acceptodds
Under review as a conference paper at ICLR 2027

Feedback-Driven Prompt Refinement for Image Editing via Self-Distillation

Abstract

Instruction-guided image editing enables users to modify images through natural-language instructions, yet editing quality remains sensitive to the prompts used to control the editor, creating a gap between user intent and model behavior. Existing iterative refinement methods often explicitly evaluate intermediate editing results before generating revised instructions, coupling evaluation and refinement within the deployed policy. Such intermediate assessments can misguide subsequent refinement when inaccurate, while also introducing additional supervision and inference requirements. We propose , a feedback-driven prompt-refinement framework that learns refinement decisions directly from editing outcomes. FeedEdit formulates prompt refinement as an adaptive stop-or-refine decision problem, where evaluator feedback is used during training to teach the policy whether to stop or how to refine, eliminating the need to explicitly generate an intermediate evaluation at inference. Our key insight is that outcome feedback can supervise the policy in two complementary ways: *causal feedback* evaluates the current edit and guides the construction of stopping and refinement targets for offline learning, while *hindsight feedback* evaluates the outcome of a policy-sampled action and provides token-level supervision for online self-distillation. On ImgEdit-Bench and GEdit-Bench-EN, FeedEdit achieves overall scores of 4.54 and 8.13 with FLUX.2-klein-9B, outperforming strong state-of-the-art refinement methods. Further analyses show that FeedEdit learns effective adaptive stop-or-refine behavior and transfers across editors without retraining. We plan to release the dataset, model checkpoints, and code.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Image Editing, Prompt Refinement, Learning from Feedback, Visual Large Language Models, Self-Distillation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_k9LnOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces FeedEdit, a feedback driven prompt refinement framework for instruction guided image editing. Instead of explicitly generating an eval of intermediate editing results at inference time, it learns when to stop and how to refine prompts using evaluator feedback during training, combining offline learning from causal feedback with online self distillation from hindsight feedback.

Critical strengths
  • The paper is easy to follow and code release statement is available.
  • The paper reports some gains of Flux-Klein-9B with FeedEdit on two image editing benchmarks over some existing prompt refinement frameworks.
Critical weaknesses
  • I find novelty to be limited as compared to a closely related prior work: How to Train Your Advisor: Steering Black Box LLMs with Advisor Models (Asawa et al., 2026, ICML 2026). The core idea of FeedEdit appears closely related to this paper. Both methods train a separate policy to generate instance specific textual guidance for a frozen downstream model, use feedback from the downstream model's outputs to improve this policy and support iterative refinement and transfer across downstream models. Importantly, Advisor Models already considers settings where the advisor observes an initial model output before suggesting corrections, making the conceptual overlap particularly relevant. While FeedEdit introduces a different training strategy based on offline supervision and online self distillation, along with an explicit stopping mechanism for image editing, the paper does not sufficiently distinguish these contributions from the broader framework established by Advisor Models. The absence of a discussion or comparison with this closely related work makes it difficult to assess the methodological novelty of FeedEdit. A clearer comparison is needed to establish whether the proposed approach offers substantive advantages beyond adapting an existing feedback driven guidance framework to image editing. Perhaps using a similar setting to that of Advisor Model and adapting that as-is to the Image Editing use case could have given similar improvements. So, it's unclear since we do not see such comparison baselines.
  • The proposed framework depends heavily on evaluator generated scores and textual feedback during both offline trajectory construction and online self distillation. These signals determine which refinement trajectories are retained, when the model should stop and how the policy is updated. However, the paper does not sufficiently investigate how inaccurate or inconsistent evaluator feedback affects the learned refinement behavior. This is particularly concerning because one of the main motivations is to avoid unreliable intermediate evaluations at inference time, while the training process still relies extensively on such judgments. Also, even though the authors emphasize inference efficiency by eliminating explicit evaluation generation, they do not provide a comprehensive comparison of actual inference latency, average editor calls or computational costs. As a consequence, the robustness of the learned policy and the practical efficiency advantages of the proposed framework remain insufficiently established. Finally, the authors used just one model Flux Klein 9B to report results and effectiveness of the approach which by itself is insufficient (see my question below for further details on this).
Questions for authors
  • The improvement from the causal feedback SFT checkpoint to the full FeedEdit model is relatively small, increasing from 4.51 to 4.54 on ImgEdit-Bench and from 8.09 to 8.13 on GEdit-Bench. Could the authors provide statistical significance analysis to establish whether these improvements are consistent?
  • As pointed out in the weaknesses, since both offline trajectory construction and online self distillation depend on evaluator generated scores and textual feedback, how sensitive is FeedEdit to inaccurate or inconsistent evaluator judgments? Could the authors evaluate the method using a different evaluator or controlled perturbations of the feedback to demonstrate robustness?
  • The reported efficiency analysis mainly considers prompt length and whether evaluation text is generated. Could the authors provide actual inference latency, average number of editor calls and total computational cost per sample compared with EditThinker and other iterative refinement baselines?
  • The idea of prompt rewriting or refinement makes more sense when the editing model is black-box. The authors used just one model variant Flux Klein 9B in this case. I feel the base model could have been updated since weights are open so the utility of the approach is not established well. How about black-box models like gpt-image-2, nanobanana where prompt refinement actually makes sense since we can't access their weights? Also, how does this approach works with smaller variants like Flux Klein 4B where editing performance can actually be improved with such prompt refinement strategies?
  • The offline stopping targets are constructed using a fixed threshold on the sum of instruction compliance and content preservation scores. How sensitive is the learned stopping behavior to this threshold? Particularly, could the authors clarify whether the policy sometimes stops prematurely on edits that still contain significant instruction violations?
  • FeedEdit achieves 4.09 on FLUX.1 Kontext dev compared with 4.08 for EditThinker. Given the very small difference, a more detailed analysis of cross editor transfer, including whether the improvement is consistent across editing categories is needed.
  • In the main figure 1, I am not sure if the bird's fur color is changed to light brown from red, it seems yellow to me. Moreover, in Figure 3, first row, there's no reason to think that Qwen3-VL-8B Instruct output is bad, the instruction was to "Change the garden environment in the picture to a snowy landscape." and the foreground can very well change as well. Nothing in the prompt specified anything particular to the background and so background preservation may not be required on this example. Was there any human study conducted?
7 comments
AuthorsOpenReview ↗

Response (1/5): W1 (Relation to Advisor Models)

We thank the reviewer for the detailed review and for pointing us to Advisor Models. We now cite it in the related work (Section 5) and discuss it in Appendix A.4. Table, section, and appendix numbers refer to the revised PDF.

