acceptodds
Under review as a conference paper at ICLR 2027

REVE: Region-Aware Velocity Extrapolation for Accelerated Image Editing

Abstract

Instruction-based image editing models produce high-quality edits but often incur substantial inference latency due to repeated evaluations of diffusion transformers. Existing training-free acceleration methods reduce this inference cost by reusing cached velocities or rescaling their magnitudes at skipped timesteps, but these approximations often compromise fine detail. We observe that velocity extrapolation improves fine detail in edit regions but introduces artifacts in non-edit regions, while velocity reuse helps preserve non-edit content. Building on this insight, we propose REgion-aware Velocity Extrapolation (REVE), a training-free framework that approximates velocities at skipped timesteps in a region-aware and temporally adaptive manner. REVE distinguishes edit from non-edit regions by measuring agreement between predicted velocities and reference velocities that reconstruct the reference image. At skipped timesteps, REVE reuses velocities in non-edit regions and extrapolates velocities in edit regions, with a time-dependent weight that progressively strengthens extrapolation as denoising advances from structure formation to detail refinement. Experiments with Step1X-Edit, Boogu-Image-0.1-Edit, and LongCat-Image-Edit on GEdit-Bench and GIE-Bench show that REVE achieves – speedups, matching or exceeding those of the compared training-free acceleration methods. Among these methods, REVE achieves the best semantic consistency, perceptual quality, and detail fidelity, together with the highest pixel-level and perceptual similarity to the unaccelerated results.

Primary area:
generative models
Keywords:
Instruction-based image editing, training-free acceleration, flow matching, region-aware velocity extrapolation

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_W5CaOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper introduces REVE, a training-free acceleration framework for instruction-based image editing. The method identifies edit and non-edit regions through agreement between predicted velocities and reference-image reconstruction velocities. At skipped timesteps, it reuses cached velocities in non-edit regions and applies first-order velocity extrapolation in edit regions, with an extrapolation weight that increases throughout denoising. Experiments on three editing models and two benchmarks report approximately 2.54–2.73× speedups while preserving output fidelity better than the evaluated acceleration baselines.

Critical strengths
  1. Clear motivation and practical design. The distinction between detail refinement in edit regions and preservation in non-edit regions provides a compelling motivation for region-dependent velocity approximation. The resulting method is simple, training-free, and compatible with existing editing models.
  2. Consistent results across models and benchmarks. Evaluation covers three editing backbones and two benchmarks using shared principal hyperparameters. REVE consistently improves fidelity to unaccelerated outputs over the reported acceleration baselines while achieving comparable or higher speedups.
  3. Informative mechanism-oriented ablations. The paper compares velocity reuse, global extrapolation, region-specific extrapolation, and time-weighted extrapolation. These results support the importance of restricting extrapolation to edit regions. Direction perturbation and extrapolation-order experiments provide additional evidence beyond a standard component ablation.
  4. Compatibility with adaptive skipping. Integration with TeaCache demonstrates that the proposed approximation strategy can complement an existing adaptive evaluation schedule, rather than being useful only under the default fixed schedule.
Critical weaknesses
  1. The contributions of region partitioning and velocity approximation are not fully disentangled. REVE adopts RegionE’s spatial acceleration and KV-caching framework while changing both region identification and skipped-step velocity approximation. The existing ablations primarily examine extrapolation choices. A controlled comparison that independently swaps the masks and approximation rules under identical evaluation schedules would more clearly establish the contribution of each proposed component.
  2. The quality–efficiency comparison is concentrated on individual operating points. The main tables demonstrate favorable results at the selected configurations, but do not establish performance across a broader range of acceleration strengths. Matched-latency comparisons and quality–latency curves would strengthen the claim. A simple reduced-step sampling baseline is also missing, despite being a natural alternative for reducing model evaluations.
  3. The proposed Detail Fidelity metric has limited specificity. DF measures agreement in mean gradient magnitude within the edit region. Different textures, misplaced edges, and some artifacts can have similar mean gradient magnitudes, so a lower DF does not necessarily indicate more accurate detail reconstruction. PSNR, SSIM, and LPIPS provide useful complementary evidence, but the stronger claims about fine-detail preservation would benefit from localized perceptual evaluation or human assessment.
  4. Region-identification reliability and statistical uncertainty are insufficiently characterized. Mask visualizations and threshold sensitivity studies are useful, but quantitative localization accuracy and failure analysis remain limited. Some SC/PQ improvements over RegionE are small—for example, PQ increases by 0.019 on Step1X-Edit with GEdit-Bench—and no confidence intervals or repeated judging results are provided. This makes the robustness of the semantic and perceptual improvements harder to assess.
Questions for authors
  1. Under identical masks, evaluation schedules, and cache updates, how much improvement comes from the proposed velocity approximation? Conversely, what improvement comes from replacing only the region-partitioning criterion?
  2. Are matched-latency results or quality–latency curves available for REVE and RegionE? How does REVE compare with simply reducing the number of sampling steps at a similar computational budget?
  3. Are the subject-replacement samples used to select the mask threshold disjoint from the samples used for the final benchmark results? How are pseudo ground-truth edit masks constructed for global stylization, object movement, and object addition?
  4. Are the SC/PQ improvements stable under paired uncertainty analysis or repeated VLM evaluation? Is there supporting human evaluation of edit-region detail and non-edit-region preservation?
4 comments
AuthorsOpenReview ↗

