acceptodds
Under review as a conference paper at ICLR 2027

PrefLUT: Reusable and Refinable Personalized Color Editing from Pairwise Preferences

26 views

Abstract

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.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Color Grading, Color Editing, Image Enhancement, User Preference Modeling

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_8HM3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates personalized photo color grading based on pairwise preferences and introduces the PrefLUT method. Users provide pairs of "preferred" and "non-preferred" images, which the system encodes into a reusable user profile. Subsequently, for any new image, the model utilizes this profile to predict a 3D LUT for image editing, eliminating the need for additional optimization or fine-tuning for individual users. Additionally, the paper proposes the PCVP to evaluate the effectiveness of personalized image enhancement and color grading. Comprising five controlled tests that assess quality improvements based on target user preferences and specific query images, this verification protocol establishes a rigorous standard for evaluation in the field of personalization technology.

Critical strengths

Unlike previous methods requiring per-user fine-tuning, PrefLUT introduces an approach that converges to individual user preferences without fine-tuning, thereby enhancing personalization capabilities. Additionally, the PCVP mechanism was designed to verify the differences between individual and group preferences.

Critical weaknesses
  1. Lack of user studies, incomplete description of the deployment workflow, and a lack of direct evidence supporting core claims. The paper's key selling point is a personalized editing framework that is reusable and continuously updateable via new feedback; however, it fails to explain how user preferences are generated in practice. For instance, where do the candidate images presented for binary choice come from, for example, random LUTs, preset filters, or the model's own outputs? If candidates are model-generated, are user preferences constrained by the model's output range? More critically, all "feedback-based update" experiments (Fig. 4, Table F1) rely on offline simulations using pre-collected PPSD data; there is no actual user participation in the "selection → profile update → re-editing" closed loop. The feedback pathway involving "user-confirmed manual adjustments" mentioned in the paper lacks any experimental validation. Furthermore, all conclusions regarding personalization rely on proxy metrics like CQS without any human evaluation. Given the marginal gains observed in PCVP (e.g., ΔE00-CQS improvements of ~0.002–0.007), user studies are essential to determine whether these differences are perceptually distinguishable. For a study centered on "personalization," this constitutes a significant evidentiary gap.

  2. Personalization claims lack independent verification and are validated on only a single dataset. The primary evidence that the model "truly utilizes user preferences" is the authors' own PCVP metric. Since the PrefLUT training objective , Wrong-User Contrast Loss, directly corresponds to the "Wrong-User" test within PCVP, the claim that the method "uniquely passes PCVP" carries limited weight. It is recommended to include simple yet robust baselines independent of PCVP, such as applying the average changes in brightness, saturation, and color temperature (derived from the "preferred vs. non-preferred" image pairs) to new images, or fitting a single average LUT for each user. If PrefLUT offers only marginal advantages over these methods, the necessity of its complex structure is questionable. Regarding datasets, all personalization experiments were conducted solely on PPSD. The experiments on FiveK and PPR10K utilized data from the target experts directly for training; as such, they represent supervised enhancement for fixed targets rather than learning new users from limited preference data, and thus cannot serve as a secondary personalization test set. It is recommended to adopt an "unseen-expert" protocol by treating the five experts as distinct users and reserving experts not involved in training for testing purposes. The published baselines in Table E2(A) employ this protocol, allowing for direct comparison.

Questions for authors

Mentioned in Weaknesses. User studies are essential for this work, and more verification protocols & datasets are needed.

1 comment

Thank you for recognizing PrefLUT's personalization and PCVP. We apologize for the late reply, as additional experiments and the user study took time.

(1) User preferences in practice (W1): A user-confirmed edit and its pre-edit image form a preferred/non-preferred pair, so explicit binary selection is optional. This is mentioned in lines 77-78 and 190-197 and reflected in our new user study. Preference pairs need not come from a model. PPSD experiments use pre-collected preferred/non-preferred pairs. In our user study, participants make their own choices on candidates from our PPSD validation-user split and provide manual feedback by adjusting PrefLUT's output (Appendix G, lines 1921-1924 and 1995-2013). Thus, preference collection is not restricted to what a model generates.

