acceptodds
Withdrawn from ICLR 2027

Edit-GRPO: A Locality-Preserving Policy Optimization Framework for Image Editing

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

A fundamental challenge in image editing lies in preserving spatial locality: edits should improve targeted content without inadvertently altering surrounding regions. However, most optimization-based editing approaches treat images as holistic entities, causing global policy updates that undermine locality and introduce undesired context changes. We observe that this issue stems from a mismatch between localized editing intent and globally applied optimization signals. Motivated by this insight, we propose Edit-GRPO, preserving Locality while optimizing image editing, a locality-preserving policy optimization framework that explicitly decouples editing and preservation objectives. By assigning region-specific optimization signals to edit and non-edit areas, Edit-GRPO aligns policy updates with the spatial structure of editing tasks, enabling localized improvements while maintaining global visual coherence. This design effectively suppresses common artifacts such as context distortion and boundary inconsistency. Extensive experiments across diverse image editing scenarios demonstrate that Edit-GRPO significantly improves locality preservation while maintaining strong editing performance compared to existing optimization-based methods, validating the generality and effectiveness of the proposed framework.

Primary area:
generative models
Keywords:
Image Generation, Image Editing, Reinforcement Learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_dRzBOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Edit-GRPO, a region-aware RL framework for instruction-guided image editing that separates editing and preservation regions using masks and applies different rewards and GRPO optimization signals to each region. Experiments on FLUX.1-Kontext and Qwen-Image-Edit show improved background preservation while maintaining well editing performance.

Critical strengths
  1. The paper is well-written and the motivation is clear. Also, the proposed method directly targets the identified problem of the conflict between editing and background preservation. The overall design is intuitive and reasonable.

  2. It has relatively complete experiments on the two popular image-editing backbones, FLUX and Qwen. The ablation studies are also meaningful, including the comparison between semantic-only, preservation-only, combined, and decoupled optimization, which helps demonstrate the effectiveness of the proposed design.

Critical weaknesses
  1. Limited novelty. The proposed method is essentially a straightforward extension of existing GRPO-based optimization by introducing region masks and applying separate rewards to editing and preservation regions. While the motivation is reasonable, I find the technical contribution limited, insufficient to support a standalone contribution to image editing.

  2. Although the experiments are relatively comprehensive, I found several presentation flaws. First, Tables 1 and 2 include three additional baselines without clearly explaining why they were selected. All their locality preservation metrics are missing. Second, while Figure 4 provides qualitative comparisons, I believe the visual improvements are quite minor. I am not fully convinced that the improvements are good enough to justify the proposed method.

Questions for authors
  1. One of the main motivations of this paper is that global optimization can lead to poor locality preservation in image editing. But the current visual examples are insufficient to establish how common or significant this problem actually is (Fig 1 doesn't provide any useful information, and Fig 2 only show a single example). Can the authors provide a more rigorous analysis of why global GRPO causes this interference? For example, analysis of the optimization signals or gradient behavior in editing and non-editing regions would help establish whether the claimed conflict actually exists and how the proposed method addresses it.

  2. The proposed method shares similar motivations and design ideas with recent region-aware RL methods such as RC-GRPO-Editing and SpatialFlow-GRPO. Can the authors clarify what are the unique advantages of the proposed region-decoupled optimization strategy, beyond applying different rewards and advantages to different spatial regions?

  3. Tables 1 and 2 include InstructPix2Pix, OmniGen, and UniWorld-V1, but their locality preservation metrics are entirely missing. Can the authors clarify where these scores come from, why these particular baselines were selected, and why their preservation metrics were not evaluated?

Reviewer_SbU3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses a kind of core problem in instruction-guided image editing: making the requested change while preserving unrelated regions. It identifies edit and non-edit regions using automatically generated masks, assigns semantic-following rewards to the edit region and preservation rewards to the remaining content, and applies the resulting optimization signals separately at the image-token level. Experiments on FLUX.1 Kontext and Qwen-Image-Edit use GEdit-Bench and ImgEdit to assess editing quality and locality preservation. The paper reports improved preservation while maintaining competitive instruction-following performance, supported by ablations of its rewards and optimization objective.

Critical strengths
  1. The method separates edit and preservation regions, assigns them different rewards, and carries that separation into policy optimization. Generating masks from both the source and edited images is a practical way to handle changes in the target’s shape or extent.

  2. The evaluation part is across models and edit types, providing relatively comprehensive results. Results on both FLUX.1 Kontext and Qwen-Image-Edit, using GEdit-Bench and ImgEdit, help show that the approach is not tied to one editor or one collection of instructions.

