acceptodds
Under review as a conference paper at ICLR 2027

MatchEdit: Matched-State Prompt Directions for Inversion-Free Flow-Based Image Editing

Abstract

Inversion-free flow-based image editing typically constructs an editing direction by computing the difference between source and target velocity predictions. However, these predictions are evaluated at different latent states and under different prompts. Consequently, this difference captures not only the desired prompt change, but also the change in latent state. Through a controlled state-prompt decomposition, we show that this state effect strengthens target semantics, but also increases source deviation and unintended changes outside the edit region. To address this issue, we introduce MatchEdit, an inversion-free image editing method based on matched-state prompt comparisons. By comparing the source and target prompts at the same latent state, MatchEdit removes the additive state effect from the base editing direction and yields a more spatially selective edit. We then adaptively enhance target semantics using a strictly bounded correction guided by the spatial response of the edit. MatchEdit achieves state-of-the-art source preservation on PIE-Bench while maintaining competitive semantic alignment, and further generalizes to long-prompt editing on DNA-Bench and sequential multi-turn editing on MSE-Bench.

Primary area:
generative models
Keywords:
Rectified Flow Models, Image Editing, Inversion-Free Editing

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_HTh8OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes MatchEdit, a training-free and inversion-free image editing method built on pretrained flow models. Through a controlled state-prompt decomposition, this paper analyzes how state-prompt changes affect the conventional FlowEdit direction. MatchEdit combines prompt contrasts evaluated at matched states and introduces spatially adaptive, bounded target guidance to balance source preservation and editing strength. Experiments on PIE-Bench, DNA-Bench, and MSE-Bench show promising results.

Critical strengths
  1. The controlled state-prompt analysis provides useful insights into how the additive state effect influences semantic alignment, source preservation, and spatial selectivity of inversion-free image editing.

  2. The method shows consistently strong source preservation with competitive semantic alignment across PIE-Bench, DNA-Bench, and sequential multi-turn editing on MSE-Bench.

Critical weaknesses
  1. The technical contribution is not clearly distinguished from prior work. WSSC [1], CVC [2], and FlowSlider [3] have similar matched-state prompt comparisons or velocity decompositions. Experimental comparisons with WSSC and CVC are missing, while FlowSlider should at least be discussed. These comparisons and a more explicit related-work discussion are necessary to clarify MatchEdit’s novelty.

  2. The connection between the analysis in Section 4 and the design in Section 5 is not fully convincing. The analysis shows that the state effect contributes to the preservation-editability trade-off. However, it does not establish why this component should be removed entirely from the base direction rather than selectively retained through a spatial gate. The subsequent choice to recover editing strength using target CFG is also insufficiently motivated.

  3. The experiments are limited to SD3.5. It is unclear whether the state-prompt analysis and editing improvements also hold on other backbones, such as FLUX. This is important for establishing whether the findings generalize across flow models or depend on the tested backbone.

[1] Wavelet-Guided Semantic Signal Compensation for Inversion-Free Image Editing. ECCV 2026.

[2] On Exact Editing of Flow-Based Diffusion Models. arXiv 2025.

[3] FlowSlider: Training-Free Continuous Image Editing via Fidelity-Steering Decomposition. arXiv 2026.

Questions for authors
  1. MatchEdit requires additional velocity predictions, but Table 9 reports only a small runtime increase over FlowEdit (5.28 vs. 5.63 seconds). How are model evaluations counted, and are the sampling and timing settings consistent across methods?

  2. What is the purpose of the additional τi\tau_i ϵi\epsilon_i term in Eq. 19? This term is absent from the standard FlowEdit update. Is it also included in the baseline and ablation experiments?

6 comments
AuthorsOpenReview ↗

Response to Reviewer HTh8

Thank you for recognizing (i) the insightful state-prompt analysis and (ii) competitive performance across benchmarks. Our responses are provided below.

W1: Distinction with concurrent work

A1: Thank you for bringing these concurrent works to our attention. Upon careful examination, we find that WSSC (ECCV, 2026), CVC (arXiv, 2025), and FlowSlider (arXiv, 2026) follow different technical approaches from MatchEdit. In particular,

  • WSSC adds Haar wavelet-based low-frequency compensation to the original editing direction, whereas MatchEdit constructs the editing direction from matched-state prompt contrasts.
  • CVC focuses on posterior-based velocity correction to reduce trajectory errors, whereas MatchEdit explicitly analyzes state-prompt decomposition to construct an editing direction.
  • FlowSlider globally adjusts fidelity and steering components, whereas MatchEdit retains state-prompt interactions and applies spatially adaptive, norm-bounded semantic correction.