(2) Selection → profile update → re-editing (W1): Our new user study completes this process (Appendix G.2, Table G1, lines 1987-2013 and 1999-2008). Users make new choices, PrefLUT updates their profiles with frozen weights, and users judge edits on unseen photographs. Users prefer the updated profiles' outputs over their initial profiles' outputs. We also add matched feedback accumulation with UPE (Appendix F.1, Table F1, lines 1707-1721 and 1729-1774), where PrefLUT shows stronger overall refinement gains and benefits earlier from new feedback. These complementary evaluations show that new choices improve later editing on other images, validating both the practical feedback loop and its effectiveness.

(3) User-confirmed manual adjustments (W1): We add a user study in which participants adjust exposure, contrast, saturation, and temperature and confirm their preferred result (Appendix G.2, Table G1, lines 1987-2013 and 1999-2008). The edited/pre-edit pair updates the profile as defined in Section 3.1 (lines 190-197). On unseen images, manually refined PrefLUT achieves 74.2% wins, 23.3% ties, and 2.5% losses against the initial profile, with all network weights frozen. Thus, users' manual edits improve later automatic editing, providing direct validation of this feedback route.

(4) Perceptually distinguishable differences (W1): We add a blinded user study with 10 participants and 600 judgments across five comparisons, 120 per comparison (Appendix G, Table G1, lines 1921-1939 and 1999-2008). PrefLUT achieves 50.8% wins, 30.0% ties, and 19.2% losses against UPE, and 45.0% wins, 20.0% ties, and 35.0% losses against StarEnhancer. Personal profiles achieve 41.7% wins, 39.2% ties, and 19.2% losses against training-mean profiles. Users more often prefer PrefLUT over both baselines and their personal profile over a population average. This directly supports perceptible personalization benefits without relying on CQS alone.

(5) Independent verification and Wrong-User Contrast Loss (W2): We add an experiment removing the Wrong-User Contrast Loss throughout PrefLUT's entire training chain (Appendix F.2.1, Table F3, lines 1837-1849). PrefLUT still passes all five PCVP tests on all four metrics, exceeding all compared baselines in tests passed. Therefore, full PCVP does not depend on training with this loss.

(6) Average changes and a single average LUT (W2): Following your suggestion, we add average color adjustments and fixed per-user LUTs fitted only on reference pairs (C.2.1, Table E5, lines 1125-1131 and 1621-1644). PrefLUT leads all four CQS and all four direct metrics. Its PSNR is 22.093 dB, versus 20.732 for simple color adjustment and 18.043 for the fixed LUT fitted with the shared PPSD objective. LPIPS also improves to 0.08880 from 0.10095 and 0.25750, respectively. Together with the aggregation and feature-difference ablations (Table 3, lines 433-442), these gains support learning how each user's preferences should guide each image, rather than applying one fixed edit.

(7) Supervised enhancement for fixed targets and an unseen-expert protocol (W2): FiveK and PPR10K provide retouched targets from a small number of experts. Following your suggestion, we add FiveK evaluation that trains on four experts and tests the held-out expert, repeated for all five experts, with 16 references and 500 unseen images (E.2.3, Table E4, lines 1561-1588). Correct profiles outperform wrong-expert and reversed profiles on all five averaged metrics. The training-expert mean is slightly stronger in average fidelity than the held-out expert profile. These expert-target datasets do not directly collect users' pairwise preferences, and the same expert can adjust editing style according to each image's content and intended appearance. PPSD instead directly records many users' preference choices, better reflecting real-world use. PrefLUT demonstrates effective personalization on this dataset. It passes all five PCVP tests on users excluded from training, and our user study further supports its ability to match real participants' preferences. We interpret these findings as support for learning personalization from diverse users' expressed preferences.

Reviewer_rNmjOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper presents PrefLUT, a framework for personalized color editing that learns a reusable user profile from preferred/non-preferred image pairs. Given a new image, the model combines this profile with query features to predict a 3D LUT and its editing strength. The profile can be updated with additional feedback without fine-tuning the network.

The authors also propose the Preference-Conditioning Verification Protocol (PCVP) to evaluate whether the editing results actually benefit from the intended user's preferences and the current query image. Experiments on PPSD show competitive editing quality and positive results under the proposed tests. The paper further evaluates the framework on general-purpose image enhancement tasks.

A notable advantage is its efficiency. The compact configuration stores each quantized user profile in just 260 bytes, with a reported editing latency of 1.365 ms per 512 × 512 image on an RTX 5090.

Critical strengths

1. A practical approach to personalized LUT prediction.

The overall design is straightforward and practical. Encoding preference pairs and combining them with query features to predict an explicit LUT seems well suited for personalized application. The LUT can be exported and applied to full-resolution images, while the cached profile avoids repeatedly encoding the reference images.

