acceptodds
Under review as a conference paper at ICLR 2027

GRAPE: Rethinking the Velocity Field and Integration for Inversion-Free Flow Editing

Abstract

Recent inversion-free image editing methods based on pretrained flow models directly modify source images using differences between source- and target-conditional velocities. However, this process often introduces changes beyond the requested attributes. We revisit these editing dynamics in terms of how the editing velocity is constructed and integrated across noise levels. Since the source- and target-conditional velocities are evaluated at distinct noisy states, their difference mixes the response to the condition change with a denoising response to the state difference, which can affect content that should be preserved. Based on this observation, we construct a gradient-aligned field that evaluates both conditions at a shared noisy observation of the current edit. Under population-optimal velocities, this field is proportional to the target-to-source log-density-ratio gradient. We further find that standard noise-schedule integration concentrates editing progress in late, low-noise steps, revealing a mismatch between noise-schedule progression and clean-state progress. This motivates progress-adaptive integration, which measures and compensates for the mismatch between noise-schedule progression and clean-state progress. We introduce GRAPE (Gradient-Aligned, Progress-adaptive Editing), a training-free and inversion-free method that combines gradient-aligned field construction with progress-adaptive integration. Across three flow backbones on PIE-Bench, GRAPE consistently improves the trade-off between target alignment and source preservation and retains this advantage under substantially reduced inference budgets.

Primary area:
generative models
Keywords:
text-guided image editing, flow matching, rectified flow, inversion-free editing, training-free image editing, generative models

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_oKuYOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces GRAPE, a training-free and inversion-free editing method, a method that improves FlowEdit by combining gradient-aligned field that evaluates the source and target branches of the current edit state at a shared noisy observation, which is proportional to the target-to-source log-density-ratio gradient. The paper also proposes progress adaptive integration, which compensates the mismatch between noise-schedule progression and clean-state progress. The method is evaluated on 3 different backbones on the PIE-Bench dataset, and demonstrates that they achieve a better source-alignment and text-adherence tradeoff, under reduced inference budget.

Critical strengths
  • The paper is written clearly.
  • The qualitative results are quite impressive and emphasize the method's ability to edit only the intended areas.
  • Extensive ablations have been conducted.
Critical weaknesses
  • Comparison against non-inversion-free methods is missing, such as FireFlow [1] and UniEdit [2], as well as against inversion-free methods such as FlowCycle [3].
  • A user study is missing.
  • The qualitative comparisons seem to be compressed and have artifacts, for example the cat -> tiger example (3rd row) in Figure 6.
  • Evaluations on datasets with higher dimensions, such as provided by FlowEdit to show that the method and hyperparameters are robust to other dimensions as well would be beneficial.
  • As demonstrated in the FlowEdit paper, as NeditN_{edit} is closer to TT, the less fidelity to the source image. It would be beneficial to assess the method’s ability to preserve the source even with higher NeditN_{edit} value which is close to TT to achieve more difficult edits, such as style transfer, changes in color palette, etc.

Minor Issues

  • The caption of Figure 3 is not clear - please describe it in more detail.
  • In Eq. 8 the motivation is not clear why replacing the unknown xtgtx_{tgt} with did_i and not with the estimated x^tgt\hat{x}_{tgt}.
  • In the related work non-inversion methods is missing, e.g. FlowOpt [4] and SONIC [5].

Final Note

Despite these weaknesses, I find the paper overall good. I would be willing to raise my score if the authors address my concerns.

[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] Wang, Yanghao, et al. "FlowCycle: Target-Aware Image Editing via Cycle-Consistent Constraints." (ECCV’26)

[4] Ronai, Or, et al. "FlowOpt: Fast Optimization Through Whole Flow Processes for Training-Free Editing." (arXiv’25)

[5] Baek, Seungyeon, et al. "SONIC: spectral optimization of noise for inpainting with consistency." (ECCV'26)

Questions for authors

See weaknesses.

3 comments
AuthorsOpenReview ↗

Response to Reviewer oKuY

We appreciate the reviewer's constructive feedback. The suggestions helped us strengthen the empirical evidence through additional method comparisons, a human preference study, higher-resolution evaluation, and experiments with higher starting noise levels.

