acceptodds
Withdrawn from ICLR 2027

DecFlowEdit: Self-Localized Flow-based Image Editing via Guidance Decoupling

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Flow-based image editing (FlowEdit) enables inversion-free semantic changes through the difference between source and target velocities. In this paper, we observe that FlowEdit’s default classifier-free guidance (CFG) configuration, with asymmetric source and target scales, causes substantial background leakage. Matching these guidance scales, for example by removing CFG, improves edit-relevant localization but severely degrades editability. To get the best of both worlds, we propose DecFlowEdit, which decouples the optimal guidance scales for localization and for editing in flow-based generative models. In particular, DecFlowEdit first extracts an edit-relevant prior by temporally aggregating velocity differences evaluated without CFG, and then uses this prior to reweight the original updates under default CFG. Our method remains training-free and inversion-free, requiring neither external spatial masks nor attention manipulation. Experiments on PIE-Bench across FLUX, SD3, and SD3.5 show that DecFlowEdit improves background preservation, reducing structure distance by approximately 61–73% and background LPIPS by 68–80% relative to FlowEdit at comparable editing fidelity.

Primary area:
generative models
Keywords:
Image editing, Flow Matching, Generative Models, Diffusion Models

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_hVdHOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper introduces DecFlowEdit, a training-free and inversion-free image editing method, that is based on the FlowEdit paper. The authors identified that there is a coupling between guidance strength, editing localization, and effective semantic changes. Therefore, they propose to use an estimation of the intended editing region using the velocity difference between the target and the source without the CFG, and leveraging the original CFG parameters as the editing itself, i.e. the estimated editing region is a mask multiplied elementwise by the original velocity difference of FlowEdit. The authors show their method’s effectiveness on the PIE-Bench leveraging 3 models (SD3, SD3.5 and FLUX), especially for source image preservation.

Critical strengths
  • The results obtained are nice.
  • The observation of the CFG is simple and nice.
  • The presentation is clear.
Critical weaknesses
  • The method uses additional NFEs than other methods. While the trend is to reduce the computations (such as FireFlow [1], UniEdit [2], and FlowAlign [3]), the authors suggest increasing the total computations to achieve better results. It would be interesting to see a matched-budget NFEs comparison between all methods.
  • A user study would be beneficial to assess the method.
  • The observation regarding the CFG is nice and simple. However, by reading the Abstract and the Introduction I expected that the method would also be simple and utilize as a mask only the velocities difference, while in practice the method utilizes some other ``engineering’’ parts and hyperparameters (such as T\mathcal{T} τ\tau and ss in Eqs. 2 and 4).
  • It would be interesting to see the sensitivity of the hyperparameters for other resolutions than 512×512512 \times 512 (PIE-Bench), e.g. FlowEdit benchmark (1024×10241024 \times 1024).
  • Many competing methods are missing in the quantitative comparison (Table 1) on FLUX model and in the qualitative comparison, such as UniEdit [2], FireFlow [1], etc.
  • There are only 4 qualitative comparisons (Figure 1). To assess the method ability and versatility to edit other types of editing, e.g. style modification.
  • Many other methods already utilize masks in the FlowEdit algorithm such as [4], [5], [6], reducing the novelty of the DecFlowEdit. A comparison to inversion-free methods such as [4-6] would be beneficial to assess the method.

The manuscript presents very solid and visually compelling results, demonstrating that the proposed modifications to FlowEdit lead to practical improvements. However, the conceptual and technical delta relative to FlowEdit remains primarily incremental. Because the contribution is essentially an effective refinement rather than a foundational new framework or a substantial technical leap, to my opinion, it may not meet the bar for the main conference track. I think that the strong empirical findings and practical insights would make this a good contribution to a workshop.

[1] Deng, Yingying, et al. "Fireflow: Fast inversion of rectified flow for image semantic editing." (ICML’25)

[2] Jiao, Guanlong, et al. "Uniedit-flow: Unleashing inversion and editing in the era of flow models." (ICLR’26)

[3] Kim, Jeongsol, et al. "Flowalign: Trajectory-regularized, inversion-free flow-based image editing." (ICLR’26)

[4] Cui, Haowang, et al. "SAM-Flow: Source-Anchored Masked Flow for Training-Free Image Editing." (arXiv’06.26)

[5] Chen, Jili, et al. "SMART-Edit: Spectral Mutual Attention Refinement and fusion for precise training-free image editing." Pattern Recognition (2026)

[6] Vu, Trong-Tai Dam, and Vinh-Tiep Nguyen. "MaskFlow: Attention-Guided Localized Editing for Rectified Flow Models." 2026 International Conference on Multimedia Analysis and Pattern Recognition (MAPR). IEEE, 2026.

Questions for authors

See weaknesses.

Reviewer_VxpaOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper focuses on the undesired effects that FlowEdit has on the background and editing-irrelevant regions. This paper argues that asynchronous classifier-free guidance (CFG) prevents the precise localization of editing-related regions. To resolve this, this paper proposes calculating the differences between velocities derived from various conditions without the use of CFG, and aggregating these differences to obtain a mask for the editing-related regions. By employing this mask as a weight for the editing velocity of FlowEdit, this paper introduces DecFlowEdit and demonstrates its effectiveness and mask's accuracy using various models including FLUX, SD3, and SD3.5 on PIE-Bench.

