acceptodds
Under review as a conference paper at ICLR 2027

SpurFlow: Spatially-Adaptive Residual Steering for Few-Step Image Editing

Abstract

Recent diffusion distillation methods enable high-quality image generation with only a few sampling steps, yet extending them to image editing remains challenging. We argue that this difficulty arises from a mismatch between conventional distillation objectives and the asymmetric nature of image editing. Unlike generation, most editing tasks preserve the majority of the reference image while modifying only sparse instruction-specified regions. Under full-field matching, preserved tokens dominate both spatially and in gradient contribution, diluting sparse yet critical edit signals. To address this, we propose SpurFlow, an editing-oriented distillation framework that separates reference-preserving content from edit-specific updates. SpurFlow learns token-wise reference coefficients to adaptively control reference preservation across spatial locations while steering residual learning toward regions requiring edits. These coefficients further guide an edit-aware adversarial objective that concentrates perceptual supervision on edited regions. Extensive experiments across multiple benchmarks and foundation models demonstrate strong few-step editing performance. With only two sampling steps, SpurFlow closely approaches the full-step FLUX.1-Kontext teacher, surpasses the Qwen-Image-Edit-2511 teacher on both benchmark overall scores, and approaches the performance of the officially distilled 4-step FLUX.2-Klein model.

Primary area:
generative models
Keywords:
Image Editing, Flow Matching Distillation

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_3ZieOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes SpurFlow, a few-step image editing distillation framework that addresses the imbalance between preserved and edited regions. It introduces learnable token-wise reference coefficients and residual velocity prediction to adaptively preserve source content, along with localized adversarial supervision to enhance editing quality.

Critical strengths
  1. The paper identifies the gradient imbalance between preserved and edited regions as an important limitation of conventional distillation methods for image editing.
  2. The spatially adaptive reference coefficients and localized adversarial supervision provide a practical approach to improving few-step image editing without explicit edit-mask annotations.
Critical weaknesses
  1. This work identifies gradient dominance by preserved regions as a key limitation of conventional distillation, but does not directly demonstrate how SpurFlow redistributes gradient contributions between preserved and edited regions. Quantitative gradient analysis after introducing the proposed reference coefficients would support the central motivation.
  2. Table 2 shows that learnable reference coefficients alone provide only modest improvements. The author should support effectiveness of the proposed velocity parameterization and the improvements stem from localized adversarial supervision rather than the core distillation mechanism.
Questions for authors
  1. Could the authors provide quantitative gradient analysis to demonstrate how SpurFlow redistributes gradient contributions between preserved and edited regions?
  2. Could the authors disentangle the contributions of the proposed velocity parameterization and localized adversarial supervision?
4 comments
AuthorsOpenReview ↗

Response to Reviewer 3Zie (1/2)

We sincerely thank Reviewer 3Zie for carefully reading our paper and providing constructive feedback. We address the reviewer's concerns and questions point by point below.

Gradient Redistribution (Answer to W1 and Q1)

We address this concern through the analytical derivation already provided in Appendix B and additional empirical gradient analysis.

(1) Analytical Gradient Derivation in Appendix B

We thank the reviewer for this suggestion and kindly invite the reviewer to revisit Appendix B, which provides an exact token-wise derivation of gradients with respect to the shared trainable parameters. Eqs. (18)–(22) decompose the gradient into reference-related and remaining-mismatch contributions, showing how the reference-related contribution is attenuated as the coefficients approach full reference reuse. Eqs. (23)–(24) further show that this contribution vanishes for ideally preserved tokens, while unresolved edits retain their residual target-error signal. This explains how our formulation can reduce preservation-related gradients and increase the relative influence of edited regions. We will highlight this analysis more clearly in the main text.

(2) Additional Empirical Gradient Analysis

To complement the derivation, we compare regional backbone gradients between SpurFlow and ArcFlow on FLUX.1-Kontext. SpurFlow uses a 20k-step warmup without GAN supervision, followed by the mask-guided GAN stage. We measure gradients at both checkpoints with respect to the shared backbone only, excluding the GAN loss and all discriminator parameters.

To ensure a consistent comparison, all three configurations use the same pool of 32k training pairs and the same token partition. We generate the edit masks once using SpurFlow and reuse them for ArcFlow and both SpurFlow checkpoints. Measurements cover 5000 gradient-measurement steps on 32 GPUs, with one pair per GPU per step. All regional losses use the same whole-image normalization.