W1. Novelty relative to Advisor Models

FeedEdit shares the frozen-model steering setup but differs in decoupling, training signal, and learned stopping; Advisor Models adapted to our task is significantly below FeedEdit in all four variants.

What is shared. We agree that both works train a separate textual policy that steers a frozen downstream model, including settings where the policy observes an existing output before suggesting a correction. We do not claim this paradigm as our contribution; within image editing it is already instantiated by iterative refiners such as EditThinker.

What is different. While FeedEdit shares this setup, it differs from Advisor Models in three concrete ways:

  1. Evaluation–refinement decoupling. The deployed policy generates no critique and receives no evaluator input; it maps the editing context directly to a stop-or-refine action. Evaluation information is retained in learning and removed from the deployed generation path.
  2. Training signal. Advisor Models optimize the advice policy with reinforcement learning (GRPO) on rewards derived from downstream outcomes. FeedEdit instead uses the content of the evaluation, i.e., textual feedback together with scores, as privileged supervision: causal feedback yields execution-verified stop and refinement targets offline, and hindsight feedback conditions a self-teacher that gives token-level supervision online.
  3. Learned stopping. The policy explicitly learns when not to act, from verified stop targets. This matters in image editing, where an unnecessary extra edit can damage content that was already correct.

Direct comparison: Advisor Models on our task. As suggested, we adapted Advisor Models (3-step) to image editing, with the same FLUX.2-klein-9B editor, test sets, and judges as our controlled comparison.

  • Inference. The advisor sees the instruction, the source image, and the first edit, and writes advice. The editor then re-edits the source with the instruction plus the advice ("append"). This is exactly 2 editor calls, following the three steps of Advisor Models (Initial Attempt, Advice Generation, and Final Revision). We also test a "rewrite" variant in which the advisor's output replaces the prompt.
  • Training. We use the same base model as FeedEdit (Qwen3-VL-8B-Instruct) and follow the GRPO training recipe of Advisor Models, with a KL penalty of 0.001 and the scalar reward √(C·P)/5 from our evaluator. Textual feedback is not used.
  • Budget. The training budget is matched to FeedEdit's online training stage. We use refinement samples to train the Advisor Model following its three-step framework.
MethodImgEditGEdit-ENΔ\Delta vs. FeedEdit [95% CI] (ImgEdit / GEdit-EN)
Editor only4.2797.898—
Advisor, untrained, append4.3007.826−0.241 [−0.303, −0.181] / −0.309 [−0.429, −0.193]
Advisor, untrained, rewrite4.3647.881−0.177 [−0.235, −0.120] / −0.254 [−0.384, −0.125]
Advisor, trained, rewrite4.2577.877−0.284 [−0.351, −0.218] / −0.258 [−0.381, −0.135]
Advisor, trained, append4.3967.925−0.145 [−0.202, −0.089] / −0.210 [−0.338, −0.086]
FeedEdit4.5418.135—
  • Every variant is significantly below FeedEdit on both benchmarks, including the best one (trained, append).
  • FeedEdit also uses fewer editor calls: 1.78 / 1.66 on average, against a fixed 2.

Adapting Advisor Models to image editing therefore does not yield similar improvements. Unlike an LLM that can flexibly interpret high-level textual hints, an image editor is conditioned directly on the exact wording of the prompt, making imprecise or over-specified advice prone to degrading the edited results.

In summary, FeedEdit shares the frozen-model steering setup with Advisor Models but contributes evaluation–refinement decoupling and a different learning mechanism, and the comparison above shows that these differences, rather than the shared setup, account for the gains. This discussion and the comparison are now in Appendix A.4 (Table 9).

AuthorsOpenReview ↗

Response (2/5): W2 + Q2 (Reliance on Evaluator Feedback and Robustness), Q5 (Stopping Threshold)

W2 (first part) + Q2. Training relies on evaluator feedback; sensitivity to inaccurate or inconsistent judgments

Training feedback is verified by executing every candidate; the online gain survives a different training evaluator, and a blind human study prefers FeedEdit.

Is training-time reliance inconsistent with our motivation? We believe not, because evaluation plays fundamentally different roles at inference and in training:

  • At inference, an evaluate-then-refine policy acts on a single self-generated critique: one erroneous judgment directly misdirects the next edit, with no verification mechanism.
  • In training, feedback is aggregated over many samples and is never trusted blindly. In offline construction, every candidate refinement is actually executed and is retained only if the total score improves by at least 2 points with no dimension decreasing, which filters out noisy scores and unhelpful critiques via observed outcomes. In online self-distillation, the feedback describes the outcome of an action that was actually executed rather than an unverified guess about the next edit.

Decoupling also lets us employ a stronger dedicated evaluator in training without incurring its latency or failure modes at deployment. Our training does not require every individual judgment to be accurate.

Robustness evidence:

  1. Alternative evaluator. (a) Replacing GPT-5.6-Luna with Qwen3-VL-235B as the online-training evaluator changes final scores by +0.001 (ImgEdit) and −0.026 (GEdit-EN), so the online stage still improves over the SFT checkpoint. (b) Re-scoring all five methods on the same editor (editor only, Qwen3-VL-8B, EditThinker, FeedEdit-SFT, and FeedEdit) with Qwen3-VL-235B as the benchmark judge yields rankings closely aligned with the official judges (Kendall's τ\tau = 1.0, 95% CI [0.6, 1.0] on ImgEdit; τ\tau = 0.8, 95% CI [0.4, 1.0] on GEdit-EN), with FeedEdit ranking first on both benchmarks under either judge.
  2. Controlled feedback perturbation. Swapping 50% of the hindsight textual feedback and scores with those of other samples during online training lowers GEdit-EN by 0.035, with ImgEdit at +0.001, so corrupted feedback does make the policy worse. Stopping behavior shifts less (ImgEdit / GEdit-EN): the stop rate by +0.0 / −2.2 pp, and the over-optimization rate (already satisfactory edits that the policy still refines) by +1.2 / +0.2 pp. This is consistent with noisy feedback primarily affecting the direction of prompt rewriting rather than the decision to continue or stop.
  3. Human validation. In a blind study of 600 pairwise comparisons, human annotators significantly prefer FeedEdit over all three baselines (Table 3), indicating that the gains are not an artifact of evaluator-specific preferences.

Q5. Sensitivity to the stopping threshold; does the policy stop prematurely?

With our threshold, premature stops occur on 11.8% of clearly failed first edits, against 27.5% and 76.5% with lower thresholds.