Critical strengths
  1. The method is simple and plug-and-play. DecFlowEdit yields obvious improvements in preservation metrics such as structure distance, LPIPS, and PSNR, without relying on external segmentation model or masks.
  2. The experimental setup is clean and the visual analysis is clear. The paper provides an visualization comparison of the relationship between CFG and editing effectiveness and localization, while also presenting direct visualizations of the masks and results achieved by the proposed method.
Critical weaknesses
  1. The method lacks sufficient technical novelty. The construction of the spatial prior in this paper, which involving the normalization, aggregation, smoothing, and thresholding of prediction differences, is essentially a form of post-processing applied to existing signals, rather than the introduction of a novel mechanism for generative models. Its two core components have closely related existing approaches: UniEdit-Flow [1] constructs a spatial mask by directly utilizing the difference between source-conditioned and target-conditioned velocities within the same latent, while DiffEdit [2] automatically localizes editing-related regions by aggregating the differences of denoising predictions derived from various text conditions. DecFlowEdit seems to simply emphasizes predicting without CFG in localization, making its technical contribution remains insufficiently clear. Furthermore, obtaining the spatial prior appears to require a significant number of additional NFEs, which is likely to impact the computational efficiency.
  2. There is a lack of analysis regarding the proposed arguments. The spatial prior proposed by DecFlowEdit could be applied to many existing image editing methods, it doesn't explicitly rely on the specific design of FlowEdit, nor is it inherently dependent on CFG analysis. Consequently, it remains unclear whether the observed performance improvements come from the proposed "CFG decoupling" insight, or if they simply indicate again that a well-defined editing-related mask enhances background preservation as many existing work. Furthermore, the paper’s analysis of the relationship between CFG scales and localization remains largely at the level of visualization, failing to provide an explanation for this phenomenon.

[1] UniEdit-Flow: Unleashing Inversion and Editing in the Era of Flow Models.
[2] DiffEdit: Diffusion-based Semantic Image Editing with Mask Generation.

Questions for authors
  1. Compared to existing work, what is the novel technical component in the spatial-prior construction presented in this paper?
  2. This paper should compare the efficiency of various methods and examine whether the performance gains resulting from the additional computational cost are showing a better trade-off or just marginal. Meanwhile, such a spatial prior may also improve other existing methods instead of FlowEdit, does it possess generalizability?
  3. Existing work has already provided some specific explanations for similar undesired impacts come from the velocity differences used in FlowEdit. For example, CVC [3] explicitly breaks down the effects of different source and target guidance scales, arguing that different CFG scales leave behind residual unconditional components that cannot be dropped out. FlowSlider [4] decomposes the FlowEdit velocity into a same-state cross-prompt steering term and a cross-state fidelity term to decouple the influence of unconditional components. Could the authors provide further analysis and explanation regarding the CFG-localization trade-off observed in this paper?

[3] On Exact Editing of Flow-Based Diffusion Models.
[4] FlowSlider: Training-Free Continuous Image Editing via Fidelity-Steering Decomposition.

Reviewer_7ZvUOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DecFlowEdit, which extracts spatial priors from the source-target velocity difference under CFG-free conditions and then uses these priors to constrain FlowEdit updates. The method requires no extra training, inversion, or external segmentation models. On FLUX, SD3, and SD3.5, the method achieves better improvements in background preservation and structural consistency.

Critical strengths
  1. The proposed method is training-free and inversion-free, which has practical value.

  2. The results look better in FLUX, SD3, and SD3.5, which show better generalization.

  3. The ablation studies are comprehensive. The authors compared different CFG configurations, various spatial priors, spatial weighting, and uniform scaling, which support the effectiveness of the method.

Critical weaknesses
  1. Limitation of technical novelty. The paper first calculates the velocity without CFG and weights the original FlowEdit updates. This is strongly related to existing works based on predicting velocity differences and mask guided Editing[1][2][3]. The paper only combine them, which likes a engineering attempt.

  2. The CLIP score declines in the quantitative results. Although I think the CLIP scores can not judge the performance of the editing task only, the paper does not show any user studies. Therefore, I think it cannot be fully confirmed that its overall editing quality has improved.

  3. The method has a limited scope. Appendix G notes that for object addition tasks, the spatial prior may fail to accurately locate where new objects should appear, and for global style transfer, the prior may also fail to cover regions requiring holistic modification. This indicates that the method is mainly suitable for local editing with clear spatial anchors, which limits the application scope of the method.

  4. The computational cost is unclear. The full method adds sampling steps approximately twice those of the original FlowEdit, and the short version remains around 1.25 times. The paper mainly reports the ratio of sampling steps and lacks end-to-end inference time comparisons for the full method.

References:

[1] UniEdit-Flow: Unleashing Inversion and Editing in the Era of Flow Models. ICLR 2026

[2] FlowEdit: Inversion-Free Text-Based Editing Using Pre-Trained Flow Models. ICCV 2025

[3] DiffEdit: Diffusion-Based Semantic Image Editing with Mask Guidance. ICLR 2023

Questions for authors

Most of my questions are listed in the weaknesses, and:

  1. I would like to see some user studies to prove the performance of the proposed method.

  2. I would like to see the end-to-end inference time to show the efficiency of the proposed method.

  3. Some baseline results in Tab 1 of the paper are taken from existing works. I think the authors should clarify the experimental details of the baseline. If the experimental setups differ, how can the authors ensure that performance differences are mainly due to the proposed methods?

  4. The experiments are mainly based on PIE-Bench. Could the authors provide results on other datasets?

  5. The method uses fixed thresholds and softness, while different backbone models adopt different aggregation time windows. How did the authors select these parameters? Were they determined on an independent validation set?

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

Related papers

Discussion (0)

Sign in to comment.