Let ϕ\phi denote the shared backbone parameters. For each region (r = pres or edit), we compute the average per-pair gradient norm:

Gr=1N∑n=1N∥∇ϕLn,r∥2.G_r = \frac{1}{N}\sum_{n=1}^{N}\Vert\nabla_{\phi}\mathcal{L}_{n,r}\Vert_2.

Here, N is the number of pair-level gradient measurements, and the regional loss is computed for each measured training pair. Relative edit gradient magnitude (%) is defined as:

Redit=100×GeditGpres+Gedit.R_{\mathrm{edit}} = 100\times\frac{G_{\mathrm{edit}}}{G_{\mathrm{pres}}+G_{\mathrm{edit}}}.

The table below reports regional backbone gradient norms on the same training-pair pool. Masks are generated once by SpurFlow and reused across all rows. Reported gradients exclude the GAN loss and discriminator parameters. Each entry averages per-pair backbone gradient norms.

MethodGpresGeditRelative edit magnitude (%)
SpurFlow, Warmup2.033.1460.74
SpurFlow, GAN5.4211.4667.89
ArcFlow1.842.0352.45

SpurFlow's relative edit gradient magnitude is 60.74% after warmup and increases to 67.89% at the GAN-stage checkpoint, compared with 52.45% for ArcFlow. These results indicate greater relative gradient emphasis on the predicted edit regions under the shared evaluation protocol.

AuthorsOpenReview ↗

Response to Reviewer 3Zie (2/2)

Component Contributions (Answer to W2 and Q2)

We thank the reviewer for raising this question and kindly invite the reviewer to revisit Sec. 5.2 and Table 2 for the existing component ablations. For clarity, we summarize them below, include the previously omitted DMD2 [1] baseline, and provide additional ablations on FLUX.2-Klein.

Experimental setup

We evaluate six configurations to distinguish the contributions of the velocity parameterization and localized adversarial supervision. The configuration IDs below are shared across both backbones. Random-Crop GAN and α-Guided GAN both retain global adversarial supervision.

IDConfiguration
ADMD2 (full-field velocity)
BReference formulation (fixed coefficients)
CB + Learnable token-wise coefficients
DC + Global GAN
EC + Random-Crop GAN
OursC + α-Guided GAN (SpurFlow)

In the following tables, Action through Average are ImgEdit-Bench scores; Q-SC, Q-PQ, and Q-O are GEdit-Bench v2 scores. Higher scores are better; lower NFE is better. Score pairs in the subsequent analysis denote ImgEdit-Bench average / GEdit-Bench Q-O.

Component analysis on FLUX.1-Kontext

IDNFEActionAddAdjustBackgroundComposeExtractRemoveReplaceStyleAverageQ-SCQ-PQQ-O
A43.923.953.693.053.091.812.313.263.963.175.747.275.39
B23.953.723.853.552.832.303.254.024.013.496.017.655.72
C24.133.873.913.692.902.263.584.103.973.596.037.705.76
D24.043.963.873.663.382.413.444.183.983.636.017.805.78
E23.713.814.123.763.032.053.844.193.943.625.997.795.77
Ours24.274.084.194.033.492.523.554.174.023.816.108.035.93

Component analysis on FLUX.2-Klein

IDNFEActionAddAdjustBackgroundComposeExtractRemoveReplaceStyleAverageQ-SCQ-PQQ-O
A23.824.123.703.563.012.192.403.654.563.467.177.406.73
B24.364.394.374.253.532.314.374.404.774.047.427.987.06
C24.554.314.454.313.622.264.614.424.804.117.488.067.14
D24.524.284.334.303.652.284.084.454.724.047.428.137.23
E24.564.344.264.343.321.982.734.064.733.787.198.096.95
Ours24.574.374.644.523.652.234.524.564.854.217.608.087.31

(1) Velocity parameterization.

Rows B and C evaluate the reference formulation without adversarial supervision. On FLUX.1-Kontext, introducing the reference term with fixed coefficients improves scores from 3.17/5.39 to 3.49/5.72, while reducing NFE from 4 to 2. Learning the token-wise coefficients further improves them to 3.59/5.76. On FLUX.2-Klein, the corresponding progression is 3.46/6.73 → 4.04/7.06 → 4.11/7.14. These results demonstrate improvements from the reference-based parameterization before localized adversarial supervision is introduced.