1. Additional methods (W1)

We evaluated FireFlow, UniEdit-Flow, and FlowCycle on all 700 PIE-Bench examples using settings supported by their official implementations. We include GRAPE at comparable inference budgets and report CLIP-T explicitly, since matching NFE does not necessarily match editing strength. SD is in units of 10−310^{-3}.

BackboneMethodNFECLIP-TSD
SD3FlowCycle13,39826.2012.78
SD3UniEdit-Flow2826.5221.51
SD3GRAPE2726.608.78
FLUXFireFlow1826.1226.18
FLUXGRAPE1826.5817.33
FLUXUniEdit-Flow2825.7710.61
FLUXGRAPE2826.1611.95

On SD3, GRAPE achieves higher CLIP-T and lower SD than UniEdit-Flow at similar NFE. FlowCycle's official configuration requires 13,398 evaluations, so this is not an NFE-matched comparison. On FLUX at 18 NFE, GRAPE achieves higher CLIP-T and lower SD than FireFlow. At 28 NFE, GRAPE has 0.39 higher CLIP-T and higher EditHF preservation and visual-quality scores than UniEdit-Flow, but 1.34 higher SD. This does not establish a preservation ranking at matched editing strength. Appendix E.3 and Tables 10–11 report the additional metrics and configurations.

2. Human evaluation (W2)

Twenty-one non-author participants each assessed 20 anonymized pairs, yielding 420 judgments. They viewed the source image, instruction, and two outputs, and selected the result that better performs the requested edit while preserving unrelated content, or "Not sure." Cases were selected with method identities hidden and restricted to those for which all compared methods received a VLM instruction-following score of at least 4.

Overall, 70.5% of responses preferred GRAPE, 17.9% preferred the baseline, and 11.7% were neutral. GRAPE preference rates were 72.4% against FlowEdit, 57.1% against FlowAlign, and 76.2% against both DRFS and DNAEdit. Section 4.2 describes the protocol, and Figure 7 reports the distributions including neutral responses.

3. Figure 6 image quality (W3)

We inspected the original outputs for the cat-to-tiger example but could not identify the specific artifact mentioned. To preserve image quality in the PDF, we reconstructed Figure 6 from the original outputs and inserted it as a high-resolution PNG. If a particular region remains problematic in the revised figure, we would appreciate clarification so that we can inspect it directly.

4. Higher-resolution evaluation (W4)

We evaluated the unmodified real-image benchmark released by FlowEdit, containing 77 images and 280 prompt pairs at 1024×10241024\times1024. We retained the SD3 editing settings used for PIE-Bench and selected target CFG to obtain similar CLIP-T. Since masks are unavailable, preservation metrics use the full image. SD and LPIPS are in units of 10−310^{-3}, and SSIM is multiplied by 100.

MethodCLIP-TSDPSNRLPIPSSSIM
FlowEdit29.128.4823.4477.185.43
FlowAlign29.198.6524.2571.486.46
DRFS29.1210.3122.62100.183.11
DNAEdit28.9624.2719.65220.275.41
GRAPE29.136.0224.8969.286.30

GRAPE achieves the lowest SD and LPIPS and the highest PSNR, while FlowAlign has slightly higher SSIM. DNAEdit has lower CLIP-T at the reported operating point. Section 4.2 and Figure 8 present the guidance-sweep comparisons, and Figure 24 provides qualitative examples.

5. Higher starting noise levels (W5)

On SD3, we fixed T=50T=50 and increased nmax⁡n_{\max} from 33 to 50. The table reports selected operating points, with CLIP-T shown to make differences in editing strength explicit. SD is in units of 10−310^{-3}.

nmax⁡n_{\max}FlowEdit CLIP-TFlowEdit SDGRAPE CLIP-TGRAPE SD
3326.4810.9926.576.54
4026.1913.0326.888.19
4526.8832.0227.039.88
5027.6485.6027.6619.12

GRAPE has lower SD with comparable or higher CLIP-T at every reported setting. At nmax⁡=T=50n_{\max}=T=50, CLIP-T is nearly identical, yet SD is 85.60 for FlowEdit and 19.12 for GRAPE. Appendix E.4 and Table 12 report this analysis.