Response to Reviewer W5Ca

We thank the reviewer for the detailed and constructive comments. All new results are included in the revised paper, with revisions highlighted in blue.

1. Disentangling region partitioning and velocity approximation (W1, Q1). We replace our region partitioning and velocity approximation with their RegionE counterparts, one at a time, on Step1X-Edit with GEdit-Bench (Table 3, Section 5.3). All variants share identical evaluation schedules, cache updates, and latency.

Region Part.Velocity Approx.SC ↑PQ ↑DF ↓PSNR ↑SSIM ↑LPIPS ↓
REVERegionE8.2068.0250.07031.730.9420.056
RegionEREVE8.1888.0260.04731.810.9460.051
REVEREVE8.2388.0500.04631.790.9470.049

With identical masks (rows 1 and 3), our velocity approximation reduces DF from 0.070 to 0.046 and LPIPS from 0.056 to 0.049. With an identical approximation (rows 2 and 3), replacing only the partitioning criterion with ours raises SC from 8.188 to 8.238 and PQ from 8.026 to 8.050. The velocity approximation thus mainly improves detail fidelity, while the region partitioning mainly improves instruction following and perceptual quality.

2. Quality–latency curves and reduced-step sampling (W2, Q2). We provide quality–latency curves for REVE, RegionE, and the unaccelerated model with fewer sampling steps on a 330-sample subset of GEdit-Bench (Figure 5, Section 5.2). We vary the number of edit-region evaluations for REVE, the skipping threshold for RegionE, and the number of sampling steps for the unaccelerated model. REVE achieves lower DF and LPIPS than RegionE at all evaluated speedups, with comparable or higher SC and PQ, and its advantage grows as the speedup increases. Compared with reducing the sampling steps, although the 10-step baseline achieves slightly higher SC at a 2.80× speedup, REVE achieves higher PQ and reduces DF and LPIPS to about one third of those of the baseline, even at a higher speedup of 2.88×.

3. Specificity of DF (W3). We agree that DF does not verify whether details are reproduced at the correct locations, and we now state this in App. C. We therefore complement DF with PSNR, SSIM, and LPIPS computed within the pseudo ground-truth edit region (Table 5, App. C), where LPIPS provides a localized perceptual evaluation.

