acceptodds
Withdrawn from ICLR 2027

Reward-Guided Noise-Level Trajectories and Task-Aware Evaluation for Image Editing

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Instruction-guided diffusion and flow-based image editing requires selecting a noise-level trajectory whose effectiveness can vary with the source image and instruction. Selecting such trajectories requires a reliable editing reward, yet the evidence needed to assess quality varies across tasks despite the shared goals of faithful editing, content preservation, and visual integrity. We address these two challenges, input-adaptive trajectory selection and task-aware reward construction. Amortized Trajectory Selection (ATS) learns a lightweight policy optimized for a chosen editing reward to predict a complete trajectory from each source–instruction pair before execution, without test-time optimization or online reward queries. To our knowledge, ATS is the first image-editing method to learn such direct per-input trajectory prediction without test-time optimization. Geometry-Normalized Trajectory Selection (GeoTS) complements ATS by making the sampling trajectory itself a structured source of candidate variation without training a selector or iteratively optimizing individual trajectories. To provide a reliable reward across editing tasks, we introduce EditCal, which selects and combines complementary signals from existing evaluation models according to the task and criterion, calibrates them to shared human-anchored ordinal semantics, and integrates them into an overall reward with criterion-specific failure diagnostics. On FLUX.1-Kontext and FlowEdit, ATS improves mean reward over native schedules across all four evaluated objectives and retains its gains at unseen sampling-step counts without retraining. Across three public human-judgment benchmarks, EditCal improves evaluation agreement and failure detection.

Primary area:
generative models
Keywords:
Flow model, Image editing, trajectory 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_Hb3COpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper leverages the noise schedule as a control interface for image editing models. It introduces three contributions: ATS trains a lightweight policy that predicts a two-parameter schedule for each data pair; GeoTS performs best-of-N selection over geometry-normalized perturbations of a reference schedule; EditCal combines existing evaluators into a calibrated editing reward.

Critical strengths
  1. The statistics of experimental results is careful with paired bootstrap CIs.
  2. EditCal's calibrated, per-criterion failure probabilities are an interpretable alternative to scalar editing rewards.
Critical weaknesses
  1. ATS is neither new nor clearly effective. Its schedule family is simply FLUX's built-in time-shift function, and a single input-independent schedule recovers most of the gain. The remaining improvement is tiny and often not significant on Kontext.
  2. The other contributions do not hold up. GeoTS and EditCal offer little methodological novelty. GeoTS is worse than simple noise resampling on Kontext at equal cost (Table 12 in the appendix). EditCal is a calibrated ensemble of its strongest baselines, with small and untested gains.
Questions for authors
  1. Can you report the end-to-end inference time and compute cost? This should include the policy forward pass for ATS and the B full edits and reward evaluations for GeoTS?
  2. Is the improvement perceptible in a human user study?
Reviewer_X4umOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies task-conditioned noise-level trajectories for image editing and introduces a corresponding task-aware evaluation framework. ATS predicts a two-parameter monotone trajectory from the source image and editing instruction before running a frozen editor. It is trained offline through reward-weighted imitation of trajectories from a candidate bank. The same continuous trajectory can be discretized at different inference budgets.

The paper also introduces GeoTS, an optional local trajectory search method using a geometry-based normalization of proposals, and EditCal, a reward model that routes task-specific evidence into calibrated ordinal judgments of Edit, Preserve, and Clean quality. Experiments evaluate two FLUX-based editing settings, several reward objectives, matched-noise comparisons, transfer across numbers of inference updates, and external human-annotated benchmarks for EditCal. The central finding is that a small task-conditioned schedule predictor can improve editing rewards over native and learned global trajectories without per-input optimization at inference time.

Critical strengths

The principal ATS claim is supported by relevant and reasonably controlled experiments. At 12 updates, the EditCal-trained policy improves the reported EditCal score from 0.659 to 0.679 on Kontext and from 0.505 to 0.542 on FlowEdit, with positive paired intervals. 1. These comparisons use the same noise, update count, and a single output per method. The paper also includes a learned global policy trained under the same general procedure: Table 14 reports task-conditioned improvements of 0.0079 on Kontext and 0.0146 on FlowEdit, with positive paired intervals. This directly tests whether conditioning contributes beyond learning a generally better schedule. Transfer to other update counts and corroborating fixed-output evaluations provide additional support. Together, these experiments establish a meaningful, practically useful contribution for the main method.

  1. EditCal offers an interpretable evaluation design with substantial external validation. Separating editing success, preservation, and cleanliness is useful for image editing, where a single similarity or aesthetic score can hide important failure modes. Source-level cross-fitting and exclusion of the public evaluation benchmarks from calibration fitting strengthen the evaluation. Tables 4 and 5 report the highest scores on nine of twelve agreement endpoints and eight of nine failure-detection endpoints, respectively. The preservation and cleanliness interventions are also informative. Although not every component’s incremental value is established equally strongly, the complete evaluator has a credible empirical case.
Critical weaknesses
  1. The evidence for improved human-perceived editing quality is narrower than the evidence for improved reward values. The trajectory evaluation uses 150 held-out tasks and two FLUX-based editing settings, with two noise repetitions. The human annotations used to develop and validate EditCal do not constitute a direct blinded comparison of ATS outputs against native schedules. In addition, EditScore and EditReward are constituent features of the optimized EditCal reward, so their improvements are not fully independent validation. ImageReward and the additional VLM judges provide useful corroboration, although the results are not uniformly conclusive; for example, the FlowEdit Qwen-judge improvement has an interval crossing zero. Appendix A.1 also makes clear that the reported intervals are conditional on the trained policies rather than reflecting policy-retraining variability. These points limit the breadth of the quality and generalization claims, even though the main reward-guided improvement is supported.