(2) Localized adversarial supervision.

Configurations C, D, E, and Ours retain the same reference-based parameterization and compare no GAN, Global GAN, Random-Crop GAN, and our coefficient-guided GAN at 2 NFE. Adding our localized GAN improves scores from 3.59/5.76 to 3.81/5.93 on FLUX.1-Kontext and from 4.11/7.14 to 4.21/7.31 on FLUX.2-Klein. It also outperforms both Global GAN and Random-Crop GAN on these two metrics across both backbones, supporting the benefit of directing adversarial supervision toward edited regions.

Overall, the reference-based velocity parameterization benefits both backbones, with larger gains on FLUX.2-Klein. Localized adversarial supervision provides further gains on both: its ImgEdit improvement is more pronounced on FLUX.1-Kontext (+0.22), while remaining beneficial on FLUX.2-Klein (+0.10), and both backbones gain 0.17 in Q-O. These results support the complementary contributions of the two designs.

We have revised the paper accordingly. The additional gradient analysis is included in Appendix G, “Regional Gradient Analysis,” and the additional FLUX.2-Klein ablations are included in the appendix section “Component Ablations on FLUX.2-Klein.” Changes in the revised manuscript are highlighted in blue.

[1] Yin et al. Improved Distribution Matching Distillation for Fast Image Synthesis. NeurIPS, 2024.

I thank the authors for the detailed clarifications and additional experiments. The Appendix G gradient analysis and the new component ablations do make the mechanism and the individual contributions clearer. I will raise my score to 4.

Dear Reviewer,

We thank the reviewer for taking the time to revisit our analysis and additional experiments. We appreciate the constructive feedback throughout the discussion, which has helped us clarify our method and strengthen the paper.

Thank you again for your time and consideration.

Best regards, Authors

Reviewer_d6uiOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

SpurFlow targets few-step distillation for instruction-based editing. The observation is that full-field velocity matching is dominated by preserved tokens — the authors measure roughly an 84%/16% split of gradient contribution between preserved and edited regions — which dilutes the sparse edit signal. They parameterize the student velocity as a token-wise learnable reference coefficient lambda^(j) times the reference latent plus a residual predictor, and reuse the learned lambda to guide a localized adversarial objective that crops DINOv3 features around predicted edit regions. Two-NFE students are distilled from FLUX.1-Kontext-dev, Qwen-Image-Edit-2511, and FLUX.2-Klein-9B-Base, evaluated on ImgEdit-Bench and GEdit-Bench v2, with a localization eval (mask IoU vs. human annotations) on a 507-sample PIE-Bench subset.

Critical strengths
  • The motivation is well-grounded and, as far as I can tell, original. The 84%/16% gradient-split measurement concretely explains why editing distillation behaves differently from generation distillation, and it makes the design choices feel earned rather than ad hoc.
  • The gains hold across three heterogeneous backbones, and the 2-step Qwen student (4.48 / 7.94) beats released 4-step models: Qwen-Lightning (4.40 / 7.74) and TeleStyleV2-DMD (4.44 / 7.76) , and edges its own multi-step teacher. That is a real efficiency result, not just a marginal win.
  • The learned localization actually earns its keep. In Table 2, lambda-guided crop GAN reaches 3.81 ImgEdit / 5.93 Q-O, clearly above both Global GAN (3.63 / 5.78) and Random-Crop GAN (3.62 / 5.77). So the win is not just "more adversarial data on crops" : where you crop matters, and the learned lambda picks better regions. The data-only SFT control is a nice touch that separates data-composition effects from the objective.
Critical weaknesses
  1. The localization comparison in Table 3 is not apples-to-apples. SpurFlow's lambda is learned during distillation with paired (reference, target) supervision, whereas RegionE and RWTE localize at inference time with no paired data, and RWTE is the authors' own reproduction. The IoU lead (0.496 vs. 0.455 / 0.384) therefore mixes the method with a supervision advantage.

  2. The crop control is incomplete in one direction. Random-crop is a weak strawman; an informative-but-unlearned crop (attention-, gradient-, or SAM-guided) would isolate whether learning lambda specifically helps or whether any sensible localization would. Relatedly, the learnable coefficient on its own is a modest step (3.49 -> 3.59 ImgEdit), with most of the jump coming from the adversarial stage.

