acceptodds
Withdrawn from ICLR 2027

FlowFocus: Focused Trajectory Regularization for Inversion-Free Image Editing

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Inversion-free image editing offers an efficient alternative to inversion-based approaches, yet achieving substantial semantic changes while preserving source content unrelated to the desired edit remains challenging. Existing trajectory-regularized methods improve source consistency, but typically apply preservation constraints uniformly across the latent representation, overlooking that different spatial regions may require different degrees of modification. In this work, we propose FlowFocus, an inversion-free image editing framework with focused trajectory regularization. Our key insight is that the editing velocity can be decomposed into a counterfactual semantic response induced by changing the textual condition and a state-induced response arising from latent variation. Their relative strengths provide a model-intrinsic and time-varying signal that indicates where semantic modification is required. Based on this, FlowFocus constructs a continuous counterfactual edit field and incorporates it into a trajectory-level optimal-control objective, replacing global scalar regularization with spatially and temporally adaptive control. This formulation strengthens source preservation where semantic changes are unnecessary while retaining sufficient flexibility in regions relevant to the target edit, without relying on external masks, auxiliary perception models, or additional supervision. Experiments on the PIE benchmark demonstrate consistent improvements in source preservation and editing quality over state-of-the-art methods. The code will be released after publication.

Primary area:
generative models
Keywords:
Image editing, Inversion-free, Flow model

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_JWXhOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper introduces FlowFocus, an inversion-free image editing framework that uses focused trajectory regularization to balance editing and source preservation (unedited region). Dealing with the current limitation of uniform regularization in the current inversion-free paradigm, it decomposes the editing velocity into semantic and state-induced responses, and uses their relative strengths to construct a continuous edit field for spatially and temporally adaptive control. Experiments on SD3 and SD3.5 report an improved balance between source preservation and text-image alignment, with ablations examine the effects of spatial adaptation and updating the edit field over time. Overall, it provides a spatially and temporally adaptive extension of existing global trajectory regularization for inversion-free image editing.