6. Other revisions

We expanded Figure 3's caption to explain the progress measure and normalization. We clarified Eq. (8) by distinguishing the unknown target endpoint xtgtx_{\mathrm{tgt}} from its estimate x^tgt,i\widehat{x}_{\mathrm{tgt},i} and using the latter to construct the progress reference. We also added FlowOpt and SONIC to Related Work in Appendix A.

I appreciate the authors detailed response and clarifications, which resolved my concerns. I'm quite positive about this paper, therefore I'd raise my score to Weak Acceptance.

AuthorsOpenReview ↗

Thank you for the response

Thank you for following up. We are glad that our response addressed your concerns. Your suggestions on additional baselines and robustness tests helped us strengthen the evaluation and clarify where the method works well.

Reviewer_9cQaOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper reanalysis the construction and integration of the editing field of inversion-free flow-based image editing. It argues that since FlowEdit computes source- and target-conditioned velocities across different noisy states, the resulting velocity difference captures both the change in condition and in latent state, which ultimately causing unintended changes in editing-irrelevant regions. To address this issue, this work proposes using the source-target velocity difference based on a shared noisy observation derived from the current clean edited state, thereby constructing a gradient-aligned field. Furthermore, this paper introduces the progress-adaptive integration, which dynamically increases the step size based on the lag of the clean-state editing progress relative to the noise schedule. Experiments conducted on various models including SD3, FLUX.1-dev, and Qwen-Image demonstrate the effectiveness of the proposed method.

Critical strengths
  1. The methodological design is concise and clean, while the proposed research question is quite clear.
  2. The experiments and analysis are generally comprehensive, and the method is efficient and effective compared with existing inversion-free approaches.
Critical weaknesses
  1. Adaptive integration is one of the two core contributions of the paper. However, there is a lack of direct analysis regarding how it works. In particular, the term corresponding to the "adaptive" aspect of Δℓi\Delta \ell_i, or its size relative to ℓi\ell_i remains unclear throughout the editing process. Consequently, it is difficult to intuitively know how this mechanism leads to a more uniform accumulation of target-relative editing progress as illustrated in Figure 3.
  2. Compared to the GA field, GRAPE not only utilizes adaptive integration but also replaces the sis_i terms used in the GA field with pip_i. I think the rationale for this change requires a clearer explanation. Furthermore, the respective contributions of this state modification and the adaptive integration to the performance improvements of GRAPE need to be analyzed, otherwise it's hard to say how effective the proposed adaptive integration is.
Questions for authors
  1. When and how much does adaptive integration modify the editing step? What does the trend of these step modifications reveal about the underlying dynamics of the editing process? Could the authors show how the Δℓi\Delta\ell_i value evolves across the editing steps?
  2. Should the uiu_i in Eq. (9) actually be uiGAu_i^{GA}? This paper uses a lot of distinct symbols, so it is recommended that the authors review the entire manuscript to ensure clarity and consistency in notation.
  3. It is recommended to separately demonstrate how each of the proposed operations (removing source correction, modifying the steering evaluation state, and incorporating adaptive integration) contributes to the improvement of editing. If the authors are permitted to submit a revised version of the paper during the further respond stage, I think expanding the kinds of velocity field heatmaps in Figure 13 would make these points much more intuitive.
3 comments
AuthorsOpenReview ↗

Response to Reviewer 9cQa

Thank you for requesting a more direct analysis of PAI and the individual components. We apologize for the unclear description of the evaluation states. Both the GA-field baseline and GRAPE evaluate source and target velocities at the shared observation sis_i. GRAPE does not replace sis_i with FlowEdit's target-side state pip_i. PAI changes only the application length of the GA direction at each step. Subsequent editing states can consequently differ, but the shared-observation construction is unchanged. We clarified this relationship in Section 3.1.2.