Threshold sensitivity. We trained FeedEdit-SFT with stop pairs constructed at lower thresholds and evaluated them on the ImgEdit subset where the first edit clearly failed (Turn-1 score < 3, n = 102):

Stop thresholdPremature stops on failed first editsScore change on these samples
S > 8 (ours)11.8%—
S > 727.5%−0.48
S > 576.5%−1.35

Lower thresholds teach the policy to terminate on edits with clear instruction violations, degrading final quality accordingly. With S > 8, premature stops occur on only 11.8% of failed first edits (vs. 27.5% at S > 7 and 76.5% at S > 5).

Why S > 8? With integer scores in 1–5, S = C + P > 8 requires both dimensions to be at least 4 and at least one to be 5; indeed, 70% of our stop targets (18,182 of 25,967) are perfect (5, 5) states. The only stricter choice is to require (5, 5) exclusively, so S > 8 enforces high quality while retaining enough stop pairs to balance the refinement pairs during training.

Stopping quality on benchmarks. Evaluated by the official benchmark judges (Tables 12–13 in the appendix), the edits on which FeedEdit chooses to stop are sharply separated in quality from those it chooses to refine, more so than for EditThinker on both benchmarks:

Turn-1 scoreStop subsetContinue subsetFeedEdit gapEditThinker gap
ImgEdit (/5)4.723.35+1.37+0.78
GEdit-EN (/10)8.286.78+1.50+0.97
AuthorsOpenReview ↗

Response (3/5): W2 + Q3 (Latency, Editor Calls, Cost), Q1 (Statistical Significance)

W2 (second part) + Q3. Actual inference latency, editor calls, and computational cost

Versus EditThinker, FeedEdit generates 87–91% fewer tokens and is 1.8× / 4.3× faster end to end (ImgEdit / GEdit-EN), at higher quality.

We now report all three quantities (Table 2). Each cell: average editor calls / total generated tokens over all iterations / estimated latency (s), all with the FLUX.2-klein-9B editor. Bold marks the lower value between EditThinker and FeedEdit.

MethodImgEditGEdit-EN
Editor only1.00 / 0 / 1.791.00 / 0 / 0.46
Qwen3-VL-8B-Instruct (untrained)1.26 / 25 / 2.691.78 / 48 / 1.26
EditThinker1.57 / 363 / 5.641.89 / 437 / 3.97
FeedEdit1.78 / 46 / 3.851.66 / 40 / 1.14
  • Editor calls. FeedEdit averages fewer than two editor calls (1.78 / 1.66), well below the budget of 5. This is because FeedEdit can recognize when an edited result already satisfies the editing requirements.
  • Generated tokens and computational cost. FeedEdit produces 87.3% / 90.8% fewer tokens than EditThinker, since it decodes no evaluation or reasoning text. This cuts the refiner's autoregressive decoding by roughly 8× / 11× relative to EditThinker.
  • Latency. Estimated latency is 31.7% / 71.3% lower than EditThinker's.
  • Wall-clock validation. Direct end-to-end latency measurements on 30 samples per benchmark show the same ordering, with an even larger gap: on ImgEdit, FeedEdit takes 5.33 ± 4.44 s vs. 9.44 ± 4.43 s for EditThinker (43.5% lower, 1.8× faster); on GEdit-EN, 1.14 ± 0.77 s vs. 4.88 ± 2.76 s (76.6% lower, 4.3× faster).

Estimates are for a single H800 GPU, obtained from each method's measured prefill time with its own prompts and images, a decoding speed of about 148 tokens/s, the time of one editor call, and per-sample call and token counts. FeedEdit therefore achieves higher quality than EditThinker (4.54 vs. 4.50; 8.13 vs. 8.07) at substantially lower cost. We have added these results to the main text (Table 2, Section 4.2), with details in Appendix A.5.

Q1. Statistical significance of the SFT → full FeedEdit improvement

Over three seeds, the gain is significant against both SFT (p=0.021p = 0.021) and matched-budget GRPO (p=0.007p = 0.007).

We repeated the online stage with two additional seeds (three in total) for both FeedEdit and GRPO under the same setting. P-values are from per-setting paired permutation tests, combined with Stouffer's method (Table 11):

Mean difference over 3 seedsKlein / ImgEditKlein / GEdit-ENKontext / ImgEditCombined ppStandardized Δ\Delta [95% CI]
FeedEdit − FeedEdit-SFT+0.015+0.047+0.0440.021+0.030 SD [+0.004, +0.056]
FeedEdit − GRPO (matched budget)+0.013+0.036+0.0410.007+0.026 SD [+0.007, +0.045]

Standardized Δ\Delta is each setting's gain divided by the per-sample score SD, averaged over the three settings; its 95% CI is from 10,000 bootstrap resamples of the test items (the two ImgEdit settings resampled jointly).

On average over three seeds, FeedEdit outperforms both the SFT checkpoint and matched-budget GRPO across all three evaluation settings (combined p=0.021p = 0.021 and p=0.007p = 0.007, respectively).

Interpreting the magnitude. The online stage starts from an already strong SFT policy on benchmarks with limited headroom: on ImgEdit, 55% of samples already receive the maximum score with the editor-only baseline (ceiling effect), and the offline SFT stage has already captured a large portion of the available room for improvement. Within the remaining headroom, online training adds no inference cost, requires no new large-scale offline trajectory construction, and under the same training budget achieves a significantly larger improvement than scalar-reward GRPO. We now report the multi-seed results (Appendix A.6, Table 11) and characterize the gain as modest in absolute magnitude yet statistically significant (Section 4.2).

Reviewer_oxK9OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents FeedEdit, a feedback-driven prompt-refinement framework for instruction-guided image editing. FeedEdit formulates prompt refinement as an adaptive stop-or-refine decision problem.

Evaluator feedback is used during training to teach the policy whether to stop or how to generate a refined prompt, eliminating the need to explicitly generate an intermediate evaluation at inference time. The framework uses causal feedback to construct offline stopping and refinement targets, and hindsight feedback to provide privileged supervision for online self-distillation.

During inference, the deployed policy operates only on the feedback-free editing context and outputs either a stop decision or a refined prompt.

The authors evaluate a causal-feedback SFT checkpoint, Vanilla GRPO with scalar rewards, and hindsight-feedback self-distillation on ImgEdit-Bench and GEdit-Bench-EN. On the FLUX.2-klein-9B editor, the hindsight variant improves the overall scores from 4.51 to 4.54 on ImgEdit-Bench and from 8.09 to 8.13 on GEdit-Bench-EN. The paper also reports cross-editor transfer to FLUX.1-Kontext-dev without retraining, evaluated on ImgEdit-Bench.