  3. The GPT-5.5 mask/quality scores carry part of the same localization claim but are not validated against human judgment, so it is unclear how much they reflect localization quality vs. judge bias. (Minor: the FLUX.2-Klein framing is a bit generous, on GEdit-Bench v2 Q-O the 2-step student is 7.31 vs. the official 4-step 7.58, which is a non-trivial gap.)

Questions for authors
  1. Can you give RegionE/RWTE the same paired-data advantage, or evaluate lambda in an inference-only setting, so the Table 3 IoU comparison is fair?

  2. How does lambda-guided cropping compare against an informative-but-unlearned crop (attention/gradient/SAM)? This is the control that would show learning the localization is what matters.

4 comments
AuthorsOpenReview ↗

Response to Reviewer d6ui (1/2)

We thank Reviewer d6ui for carefully evaluating our work and providing valuable feedback. We address the reviewer’s concerns and questions point by point below.

Fairness of Localization Evaluation (Answer to W1)

We thank the reviewer for carefully considering the fairness of this evaluation. We believe Table 3 provides a fair and meaningful practical comparison of localization performance and inference efficiency. We clarify three points below.

(1) RegionE and RWTE also benefit from the teacher’s learned editing priors.

In the main paper, we apply RegionE and RWTE to the pretrained teacher model, which has already acquired substantial editing knowledge through large-scale reference–target supervision. Their inference-time localization therefore draws on the editing priors learned during the teacher’s training.

(2) SpurFlow requires fewer inference steps for localization.

We acknowledge that our α head is learned during distillation. SpurFlow produces localization within a two-step pipeline, whereas RegionE and RWTE extract masks at the tenth denoising step of a standard 28-step trajectory. Their localization thus requires more denoising steps.

(3) Our localization is learned without explicit mask supervision.

The α head receives neither ground-truth edit masks nor a direct mask-supervision loss; its localization behavior emerges through the distillation objective. We nevertheless acknowledge that paired image supervision can indirectly influence this behavior, so the IoU differences should not be attributed solely to the localization mechanism.

Localization Comparison with Matched Data Exposure (Answer to Q1)

(1) Matching paired-data exposure.

To further address the reviewer’s concern, we additionally evaluate RegionE and RWTE on the teacher after SFT on the same paired PICO data used for SpurFlow’s distillation. All methods therefore benefit from the same paired PICO dataset, ensuring a fair comparison in terms of access to these paired data. As shown below, SpurFlow retains higher overall IoU than both baselines under this control.

MethodRandomReplace Obj.Delete Obj.Attr. ContentAttr. PoseAttr. ColorAttr. MaterialBackgroundOverall
RegionE (SFT)0.47180.39050.38450.29720.34660.60400.39100.56670.4406
RWTE (SFT)0.35670.37770.28370.32170.31810.40600.46240.58380.3885
SpurFlow (ours)0.53510.45480.39400.37640.31820.63060.49880.66950.4847

(2) Applicability of inference-only localization.

Our coefficients λ are learned during distillation without mask annotations. They therefore require training and cannot directly operate as a training-free, inference-only localization method. Removing this learning stage would leave the coefficient head untrained. We consequently address the fairness concern through the paired-data control above, while retaining our learned localization formulation.

AuthorsOpenReview ↗

Response to Reviewer d6ui (2/2)

Alternative Guidance for Localized GAN (Answer to W2 and Q2)

We thank the reviewer for this helpful suggestion. We have therefore conducted additional experiments with informative alternative crop-guidance signals to assess the benefit of our learned localization.

(1) Attention-, gradient-, and image-difference-guided crops.

We evaluate RWTE [1] as an attention-based alternative, Grad-ECLIP [2] as a gradient-based alternative, and RegionE [3] as an image-difference-based alternative. For Grad-ECLIP, we adapt its image–text relevance maps to select local crops. These alternatives compute localization without training an additional crop-selection module. All guided variants retain our learnable reference formulation and global adversarial supervision, replacing only the local crop-guidance signal under otherwise identical training settings.

The table below compares these alternatives on FLUX.1-Kontext.