1. PAI behavior (W1, Q1)

We recorded the requested multiplier riraw=1+Δℓi/ℓir_i^{\mathrm{raw}}=1+\Delta\ell_i/\ell_i and applied multiplier ri=min⁡(riraw,M)r_i=\min(r_i^{\mathrm{raw}},M). Thus, riraw−1r_i^{\mathrm{raw}}-1 directly measures the requested extension relative to the nominal step. To examine the full noise range while retaining 33 editing steps, we used T=nmax⁡=33T=n_{\max}=33 on all 700 SD3 PIE-Bench examples, starting at σ=1\sigma=1, with target CFG 7.5 and M=3M=3. Here D=zfinal−xsrcD=z_{\mathrm{final}}-x_{\mathrm{src}}, and "Opposite" denotes cos⁡(di,D)<0\cos(d_i,D)<0.

Noise rangeAt capRequested median [10%, 90%]Applied meanOppositeMedian cos⁡(uiGA,di)\cos(u_i^{\mathrm{GA}},d_i)
1.00–0.8573.1%5.1 [1.8, 10.4]2.730.05%0.87
0.85–0.7086.3%9.1 [2.0, 15.8]2.800%0.74
0.70–0.4072.6%8.0 [1.0, 22.4]2.550%0.47
0.40–067.3%8.3 [1.0, 27.2]2.420%0.16
All steps74.7%6.8 [1.0, 17.5]2.640.02%0.69

The median requested multiplier exceeds one in every range, indicating a remaining projected deficit after the nominal update. The cap bounds the actual extension. Across image-step pairs, 74.7% reach the cap, 13.8% receive a partial extension, and 11.5% receive none. The fraction without extension increases from 2.5% in the highest noise range to 25.6% in the lowest. PAI therefore extends the GA direction more frequently at higher noise levels in this setting, consistent with the earlier accumulation of editing progress in Figure 3. We report these statistics by noise range in Appendix E.6, Table 16.

Figure 10 further shows that PAI advances editing at the same target CFG. At comparable CLIP-T, GRAPE preserves source structure better than the GA-field baseline over most of the evaluated range.

2. Component contributions (W2, Q3)

Removing the source-field correction from FlowEdit leaves steering evaluated at pip_i. Moving that evaluation to sis_i gives the GA field, and adding PAI gives GRAPE. We compared these four stages on all 700 SD3 PIE-Bench examples using T=50,nmax⁡=33T=50,n_{\max}=33. Source CFG is fixed at 3.5, and target CFG is selected separately to obtain comparable CLIP-T, ranging from 26.48 to 26.57. SD and LPIPS are in units of 10−310^{-3}, and SSIM is multiplied by 100.

MethodCLIP-TSDPSNRLPIPSSSIMTarget CFG
FlowEdit26.4810.9926.9945.190.387.5
Steering only at pip_i26.538.5228.2243.890.4015
GA field at sis_i26.507.9728.2339.690.6815
GRAPE (GA + PAI)26.576.5429.7630.092.065.5

At comparable editing strength, the successive operations reduce SD by approximately 22%, 6%, and 18%. The other preservation metrics also improve sequentially. Section 4.4 reports this analysis, and the new qualitative ablation in Figure 9 shows the corresponding edits and accumulated field responses. PAI retains the GA field construction, while its changed integration produces different subsequent states and responses.

3. Notation (Q2)

The uiu_i in Eq. (9) denotes uiGAu_i^{\mathrm{GA}}, as the reviewer noted. We corrected this notation and aligned the GA direction's sign in Algorithm 1 with Eq. (3) and the implementation.

Reviewer_9cQaOpenReview ↗

Response to Authors's Rebuttal

Thanks for authors' careful response. I think my concerns have been fully solved, I'll keep my score towards acceptance.

AuthorsOpenReview ↗

Thank you for the response

Thank you for your thoughtful review and for confirming that your concerns have been resolved. Your suggestions on the component decomposition and the analysis of PAI have substantially improved the paper.

Reviewer_t4myOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper revisits field construction and numerical integration in inversion-free flow editing. Existing methods such as FlowEdit evaluate source- and target-conditioned velocities at different noisy states, so their difference mixes condition changes with state-induced variation. GRAPE instead evaluates both conditions at a shared noisy observation of the current edit. Under population-optimal velocities, this Gradient-Aligned (GA) field is proportional to the gradient of the target-to-source log-density ratio.