  2. The incremental contributions of the more elaborate secondary components are not isolated as convincingly as ATS itself. GeoTS is compared against same-budget noise resampling, which is an appropriate control, but the paper does not isolate geometry normalization against otherwise comparable unnormalized trajectory proposals. The radius sweep explicitly does not resolve that attribution question. Moreover, the comparison with noise resampling is editor-dependent. For EditCal, the independence ablation in Table 16 does isolate the copula contribution, but the gains of the full model over the no-copula version are only 0.0027, 0.0006, and 0.0016 on the three reported endpoints, without uncertainty. Thus, the evidence for the complete calibration pipeline and task-conditioned schedule learning is stronger than the evidence that these additional modeling choices are necessary or consequential. This limits the demonstrated significance of the secondary contributions.

Questions for authors
  1. [Assessment-relevant validation] Is there an existing blinded human comparison of the fixed ATS and native-schedule outputs, or trajectory evaluation on an additional independently sourced task set? A source-balanced audit of the current outputs that separately measures instruction execution and preservation would materially strengthen the interpretation of the reward gains.

  2. [Assessment-relevant validation] How does GeoTS compare with unnormalized shift/sharpness trajectory proposals under the same candidate budget, reference noise, selection reward, and radius-tuning protocol? This would help determine whether geometry normalization contributes beyond searching a useful low-dimensional trajectory space.

  3. [Assessment-relevant validation] Please provide source-paired uncertainty for EditCal’s improvements over the strongest relevant alternatives in Table 4 and for the full-versus-independent-copula comparison in Table 16. Are the small copula increments reliable enough to support a distinct empirical contribution?

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

This paper studies noise-level schedules as a control interface for frozen image editors. It introduces Amortized Trajectory Selection (ATS), which predicts a continuous, two-parameter schedule from the source image and editing instruction; Geometry-Normalized Trajectory Selection (GeoTS), which explores a structured neighborhood around a reference schedule; and EditCal, which combines task-dependent evaluation evidence into calibrated criterion-level distributions and an overall reward. The paper presents encouraging results for learned schedule prediction and transfer across sampling-step counts. However, the downstream benefits of EditCal’s additional complexity and the practical advantage of GeoTS over simpler candidate-generation strategies require stronger evidence.

Critical strengths
  1. ATS controls a frozen editor through a compact schedule representation, avoiding per-input optimization and reward queries at inference time. Separating the continuous schedule from its discretization also provides a natural mechanism for transfer across sampling-step counts.

  2. The task-aware integration of instruction fulfillment, content preservation, and visual integrity provides interpretable criterion-level diagnostics, helping distinguish plausible-looking edits from those that alter protected content.

  3. The paper presents an extensive experimental evaluation across two frozen editors, four reward objectives, multiple sampling-step counts, and three public human-annotated benchmarks. Beyond improvements under the optimized reward, Table 2 evaluates the same generated outputs using six metrics, including external vision-language judges, providing broader evidence of editing quality. Comparisons with noise resampling under matched candidate budgets, together with input-conditioning ablations, strengthen the empirical analysis and help clarify where the proposed components are somehow effective.

Critical weaknesses
  1. The downstream advantage of EditCal over existing reward models is not directly established. Table 4 provides encouraging evidence of agreement with human judgments, but this does not directly establish EditCal’s effectiveness as supervision for ATS. Table 1 trains ATS with different rewards but evaluates each policy under its own objective, while Table 2 evaluates only EditCal-trained ATS. A controlled comparison should train ATS using EditCal, existing reward models, and simpler fusion baselines under matched settings, then evaluate all policies on the same held-out tasks using common independent criteria. This would directly test whether EditCal leads to better editing outcomes.

  2. The necessity of EditCal’s individual components requires stronger downstream validation. Table 16 provides useful evaluator-level ablations, but does not establish whether the individual components improve ATS training. In particular, removing the copula while retaining the calibrated marginals and expected-geometric readout reduces Overall agreement by only 0.0027, 0.0006, and 0.0016 across the three benchmarks. It remains unclear whether these small differences translate into meaningful editing improvements. A downstream comparison between ATS trained with full EditCal and the independent-joint variant would be particularly informative. Similar comparisons for annotation alignment and ordinal calibration would further clarify their practical contributions.

  3. GeoTS does not demonstrate a consistent advantage under equal candidate budgets. Table 3 provides an appropriate comparison against noise resampling. However, GeoTS underperforms on five of six metrics for Kontext. For FlowEdit, it improves four metrics but underperforms on ImageReward and the GPT judge. While the authors acknowledge that the benefits are editor-dependent, these mixed results leave the practical advantage over noise resampling uncertain.

  4. The failure-detection evaluation has limited coverage. The Preserve results in Table 5 rely on 65 matched failure-success pairs constructed by inducing severe preservation failures through target-image descriptions and pairing them with successful correct-instruction controls. This panel may not represent subtle preservation errors in ordinary editor outputs. Broader tests covering identity drift, unintended object removal, and spatial-relation changes would strengthen the evidence. Targeted tests involving count and text errors would also help determine whether the frozen evidence models provide sufficient signals for these cases.

Questions for authors
  1. Under a common blinded human evaluation, how does EditCal-trained ATS compare with ATS trained using EditScore, EditReward, or simpler evidence fusion?
  2. Do EditCal ablations materially change downstream ATS performance, beyond changing evaluator-agreement metrics?
  3. Does geometry normalization outperform simpler schedule perturbations under matched candidate budgets?
  4. How reliably does EditCal detect subtle count, identity, and spatial-relation errors in outputs generated by ATS and GeoTS?

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

Related papers

Discussion (0)

Sign in to comment.