We will cite these concurrent works and discuss them in the revision.

We emphasis that different from these concurrent work, MatchEdit is theoretically grounded with our state-prompt decomposition.

We added WSSC's reported SD3.5 results to the PIE-Bench main table. MatchEdit achieves better overall performance.

MethodStructure (×10³) ↓PSNR ↑LPIPS (×10³) ↓MSE (×10⁴) ↓SSIM (×10²) ↑CLIP Whole ↑CLIP Edited ↑Avg. Rank ↓
WSSC17.4127.4641.0926.4691.0726.4323.892.29
MatchEdit7.4529.0246.1519.1490.3226.0322.712.00

W2: Excluding the State Effect and Choosing Target CFG over Spatially Gated State Effect

A2: Although the State Effect enhances target semantics, its global inclusion in the base direction introduces a prompt-independent shift. This can cause excessive changes in non-target regions. We therefore exclude the explicit State Effect while retaining state–prompt interactions.

We compared spatially gated State Effect with Target CFG on all 700 PIE-Bench images. Our Target CFG reduces Structure by 40.8% and reduces LPIPS by 20.5%, with minimal changes in CLIP Whole and CLIP Edited (26.058 → 26.029; 22.722 → 22.710).

These results show that spatially gated State Effect still introduces greater background changes than Target CFG.

W3: Generalize to FLUX

A3: We extended MatchEdit to FLUX.1-dev and evaluated it on all 700 PIE-Bench samples. Results are reported below.

MethodStructure (×10³) ↓PSNR ↑LPIPS (×10³) ↓MSE (×10⁴) ↓SSIM (×10²) ↑CLIP Whole ↑CLIP Edited ↑Avg. Rank ↓
RF-Inversion48.7619.51195.85155.7468.9525.1122.504.71
StableFlow19.2423.0476.9484.8587.2224.3021.283.00
RF-Edit27.7023.22131.1875.0081.4425.2222.403.57
DCEdit22.3625.4194.1748.0985.6025.4722.712.43
MatchEdit-FLUX16.7025.8983.4839.8186.0426.3622.771.29

Q1: Computational Cost

A1: Thanks for your helpful feedback. We apologize that, due to an oversight before the deadline, our original timing protocol was inconsistent. We re-run all methods using a consistent protocol. Runtime covers the full per-image pipeline of each method. We updated the manuscript and will release the code for reproducibility.

MethodTime/Image (s)
FlowEdit2.807 ± 0.016
DNAEdit2.613 ± 0.012
FSI-Edit10.603 ± 0.024
FIA-Edit3.639 ± 0.013
MatchEdit5.596 ± 0.014

Q2: Purpose of τiϵi\tau_i\epsilon_i in Eq. (19)

A2: We initially adopted the term from FIA-Edit in our implementation. It introduces a noise-dependent perturbation into the Euler update. It was consistently included in all ablations. We evaluated its effect on all 700 PIE-Bench images by removing only this term, with other settings unchanged. Interestingly, the results show minor changes: Structure Distance (0.007445→0.007041), MSE (0.001914→0.001821), and SSIM (0.903189→0.909547), while changing Edited CLIP by only 0.011. Thus, this term is nonessential to MatchEdit's performance, and can be removed.

Thank you again for your thoughtful and insightful comments. We understand the considerable time and effort involved in the reviewing process and sincerely appreciate your constructive suggestions.

Following your suggestions, we have:

  • clarified the differences in technical approaches from concurrent works
  • explained the exclusion of the state effect and evaluated its spatially gated alternative
  • extended MatchEdit to FLUX
  • re-evaluated the computational cost and corrected the corresponding results
  • examined the effect of the term in Eq. (19)

Given the limited rebuttal period, we sincerely hope to have further discussions if you have any remaining questions or concerns. We would greatly appreciate any further feedback you may have, as it would help us further improve our work.

Thank you again for your time and valuable feedback.

Thank you for the responses. My concerns about excluding the state effect and using target CFG, generalization to FLUX, and the additional noise term have been addressed.

The corrected timings and the fewer-step experiments provided to Reviewer EWp9 clarify the computational cost and show a reasonable efficiency–quality trade-off. However, the original timing error still affects my confidence in the experimental evaluation, particularly since runtime close to FlowEdit was identified as a strength by another reviewer zStQ.

My main concern remains the novelty and comparison with prior work. The key question is whether WSSC and CVC already use matched-state prompt comparisons or separate state and prompt changes, and what new conclusions or insights MatchEdit provides beyond them. Pointing to differences in subsequent compensation or correction modules does not sufficiently answer this question.

