acceptodds
Withdrawn from ICLR 2027

CAPA: A Context-Aware Agent for Continual Personalized Photo Retouching

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

A continual photo-retouching agent must use personal feedback across images while preserving scene-specific preferences and temporary intent. We introduce Context-Aware Preference Agent (CAPA), which connects feedback use, acquisition, and retention through hierarchical preference states and an axis-level evidence ledger. Candidate contrasts represent each choice relative to the alternatives shown. Responses adapt the current session immediately, while the ledger tracks image, session, and scope information for persistent promotion. Remaining-axis masks let eligible coordinates enter long-term memory while preserving the others for later promotion; each coordinate is consumed once. CAPA with Decision Value (CAPA-DV) uses personal response probabilities and semantic candidate partitions to select questions and follow-ups within an interaction budget. In chronological replay of data from 31 participants, hierarchical memory achieves 44.23% candidate-selection Top-1 accuracy versus 40.26% for flat memory using the same history. In fixed-prefix replay with choice-consistent answers, one-step CAPA-DV reaches 80.14% versus 67.75% for fixed coverage, using 326 versus 452 question exposures. Synthetic tests further characterize answer-noise sensitivity and retention under temporary perturbations. CAPA links current editing intent to comparative evidence and its persistent reuse in a continual personalization workflow.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Personalized Photo Editing, Multimodal Agents, Preference Learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_j5tbOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents CAPA, a context-aware agent for continual personalized photo retouching. It combines global, scene-specific, and session-level preferences with an evidence ledger that selectively promotes comparative feedback into long-term memory. CAPA-DV selects clarification questions based on their expected decision value under an interaction budget. Offline replay and synthetic experiments demonstrate improved preference prediction and query efficiency.

Critical strengths

The distinction between long-term preferences and temporary requests is well motivated, and tracking evidence separately for each editing axis allows the system to retain useful feedback while waiting for more support on other axes. The clarification strategy is also practical, focusing on questions that help distinguish the most likely candidates. In replay experiments, it achieves higher selection accuracy with fewer questions than fixed coverage.

Critical weaknesses

The method involves many interacting components and hand-set rules, but the ablations do not make clear whether this complexity is needed over a simpler hierarchical memory. The evaluation also leaves a gap between the reported results and real use. Clarification answers are inferred from users’ final choices rather than collected from actual interactions, and the coverage baselines can ask redundant questions that CAPA-DV excludes. These results therefore offer limited evidence that the decision-value criterion itself would improve a real user’s editing experience.

Questions for authors
  1. How much of the improvement remains with a simpler hierarchical memory, and which additional components are necessary to achieve the reported gains?
  2. Would CAPA-DV still outperform the baselines if users answered the clarification questions directly and all methods excluded redundant questions?
Reviewer_8E48OpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes CAPA for continual personalized photo retouching. It combines hierarchical preference states with an axis-level evidence ledger to separate session adaptation from long-term retention. Its clarification policy, CAPA-DV, selects questions over rendered candidates based on expected decision gains and interaction costs, with a budgeted extension for multi-step planning.

Critical strengths
  1. Meaningful problem. Distinguishing users’ long-term preferences from short-term intent based on their past feedback is a practical challenge that has received limited formal treatment in personalized photo retouching.
  2. Detailed reproducibility information. The appendix provides detailed specifications of the equations, thresholds, simulators, and evaluation protocols.
Critical weaknesses
  1. Reliance on externally supplied scope labels. Scene labels are externally supplied, and temporary-intent flags come from questionnaires. Although determining the scope of user feedback is a central challenge highlighted in the introduction, the proposed method assumes that this information is available rather than inferring it. This limits its applicability to real-world interactions, where such labels may be unavailable or unreliable. Performance with missing, noisy, or automatically inferred labels remains unclear.

