acceptodds
Under review as a conference paper at ICLR 2027

EditGRN: Inversion-Free Image Editing by Source–Target Generative Synchronization

Abstract

Text-guided image editing faces a persistent trade-off between preserving source content and realizing coherent target semantics. We explore this challenge through training- and inversion-free image editing based on discrete generative models. Existing inversion-free discrete editors primarily rely on post-hoc probability correction, creating a disconnect between local token-level adjustments and globally coherent semantic generation. To bridge this gap, we introduce EditGRN, a training- and inversion-free framework with trajectory-wide source-target generative synchronization in discrete space. EditGRN adopts a dual-branch design—the source branch constructs source references and the target branch generates the target image—enabling trajectory-wide synchronization between the two. Exploiting GRN's robustness to random binary corruption, the source branch decouples source-reference construction from any specific generation trajectory by directly corrupting source tokens at varying levels across the full trajectory. The target branch synchronizes with this same corruption schedule, fusing the corresponding source context with editable representations intra-model. By synchronizing source context with target generation along the entire trajectory, our approach faithfully preserves the background while producing high-quality edits that blend seamlessly with the surrounding scene through intra-model fusion. PIE-Bench show state-of-the-art performance on background preservation and text alignment across diverse editing tasks against continuous and discrete baselines. The code will be released.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Image Editing, Discrete Generative Models

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_tTt7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper tries to solve the text-based image editing task and proposes EditGRN that leverages the external GRN (Generative Refinement Networks) to refine the sampling process by mitigating the trajectory inconsistency. The experimental results with extensive ablatioons support its effectiveness for this task.

Critical strengths
  1. This paper shows significant improvement in both user-provided and attention-derived mask settings.
  2. Ablations across Synchronized Corruption Schedule, Protected Region Writeback, Synchronized Reference, Trajectory-Decoupled Source, and Initial Retention Ratio support the method design with more mechanistic analysis.
  3. External refinement incorporation is new in this area and aligns quite well with the recent hot agent tool style.
Critical weaknesses
  1. Lack of implementation details for baseline reproductions. For example, in Table 1, FlowEdit shows non-competitive performance compared with most methods. However, from the community's experience, FlowEdit is a strong baseline and usually shows at least competitive performance in most of the text-based image editing datasets. Moreover, FlowEdit does not provide results on PIE-Bench; thus, a reproduction implementation is more important to ensure the comparisons are fair.
  2. Lack of ablation on GRN models.
  3. The external, pretrained GRN incorporation looks like an engineering trick, while it serves as an inconsistency mitigation tool, making the method show limited generalization.
Questions for authors
  1. The GRN ablations and baselines reproduction problems are quite important for this submission; please add an explanation and discussion of these two parts.
  2. An external tool can be incorporated; what's the advantage of this image editing method compared with instruction-based VLM editing models like Qwen-Image?
Reviewer_rnJ6OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents EditGRN, a text-guided image editing framework built on a pretrained Generative Refinement Network (GRN), without additional training or explicit inversion. A source branch directly corrupts the encoded source image at each refinement stage and extracts source KV features. A target branch shares the corruption schedule, uses source KV from the protected region to guide generation within the edit region, and preserves protected tokens through source-token writeback. The paper evaluates the method on PIE-Bench using background-preservation and text-alignment metrics. Ablation studies examine source KV injection, shared corruption, protected-region writeback, and related design choices.

Critical strengths

1.A useful adaptation of GRN for image editing. While reusing pretrained features for editing is not new, adapting this approach to GRN’s binary refinement process is valuable. Direct source-reference construction and shared corruption enable editing without additional training or inversion.

2.Promising experimental results. The reported PIE-Bench results show strong background preservation and text alignment. The source-KV ablation also suggests that the editing design adds value beyond the pretrained backbone alone.

Critical weaknesses

1.Some ablations need clarification. The background-writeback ablation lacks implementation details, and the synchronization ablation does not separate the effects of shared masks and random values. The Flow comparison also needs the original inversion-based editor’s results on the same subset. These gaps limit the conclusions about which mechanisms drive the improvements.

2.The evaluation needs stronger support for some claims. Baseline sources and task coverage need clarification to establish fair comparisons. Evidence for natural integration is mainly qualitative, with limited discussion of edits requiring changes outside the mask, such as shadows or reflections.

3.Table 2 and Figure 7 support synchronization, but does the Unsynchronized Schedule ablation keep the same retention-ratio schedule for both branches? If only the masks and random values differ, why attribute the performance drop to mismatched corruption levels? Testing mask sharing and random-value sharing separately, with the retention-ratio schedule fixed, would clarify their contributions.

Questions for authors

1.Algorithm 1 (lines 229–234) initializes the protected region with source tokens and predicts only the edit region. What happens to the protected tokens in the writeback ablation at lines 451–455? If they are simply carried over unchanged, why does background preservation worsen in Table 2? Does this ablation also change how tokens outside the edit region are corrupted or predicted?

2.The paper states that edits “blend seamlessly with the surrounding scene” (lines 049–050), while “protected tokens remain anchored to the source” (line 177). How does the method handle edits that require changes to shadows or reflections outside the mask? Must the mask be expanded to cover these changes?

3.The paper states that “all metrics degrade sharply” after replacing inversion with direct source noise injection (lines 465–471). However, Table 2 and Figure 8 compare this Flow variant with EditGRN, not with the same Flow editor using inversion. Please provide that baseline on the same test subset, with other settings unchanged. Performing worse than EditGRN alone does not show that replacing inversion caused the performance drop.

4.Which methods were rerun for Table 1, and which results were taken from prior papers? KV-Edit states that “we exclude style transfer tasks from PIE-Bench” (Section 4.1), whereas this paper includes style editing (lines 327–330). Was KV-Edit reevaluated on the same samples and tasks as EditGRN? Please clarify whether the reported scores are directly comparable.

5.In Figure 2, the token patterns inside the purple box of z_e^(t_s) appear to differ from the corresponding region of z_s^(t_s). According to Equation (6), should these regions be identical at initialization?

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

The paper proposes EditGRN, which applies a pretrained GRN for training-free and inversion-free image editing. A dual-branch design is used, where the source branch constructs source references by directly corrupting source tokens at each refinement step, and the target branch uses the same retention maps and random bits to generate tokens in editable regions. Such a design is used to preserve source tokens outside the editable regions.

Critical strengths
  1. The paper proposes a training-free and inversion-free image editing method based on GRN.

  2. A dual-branch design synchronised under the same GRN refinement schedule is proposed to preserve protected regions.

Critical weaknesses
  1. Among the related works, KV-Edit is the most relevant to this paper. It already uses background KV reuse, foreground-only queries, and is inversion-free. The paper should provide a more detailed discussion and comparison beyond the application of those techniques to GRN.

  2. The paper claims a difference between directly corrupting GRN tokens and methods requiring a particular generative path in Section 4.2. However, diffusion models can also directly sample noisy states at arbitrary timesteps as defined by DDPM, and FlowEdit already constructs directly noised source states and uses shared noise. Thus, the discussion should be more about providing useful editing features than about the ability of direct construction.

Questions for authors

Please address the concerns mentioned in the critical weaknesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.