The WSSC comparison does not establish a clear overall advantage. WSSC achieves higher CLIP scores and better LPIPS and SSIM, while MatchEdit improves Structure, PSNR, and MSE. The better average rank depends on the baseline set and the resulting rank gaps, and does not resolve this preservation–alignment trade-off. Under the same criterion, CVC’s reported FLUX results would also be favored on the shared metrics, with better scores on all reported source-preservation metrics. The additional experiments addressing W2 and W3 are more convincing because they improve preservation while maintaining or improving semantic alignment.

Reviewer_EWp9OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper analyzes the limitation in FlowEdit that the conventional source-target velocity difference conflates changes in the latent state with changes in the prompt. By decomposing and recombining four types of state-prompt velocities, this work reformulates the conventional joint state-prompt direction as a combination of prompt effect, state effect, and interaction term. Then this work constructs a matched-state-prompt direction to eliminate the impact from different states. Furthermore, the paper validates the spatial selectivity of various editing directions and introduces the direction-aware bounded semantic correction, which enhances target semantics while preserving the source structure. The paper conducts comprehensive experiments on the PIE-Bench, DNA-Bench, and MSE-Bench, showing its effectiveness across various existing approaches.

Critical strengths
  1. The problem analysis are quite clear. The decomposition of conventional editing velocity into state and prompt components is highly valuable, and the explanation of the relationships between JSPD, state effect, and matched-state direction is quite insightful.
  2. The experiments are comprehensive, while the paper is clearly written and the validation results across multiple benchmarks are highly convincing.
Critical weaknesses
  1. The matched-state construction introduces additional velocity evaluations, it needs additional source-state-target-prompt and target-state-source-prompt velocities. For SD 3.5 which hasn't been CFG distilled, this would increase the single-step NFE from 4 to 6 I think. However, Table 9 in the appendix suggests that this extra computational overhead does not obviously affect the inference latency of MatchEdit, this point may need further explanation.
  2. The method involves a large number of manually set hyper-parameters, and the core parameter α\alpha is not a straightforward, easily interpretable trade-off knob that can be adjusted intuitively. Beyond α\alpha, the appendix introduces several additional hyper-parameters used within the mapping functions: k,b,βg,τg,pg,ρbg,ρeditk, b, \beta_g, \tau_g, p_g, \rho_{bg}, \rho_{edit}, and pρp_{\rho}. Given that inversion-free, flow-based editing already requires hyper-parameters such as the step number and two CFG scales, the inclusion of so many settings raises a question that might this complexity make it confusing for users practical application?
Questions for authors
  1. Please clarify the detailed computational cost of the editing process of MatchEdit.
  2. Could the authors provide further clarification on the role of α\alpha? As the results show, α\alpha does not appear to be a simple interpolation coefficient between preservation and editability. Although the appendix includes the ablation study on α\alpha which analyzes its impact on PIE-Bench metrics, the visual results suggest that it seems to be hard to tune α\alpha for a desired editing. From my perspective, Figure 6 in the appendix indicates that a smaller α\alpha is sufficient to achieve the desired edit, though it may alter the source background or structural. As α\alpha increases, the preservation of the background and structure initially improves, but subsequently degrades due to overexposure or excessive sharpening. When α\alpha approaches 1, the image structure closely resembles the source, but the content accuracy suffers. Given that α\alpha does not function as a direct control for the editing trade-off, the paper should provide a more detailed explanation of its effects.
  3. Apart from α\alpha, how are the other hyper-parameters determined? Are they fixed after being tuned on benchmarks? Are they used constantly for all three benchmarks? If so, the authors should discuss their stability across benchmarks; if not, the tuning protocol should be clearly showed. Also, does the performance improvement truly come from the design methodology or from these meticulously designed tricks? The reason why MatchEdit needs these complexly designed tools should be discussed.
8 comments
AuthorsOpenReview ↗

Response to Reviewer EWp9

Thank you for recognizing (i) the clear problem analysis and valuable state-prompt decomposition, (ii) insightful methodological interpretation, and (iii) comprehensive experiments with highly convincing results. We are also grateful for your constructive suggestions. Below, we address your concerns in detail.

W1 & Q1: Computational Cost and Inference Latency

A1: Thank you for identifying this issue. We apologize that, due to an oversight before the deadline, our original evaluation scripts used inconsistent timing boundaries. We have re-run all methods under a unified protocol with two warm-up runs and measurements over 10 images. Runtime covers the full per-image pipeline of each method. We report the mean ± standard deviation below.