2. A meaningful evaluation protocol for personalization.

I particularly appreciate the proposed PCVP protocol. Standard fidelity metrics alone cannot tell whether a model actually benefits from user-specific preferences or simply applies a generally effective enhancement. By testing different users, reversed preferences, mismatched pairs, population-level preferences, and incorrect query conditions, PCVP provides a more convincing way to evaluate personalization.

I think this is a useful contribution beyond the specific PrefLUT architecture, especially for methods that build reusable user representations from a set of reference images.

3. Promising efficiency and deployment potential.

The small profile size, low inference latency, and separation between profile construction and query-time editing are appealing for practical applications. The profile-reuse benchmark also gives concrete evidence that caching user preferences can reduce repeated computation.

Critical weaknesses

1. The choice of a global LUT needs stronger justification for personalized editing.

The introduction points out that existing PPSD baselines do not explicitly produce LUTs, but I am not fully convinced that this alone establishes a meaningful methodological gap. The paper explains the efficiency and exportability advantages of LUTs, but gives less attention to why a global LUT is a suitable representation for personalized preferences.

  • A global 3D LUT has an inherent limitation: pixels with the same RGB values receive the same transformation, regardless of their location or semantic content. For example, it cannot independently adjust a blue sky and a similarly colored garment. This can be restrictive in personalized retouching, where users may want to enhance a particular subject while leaving other regions unchanged.

  • This limitation is particularly concerning given that PPSD, the primary dataset used for personalized training, contains many examples of localized color adjustments, yet the proposed model relies solely on a global LUT.

Recent methods have explored more flexible alternatives. BPAM [1] combines bilateral grids with pixel-adaptive MLPs, while SVDLUT [2] explores a lightweight spatially aware LUT representation. These methods highlight the trade-off between efficiency and spatial flexibility, which I think deserves more discussion here.

Although the authors introduce an additional spatial residual for FiveK and PPR10K, the main PPSD model still relies on a global LUT. It would be helpful to understand how much this restriction affects personalized editing in practice. For example, the authors could evaluate cases requiring region-dependent adjustments or compare global and spatially adaptive editing backends under the same preference-conditioning framework.

2. The novelty of the reusable user profile over existing reference-conditioned methods is not entirely clear.

My main concern is that the proposed profile mechanism seems conceptually close to existing preference-embedding approaches, particularly UPE [3] .

At a high level, the two approaches follow a similar pattern:

  • UPE: Preference pairs → Preference embedding → Query-conditioned image decoder → Edited image.
  • PrefLUT (this work): Preference pairs → User profile ( ?, not proved, same to neural embedding)→ Query-conditioned LUT predictor → LUT → Edited image.

The paper emphasizes in the Introduction (Line 52) that "The framework divides personalized color editing into profile construction and query-time editing." However, I find this distinction somewhat misleading when considering how the model is actually trained.

As described in Appendix C.1, the Preference Set Aggregator and LUT prediction heads are trained jointly using losses computed on the edited query images, while the pretrained image encoders and LUT decoder remain frozen. In other words, profile construction and LUT prediction are not learned as two independent stages, but are optimized together through the same editing objective. The separation mainly comes from the inference procedure, where the user profile can be computed once and cached for subsequent queries. This is a useful engineering choice, but not necessarily a new preference-learning formulation.

From this perspective (Jointly-training), PrefLUT appears to follow the same general conditioning paradigm as UPE or other preset-based baselines mentioned in the paragraph 'Learned LUTs and Reusable Color Transforms.' of the Related works of the paper, with the main difference being the use of an explicit LUT instead of directly re-coloring the image. Similarly, the proposed profile refinement simply rebuilds the representation from additional preference pairs without updating the network weights, which could also be achieved with a conventional reference-conditioned model.

I would therefore like to see a clearer explanation of what distinguishes the proposed profile-learning mechanism from existing preference embeddings. A controlled comparison using the same preference encoder and aggregation module but different editing backends (e.g., spatially adaptive methods such as BPAM and SVDLUT ) would help isolate the contribution of LUT prediction. It would also be useful to compare against a cached preference-embedding baseline (or preset-based baseline, e.g. Neural Preset [4] )and examine whether profiles constructed from different reference sets of the same user lead to consistent editing results.

