate published unseen-expert protocols from PrefLUT’s expert-target evaluations, so the tables do not provide a shared ranking.FiveK evaluates Expert C on 500 held-out images; PPR10K evaluates a shared model for three experts on 2,286 images per expert, with differing image resolutions and protocols.Source: §E.2 · p. 24, §E.2.1 · p. 25, §E.2.2 · p. 25

F.1 PROFILE REFINEMENT

  • The refinement experiment compares cumulative feedback against a frozen profile built from four pairs, with the primary comparison at 16 accumulated pairs.Profiles are rebuilt from retained ordered pairs with network weights frozen; Table F1 reports paired gains, but the supplied passages do not state their values.Source: §F.1 · pp. 25–26, Table F1 · p. 26

F.2 CORE ABLATIONS

Core ablations identify components associated with PCVP gains, while profile-width tests find a compact quality-retaining configuration; the reported PPSD estimates have user-split and checkpoint-selection caveats.

F.2 CORE ABLATIONS

  • Removing learned set aggregation yields paired intervals entirely below zero on all four Fixed Wrong-User Profile PCVP metrics; three other component removals do so on three metrics.The other removals are Explicit Feature Difference, Query Feature, and Wrong-User Contrast Loss; each comparison uses a matched full model.Source: Table F2 · p. 27, §F.2 · pp. 27–28

F.3 PROFILE WIDTH

  • 256 dimensions is the smallest tested profile width meeting the quality-retention criterion, requiring 260 bytes; 512 dimensions gives comparable quality at nearly double storage.Retention requires the lower 95% interval bound to exceed −1% of reference CQS on every metric.Source: §F.3 · p. 28, Table F3 · p. 28

G LIMITATIONS

  • PPSD lacks an official user split and episode generator, and validation users select checkpoints, so reported estimates may be optimistic.PPSD tests personalized conditioning; FiveK and PPR10K evaluate fixed expert targets, and spatially varying retouching requires an additional component.Source: G LIMITATIONS · p. 28

H EXTENDED QUALITATIVE EVALUATION

The qualitative comparisons show preferred-target-directed edits, visible sensitivity to conditioning controls, and profile-refinement changes as feedback accumulates.

Personalized Editing

  • PrefLUT adjusts brightness and color toward preferred targets across animal, food, and outdoor scenes, with visible differences in saturation and scene-detail preservation.Preferred targets include both brighter and darker appearances, so edit direction varies across examples.Source: H EXTENDED QUALITATIVE EVALUATION · p. 28, Figure H1 · p. 29

PCVP

Profile Refinement

Related papers

Loading 2609.34133v1…