Critical strengths

S1. Compared with methods that use a fixed refinement schedule, FeedEdit makes an adaptive stop-or-refine decision from the current editing context.This has practical significance because it can avoid unnecessary editor calls and reduce the risk of over-modifying non-target content.

S2. The framework uses evaluator feedback as privileged information during training while keeping the deployed policy feedback-free at inference time. This design allows the policy to learn feedback-guided refinement behavior without requiring explicit evaluation generation or evaluator feedback as an additional inference input, which simplifies deployment.

Critical weaknesses

W1. Both the offline and online training phases rely predominantly on GPT-5.6-Luna as the evaluator. Although downstream benchmarks utilize GPT-4o and GPT-4.1, the manuscript lacks rigorous cross-evaluator robustness analysis, such as assessing evaluator agreement, score consistency, and evaluator bias. Consequently, there remains a risk of adaptation to the specific preferences of the training evaluator. W2. The performance improvements attributed to hindsight-feedback self-distillation appear relatively small, increasing from 4.51 to 4.54 on ImgEdit-Bench and from 8.09 to 8.13 on GEdit-Bench-EN. Furthermore, since confidence intervals and results across multiple runs are not reported, the statistical significance and robustness remain uncertain. W3. The baselines use different image editors and experimental settings, making the overall comparison less consistent and complicating attribution of the reported gains, despite the more controlled FLUX.2-klein-9B block. W4. The online self-distillation framework lacks targeted ablations, such as removing the successful refinement demonstration, varying the teacher’s hindsight-feedback inputs, or testing alternative distillation objectives. Consequently, the respective contributions of the feedback-conditioned teacher and feedback-free student remain difficult to disentangle.

Questions for authors

as stated in Critical Weaknesses

3 comments
AuthorsOpenReview ↗

Response (1/2): W1 (Cross-Evaluator Robustness), W2 (Statistical Significance)

We thank the reviewer for the positive assessment and the constructive suggestions. Table, section, and appendix numbers refer to the revised PDF.

W1. Reliance on GPT-5.6-Luna; cross-evaluator robustness

FeedEdit ranks first under a judge from a different model family and in a blind human study, and still improves when trained with a different evaluator. We conducted four new analyses.

(1) Re-evaluation with a judge from a different model family. In addition to the official benchmark judges (GPT-4o, GPT-4.1), we re-scored the outputs of all five methods on the same editor (editor only, Qwen3-VL-8B, EditThinker, FeedEdit-SFT, and FeedEdit) using the open-source Qwen3-VL-235B. System-level rank correlation with the official judges is high on both benchmarks: Kendall's τ\tau = 1.0 (95% CI [0.6, 1.0]) and Spearman's ρ\rho = 1.0 (95% CI [0.7, 1.0]) on ImgEdit, and Kendall's τ\tau = 0.8 (95% CI [0.4, 1.0]) and Spearman's ρ\rho = 0.9 (95% CI [0.5, 1.0]) on GEdit-EN. FeedEdit ranks first on both benchmarks under either judge.

(2) Changing the training evaluator. We replaced GPT-5.6-Luna with Qwen3-VL-235B as the online-training evaluator and retrained FeedEdit:

BenchmarkScore difference (Qwen-trained − Luna-trained)
ImgEdit+0.001
GEdit-EN−0.026

The online stage still improves over the SFT checkpoint with this evaluator, so the method is not tied to one training evaluator.

(3) Evaluator bias. We find that the size of FeedEdit's advantage depends on the judge; it is largest under GPT-5.6-Luna, the training evaluator. FeedEdit, however, ranks first under every judge. Evaluator choice thus affects the estimated margin but not the comparative conclusion.

(4) Blind human evaluation (Table 3). Three annotators completed 600 blind pairwise comparisons. In overall preference, FeedEdit is significantly preferred over all three baselines:

FeedEdit vs.Win %Tie %Loss %pp
Editor only43.643.013.41.5e-8
Qwen3-VL-8B-Instruct35.347.717.00.0004
EditThinker25.561.113.40.0088

Together, these results indicate that the improvements are not an artifact of adapting to the training evaluator's preferences.

W2. Small hindsight gains; no confidence intervals or multiple runs

Over three seeds, the gain is significant against both SFT (p=0.021p = 0.021) and matched-budget GRPO (p=0.007p = 0.007).

Multiple runs and significance (Table 11). We repeated online training with two additional seeds for both FeedEdit and GRPO (three seeds each):

Mean difference over 3 seedsKlein / ImgEditKlein / GEdit-ENKontext / ImgEditCombined ppStandardized Δ\Delta [95% CI]
FeedEdit − FeedEdit-SFT+0.015+0.047+0.0440.021+0.030 SD [+0.004, +0.056]
FeedEdit − GRPO (matched budget)+0.013+0.036+0.0410.007+0.026 SD [+0.007, +0.045]

P-values are from per-setting paired permutation tests combined with Stouffer's method. Standardized Δ\Delta is each setting's gain divided by the per-sample score SD, averaged over the three settings; its 95% CI is from 10,000 bootstrap resamples of the test items (the two ImgEdit settings resampled jointly). The gain is positive in every setting.

Interpreting the magnitude. The online stage starts from an already strong SFT policy on benchmarks with limited headroom: on ImgEdit, 55% of samples already receive the maximum score with the editor-only baseline (ceiling effect). Within the remaining headroom, online training adds no inference cost, requires no new large-scale offline trajectory construction, and under the same training budget achieves a significantly larger improvement than scalar-reward GRPO (p=0.007p = 0.007). We now report the multi-seed results (Appendix A.6, Table 11) and characterize the gain as modest in absolute magnitude yet statistically significant (Section 4.2).

AuthorsOpenReview ↗

Response (2/2): W3 (Unified-Editor Baselines), W4 (Online Self-Distillation Ablations)

W3. Baselines use different editors, complicating attribution

All refiners are now compared on one frozen editor, and FeedEdit has the best overall score.

We agree that the controlled FLUX.2-klein-9B block is the appropriate basis for attribution, and we have extended it. DIM-4.6B-Edit and ReasonEdit-S were originally reported with their own editors; we now pair their instruction refiners with the same frozen FLUX.2-klein-9B editor and run them under the same experimental setting as all other methods:

Refiner on FLUX.2-klein-9BImgEditGEdit-ENFeedEdit advantage
Editor only4.2797.898+0.262 / +0.237
+ ReasonEdit-S refiner4.2117.944+0.330 / +0.191
+ DIM refiner4.0887.398+0.453 / +0.737
+ FeedEdit4.5418.135—

With this addition, every refiner baseline except MIRA (not publicly released) is compared under one editor and one experimental setting, and FeedEdit outperforms all of them in overall score on both benchmarks. In the revision, Table 1 compares all refiners on the same frozen FLUX.2-klein-9B editor; native-editor results and MIRA are in Appendix A.3 (Table 7).