3. The writing of this paper is hard to follow than necessary

The main text focuses on profile construction and query-time editing, while several important details about pretraining, frozen modules, and jointly optimized components are deferred to Appendix C.1. These details are central to understanding how the user profile is learned, so I think at least a concise training overview belongs in the main paper.

Figure 2 also feels somewhat overloaded. It includes many intermediate tensors, projections, and module labels, making it difficult to identify which components are pretrained, frozen, or optimized during personalized training. A simpler figure, perhaps with a small table summarizing the training stages, would make the method much easier to follow.

The mathematical presentation could also be simplified. Most symbols such as channel width, trilinear function, encoding features and etc. are not neccessary to mention in the main text. The main-text presentation could then focus on the substantive mechanisms rather than repeatedly restating the same sequence of operations.

Overall, I think the paper would benefit from spending less space describing the inference pipeline repeatedly and more space explaining what is actually learned during training and how it differs from existing preference-conditioned editing methods.

References

[1] Lou, J., Zhao, X., Shi, K., and Gu, S. "Learning Pixel-adaptive Multi-layer Perceptrons for Real-time Image Enhancement." ICCV, 2025. Paper

[2] Kim, W., Lee, K., and Cho, N. I. "Lightweight and Fast Real-time Image Enhancement via Decomposition of the Spatial-aware Lookup Tables." ICCV, 2025. Paper

[3] Kim, J., Yoo, J., and Kim, S. J. "Learning Personalized Photographic Style from Pairwise User Preferences." CVPR, 2026. Paper

[4] Ke, Z., Liu, Y., Zhu, L., Zhao, N., and Lau, R. W. H. "Neural Preset for Color Style Transfer." CVPR, 2023. Paper

Questions for authors
  1. Global LUT limitation. Given the spatially varying edits in training dataset (PPSD), could the authors compare the current global LUT with other 4D LUT baselines under the same preference-conditioning framework? This would help justify the representation choice.

  2. Methodological novelty. Since profile construction and LUT prediction are jointly optimized, what fundamentally distinguishes PrefLUT from existing preference-embedding methods such as UPE? Controlled comparisons with cached preference embedding and simpler preset-based baselines would be valuable.

My main concerns are the global LUT's expressiveness and the methodological novelty. Convincing evidence addressing these two points could lead me to reconsider my score.

1 comment

Thank you for recognizing PCVP's contribution and the deployment value of explicit LUTs. We apologize for the late reply, as additional experiments and the user study took time.

(1) Global LUT representation and localized adjustments in PPSD (W1): PrefLUT's global LUT leads all four direct metrics and three of four CQS metrics in the main PPSD comparison (Table 1), and passes all five PCVP tests (Table 2). These results show that global LUTs can produce accurate edits that respond to individual preferences. We also agree that spatially aware editing has potential and plan to extend PrefLUT in future work. However, PrefLUT already achieves the highest CQS and lowest editing latency among the tested backends in the shared-profile comparison below.

(2) Spatially adaptive editing backends and 4D LUTs (W1-W2, Q1): We add BPAM, SVDLUT, SA-LUT, BGrid, and LLF-LUT++, all using the same frozen PrefLUT preference encoder and aggregator (Table 4, lines 487-496 and 499-503). PrefLUT leads all four CQS metrics and has the lowest latency among these backends. SA-LUT uses per-pixel context and a 4D LUT, as requested. Compared with SA-LUT, PrefLUT leads all four CQS metrics and is 13.61 times faster (1.210 vs. 16.473 ms/image). Thus, PrefLUT combines better personalized editing quality with substantially faster inference under the same preference interface. Adaptation is detailed in C.3.4 (lines 1261-1277).

(3) Novelty over existing preference embeddings (W2, Q2): UPE guides a neural image decoder with a preference embedding. PrefLUT predicts an image-dependent LUT and edit strength from its profile (Sections 2-3, lines 125-128 and 200-215). We add matched feedback accumulation with UPE (Appendix F.1, Table F1, lines 1707-1721 and 1729-1774). PrefLUT shows larger observed gains across all CQS metrics in the primary comparison and consistently benefits from correct preference directions. These results show that its explicit LUT framework supports effective, ongoing preference updates with frozen network weights, alongside reusable and manually adjustable color transforms.