ConfigurationNFEActionAddAdjustBackgroundComposeExtractRemoveReplaceStyleAverageQ-SCQ-PQQ-O
Global GAN24.043.963.873.663.382.413.444.183.983.636.017.805.78
Random-Crop GAN23.713.814.123.763.032.053.844.193.943.625.997.795.77
RWTE-Guided GAN24.083.904.053.753.302.503.354.204.023.666.157.825.90
RegionE-Guided GAN24.083.943.943.823.172.173.124.154.053.586.137.825.88
Grad-ECLIP-Guided GAN23.863.884.123.773.202.763.454.114.023.706.087.815.84
α-Guided GAN (ours)24.274.084.194.033.492.523.554.174.023.816.108.035.93

Our coefficient-guided GAN achieves the highest ImgEdit-Bench average (3.81), GEdit-Bench Q-PQ (8.03), and Q-O (5.93), while RWTE achieves higher Q-SC (6.15 vs. 6.10). These gains are consistent with α learning the teacher’s implicit edit regions during distillation, aligning crop selection and localized adversarial supervision with the teacher’s editing behavior.

(2) Applicability of SAM-guided crops.

Original SAM [4] can generate candidate masks, but selecting those relevant to an editing instruction requires an additional prompting and mask-selection strategy. Our data provide no oracle edit-region prompts, and object masks need not match the full region affected by an edit. Our coefficients provide edit-aware guidance directly during distillation, without an external segmentation model or mask annotations.

Human Evaluation on Prolific (Answer to W3)

We thank the reviewer for raising this concern. To complement the automated evaluation, we conducted a blinded mask-preference study on Prolific, comparing SpurFlow with RegionE [3] and RWTE [1] on 50 randomly selected PIE-Bench examples spanning nine editing categories. Participants viewed the source image, editing instruction, and two mask overlays, and assessed coverage of required edits and exclusion of unrelated regions. Method names were hidden and presentation order randomized. Fifteen participants provided 300 judgments, with three independent judgments per comparison.

SpurFlow achieved human preference win rates of 90.54% against RegionE and 97.33% against RWTE on this subset. Win rates are computed as Win / (Win + Tie + Loss), excluding “cannot judge” responses. Counts below are from SpurFlow’s perspective; ties include “both good” and “both poor,” while Unsure denotes “cannot judge.”

ComparisonWinTieLossUnsure
vs. RegionE134952
vs. RWTE146400

These results provide complementary human evidence of localization quality on the evaluated subset.

FLUX.2-Klein Comparison (Answer to W3)

We thank the reviewer for pointing this out. We agree that our wording was overly optimistic: although our student uses only two steps, its GEdit-Bench v2 Q-O score remains lower than that of the official four-step model (7.31 vs. 7.58). We have revised the relevant text to present this result as a quality–efficiency trade-off, with the changes marked in blue in the revised PDF.

[1] He et al. Rethinking Where to Edit: Task-Aware Localization for Instruction-Based Image Editing. ACM Multimedia, 2026.

[2] Zhao et al. Grad-ECLIP: Gradient-Based Visual and Textual Explanations for CLIP. IEEE TPAMI, 2026.

[3] Chen et al. RegionE: Adaptive Region-Aware Generation for Efficient Image Editing. ICLR, 2026.

[4] Kirillov et al. Segment Anything. ICCV, 2023.

Thanks for the detailed response and additional experiments. The matched-data comparison and alternative crop-guidance results address my main concerns and strengthen the paper. The human study is also helpful, though the agreement between GPT-based scores and human judgments remains unclear. Overall, I appreciate the clarifications and will keep my score at 3.

Dear Reviewer,

We thank the reviewer for revisiting our work and for the constructive feedback throughout the discussion.

We acknowledge that the human-study protocol was not sufficiently clear. We have expanded Appendix H with the full questionnaire and a tabulated summary of the study design. We also kindly invite the reviewer to consult Appendix E.4 for the complete GPT evaluation prompt and procedure. These additions clarify both evaluation protocols and provide the details needed to reproduce them.

Thank you again for your time and consideration.