W4. Targeted ablations of online self-distillation

Textual feedback drives corrections, demonstrations curb over-refinement, and distillation beats GRPO.

We ran three new ablations, all initialized from the same SFT checkpoint with identical training configurations. The over-optimization rate is the fraction of samples whose current edit is already satisfactory yet the policy continues to refine. Rows report changes relative to the full FeedEdit setting, whose absolute values are in the bottom row (ImgEdit / GEdit-EN):

AblationScore Δ\DeltaFirst-edit stop rateOver-optimization rate
No successful demonstration−0.029 / +0.016−5.6 / −6.3 pp+2.8 / +2.2 pp
No textual feedback−0.014 / −0.016+5.9 / +6.8 pp−0.6 / −2.3 pp
50% of feedback and scores swapped with other samples+0.001 / −0.035+0.0 / −2.2 pp+1.2 / +0.2 pp
FeedEdit4.541 / 8.13567.5% / 71.9%6.2% / 8.6%

Demonstrations and textual feedback are complementary. Removing textual feedback causes the policy to stop earlier (+5.9 / +6.8 pp) and lowers scores on both benchmarks, indicating that feedback drives necessary corrective refinements. Conversely, removing successful demonstrations deprives the teacher of a concrete example of what action succeeds in the same context. As a result, the policy continues modifying already adequate outputs more often, raising over-optimization on both benchmarks (+2.8 / +2.2 pp). While extra refinement steps can improve under-edited cases, they can also degrade other samples through over-editing. Combining both signals allows textual feedback to trigger targeted corrections when needed, while successful demonstrations help establish an appropriate stopping boundary.

Teacher input quality mainly affects rewriting direction. Corrupting half of the hindsight feedback lowers GEdit-EN by 0.035, whereas stopping behavior shifts only slightly (stop rate +0.0 / −2.2 pp; over-optimization +1.2 / +0.2 pp). This is consistent with mismatched feedback breaking the correspondence between an action and its outcome, primarily impairing the direction of rewriting rather than the decision to continue.

Alternative online objective. We also replaced the distillation objective with scalar-reward policy optimization (GRPO), keeping the SFT initialization, data, and training budget the same. Over three seeds, FeedEdit is ahead of GRPO by +0.013 on Klein / ImgEdit, +0.036 on Klein / GEdit-EN, and +0.041 on Kontext / ImgEdit (combined p=0.007p = 0.007; Table 11; see W2 in Part 1/2). Distilling from the feedback-conditioned teacher is thus more effective than optimizing the evaluator's scalar score directly.

Overall, the feedback-conditioned teacher determines what is learned (when to continue and in which direction to rewrite), and the feedback-free student reliably absorbs it. These ablations are now in the paper (Section 4.3; Appendix A.11, Table 16), and we thank the reviewer for suggesting them.

We hope these additions resolve the remaining concerns and strengthen your support for the paper. We would be happy to provide any further clarification.

Dear Reviewer oxK9,

Thank you again for your thoughtful review and constructive suggestions. We have carefully considered your feedback and made every effort to address each concern, with new experiments in the revised paper:

  • W1 (evaluator robustness): training with a different evaluator, re-scoring with a third judge, and a blind human study in which annotators prefer FeedEdit over all three baselines.
  • W2 (statistical reliability): three random seeds with significance tests and confidence intervals for the online gain.
  • W3 (consistent comparison): all baselines, including the DIM and ReasonEdit-S refiners, evaluated with the same editor and setting.
  • W4 (ablations of online self-distillation): removing the successful demonstration, removing the textual feedback, and corrupting the feedback.

As the discussion deadline is approaching, we would greatly appreciate it if you could take a look at our responses and let us know whether any questions remain. We would be happy to provide further clarification before the discussion period ends. If our responses have addressed your concerns, we would be grateful if you would consider updating your assessment.

Thank you again for your time and consideration.

Best regards,

The Authors

Reviewer_5xQQOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents FeedEdit, an adaptive prompt refinement framework for image editing with a frozen editor. Given the source image, original instruction, current prompt, and edited result, the policy directly predicts whether to stop or produce a revised prompt. Training consists of offline supervised learning from feedback-guided trajectories, followed by online self-distillation using outcome feedback as privileged teacher information. At inference, the policy operates without external evaluator feedback or explicit critique generation. Experiments with FLUX.2-klein-9B achieve overall scores of 4.54 on ImgEdit-Bench and 8.13 on GEdit-Bench-EN, compared with 4.28 and 7.90 for the base editor.

Critical strengths
  • Learning when to stop and how to refine prompts is useful for improving image editing without modifying the underlying editor. The direct stop-or-refine interface is simple at inference.
  • Refinement targets are retained based on improvements in executed editing results rather than prompt quality alone. Combining verified refinement examples with stopping examples provides a sensible training signal for adaptive behavior.
  • The method improves automatic scores on both evaluated benchmarks. The feedback ablation and cross-editor evaluation provide additional evidence that the training pipeline is useful.
Critical weaknesses
  • The method combines established iterative prompt refinement with SDPO. The main contribution therefore appears to be their adaptation and integration into a direct stop-or-refine policy. A matched comparison with and without explicit critique generation is needed to isolate the contribution of the direct stop-or-refine design.
  • Both training and evaluation rely on model-based judges with related criteria, which may share systematic preferences or blind spots. Higher automatic scores could therefore partly reflect alignment with evaluator preferences rather than better editing quality. Blinded human evaluation is needed to validate the reported improvements.
  • The quality–cost trade-off is unclear. Final-prompt length does not capture the cost of up to five editor calls; average editor calls and end-to-end latency are needed. Moreover, hindsight training improves the SFT checkpoint only from 4.51 to 4.54 and from 8.09 to 8.13. It is unclear whether these gains are reliable and justify the additional training complexity.
Questions for authors
  • How does FeedEdit compare with strong one-shot prompt enhancement and repeated sampling under comparable budgets?
  • Are the hindsight gains consistent across training settings? How much comes from textual feedback versus successful demonstrations?
  • Does refinement mainly adapt wording to the editor or correct observed errors? An ablation removing the intermediate image would help.
5 comments
AuthorsOpenReview ↗

Response (1/3): W1 (Contribution and Matched Comparison), W2 (Evaluator Bias and Human Study)

We thank the reviewer for the thoughtful and constructive review. We address W1–W2 here, W3 in Part 2, and Q1–Q3 in Part 3. Table, section, and appendix numbers refer to the revised PDF.