(4) Pretraining, frozen modules, and joint optimization (W2-W3): The two stages describe deployment, not independent training. We clarify this distinction and add the training overview to Section 3.3 (lines 142-152 and 227-239). The image encoders and LUT decoder are pretrained and frozen. Losses on edited images jointly train the newly initialized preference aggregator, LUT latent-vector head, and edit-strength head. Gradients still pass through the frozen decoder to train the aggregator and latent-vector head. Training uses images separate from the reference set, so the profile learns preferences useful for editing new images.

(5) Cached preference embeddings and Neural Preset (W2, Q2): We add UPE with its preference embedding stored for reuse, Neural Preset, simple color adjustments, fixed per-user LUTs, and D-LUT (Appendices C.5/E.3, Tables E5-E6, lines 1323-1339 and 1621-1699). PrefLUT beats all these alternatives on all four direct metrics. It also leads all four CQS metrics over the presets and D-LUT, and three of four over UPE. Both PrefLUT and UPE reuse preference representations without re-encoding references. PrefLUT remains 11.62 times faster (1.210 vs. 14.071 ms/image), with 1.966 vs. 322.257 registered GFLOPs. Thus, its advantage extends beyond avoiding repeated reference encoding.

(6) Different reference sets of the same user (W2): We evaluate ten different 16-pair reference sets per user while keeping test images fixed (Appendix F.4, Table F5, lines 1877-1896). Changing reference images within one user produces smaller output changes than switching users on all three measures. Mean DeltaE00/LPIPS differences are 0.392/0.000391 within a user, versus 0.851/0.001872 when switching users. Thus, PrefLUT preserves a user's editing style across different reference samples while distinguishing different users' preferences.

(7) Figure 2 feels overloaded (W3): We simplify Figure 2 by removing redundant tensor dimensions and adding snowflake/flame markers for pretrained frozen/jointly optimized modules (lines 163-180). Readers can now identify each module's training status directly. We keep intermediate feature labels to connect each equation to the module producing its input or output. The method text follows the same module order, making the figure and mathematical description easier to read together.

(8) Mathematical presentation and repeated operations (W3): We reduce the displayed equations in Section 3 from nine to six and move the pipeline equation, Transformer recurrence, and decoder internals to Appendices A.1/A.2/A.4 (lines 740-759 and 785-790). The main text now explains why pair order encodes preference direction, why aggregation combines evidence across images, and how the user profile and current image determine the LUT and edit strength. This replaces repeated operation lists with the purpose of each mechanism.

Reviewer_SipROpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

PrefLUT is a personalized image editing method that constructs reusable user preference profiles from user-selected preferred/non-preferred image pairs and generates user-specific 3D LUTs for new input images. It encodes multiple ordered preference pairs and aggregates them using a Set Transformer to predict the LUT latent vector and edit strength based on the resulting user profile and query image features. When new preference feedback is obtained, the profile can be updated by re-aggregating the reference pairs without updating the network weights, and the generated 3D LUT can be explicitly saved and reused. Furthermore, the authors propose the Preference-Conditioning Verification Protocol (PCVP) to verify dependencies on user preferences and query images, and evaluate personalized editing and general image enhancement performance through experiments using PPSD, MIT-Adobe FiveK, and PPR10K.

Critical strengths
  1. A major strength of this paper is that it achieves very high computational efficiency while demonstrating editing performance that surpasses existing personalized image editing methods in PPSD. PrefLUT not only achieves excellent results across multiple image quality metrics but also, through lightweight user profiles and fast LUT-based inference, achieves lower computational costs and shorter inference times compared to existing methods, making it highly valuable as a practical personalized image editing method.
Critical weaknesses
  1. There are concerns regarding the limited and unclear novelty of the proposed method. The primary concern is that the framework for constructing user embeddings from pairwise preferences is similar to the existing User Preference Embedding (UPE) (Kim et al., 2026). Although the authors emphasize “reusability” and “refinability” as key contributions, UPE also appears capable of recomputing a user representation when additional preference pairs are provided, without requiring per-user optimization. Furthermore, methods for generating explicit LUTs from references or conditioning signals have already been proposed, such as D-LUT (Li et al., 2025). Therefore, it remains unclear to what extent the overall framework goes beyond a combination of existing preference-embedding and LUT-generation techniques.

  2. The performance improvement over existing methods is relatively small, and it is unclear to what extent these gains are meaningful in practical or perceptual terms. In particular, Figure 3 contains examples where the visual differences between StarEnhancer or UPE and PrefLUT appear subtle, and the paper does not sufficiently verify whether the improvements in numerical metrics are actually perceived as edits that better match users’ preferences. A user study would strengthen this claim, for example by presenting raters with a user’s reference preferences and asking which output better reflects that user’s photographic style.