Critical strengths
  1. Good Motivation with a corresponding modification: Identifies the problem of uniform regularization in inversion-free editing and proposes a targeted and intuitive modification. The proposed change follows naturally from the motivation.
  2. Broad baseline coverage and organized experiments: Compares against a broad range of baselines. Experiments are well-organized, with hard metrics and user studies included.
Critical weaknesses
  1. Unconvincing qualitative results and source preservation: Some results in the paper show noticeable artifacts and unintended changes. Specifically, "add glasses", "marble wall to flower wall" in Fig. 1, red apple and dark bird in Fig. 3 are not satisfactory. Although the method is training- and mask-free, I still expect more reliable preservation of unrelated content in these relatively straightforward editing tasks. Since improving source / background preservation is the central motivation, these visible failures directly weaken the main claim. Overall, I do not find the qualitative results convincing.
  2. Limited contribution with insufficient validation of the core design: The paper lacks targeted analysis of whether this particular design is necessary. Can this problem be addressed using a simpler signal such as the velocity-difference magnitude at each spatial location? It would be helpful to include latent visualizations showing where uniform regularization suppresses necessary edits or fails to preserve unrelated content, how the proposed response maps identify these regions over time, and whether simpler signals can or cannot achieve similar results. Without stronger mechanism-level evidence or higher editing quality, I am not convinced that this single adaptive modification constitutes a substantial standalone contribution to training-free image editing.
Questions for authors
  1. Would it be possible to provide more targeted ablations to justify the necessity of the proposed design? Visualizations of the proposed response maps and comparisons with simpler alternatives would help clarify the actual benefit of this design.
  2. Why is the proposed method only evaluated on SD3 and SD3.5? As the two most relevant works FlowAlign and FlowEdit both have results on FLUX, putting methods with different backbone models in a single table does not provide a fully reliable comparison. Would it be possible to evaluate FlowFocus on FLUX and compare it with these methods under the same backbone?
  3. The large number of baselines is somewhat distracting. Although the paper includes recent methods, it also includes many older baselines that are less informative for evaluating the proposed contribution. As this field is evolving rapidly, more targeted comparisons would be helpful, with clearer explanations of why each baseline is included rather than simply providing a large number of comparisons. ChordEdit is a good example.
1 comment
  1. The goal of FlowFocus is not to enforce pixel-level preservation of every source detail, but to achieve a better balance between faithfully realizing the requested semantic change and preserving content that is unrelated to the edit. This distinction is important for text-guided image editing, where the target prompt generally does not specify a unique correct output. In many cases, a reasonable edit also requires local contextual adaptation so that the modified content remains visually coherent with the surrounding scene. From this perspective, we believe the qualitative results demonstrate a favorable preservation--editability trade-off compared with existing methods. For example, in Fig.3, FlowFocus performs substantial target-driven changes while avoiding unnecessary modifications to source attributes such as the woman's appearance and accessories or the car color and license plate, which are altered by several competing methods. Similar behavior is observed across the diverse editing examples in Fig.1 and the additional qualitative comparisons in the Appendix. We agree that no qualitative example alone can fully characterize editing quality; therefore, these visual comparisons are complemented by the quantitative preservation/alignment metrics and the human evaluation, which separately assesses editing accuracy and source preservation.
  2. We thank the reviewer for this comment, but respectfully disagree that the proposed design constitutes only a minor modification or that its effectiveness is insufficiently validated. FlowFocus is motivated by a specific limitation of existing trajectory-regularized inversion-free editing: source-consistency regularization is applied uniformly across spatial locations, despite the fact that different regions may require very different degrees of semantic modification. Our contribution is therefore not simply to introduce an additional spatial weight, but to derive a model-intrinsic, time-varying control geometry from the conditional flow responses and incorporate it into the trajectory-level optimal-control formulation. We also believe the current empirical evidence already supports the effectiveness of this design from several complementary perspectives. The component ablation in Table2 shows a clear progression from the inversion-free baseline to global regularization, fixed focused regularization, and the full adaptive formulation, demonstrating that both spatial focusing and temporal adaptation contribute to the final performance. Fig.5 further analyzes the influence of the focused-regularization strength, while the qualitative ablation in the Appendix illustrates how different degrees of spatial adaptation affect the resulting edits. Beyond these targeted studies, the quantitative comparisons, qualitative results, category-wise analysis, and human evaluation consistently show that FlowFocus achieves a stronger balance between source preservation and target-driven editing. We nevertheless appreciate the reviewer’s request for more direct validation of the particular design choice. We believe such experiments will make clearer that the gains arise from the proposed model-derived adaptive control.
  3. We will add the experiments on FLUX backbone for a complete comparison.
  4. We thank the reviewer for this comment, but respectfully disagree that the current comparison is outdated. Our evaluation includes a broad set of representative and recent text-guided image editing methods, with particular emphasis on inversion-free rectified-flow approaches that operate under a setting comparable to FlowFocus. This includes methods such as FlowEdit, FlowAlign, FIA, DRFS, and other recent flow-based baselines (2025, 2026), and we further report results across multiple backbones to reduce the influence of model-specific advantages. Regarding the additional method suggested by the reviewer, it is designed for one-step image editing and operates under a substantially different inference setting from ours. Its model architecture, sampling procedure, and computational regime are therefore not directly aligned with the multi-step rectified-flow editing setting considered in this work. A direct numerical comparison would not provide a fully controlled or necessarily fair assessment of the proposed method.
Reviewer_TVnjOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper uses the local energy ratio between the response to a conditional change at the same state and the response to a state change under a fixed condition to construct a time-varying spatial editing field. This field modulates the strength of trajectory regularization, preserving semantic change in the edited region while strengthening source-image constraints elsewhere.

Critical strengths
  1. Relative response energy provides a spatial modulation mechanism. Converting the relative strength of the conditional response and the state response into a continuous editing field makes the source-preservation strength vary with position and time.
  2. The spatial editing field is directly connected to trajectory regularization. The method uses weights obtained from response analysis to modulate source-image constraints, requiring no additional masks.
Critical weaknesses
  1. The optimal-control derivation is not self-consistent: The positive running cost and positive terminal cost in Eq. (14) of the main text are sign-inconsistent with the control coefficient derived in Appendix A.1, undermining the foundation of the claimed contribution that a new rule is derived from optimal control.
  2. The velocity decomposition largely overlaps with CVC[1]. Eqs. (7)–(10) of CVC and Eq. (11) of this paper use the same three predictions and construct the corresponding state difference and condition difference, so the novelty of decomposition is limited.
  3. The energy ratio of editing necessity remains heuristic. A strong local response does not mean that a location needs editing. The paper does not sufficiently explain why this ratio can identify editing regions, or its principled advantage over generic spatial weighting.