MethodTime/Image (s)
FlowEdit2.807 ± 0.016
DNAEdit2.613 ± 0.012
FSI-Edit10.603 ± 0.024
FIA-Edit3.639 ± 0.013
MatchEdit5.596 ± 0.014
  • Computational Cost: MatchEdit takes approximately 2× the inference time of FlowEdit (5.596 vs. 2.807 s), with the same peak GPU memory.
  • Correction and Reproducibility: We have corrected the results in the revised manuscript and commit to releasing our code.
  • These results confirm that MatchEdit takes more runtime, but it stills acceptable for interactive editing.

W2 & Q2 & Q3: There are too many hyperparameter. Clarification on the role of α\alpha.

A2: Thank you for these valuable comments.

  • Role of α\alpha:α\alpha is not a direct preservation–editability control parameter. Instead, it balances the prompt contrasts at the source and target states. Thus, α\alpha controls the contribution of the state–prompt interaction. Since these contributions are accumulated across multiple editing steps, the improvement of visual effects are not necessarily monotonic. This explains the observations in Fig. 6. We will revise the corresponding explanation in the manuscript.
  • Hyperparameters Settings: Although our method involves several hyperparameters, we provide default values for them. These parameters are empirically chosen and the remain constant across all the results presented in the paper. We will release these hyperparameters configurations along with the code.
  • Source of Performance Gains: All results in the paper are obtained using exactly the same hyperparameter settings, without any sample-specific tuning. Therefore, we attribute the performance improvements primarily to the proposed methodological design.

Thank you again for your valuable comments and constructive suggestions. We understand the considerable time and effort involved in the reviewing process, and sincerely appreciate the effort you have devoted to our work.

Following your suggestions, we have:

  • re-evaluated the computational cost and corrected the corresponding results
  • clarified the role of α\alpha
  • explained our hyperparameter settings

Given the limited rebuttal period, we would sincerely welcome further discussion if you have any remaining questions or concerns. Your additional feedback would be valuable in helping us further improve our manuscript.

Thank you again for your time and thoughtful review.

Reviewer_EWp9OpenReview ↗

Response to Authors' Rebuttal

Thanks for authors' response.

I think the explanation of α\alpha is clear and correct, since it aligns with the figure illustration in the paper. A minor suggestion is that in the final version, authors can provide a more detailed explanation of the state-prompt interaction by combining the visualization of Figure 6, which can better avoid readers' misunderstanding.

The hyper-parameter setting is fair, which well addresses my concern.

My only remaining concern at present is the efficiency of the proposed method, since there are many recent approaches focus on not only the performance but also the efficiency, such as FireFlow [1], UniEdit-Flow [2], and ChordEdit [3]. I think it is rational to increase NFE to address the issues discovered in this paper, so if authors can further demonstrate the robustness and potential for acceleration of the method, I think the paper would be fully complete.

I apologize for my demand at this time of particular experiments that I think need to be supplemented. I think it will be pretty helpful if authors can demonstrate through the fewer-step editing experiments (e.g., modify the step setting from 50/33 to 20/13, or 15/10, or etc.) that MatchEdit actually makes image editing more accurate and effective.

[1] FireFlow: Fast Inversion of Rectified Flow for Image Semantic Editing, ICML'25
[2] UniEdit-Flow: Unleashing Inversion and Editing in the Era of Flow Models, ICLR'26
[3] ChordEdit: One-Step Low-Energy Transport for Image Editing, CVPR'26

Reviewer_zStQOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper observes that conventional inversion-free flow editing compares source- and target-conditioned velocities at different latent states, so the resulting direction mixes prompt-induced changes with state-induced variation. MatchEdit instead compares source and target prompts at matched latent states, removing the additive state effect. Since this improves preservation but weakens target semantics, the method adds a spatially localized, norm-bounded semantic correction derived from the editing direction. Experiments on PIE-Bench, DNA-Bench, and MSE-Bench show strong source preservation with competitive text alignment.

Critical strengths
  1. The idea/motivation about the state-prompt decomposition in Sec. 4.1 (i.e. Eqs. (6)-(11)) is clear, intuitive, and useful for understanding existing inversion-free editing methods.

  2. The ablations clearly show the roles of the matched-state direction and bounded semantic modulation. Figure 4 and algo. 1 in Appendix are useful for paper understanding.

  3. The preservation gains are consistent across three benchmarks, while runtime remains close to comparable SD3.5 methods.

Critical weaknesses
  1. This paper verifies the state-prompt decomposition based on SD3.5 model but lacks validation on FLUX model. It is important to add experiments on it.