Best regards,
The Authors

Reviewer_3D1kOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SpurFlow, a few-step distillation framework for instruction-based image editing. The authors point out that image editing is different from image generation, since most regions should be preserved while only a small part needs modification. Conventional full-field distillation may cause unchanged regions to dominate training gradients and weaken editing signals. To address this, SpurFlow introduces reference-aware residual prediction and learnable token-wise coefficients to control reference preservation. These coefficients also help locate editing regions without mask supervision and guide a localized adversarial loss. Experiments on three foundation models show competitive editing performance with only two sampling steps.

Critical strengths

Clear motivation and writing is clear

The reference-aware residual prediction and token-wise coefficients allow adaptive reference preservation and unsupervised edit localization. The localized adversarial loss further improves editing quality in relevant regions.

The authors evaluate SpurFlow on three foundation models and multiple benchmarks. The method outperforms several existing two-step approaches and achieves comparable or better results than multi-step teachers and four-step models in some settings.

Critical weaknesses

The method is mainly based on existing flow matching, residual prediction, and adversarial distillation techniques. The proposed components are reasonable but relatively straightforward. Also, Table 2 shows that learnable reference coefficients alone bring limited improvements, while larger gains come from localized adversarial training. The contribution of each design could be better justified.

The learned edit localization achieves an IoU of only 0.496, showing a noticeable gap from human annotations. Since the predicted regions guide adversarial supervision, localization errors may affect editing quality. However, this relationship is not sufficiently analyzed.

Although different editing categories are evaluated, more challenging cases involving multiple regions, complex semantic relationships, or large structural changes are not sufficiently studied. It remains unclear whether the method can maintain both local editing accuracy and global image consistency in these cases.

Questions for authors

na

5 comments
AuthorsOpenReview ↗

Response to Reviewer 3D1k (1/4)

We appreciate the reviewer 3D1k’s thoughtful assessment and constructive suggestions. Below, we respond to each concern and question in turn.

Component Contributions (Answer to W1)

We thank the reviewer for highlighting this point. We acknowledge that Table 2 did not clearly distinguish individual contributions, but respectfully disagree that the gains from our reference-based design are limited: they should be assessed relative to DMD2 [1], inadvertently omitted from the original ablation table. Our design consists of three components:

(1) Reference formulation: We introduce a reference term into the velocity formulation using fixed coefficients.

(2) Learnable token-wise coefficients: We replace fixed coefficients with learnable token-wise coefficients α to adaptively weight the reference term.

(3) α-Guided GAN: The learned coefficients α yield a soft edit mask for localized adversarial training.

We summarize FLUX.1-Kontext ablations and add experiments on FLUX.2-Klein. Both backbones benefit, although gains vary. Qwen ablations could not be completed within the rebuttal window due to its larger model size and training cost; we will include them in a subsequent paper version.

The following table defines the configuration IDs used across both backbones:

IDConfiguration
ADMD2 (full-field velocity)
BReference formulation (fixed coefficients)
CB + Learnable token-wise coefficients
DC + Global GAN
EC + Random-Crop GAN
OursC + α-Guided GAN (SpurFlow)

Component analysis on FLUX.1-Kontext

Component analysis of SpurFlow with FLUX.1-Kontext.

IDNFEActionAddAdjustBackgroundComposeExtractRemoveReplaceStyleAverageQ-SCQ-PQQ-O
A43.923.953.693.053.091.812.313.263.963.175.747.275.39
B23.953.723.853.552.832.303.254.024.013.496.017.655.72
C24.133.873.913.692.902.263.584.103.973.596.037.705.76
D24.043.963.873.663.382.413.444.183.983.636.017.805.78
E23.713.814.123.763.032.053.844.193.943.625.997.795.77
Ours24.274.084.194.033.492.523.554.174.023.816.108.035.93

Component analysis on FLUX.2-Klein

Component ablations of SpurFlow on FLUX.2-Klein.

IDNFEActionAddAdjustBackgroundComposeExtractRemoveReplaceStyleAverageQ-SCQ-PQQ-O
A23.824.123.703.563.012.192.403.654.563.467.177.406.73
B24.364.394.374.253.532.314.374.404.774.047.427.987.06
C24.554.314.454.313.622.264.614.424.804.117.488.067.14
D24.524.284.334.303.652.284.084.454.724.047.428.137.23
E24.564.344.264.343.321.982.734.064.733.787.198.096.95
Ours24.574.374.644.523.652.234.524.564.854.217.608.087.31

The three components (1) (2) (3) provide complementary improvements on FLUX.1-Kontext.