W1. Contribution beyond "iterative refinement + SDPO"; matched comparison with and without critique generation

Our contribution is the decoupling itself; adding evaluation generation back lowers scores (−0.057 / −0.082 on ImgEdit / GEdit-EN) with about three times as many tokens.

Positioning. SDPO is the optimization tool we use for the online stage, and we do not claim it as a contribution. Our core contribution is decoupling evaluation from refinement: evaluation is delegated to an external evaluator and used as privileged training-time information, and its guidance is distilled into a policy that directly outputs a stop-or-refine decision. Achieving this requires more than plugging SDPO into an iterative loop: the offline stage constructs stop targets and execution-verified refinement targets from causal feedback, while the online stage leverages hindsight feedback on the policy's own actions. Because evaluation is external, each module specializes in its own role—the evaluator focuses solely on critique while the policy focuses purely on refinement—allowing any feedback source (a stronger VLM, or even human feedback) to be plugged in without changing the deployed policy.

Matched comparison with vs. without critique generation. As suggested, we trained a variant supervised to generate the evaluator feedback before the refined prompt, using the same training samples and keeping every other setting identical to FeedEdit-SFT:

Variant vs. FeedEdit-SFTImgEditGEdit-ENGenerated tokens
+ explicit evaluation generation−0.057−0.0822.7× / 3.2×

Generating the critique at inference time is therefore unnecessary, and even slightly detrimental. This is likely because a self-generated critique can be inaccurate and then misdirects the refinement that follows, while adding decoding cost. Using evaluation as privileged training guidance and removing it from the output gives higher editing quality at roughly one third of the generated tokens, which isolates the benefit of the direct stop-or-refine design.

Comparison with a critique-based iterative refiner. EditThinker is the closest existing instance of iterative refinement with outcome-based training: it generates a critique before each refinement and is trained with GRPO on the scalar score of the edited result. On the same FLUX.2-klein-9B editor, FeedEdit scores higher (4.54 vs. 4.50 on ImgEdit; 8.13 vs. 8.07 on GEdit-EN) while generating 87–91% fewer tokens (Table 2), and it is preferred in the blind human study (25.5% win vs. 13.4% loss; W2 below). Within our own framework, replacing hindsight-feedback distillation with GRPO under the same data and budget is also worse over three seeds (p=0.007p = 0.007; Table 11), so, for image-editing prompt refinement, distilling the evaluator's textual feedback on the executed edit is a more effective training signal than its scalar score alone.

W2. Shared preferences among model-based judges; blinded human evaluation

FeedEdit also ranks first under a judge from another model family and is significantly preferred in a blind human study. Details:

(i) Training and benchmark judges already differ. Training uses GPT-5.6-Luna with our own two-dimension rubric, whereas the benchmarks use their official protocols with GPT-4o (ImgEdit) and GPT-4.1 (GEdit-EN). We acknowledge, however, that these judges belong to the same model family.

(ii) Re-evaluation with a judge from a different family. We re-scored all five methods on the same editor (editor only, Qwen3-VL-8B, EditThinker, FeedEdit-SFT, and FeedEdit) with the open-source Qwen3-VL-235B. Its system-level ranking agrees closely with the official judges (Kendall's τ\tau = 1.0 on ImgEdit, 0.8 on GEdit-EN), and FeedEdit remains ranked first on both benchmarks.

(iii) Blind human evaluation (Table 3). Three annotators performed 600 blind pairwise comparisons. FeedEdit is preferred significantly more often than each baseline in overall preference:

FeedEdit vs.Win %Tie %Loss %pp
Editor only43.643.013.41.5e-8
Qwen3-VL-8B-Instruct35.347.717.00.0004
EditThinker25.561.113.40.0088

Inter-annotator agreement is Fleiss' κ\kappa = 0.401 for the three-way choice; most disagreement is between a preference and a tie (reflecting differing annotator strictness/tie thresholds), two annotators choose opposite sides in only 7.2% of comparisons, and every individual annotator favors FeedEdit. The protocol and statistical test are described in Appendix A.7. These results indicate that the automatic improvements correspond to edits that humans prefer, rather than to alignment with a particular judge.

AuthorsOpenReview ↗

Response (2/3): W3 (Quality–Cost Trade-off and Reliability of the Hindsight Gain)

W3a. Quality–cost trade-off: editor calls and end-to-end latency

FeedEdit averages 1.78 / 1.66 editor calls (ImgEdit / GEdit-EN) and, versus EditThinker, generates 87–91% fewer tokens and is 1.8× / 4.3× faster end to end.

We agree that final-prompt length alone does not capture multi-turn overhead, as it excludes the number of editor calls and the generated evaluation tokens. We now report average editor calls, total generated tokens across all iterations, and latency alongside quality (full comparison in Table 2):

BenchmarkMethodScore ↑Avg. editor calls ↓Generated tokens ↓Est. latency ↓
ImgEditEditThinker4.501.573635.64 s
ImgEditFeedEdit4.541.7846 (−87.3%)3.85 s (−31.7%)
GEdit-ENEditThinker8.071.894373.97 s
GEdit-ENFeedEdit8.131.6640 (−90.8%)1.14 s (−71.3%)

Estimated latency (single H800 GPU) combines each method's measured prefill time with its own prompts and images, a decoding speed of about 148 tokens/s, the time of one editor call, and per-sample call and token counts.

  • Low call count via learned stopping. FeedEdit averages fewer than two editor calls (1.78 on ImgEdit; 1.66 on GEdit-EN), well below the budget of 5, as it can recognize when a sample already satisfies the editing instruction and avoid unnecessary iterations.
  • Better quality at lower cost. Compared with EditThinker, FeedEdit reduces generated tokens by 87.3% / 90.8% and estimated latency by 31.7% / 71.3% while achieving higher quality scores across both benchmarks.
  • Wall-clock validation. We also measured end-to-end latency directly on 30 samples per benchmark. On ImgEdit, FeedEdit takes 5.33 ± 4.44 s vs. 9.44 ± 4.43 s for EditThinker (43.5% lower, 1.8× faster); on GEdit-EN, 1.14 ± 0.77 s vs. 4.88 ± 2.76 s (76.6% lower, 4.3× faster).

We have added these results to the main text (Table 2, Section 4.2), with details in Appendix A.5.

W3b. Reliability and cost-effectiveness of hindsight online training

Over three seeds, the online gain is modest but significant (combined p=0.021p = 0.021), larger than GRPO's under the same budget, and free at inference.