  2. The algebraic decomposition does not prove that the state effect is undesirable. The paper itself shows that this component strengthens target semantics, and it may be necessary for object replacement, geometric changes, or global edits. Category-wise results are needed to clarify when removing it helps or hurts.

Questions for authors
  1. Figure 1 in main paper
  • It is better if a real image is taken as an example in Figure 1. Cartoon-style imagery may diminish a sense of persuasiveness or authority.
  • The proposed method builds upon the difference between two velocity fields in FlowEdit and extends it to the relationships among four velocity fields, explicitly disentangling the effects of the prompt and state. A conceptual illustration similar to Figure 2 in the FlowEdit paper would help clarify these relationships and better highlight the novelty of our method.
  1. In Eq. (2), please add how to get yiy_i? You can find it from FlowEdit paper.

  2. In Figure 2, the hyper-prameter α\alpha should be provided in sub-figure (b), which is more directly related to the proposed match direction.

  3. Table 10 in Appendix shows the best range of parameter α\alpha is [0.3, 0.5]. This table should be added in the main paper.

2 comments
AuthorsOpenReview ↗

Response to Reviewer zStQ

We sincerely thank you for the detailed and insightful feedback. We particularly appreciate your recognition of (i) the clear and intuitive methodological insights, (ii) well-designed ablations , and (iii) consistent preservation gains across three benchmarks. We address your specific concerns below.

W1: Generalize MatchEdit from SD3.5 to FLUX

A1: Thank you for this valuable suggestion. We have extended MatchEdit to FLUX.1-dev and evaluated it on all 700 PIE-Bench samples. The results are shown below.

MethodStructure (×10³) ↓PSNR ↑LPIPS (×10³) ↓MSE (×10⁴) ↓SSIM (×10²) ↑CLIP Whole ↑CLIP Edited ↑Avg. Rank ↓
RF-Inversion48.7619.51195.85155.7468.9525.1122.504.71 (5)
StableFlow19.2423.0476.9484.8587.2224.3021.283.00 (3)
RF-Edit27.7023.22131.1875.0081.4425.2222.403.57 (4)
DCEdit22.3625.4194.1748.0985.6025.4722.712.43 (2)
MatchEdit-FLUX16.7025.8983.4839.8186.0426.3622.771.29 (1)

MatchEdit-FLUX achieves the best overall average rank, confirming its effectiveness across different flow backbones.

W2: Category-wise analysis of the impact of removing the state effect

A2: Thank you for this insightful comment. We agree that our decomposition does not prove that the state effect is undesirable. But we emphasis that the main goal of our decomposition is to identify the role of the state effect.

Building on the state-effect analysis in Appendix A.1.2, we present category-wise results on all 700 PIE-Bench samples using Di(λs)=Pi+λsSiD_i(\lambda_s)=P_i+\lambda_s S_i. The table reports Δ=Metric⁡(λs=1)−Metric⁡(λs=0)\Delta=\operatorname{Metric}(\lambda_s=1)-\operatorname{Metric}(\lambda_s=0). Δ Structure indicates reduced source-structure preservation, while Δ CLIP Edited indicates stronger editing effectiveness.

Editing CategoryΔ Structure (×10³)Δ CLIP WholeΔ CLIP Edited
Random/mixed19.911.821.55
Object replacement25.302.242.21
Object addition19.082.062.20
Object removal19.391.300.62
Content/attribute18.171.310.86
Pose/geometric15.100.470.38
Color18.882.452.29
Material21.562.361.51
Background/global19.052.053.07
Style/global25.132.662.63

The semantic gains are more pronounced for object replacement, object addition, color, and background/style edits, whereas pose/geometric edits exhibit relatively smaller gains. These results confirm the category-dependent trade-off: removing the state effect favors source preservation while sacrificing some target-semantic strength.

Q1–Q4: Presentation Improvements

  • Q1: Replace the cartoon-style image in Fig. 1 with a real image.
  • Q2: Clarify Eq. (2) following FlowEdit.
  • Q3: Include the λs\lambda_s label in Fig. 2(b).
  • Q4: Move Table 10 from the Appendix to the main paper.

A: We sincerely appreciate your valuable suggestions. We have incorporated all the corresponding revisions into the revised manuscript.

Thank you again for your insightful and constructive feedback. We understand the considerable time and effort involved in the reviewing process and sincerely appreciate your contribution to improving our work.

Following your suggestions, we have:

  • extended MatchEdit to FLUX
  • conducted a category-wise analysis of the impact of removing the state effect
  • incorporated suggested presentation improvements

Given the limited rebuttal period, we sincerely welcome further discussion if you have any remaining concerns. We would greatly appreciate any additional feedback to help us further improve our work.

Thank you again for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.