[1] On Exact Editing of Flow-Based Diffusion Models.

Questions for authors

Please refer to Weaknesses.

1 comment

We sincerely appreciate your time and effort in evaluating our manuscript.

  1. We will correct the corresponding equations in both the main text and Appendix A.1. This correction does not change the intended role of the proposed focused control geometry: smaller rt,ir_{t,i} values yield stronger source-consistency correction, while larger rt,ir_{t,i} values keep the controlled trajectory closer to the original editing velocity. We will also remove the statement that δ=1\delta=1 recovers the coefficient used by FlowAlign. Instead, δ=1\delta=1 should be interpreted as removing the spatial adaptation, yielding a spatially uniform control geometry under the same positive-cost objective. We appreciate the reviewer for pointing out this mistake and will revise the derivation and the associated discussion accordingly.
  2. The decomposition is different from CVC. The conditional velocity correction (CVC) trys to leverage the two hyperparameters to modify the velocity path, while our method wants to use the relative stength to help the model focus on the most related regions based on our regularization.
  3. We agree that a strong local response alone does not imply that a location necessarily requires editing. Our intention is not to treat ρt,i\rho_{t,i} as a ground-truth editing mask, but as a soft, model-intrinsic measure of the \emph{relative semantic relevance} of local trajectory variations. The motivation comes from the exact decomposition of the inversion-free editing velocity. The semantic component compares source and target conditions at the same latent state, whereas the state-induced component compares different latent states under the same source condition. Thus, the two terms isolate conditional and state variation, respectively. Our ratio measures their relative energy rather than the absolute magnitude of either response. This distinguishes it from generic spatial weighting based solely on activation or velocity magnitude.
Reviewer_jhR2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper is based on the current inversion-free method, specifically FlowEdit, and proposes a focused trajectory regularization method to imporve the balance between faithfulness and consistency trade-off. By using counterfactual decomposition and builds a Focused control geometry as an optimization criterion. Based on this, it regularizes and refines the sampling process to achieve a better editing balance.

Critical strengths
  1. The motivation and proposed focused control geometry are reasonable and insightful.
  2. The experiments show significant editing performance improvement, e.g., with SD3, improves FlowEdit from 27.16 to 22.79. on the Structure Distance metric.
  3. Show generalization ability across two generative models like SD3 and SD3.5.
  4. The human preference validation and ablation study can support the mechanism design.
Critical weaknesses
  1. Figures 2(b) and (c) do not fully show the difference between conventional inversion-free editing and the proposed method. This is a potential presentation improvement. The contribution or method design should be highlighted in this Figure.
  2. Figure 2(b) is FlowEdit, and there should be more inversion-free methods, like SDEdit and FlowAlign, which are both inversion-free methods. This method is based on FlowEdit and should not be claimed as a general improvement for all inversion-free methods.
  3. Lack of Flux results.
Questions for authors
  1. Can this proposed method also show good performance on Flux models?
  2. It's better if you can also provide comparisons with VLM-based / instruction-based models like Qwen-Image or SeedDream. It's very important to. show that these text-based editing methods can be competitive with current LLM-based editing models.
1 comment

We sincerely appreciate the positive assessment and constructive comments .

  1. Regarding the presentation of Figure 2, we will modify it to better illustrate the core contributions of our proposed method.
  2. Thank you for the great suggestion on the comparison on FLUX backbone. We will complete the comparison by conducting experiments on FLUX.
  3. The LLM-based editing methods are quite different from what we are trying to focus on in this work. The inversion-free text-guided image editing methods are the most relevant ones to ours. We highlight the benefits of our model as 'without relying on external masks, auxiliary perception models, or additional supervision'.

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

Related papers

Discussion (0)

Sign in to comment.