For FLUX.1-Kontext, DMD2 predicts the full velocity field and achieves 3.17 on ImgEdit-Bench (row A). Adding the reference formulation with fixed coefficients raises the score to 3.49 (+0.32; A → B), while reducing NFE from 4 to 2. Introducing learnable token-wise coefficients further improves the score to 3.59 (+0.10; B → C). Finally, adding the α-Guided GAN increases the score to 3.81 (+0.22; C → Ours). The reference formulation and learned coefficients improve performance before localized GAN training, which provides a further gain.

Contributions (1) and (2) are particularly effective on FLUX.2-Klein, while (3) provides further gains.

The reference formulation and learnable coefficients account for most gains, while α-Guided GAN further improves image editing. Together, components (1) and (2) improve the ImgEdit-Bench average from 3.46 to 4.11 and GEdit-Bench Q-O from 6.73 to 7.14 over the DMD2 baseline. Adding α-Guided GAN further increases these scores to 4.21 and 7.31, respectively, outperforming both the global GAN and random-crop GAN on these two metrics. These results support complementary gains from effective reference utilization and localized adversarial supervision.

We have revised the PDF by updating the ablation study in the main text and adding the FLUX.2-Klein ablations to the appendix; all changes are highlighted in blue.

AuthorsOpenReview ↗

Response to Reviewer 3D1k (2/4)

Technical Contributions (Answer to W1)

We thank the reviewer for raising this concern. We would like to share the thinking behind our design and would welcome Reviewer 3D1k's perspective. We remain fully open to revising our view in light of further discussion.

We have approached this work with the understanding that meaningful research contributions can take many forms, including new theoretical derivations, model designs, and problem formulations. What we particularly value is identifying a meaningful research challenge and addressing it with a straightforward, principled, and effective solution.

In our work, we identify a challenge in few-step image editing distillation: full-field velocity distillation is hindered by gradient imbalance between edited and preserved regions. Through experiments and analysis, we found that residual prediction and an edit-focused mask for adversarial distillation effectively address this challenge. We therefore adopt these techniques to develop reference-aware velocity prediction with learnable coefficients α and use the resulting soft edit mask to guide localized adversarial distillation.

Our results support the effectiveness of this approach, with substantial improvements over strong baselines, including industry-developed methods. We acknowledge the reviewer's assessment that the components are reasonable but relatively straightforward. However, given their effectiveness in addressing the identified challenge, we believe it is unnecessary to add complexity solely for technical sophistication. We believe DMD2 reflects a similar rationale: to address the student–teacher performance gap in DMD [2], it incorporates an established GAN loss as a core contribution and achieves clear improvements. In our view, this choice is justified by its effectiveness in addressing the bottleneck. More generally, contributions can lie in identifying meaningful limitations and developing effective solutions, including combinations of existing techniques.

While our approach builds on established techniques, it introduces distinct methodological contributions in how they are formulated and connected. To our knowledge, ours is the first framework to derive edit localization from learnable reference coefficients during velocity distillation and use it to guide localized adversarial training. Specifically, our contributions include two connected designs: (1) reference-aware velocity prediction, which introduces a reference term to reuse preserved content; and (2) learning an edit mask from the coefficients α to guide adversarial distillation. This mask captures edit regions implicit in the teacher's editing behavior, is learned without mask annotations, and focuses adversarial supervision on regions requiring modification. Our ablations support both designs and show that the learned mask provides greater gains than a standard global GAN. We therefore believe our approach follows the above principle while offering distinct technical contributions supported by empirical evidence.

We welcome further discussion if the reviewer has additional suggestions or a different perspective.

AuthorsOpenReview ↗

Response to Reviewer 3D1k (3/4)

Interpreting the Localization IoU (Answer to W2)

We thank the reviewer and address this concerns from two points. First, our IoU of 0.496 is competitive for instruction-driven localization. Second, prompts may admit multiple valid regions, motivating complementary GPT-based evaluation in the paper and a new human preference study.

(1) Competitive Localization Performance.