1. Statistical reliability across seeds (Table 11). We repeated the online stage with two additional seeds (three in total) for both FeedEdit (SFT + hindsight online training) and GRPO under the same setting. P-values are from per-setting paired permutation tests, combined with Stouffer's method. Standardized Δ\Delta is each setting's gain divided by the per-sample score SD, averaged over the three settings; its 95% CI is from 10,000 bootstrap resamples of the test items (the two ImgEdit settings resampled jointly).

Mean score difference (3 seeds)Klein / ImgEditKlein / GEdit-ENKontext / ImgEditCombined ppStandardized Δ\Delta [95% CI]
FeedEdit − FeedEdit-SFT+0.015+0.047+0.0440.021+0.030 SD [+0.004, +0.056]
FeedEdit − GRPO (matched budget)+0.013+0.036+0.0410.007+0.026 SD [+0.007, +0.045]

On average over three seeds, FeedEdit outperforms both the SFT checkpoint and matched-budget GRPO in all three evaluation settings (combined p=0.021p = 0.021 and p=0.007p = 0.007, respectively).

2. Why the absolute gain is modest (ceiling effect). The online stage starts from a strong baseline: with the editor-only baseline, 55% of ImgEdit samples already achieve the maximum score, and the offline SFT stage has already captured a large portion of the available room for improvement. Because the online stage can only improve the remaining unresolved cases, a modest average gain is expected; the multi-seed tests indicate that this improvement is systematic.

3. Why it is worth the minimal complexity.

  • Zero inference overhead. The online-trained policy is a drop-in replacement for the SFT checkpoint, requiring no extra model, evaluator, or generation step at test time.
  • Lightweight on-policy training. The online stage learns directly from the current policy's own rollouts without requiring any new large-scale offline trajectory construction.
  • Higher sample efficiency than scalar-reward RL. Under the same training budget, FeedEdit's gain over SFT is far larger than GRPO's, leaving FeedEdit ahead by +0.013 / +0.036 / +0.041 (p=0.007p = 0.007). We interpret this as textual outcome feedback providing a denser training signal than a scalar reward for this task.

We have revised the main text to characterize the online gain as modest but statistically significant (Section 4.2; Appendix A.6).

AuthorsOpenReview ↗

Response (3/3): Q1 (Budget-Matched Baselines), Q2 (Consistency and Feedback vs. Demonstrations), Q3 (No-Image Ablation)

We thank the reviewer for these insightful questions. Following your suggestions, we conducted three new experiments, which we report below.

Q1. Comparison with strong one-shot prompt enhancement and repeated sampling under comparable budgets

FeedEdit improves over one-shot enhancement by +0.09 / +0.27 (ImgEdit / GEdit-EN) and over Best-of-2 by +0.257 / +0.185, while generating the fewest tokens.

  • One-shot enhancement. UnifiedThinker-7B (Table 1) is a strong one-shot baseline: it reasons explicitly before editing and emits one enhanced prompt, using the same FLUX.2-klein-9B editor.
  • Repeated sampling. We added a Best-of-2 baseline: the original instruction is executed twice with different random seeds, and Qwen3-VL-8B-Instruct selects the better output. Its fixed budget of 2 editor calls is slightly higher than FeedEdit's average (1.78 on ImgEdit, 1.66 on GEdit-EN), and it requires an additional selector at inference time.
MethodAvg. editor calls (ImgEdit / GEdit-EN)Avg. generated tokens (ImgEdit / GEdit-EN)ImgEditGEdit-EN
UnifiedThinker-7B (one-shot enhancement)1.00 / 1.00545 / 5314.457.86
Best-of-2 (repeated sampling)2.00 / 2.00154 / 156†4.2847.950
FeedEdit1.78 / 1.6646 / 404.5418.135

† Tokens generated by the selector when choosing between the two images.

FeedEdit's gains therefore do not stem from merely spending more editor calls: with a slightly lower editor-call budget than Best-of-2, refinement conditioned on the observed edit outperforms selecting among resampled outputs.

Q2. Consistency of the hindsight gain; textual feedback vs. successful demonstrations

Feedback drives corrections; demonstrations curb over-refinement.

(a) Consistency across settings. The gain is positive on average over three seeds in all three evaluation settings, covering two benchmarks and two editors (Table 11; combined p=0.021p = 0.021 vs. SFT, p=0.007p = 0.007 vs. GRPO). The gain also holds with a different online-training evaluator: replacing GPT-5.6-Luna with the open-source Qwen3-VL-235B changes final scores by +0.001 on ImgEdit and −0.026 on GEdit-EN, so the online stage still improves over the SFT checkpoint.

(b) Textual feedback vs. successful demonstrations. Starting from the same SFT checkpoint and training configuration, we re-ran online training with each teacher input removed. The full setting achieves 4.541 / 8.135 with first-edit stop rates of 67.5% / 71.9%. The over-optimization rate is the fraction of samples whose current edit is already satisfactory but which the policy nevertheless continues to refine. Changes relative to the full setting (ImgEdit / GEdit-EN):

Teacher inputScore Δ\DeltaFirst-edit stop rate Δ\DeltaOver-optimization rate Δ\Delta
No demonstration (feedback only)−0.029 / +0.016−5.6 / −6.3 pp+2.8 / +2.2 pp
No textual feedback (demonstration only)−0.014 / −0.016+5.9 / +6.8 pp−0.6 / −2.3 pp

The two signals play distinct, complementary roles:

  • Textual feedback drives error correction. Without it, the policy becomes overly conservative: it stops more often after the first edit (+5.9 / +6.8 pp) and scores drop on both benchmarks.
  • Successful demonstrations regularize against over-refinement. Without a demonstration as a concrete target, the policy more often refines outputs that are already good (+2.8 / +2.2 pp over-optimization) and makes more editor calls. The extra refinement gives a slight gain on GEdit-EN (+0.016) but a larger drop on ImgEdit (−0.029). Combining both signals gives the best balance across the two benchmarks at a higher stop rate.

For the offline stage, Table 4 independently supports the value of textual feedback: removing it lowers the trajectory score (7.746 → 7.233) and the yield of valid stop and refinement pairs.

Q3. Does refinement adapt wording to the editor, or correct observed errors?

It primarily corrects observed errors. Following your suggestion, we removed the intermediate edited image from the policy input and kept everything else unchanged.

Scores drop by 0.258 on ImgEdit (4.541 → 4.283) and 0.270 on GEdit-EN (8.135 → 7.865), i.e., to approximately the editor-only level (4.279 / 7.898), erasing essentially the entire gain of FeedEdit. If the improvement came from rephrasing instructions into editor-friendly wording, it would largely persist without the intermediate output. Instead, observing the actual edit is essential for the policy to decide whether to stop and what to fix, as illustrated in Figures 2 and 3.