Questions for authors

Please see the weaknesses.

1 comment

Thank you for recognizing our editing quality, efficiency, and utility. We apologize for the late reply, as additional experiments and the user study took time.

(1) User embeddings similar to UPE (W1): PrefLUT predicts explicit, image-conditioned LUTs, while UPE uses preference embeddings to guide a neural image decoder. Our innovation is learning and verifying a reusable preference-to-LUT framework that carries individual preferences across images and refines profiles through feedback, which, to our knowledge, has not been explored for pairwise user preferences. The profile is an embedding, but its LUT output makes edits inspectable, exportable, reusable, and manually adjustable without neural image decoding. We add this clarification in Sections 2-3 (lines 125-128 and 146-152). PrefLUT combines leading quality with high efficiency, demonstrating its practical value for deployment on edge devices. We believe that, in image color grading, deliberately increasing complexity or adopting newer architectures does not necessarily improve performance or utility.

(2) Reusability and refinability (W1): We add matched feedback accumulation with UPE (Appendix F.1, Table F1, lines 1707-1721 and 1729-1774). Both methods use the same users, feedback, and test images with frozen weights. PrefLUT achieves larger observed gains across all CQS metrics in the primary comparison and benefits earlier from additional feedback. Correct new feedback consistently improves all four metrics over reversed feedback for PrefLUT, while UPE does not show the same direction sensitivity. Thus, PrefLUT learns from which images users prefer, beyond simply processing more references. Its reusable profile supports continued improvement on unseen images without per-user network optimization.

(3) Generating explicit LUTs (W1): We add two D-LUT comparisons using the same reference pairs: one uses preferred images only, and the other uses both preferred and non-preferred images (Table E5, lines 1621-1644). Both reuse a fixed transform for each user. PrefLUT instead learns the user's preferences from ordered pairs and predicts a LUT for each new image. PrefLUT outperforms both D-LUT variants on all four CQS metrics and all four direct fidelity metrics. Thus, PrefLUT produces edits closer to users' preferred targets than fitting a fixed transform to their reference images.

(4) Combination of existing techniques (W1): PrefLUT learns which changes a user prefers across reference pairs and how to apply those preferences to new images. We add a training explanation in Section 3.3 (lines 227-239): losses on separate edited images jointly train preference aggregation and LUT prediction. Table 3 shows that replacing learned aggregation with mean pooling lowers all four CQS metrics and all four wrong-user PCVP gains (lines 433-442). LPIPS-CQS drops by 1.484 and PSNR-CQS by 2.250. Removing the explicit feature difference also lowers all eight measures. These ablations show that encoding preference direction and learning how to combine pairs both improve personalized editing.

(5) Meaningful in practical or perceptual terms (W2): We add a comparison in which PrefLUT and UPE both store and reuse user preference representations, avoiding repeated reference encoding, on the same RTX 5090 with 512 x 512 inputs (Appendix C.5 and Table E5, lines 1323-1339 and 1621-1644). PrefLUT leads UPE on all four direct metrics and three of four CQS metrics. Even with this reuse, PrefLUT is 11.62 times faster (1.210 vs. 14.071 ms/image) and uses fewer registered GFLOPs (1.966 vs. 322.257). Table 1 also reports PrefLUT's compact 260-byte quantized profiles. These results show better editing quality with lower computation and storage suitable for mobile and edge deployment.

(6) Visual differences appear subtle (W2): Although PrefLUT leads all four direct metrics and three of four CQS metrics (Table 1), its outputs can look similar to those of strong baselines. Many preferred/non-preferred image pairs in PPSD have only small color or brightness differences. Matching these preferences does not require a large visual change. StarEnhancer also receives our PPSD adaptation and the same preference evidence. The user study directly shows that users choose PrefLUT more often than UPE and StarEnhancer (Table G1, lines 1999-2008).

(7) Edits that better match users' preferences (W2): We add a blinded user study using PPSD images, with participants judging outputs against their own preferences (Appendix G, Table G1, lines 1921-1939 and 1999-2008). Participants choose PrefLUT more often than both UPE and StarEnhancer, and prefer their personal profiles over the training-user mean. These comparisons show that users perceive PrefLUT's advantages even when strong baselines produce visually similar outputs. Together with feedback accumulation, the results support better alignment with individual preferences and effective profile updates as users provide more feedback.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.