Critical weaknesses
  1. Edit-GRPO revisits the established problem of preserving non-target regions during image editing in the setting of GRPO-based post-training, but does not adequately engage with earlier region-aware solutions, including closely related GRPO methods. Its approach combines existing techniques—automatic segmentation, region-specific rewards, and spatially separated policy updates—without clearly establishing which component or interaction constitutes a substantive new contribution. As presented, the method appears primarily incremental.
  2. Thus, the central novelty claim faces a close prior-work challenge. SpatialFlow-GRPO[1] also identifies the mismatch between global rewards and spatial editing and aligns region-level advantages with latent positions. RC-GRPO-Editing[2] addresses localized editing and preservation through different GRPO mechanisms. Edit-GRPO needs a precise account of what its reward construction and optimization add beyond these works.

[1] SpatialFlow-GRPO: Where Spatial Credit Drives Image Editing [2] Region-Constrained Group Relative Policy Optimization for Flow-Based Image Editing

Questions for authors
  1. Why is the closest region-aware GRPO method absent from the experiments? Could you compare the methods on a common backbone and benchmark, or provide a controlled implementation-level comparison where model availability prevents that?
  2. In what sense is the semantic reward region-specific? Eqs. (6)–(8) appear to score the original image, edited image, and instruction without taking the edit mask as an argument. Are images cropped or masked before evaluation? If so, please specify the procedure.
  3. What is the measured effect of mask errors? The appendix includes an independent SAM2 evaluation, which is helpful. Could you additionally report a human-checked mask subset, failure rates, and performance as a function of mask quality?
  4. How many benchmark cases are removed by the locality-evaluation exclusions?
  5. Please remove a notation collision. SS denotes the language-grounded segmentation module in Eq. (5) and the set of rating tokens in Eq. (6). Please use distinct symbols.
1 comment

Thank you for clarifying the intended contribution. I do not find the designation of RC-GRPO-Editing and SpatialFlow-GRPO as “concurrent works” sufficient to explain their absence from the submitted manuscript. They were publicly available in April and June 2026, respectively, while this manuscript was submitted in late September. This does not call into question independent development, but it provided time to discuss both works and position Edit-GRPO against them. The absence of official code may make a faithful numerical reproduction difficult; it does not remove the need for a precise technical comparison.

My main concern is the comparison with SpatialFlow-GRPO. The rebuttal distinguishes Edit-GRPO by stating that it decouples optimization at both the advantage estimation and policy objective stages. However, SpatialFlow-GRPO also computes region-level advantages and uses region-aggregated policy ratios and a region-aligned loss. Considering that SpatialFlow-GRPO already uses region-level advantages and a region-aligned policy objective, the claimed “region-decoupled policy optimization” therefore closely overlaps with existing work.

While the rebuttal clarifies several details, it does not resolve my central concern about the method’s distinction from existing region-aware policy optimization. I therefore maintain my original score.

Reviewer_2xF8OpenReview ↗
Rating
2(−1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

In this paper, the authors introduce Edit-GRPO which separates semantic editing and locality preservation through region-specific rewards and region-decoupled policy optimization. It assigns the two objectives to their corresponding image regions. The method is evaluated on FLUX.1-Kontext and Qwen-ImageEdit using GEdit-Bench and ImgEdit.

Critical strengths
  • The method addresses a practical problem: improving the requested edit while preserving surrounding content. Table 3 supports decoupling the objectives: the unchanged ratio increases from 57.90% with combined optimization to 60.30%, while the overall editing score increases from 6.15 to 6.34.

  • Preservation gains appear on both editing backbones. The additional evaluation withSAM2 masks and EditReward in Appendix D provides useful evidence beyond the training reward

Critical weaknesses
  • The preservation masks depend partly on each edited output, and Appendix A.2 filters cases by mask area without reporting retained counts for each method. Methods could therefore be scored on different regions or cases. The SAM2 check reduces dependence on the training segmentation model but does not establish that these comparisons use common evaluation regions.
  • The optimization signs need clarification. Equations 14-15 define a positive clipped policy objective, whereas Algorithm 1 applies gradient descent to it.
Questions for authors

Please address the points raised in the critical weaknesses section.

1 comment
Reviewer_2xF8OpenReview ↗

Response to the Authors

Thank you for the rebuttal. After reading the authors' rebuttal and other reviewers' comments, I have concerns about the novelty of this work. Therefore, I reduce my score to a "Weak rejection."

Comments on the paper

1 comment
Area_Chair_5govOpenReview ↗

Please join the discussion before it closes

Thank you to the authors for the rebuttal and to the reviewers for engaging in the discussion. As the discussion period draws to a close, I encourage everyone to keep the conversation going. I’d appreciate a brief final assessment from each reviewer. After reading the rebuttal and the discussion so far, what is the main reason you lean toward acceptance or rejection, and what evidence supports that view? Please note which concerns the rebuttal resolved and which remaining concerns, if any, are decisive for your recommendation.

If your assessment differs from another reviewer’s, please point to the specific claim or evidence you interpret differently. Authors are welcome to clarify any factual points that remain open. This will help us identify where we agree and understand the reasons for any remaining disagreement.

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

Related papers

Discussion (0)

Sign in to comment.