  2. Insufficient experimental evidence. The evaluation primarily consists of internal ablations and simple baselines. Comparisons with relevant existing methods, such as PerTouch [1] and PALATE [2], are needed to establish the method’s added value. Moreover, the human study includes only 31 participants, and cross-session accuracy is evaluated on just 19 choices. No confidence intervals or statistical tests are reported to establish the reliability of the gains. The evidence for long-term memory benefits is particularly weak: the improvement from 31.58% to 42.11%, approximately 10.5 percentage points, corresponds to only two additional correct predictions. Furthermore, the study spans only three sessions, separated by at least 24 hours, providing limited evidence about preference changes over longer periods.

[1] PerTouch: VLM-Driven Agent for Personalized and Semantic Image Retouching

[2] Palate: Personalized aesthetic learning through adaptive taste evolution for multi-user portrait retouching

Questions for authors

See above

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

CAPA represents photo-retouching feedback as contrasts between a chosen candidate and its alternatives, maintains population/global/scene/session preference summaries, and promotes evidence through an axis-level ledger. CAPA-DV selects semantic binary questions using personal response probabilities, decision value, and a remaining interaction budget. Evaluation comprises chronological replay from 31 participants, fixed-prefix candidate identification with choice-consistent generated answers, and synthetic query/retention tests. A separately configured online editing prototype supplies qualitative examples.

Critical strengths
  1. The historical-conditioning comparison controls several important factors: candidates, scorer networks, instruction priority, event order, and observed history are shared. Hierarchical memory improves candidate Top-1 from 40.26% for flat aggregation to 44.23% (Table 1). Separating candidate likelihood from response-type prediction makes clear which gains reflect identifying a preferred edit and which reflect predicting set-level feedback. This provides useful evidence for contextual organization of the same feedback.

  2. The method gives comparative feedback an explicit operational meaning: varying axes determine recorded evidence, provenance governs promotion, and unconsumed coordinates remain available for later support (Section 3.2; Appendix A.2). Personal probabilities also improve one-step DV over uniform selection under the same DV rules, 80.14% versus 75.91% (Table 2). The synthetic tests expose the associated accuracy–cost and adaptation-delay tradeoffs instead of suggesting that more querying or stricter retention always helps.

Critical weaknesses
  1. The query comparisons do not sufficiently isolate the value of DV selection. In Table 2, fixed/random coverage can ask different axes with identical partitions, whereas DV requires a split of the remaining candidates and applies value-based stopping (Appendix C.2). The reported 12.39-point improvement over fixed coverage therefore combines filtering, ranking, and stopping. The uniform-DV control is useful but does not replace split-only coverage and entropy baselines using the same personal scorer. Entropy is evaluated only synthetically, where it achieves higher noiseless accuracy at higher cost (Table 4). A matched accuracy–cost comparison on the human replay is needed to establish the incremental query-policy contribution. The explicitly generated group answers and engineering costs further limit conclusions about actual interaction effort.

  2. The distinctive remaining-axis retention mechanism has limited direct validation. Table 1 compares the whole hierarchy; Table 3 also changes commit timing/batching, and its cross-session choice result is eight correct predictions versus six out of only 19. No paired uncertainty is reported. The synthetic comparison usefully isolates added session eligibility, but warmth and brightness always vary together, while confidence and response-type screening are inactive (Appendix D.2). It never tests a record whose eligible coordinates are promoted while other coordinates remain pending. An ablation against record-level retention under staggered axis observability, together with participant-level uncertainty, would show whether the proposed ledger adds measurable benefit beyond hierarchy and batching.

Questions for authors
  1. Assessment-changing analysis: On the existing fixed-prefix replay, how does DV compare with entropy and split-only coverage using the same scorer and feasible partitions, across matched exposure or proxy-cost levels? This would separate query ranking from redundant-question filtering and stopping.

  2. Assessment-changing analysis: What changes when pending-axis tracking is replaced by record-level promotion or discarded pending coordinates? Please include staggered axis observability and paired participant uncertainty for the historical and retention gains. For the synthetic perturbation test, does the benefit persist when the disturbance occurs at a promotion-eligible boundary rather than session three?

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

Related papers

Discussion (0)

Sign in to comment.