The paper further argues that clean-state editing progress can lag behind the noise schedule. Progress-Adaptive Integration (PAI) estimates a target displacement from the target-conditioned clean prediction and lengthens the GA update when the nominal step falls behind a schedule-aligned reference. Experiments on SD3, FLUX.1-dev, and Qwen-Image show favorable preservation-editability trade-offs and strong reduced-NFE performance.

Critical strengths
  1. The shared-state field removes a clear confounding factor in trajectory-derived editing directions and is simple to implement.

  2. The density-ratio interpretation provides a useful theoretical perspective on inversion-free editing.

  3. The paper reports full preservation-editability curves and approximately matched-CLIP operating points, rather than only isolated results.

  4. Consistent gains across three flow backbones provide meaningful evidence of generality.

  5. The reduced-budget results are compelling: on SD3, GRAPE with 30 NFEs outperforms the reported FlowEdit configuration with 132 NFEs.

  6. The field-localization experiment directly supports the claim that the GA field is more concentrated within intended edit regions.

Critical weaknesses
  1. PAI is not sufficiently distinguished from adaptive step-size or guidance scheduling.

PAI does not change the GA direction; it only increases the distance traveled along that direction, up to a factor M. It may therefore act as a sample-dependent learning-rate, guidance-strength, or time-reparameterization scheme. The paper should compare against tuned nonuniform schedules, early-step weighting, per-step CFG schedules, and baselines matched for cumulative step length or trajectory arc length. Otherwise, it is unclear whether the proposed progress reference is essential.

  1. The theoretical result does not directly describe the implemented guided velocities.

The gradient-alignment theorem assumes population-optimal conditional velocities, whereas the implementation uses approximate pretrained models and classifier-free guidance. A guided velocity does not generally correspond to the original conditional density in the stated objective. The paper should clarify the objective associated with the guided field and test alignment across CFG scales or without CFG.

  1. The progress reference may be unreliable at high noise.

PAI estimates the target endpoint from a single target-conditioned clean prediction. At high noise, this estimate may be unstable, causing the method to amplify an incorrect direction. The cap M limits but does not eliminate this risk. Statistics on step multipliers, cap activation, and failures across noise levels would be useful.

  1. The cap M requires backbone-specific tuning.

The paper uses M=3 for SD3 and Qwen-Image but M=6 for FLUX.1-dev. This raises questions about transfer to new backbones and inference budgets. The selection procedure and sensitivity across edit categories should be reported more clearly.

  1. The empirical narrative relies heavily on CLIP.

Progress lag is illustrated with a CLIP margin, and single-point comparisons are matched using CLIP-T. CLIP is unreliable for some spatial, counting, and fine-grained edits. Human edit-success evaluation or an independent VLM/region-based metric would make the conclusions more robust. EditHF is a learned evaluator, not a human study performed in this work.

  1. Practical efficiency needs direct measurement.

The NFE reduction is impressive, but end-to-end speed depends on batching, shared unconditional predictions, and architecture-specific overhead. Wall-clock latency, throughput, and peak memory on identical hardware should accompany the NFE comparison.

Questions for authors
  1. Labeling the key variables and notations, such as vsrc,vtgt,uFE,uGAv_{src}, v_{tgt}, u^{FE}, u^{GA} et al, in Figure 1 would make the methodology easier to understand.

  2. How does PAI compare with tuned nonuniform step schedules, early-step weighting, and per-step CFG schedules at matched cumulative update length? How often does PAI enlarge the step or hit the cap, and how were M=3 and M=6 selected?

  3. What density-ratio objective corresponds to the CFG-based GA field used in practice, and how does empirical alignment vary with guidance strength?

  4. Does GRAPE retain its advantage when editing strength is matched using human judgments or metrics independent of CLIP?

  5. Can the authors report end-to-end latency and memory for GRAPE and FlowEdit under the same hardware and inference budgets?