ModelBenchmarkPSNR ↑ (RegionE / REVE)SSIM ↑ (RegionE / REVE)LPIPS ↓ (RegionE / REVE)
Step1X-EditGEdit-Bench26.36 / 27.450.879 / 0.8970.076 / 0.062
Step1X-EditGIE-Bench25.80 / 26.490.856 / 0.8680.078 / 0.066
Boogu-Image-0.1-EditGEdit-Bench28.60 / 29.510.905 / 0.9160.051 / 0.042
Boogu-Image-0.1-EditGIE-Bench26.91 / 27.930.869 / 0.8840.056 / 0.048
LongCat-Image-EditGEdit-Bench26.51 / 27.390.886 / 0.9020.063 / 0.050
LongCat-Image-EditGIE-Bench25.93 / 26.970.867 / 0.8860.058 / 0.047

REVE outperforms RegionE on all three metrics for all models and benchmarks, and it also receives the highest human rating for fine detail within the edit region (see point 5 below). These results indicate that the lower DF of REVE reflects more accurate detail reconstruction.

4. Localization accuracy and failure analysis of region identification (W4). We compare the masks of REVE and RegionE with pseudo ground-truth masks on local edits from the 330-sample subset of GEdit-Bench (Table 13, App. K).

Region Part.Recall ↑Precision ↑IoU (All) ↑IoU (Small) ↑IoU (Medium) ↑IoU (Large) ↑
RegionE0.7190.6710.4590.5000.4610.440
REVE0.6840.7730.4930.6070.4940.442

REVE achieves higher precision at a slightly lower recall, indicating that its edit regions contain fewer tokens outside the edited content, and it improves IoU for all edit sizes. Our failure analysis identifies two main causes of deviation. First, the pseudo mask can be larger than the edited content; for example, when only one word on a menu is replaced, the pseudo mask covers the entire menu, whereas REVE marks only the replaced word. Second, the editing model can alter the entire image in response to a local instruction; in such cases, REVE correctly marks the changed regions, whereas the pseudo mask covers only the target object. These deviations therefore mainly reflect inaccurate pseudo masks or over-editing by the model rather than errors in region partitioning.

AuthorsOpenReview ↗

Response to Reviewer W5Ca (Continued)

5. Stability of the SC/PQ improvements and human evaluation (W4, Q4). We compute paired bootstrap 95% confidence intervals of the per-sample SC and PQ differences between REVE and RegionE (Table 7, App. E).

ModelBenchmarkΔSC [95% CI]ΔPQ [95% CI]
Step1X-EditGEdit-Bench+0.051 [+0.000, +0.109]+0.019 [−0.018, +0.054]
Step1X-EditGIE-Bench+0.026 [−0.050, +0.101]+0.034 [+0.011, +0.058]
Boogu-Image-0.1-EditGEdit-Bench+0.022 [−0.026, +0.076]+0.018 [−0.018, +0.053]
Boogu-Image-0.1-EditGIE-Bench+0.026 [−0.013, +0.068]+0.024 [+0.002, +0.045]
LongCat-Image-EditGEdit-Bench+0.040 [−0.013, +0.096]+0.051 [+0.015, +0.087]
LongCat-Image-EditGIE-Bench+0.012 [−0.030, +0.050]+0.027 [+0.007, +0.049]

The mean differences are positive in all 12 settings, and the PQ improvement is statistically significant in 4 of 6 settings. The intervals are relatively wide because both methods remain close to the unaccelerated models and receive identical scores on most samples. Nevertheless, the consistent direction across all models and benchmarks indicates that the improvements are systematic rather than incidental.

We also conduct a human evaluation on Step1X-Edit with 66 GEdit-Bench samples, 6 from each of the 11 tasks (Table 14, App. L). Given the reference image and the instruction, 24 annotators rate the anonymized outputs of all six acceleration methods, shown in random order, on a five-point scale for fine detail within the edit region (Detail) and preservation of the non-edit region (Preservation).