An IoU near 0.5 does not necessarily indicate incorrect region selection, as IoU is sensitive to spatial extent. For illustration, expanding a rectangular ground-truth region outward by 20% of its width on each horizontal side and 20% of its height on each vertical side yields an IoU of 1/(1.4²) ≈ 0.51, despite fully covering the target region. Instruction-driven localization also remains challenging: CAMILA [3] reports an IoU of 0.3819 for applicable instructions on its Context-Aware Image Editing dataset despite explicit mask supervision. Although not directly comparable across datasets, this result provides context for the difficulty of the task.

We additionally evaluate LISA-13B [4] and RePlan (Qwen2.5-VL-7B) [5], both with task-specific training, on the same 529 PIE-Bench samples without retraining. LISA predicts masks with explicit segmentation supervision, while RePlan predicts editing boxes, filled and merged into masks without SAM refinement. Higher IoU is better; bold and underline indicate the best and second-best results in each column, respectively.

MethodRandomReplace Obj.Delete Obj.Attr. ContentAttr. PoseAttr. ColorAttr. MaterialBackgroundOverall
LISA0.51290.54990̲.̲4̲2̲3̲8̲0.43950.41130̲.̲6̲0̲2̲0̲0.52020.55630.5020
RePlan0̲.̲5̲2̲5̲7̲0̲.̲4̲9̲7̲6̲0.43320̲.̲4̲2̲6̲6̲0.33760.52780.46350.62180.4792
RegionE0.47000.38300.37300.32700̲.̲3̲6̲8̲0̲0.56900.49300̲.̲6̲3̲1̲0̲0.4518
RWTE0.35900.36600.27600.32100.31800.38300.45800.58700.3835
SpurFlow (ours)0.53510.45480.39400.37640.31820.63060̲.̲4̲9̲8̲8̲0.66950̲.̲4̲8̲4̲7̲

SpurFlow ranks second overall (0.4847), close to LISA (0.5020) and above RePlan (0.4792). Our lightweight coefficient head learns the teacher's implicit edit regions during distillation without mask annotations and guides localized adversarial training.

(2) Evaluating Localization Beyond Mask Overlap.

To complement IoU and the paper’s GPT-based evaluation, we conducted a blinded human mask-preference study on Prolific. This assesses localization beyond a single annotated mask, since prompts such as “add a boat to a large lake” admit multiple valid placements.

We compare SpurFlow with RegionE and RWTE on 50 randomly selected PIE-Bench examples across nine editing categories. Participants viewed the source image, instruction, and two mask overlays to assess edit coverage and exclusion of unrelated regions. Method names were hidden and presentation order randomized. Fifteen anonymous participants provided 300 judgments, with three independent judgments per comparison. Counts below are from SpurFlow's perspective. Ties include “both good” and “both poor,” while Unsure denotes “cannot judge.”

ComparisonWinTieLossUnsure
vs. RegionE134952
vs. RWTE146400

Using Win / (Win + Tie + Loss), excluding Unsure responses, SpurFlow achieves win rates of 90.54% against RegionE and 97.33% against RWTE. These results provide complementary human evidence of localization quality on this subset.

Impact of Localization Errors on Editing Quality (Answer to W2)

We agree with the reviewer that localization errors can affect crop selection and adversarial training. We test robustness by introducing approximately 10% spatial perturbations to crop regions during adversarial training on FLUX.1-Kontext.

IDNFEActionAddAdjustBackgroundComposeExtractRemoveReplaceStyleAverageQ-SCQ-PQQ-O
D24.043.963.873.663.382.413.444.183.983.636.017.805.78
E23.713.814.123.763.032.053.844.193.943.625.997.795.77
Ours + 10% Crop Perturbation24.043.964.173.963.242.273.694.203.973.736.057.925.85
Ours24.274.084.194.033.492.523.554.174.023.816.108.035.93

The perturbations moderately reduce the ImgEdit-Bench average from 3.81 to 3.73 and GEdit-Bench Q-O from 5.93 to 5.85. Nevertheless, the perturbed variant still outperforms D (3.63 / 5.78) and E (3.62 / 5.77), indicating that the benefits of localized adversarial supervision persist under the tested perturbations.

Comments on the paper

1 comment
AuthorsOpenReview ↗

To all Reviewers

Dear Reviewers,

We sincerely thank all reviewers for their careful reading and constructive suggestions, which have helped improve the clarity and completeness of our work.

We have revised the manuscript to address the concerns and questions raised above. All changes are highlighted in blue in the revised PDF.

Best regards, Authors of Paper 15528

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.