4 comments
AuthorsOpenReview ↗

Response to Reviewer t4my (1/2)

1. PAI behavior and schedule controls (W1, W3, W4, Q2)

Thank you for suggesting direct measurements and schedule controls. PAI preserves the GA direction and adjusts its application length using the current edit's projected deficit relative to a progress reference. Unlike a predefined schedule, it evaluates this deficit and the direction's ability to reduce it separately for each image and step.

We recorded step multipliers on all 700 PIE-Bench examples using SD3. To cover the full noise range while retaining 33 editing steps, we used T=nmax⁡=33T=n_{\max}=33, starting at σ=1\sigma=1, with target CFG 7.5 and M=3M=3. Let riraw=1+Δℓi/ℓir_i^{\mathrm{raw}}=1+\Delta\ell_i/\ell_i be the requested multiplier, ri=min⁡(riraw,M)r_i=\min(r_i^{\mathrm{raw}},M) the applied multiplier, and D=zfinal−xsrcD=z_{\mathrm{final}}-x_{\mathrm{src}} the final edit displacement. "Opposite" denotes cos⁡(di,D)<0\cos(d_i,D)<0.

Noise rangeAt capRequested median [10%, 90%]Applied meanOppositeMedian cos⁡(uiGA,di)\cos(u_i^{\mathrm{GA}},d_i)
1.00–0.8573.1%5.1 [1.8, 10.4]2.730.05%0.87
0.85–0.7086.3%9.1 [2.0, 15.8]2.800%0.74
0.70–0.4072.6%8.0 [1.0, 22.4]2.550%0.47
0.40–067.3%8.3 [1.0, 27.2]2.420%0.16
All steps74.7%6.8 [1.0, 17.5]2.640.02%0.69

Across image-step pairs, 11.5% receive no extension, 13.8% receive a partial extension, and 74.7% reach the cap. The fraction without extension increases from 2.5% in the highest noise range to 25.6% in the lowest. In the highest range, did_i opposes the final displacement in 0.05% of steps. This measures consistency with the final edit, rather than endpoint accuracy. PAI also makes no extension when ⟨uiGA,di⟩≤0\langle u_i^{\mathrm{GA}},d_i\rangle\leq0. These statistics are reported in Appendix E.6, Table 16.

We disabled PAI and evaluated three types of controls using the same GA field. For nonuniform step schedules, we redistributed noise levels toward high or low noise while retaining the endpoints and 33 updates, without additional step multipliers. For early-step weighting, we used two linearly decreasing profiles and one that extends only the first half of the steps, matching their cumulative application length ∑kwkℓk\sum_k w_k\ell_k to GRAPE's mean of 2.66 times the nominal length. For per-step CFG scheduling, we retained nominal step lengths and decreased target guidance linearly from 20 to 3. The detailed control configurations and full results are provided in Appendix E.6 and Table 17.

At matched CLIP-T, interpolated along GRAPE's guidance sweep, the nonuniform schedule with denser high-noise levels has 8% higher Structure Distance, and the three early-step weighting controls have 29%, 127%, and 269% higher values. The schedule with denser low-noise levels and the CFG control yield lower CLIP-T and worse preservation than GRAPE's weakest setting. The CFG control reaches only 0.70 times GRAPE's cumulative update magnitude, so this comparison is not magnitude-matched. None of these controls reproduces GRAPE's observed trade-off. We welcome specific alternative schedules or matching procedures to evaluate under the same protocol.

Figure 15 additionally reports MM sweeps for SD3, FLUX.1-dev, and Qwen-Image. These curves contextualize the defaults of M=3M=3, 66, and 33, respectively, by showing the changes in editing strength and preservation. We do not claim that these values are optimal at every editing strength.