These experiments are now included in the revised paper. We hope they address your concerns; if so, we would be grateful if you would consider raising your score. We are glad to clarify anything further.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response (1/2): Summary, Core Contribution, and How We Addressed the Concerns

We thank the Area Chair and all reviewers for their time and constructive feedback.

Summary. We appreciate that the reviewers found the stop-or-refine formulation useful and simple at inference (5xQQ, oxK9), the execution-verified training signal sensible (5xQQ), the feedback-free deployed policy practical (oxK9), and the paper easy to follow (k9Ln, oxK9). Our key novelty is decoupling evaluation from refinement: evaluator feedback serves as privileged training-time information and is distilled into a policy that directly decides to stop or refine, without generating or receiving any evaluation at inference. In response to the reviews, the revised PDF adds multi-seed significance tests, a blind human study, cross-evaluator and feedback-perturbation analyses, cost and latency measurements, new baselines (Advisor Models [1], Best-of-2, ReasonEdit-S, DIM), ablations of online training, and results on additional editors. Overall, FeedEdit improves the editor by +0.262 / +0.237 (ImgEdit / GEdit-EN), is significantly preferred by human annotators over all three baselines, and runs 1.8× / 4.3× faster than EditThinker; without retraining, it also improves two other open-weight editors.

The headline result for each concern is listed under "How we addressed the main concerns" below, and the summary of revisions is in Part 2/2. Table, section, and appendix numbers refer to the revised PDF.

1. Core contribution: decoupling evaluation from refinement

Evaluate-then-refine methods (e.g., EditThinker) make the deployed policy generate an evaluation of the current edit before refining the prompt. This costs inference tokens, needs extra supervision, and lets an inaccurate self-assessment misdirect the next refinement. FeedEdit instead decouples evaluation from refinement: an external evaluator provides feedback as privileged training-time information, which is distilled into a policy that sees only the editing context and directly decides to stop or refine.

  • Causal feedback ("what is wrong with the current edit?"), with scores and actual execution results, yields execution-verified stop and refinement targets for offline self-distillation.
  • Hindsight feedback ("what did the policy's own action produce?") conditions a self-teacher that gives token-level supervision for online self-distillation.

Both stages are privileged self-distillation: the teacher sees the evaluator feedback, and the deployed student does not. Together they yield the full gain above, and the online stage is the on-policy extension of the offline one.

Advisor Models [1] share the frozen-model steering setup but learn from a scalar outcome reward with RL and always issue advice. FeedEdit differs in three ways: it (i) distills the content of the evaluation, not only its score, in both the offline and online stages, (ii) learns when not to act, and (iii) keeps evaluation entirely out of the deployed policy.

2. How we addressed the main concerns

Paired values are ImgEdit / GEdit-EN.

  1. Novelty and positioning (5xQQ-W1, k9Ln-W1). Adding evaluation generation back lowers scores by 0.057 / 0.082 with about three times as many tokens, and Advisor Models [1] adapted to our task is significantly below FeedEdit (best variant: −0.145 / −0.210; Table 9).
  2. Reliability of the online gain (5xQQ-W3, oxK9-W2, k9Ln-Q1). Over three seeds, the gain of the online stage is modest but significant and consistent against both the offline (SFT) checkpoint (p=0.021p = 0.021) and GRPO (p=0.007p = 0.007), with no inference cost and no additional offline trajectory construction (Table 11).
  3. Dependence on model-based evaluators (5xQQ-W2, oxK9-W1, k9Ln-W2, k9Ln-Q2). FeedEdit is significantly preferred in a blind human study (Table 3) and ranks first under a judge from another model family.
  4. Inference cost (5xQQ-W3, k9Ln-W2, k9Ln-Q3). Versus EditThinker, FeedEdit generates 87–91% fewer tokens and is 1.8× / 4.3× faster end to end, with higher scores (Table 2).
  5. Baselines (5xQQ-Q1, oxK9-W3). FeedEdit has the best overall score against the added Best-of-2, ReasonEdit-S, and DIM baselines on the same editor.
  6. What is learned (5xQQ-Q2, 5xQQ-Q3, oxK9-W4). Removing the intermediate image drops scores to the editor-only level (−0.258 / −0.270), showing that FeedEdit primarily corrects observed errors; in the online teacher, textual feedback drives corrections and demonstrations curb over-refinement.
  7. Generality across editors (k9Ln-W2, k9Ln-Q4, k9Ln-Q6). Without retraining, FeedEdit improves FLUX.1-Kontext-dev by +0.35 and the smaller FLUX.2-klein-4B by +0.341 / +0.296.

References

[1] Asawa et al. How to Train Your Advisor: Steering Black-Box LLMs with Advisor Models. ICML 2026.

AuthorsOpenReview ↗

General Response (2/2): Summary of Revisions

The new results in this response and in the per-reviewer responses are included in the revised PDF, with changes marked in blue.

  • Table 1 now compares all refiners on the same frozen FLUX.2-klein-9B editor and adds the DIM, ReasonEdit-S, and Best-of-2 baselines; native-editor results and MIRA moved to Appendix A.3 (oxK9-W3, 5xQQ-Q1).
  • Inference cost: new Table 2 and paragraph in Section 4.2; Appendix A.5 (5xQQ-W3, k9Ln-Q3).
  • Human evaluation: new Table 3 in Section 4.2; protocol in Appendix A.7 (5xQQ-W2, k9Ln-Q7).
  • Significance of the online gain: Section 4.2; Appendix A.6 (5xQQ-W3, oxK9-W2, k9Ln-Q1).
  • Evaluator robustness: Appendix A.8 (5xQQ-W2, oxK9-W1, k9Ln-Q2).
  • Ablations of online training, evaluation generation, and the intermediate image: Section 4.3; Appendix A.11–A.12 (5xQQ-W1, Q2, Q3; oxK9-W4).
  • Stop-threshold sensitivity: Appendix A.9 (k9Ln-Q5).
  • Additional editors: Section 4.4, Limitations, Appendix A.13 (k9Ln-Q4, Q6).
  • Advisor Models: cited and contrasted with FeedEdit in Section 5; discussed and compared in Appendix A.4 (K9Ln-W1).
  • Positioning: the abstract and introduction now use the same terms as Section 3 and the related work (Section 5), which already described evaluation–refinement decoupling and the offline and online self-distillation stages; the abstract also reports the speedup and the human study (5xQQ, k9Ln-W1).

If these results address your concerns, we would be grateful if you would consider updating your assessment. We thank the Area Chair and the reviewers again, and we are happy to answer any further questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.