AuthorsOpenReview ↗

Response to Reviewer t4my (2/2)

2. Objective of the guided field (W2, Q3)

Under population-optimal velocities, the guided GA field remains proportional to the gradient of a potential, with the coefficient in Proposition 1. For source and target guidance weights wsrcw_{\mathrm{src}} and wtgtw_{\mathrm{tgt}}, this potential is

(w_{\mathrm{tgt}}-w_{\mathrm{src}}) \left[\log p_{\sigma_i}(s_i\mid c_{\mathrm{tgt}}) -\log p_{\sigma_i}(s_i\mid\varnothing)\right].$$ It combines a weighted target-relative preference with a target-directed CFG term. The condition-independent state terms cancel at the shared observation for any guidance weights. Equal weights give $\widetilde F_i=wF_i$. With unequal weights, the objective changes to the generalized potential above. We clarify this distinction after Proposition 1 and derive it in Appendix B.1. Pretrained networks only approximate the ideal velocities. We additionally measured the cosine between guided and unguided GA fields along GRAPE trajectories on 100 SD3 images, fixing source guidance at 3.5. | Target guidance | 1 | 2 | 3.5 | 5.5 | 7.5 | 10 | 15 | |---|---|---|---|---|---|---|---| | Median cosine | 0.53 | 0.79 | 1.00 | 0.85 | 0.70 | 0.60 | 0.50 | With equal weights, the guided field is a positive scaling of the unguided field, including for the implemented networks. The measured cosine decreases as the weights diverge. At the default target weight of 5.5, the median is 0.85. This experiment measures directional agreement between the two implemented fields, rather than alignment with an inaccessible exact density gradient. **3. Evaluation beyond CLIP (W5, Q4)** We added a human preference study in Section 4.2 and Figure 7. We selected 20 PIE-Bench cases for which all compared methods received a VLM instruction-following score of at least 4, with method identities hidden during selection. This focuses the comparison on edit quality and preservation among results satisfying that criterion. Twenty-one non-author participants each assessed 20 anonymized pairs, yielding 420 judgments. They viewed the source image, instruction, and two outputs, and selected the result that better performs the requested edit while preserving unrelated content, or "Not sure." Approximately 70% preferred GRAPE, 18% preferred the baseline, and 12% were neutral. GRAPE was preferred in 72% of responses against FlowEdit, 57% against FlowAlign, and 76% against both DRFS and DNAEdit. Figure 7 includes neutral responses. **4. Practical efficiency (W6, Q5)** We measured both methods on one RTX A6000 with batch size one and $T=50,n_{\max}=33$ on SD3. We averaged 40 images after five warm-up images. Timing includes text encoding, VAE encoding, sampling, and VAE decoding, excluding model loading. GPU execution was synchronized, and transformer calls were counted directly. Both methods evaluate condition inputs separately. | Method | NFE | Seconds/image | Images/minute | Peak allocated GiB | |---|---|---|---|---| | FlowEdit | 132 | 4.50 | 13.3 | 16.97 | | GRAPE | 99 | 3.41 | 17.6 | 16.98 | Sharing the unconditional prediction reduces NFE from 132 to 99 and latency by 24%, with nearly unchanged peak memory. Throughput is computed from mean latency. On FLUX.1-dev, both methods use 48 NFEs and have nearly identical latency, 16.04 and 16.05 seconds. We report the timing results in Section 4.4 and the measurement procedure and memory results in Appendix D, including Table 4. **5. Figure notation (Q1)** We clarified the evaluation states in Figure 2 and expanded its caption to identify $q_i$, $p_i$, and the shared observation $s_i$.

Thanks for the detailed response from authors. My concerns have been solved and I'd like to maintain my initial score.

AuthorsOpenReview ↗

Thank you for the response

We appreciate your follow-up and are glad that our response resolved your concerns. Your suggestions helped us strengthen the analysis of PAI and clarify the connection between the theory and implementation.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions and Discussion

We appreciate the reviewers’ constructive feedback and the Area Chair’s coordination of the discussion. In the revised manuscript, we have added:

  • Direct analyses of PAI, schedule controls, and component contributions, together with clarification of the CFG-based theoretical interpretation.
  • A human preference study with 21 participants and 420 judgments.
  • Comparisons with FireFlow, UniEdit-Flow, and FlowCycle, evaluations at higher resolution and higher starting noise levels, and measured inference costs.

We also clarified notation and figure explanations. All three reviewers have now confirmed in their follow-up comments that their concerns have been resolved. We thank them for helping us strengthen the paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.