MethodDetail ↑Preservation ↑
TaylorSeer3.6823.521
TeaCache3.6723.639
RAS3.2323.593
SpotEdit3.4573.604
RegionE3.8193.665
REVE (Ours)3.8813.698

REVE receives the highest ratings for both edit-region detail and non-edit-region preservation.

6. Disjointness of threshold-selection samples (Q3). The subject-replace samples used to select τ are not disjoint from the evaluation set, since GEdit-Bench provides no separate validation split (App. G). However, the same τ = 0.82 is applied without retuning to GIE-Bench and to all three editing models.

7. Construction of pseudo ground-truth masks (Q3). We have clarified the construction of the pseudo ground-truth masks in App. C. For each instruction, the target phrase is extracted with Qwen3-VL, localized with GroundingDINO, and segmented with SAM2, with the image used for segmentation determined by the edit type. For global stylization, the edit affects the entire image, so no pseudo mask is constructed with these models. For object movement, the object is segmented in both the reference image and the unaccelerated edited image, and the union of the two segments covers both its original and new locations. For object addition, the added object appears only in the edited image and is therefore segmented in the unaccelerated edited image.

Reviewer_W5CaOpenReview ↗

Thanks for rebuttal

I believe this paper presents an interesting technique for image editing, and experiments prove its effectiveness. According to authors' feedback and other reviewers' comments, I maintain my original score, which tends to accpet.

AuthorsOpenReview ↗

Thank You for Your Follow-up

Thank you for your follow-up and for recognizing the value of our work. Your constructive comments have helped us improve the paper. We sincerely appreciate your time and effort in reviewing our work.

Reviewer_X9zCOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes REVE, a training-free framework for accelerating instruction-based image editing with flow-matching diffusion transformers. The key observation is that velocity extrapolation improves fine details in edited regions, but introduces artifacts in non-edited regions. REVE identifies edit regions through velocity agreement, reuses cached velocities in non-edit regions, and applies temporally adaptive first-order velocity extrapolation in edit regions. Experiments across three image-editing models and two benchmarks demonstrate 2.54–2.73× speedups, with improved editing quality and fidelity compared to existing training-free acceleration methods, including RegionE and TeaCache.

Critical strengths
  1. The paper identifies distinct effects of velocity reuse and extrapolation across edit and non-edit regions. The frequency-domain analysis and ablation studies provide meaningful evidence supporting region-aware velocity approximation.

  2. REVE achieves favorable quality-efficiency trade-offs across three models and two benchmarks. The paper conducts comparisons with relevant baselines, component ablations, and integration with adaptive timestep skipping. It provides relatively comprehensive empirical validation.

Critical weaknesses
  1. REVE largely builds upon RegionE's spatial acceleration, KV caching, and evaluation schedule, while introducing first-order velocity extrapolation. The individual contributions of the new region-partitioning mechanism and extrapolation strategy relative to RegionE are not sufficiently isolated.

  2. The independent validation of editing quality is insufficient. PSNR, SSIM, LPIPS, and the proposed detail-fidelity metric primarily measure similarity to unaccelerated outputs. Semantic consistency and perceptual quality rely on a single automated evaluator (Qwen3-VL-8B), the claimed perceptual improvements remain somewhat uncertain.

Questions for authors
  1. How much of REVE's improvement over RegionE comes specifically from velocity-aware region partitioning versus temporally adaptive extrapolation? This is important for assessing the paper's core novelty.

  2. Do the observed quality improvements persist under independent human evaluations, particularly for fine-detail fidelity and preservation of non-edit regions?

3 comments
AuthorsOpenReview ↗

Response to Reviewer X9zC

We thank the reviewer for the positive evaluation and the helpful suggestions. All new results are included in the revised paper, with revisions highlighted in blue.

1. Contributions of region partitioning and temporally adaptive extrapolation (W1, Q1). To isolate each contribution relative to RegionE, we replace our region partitioning and velocity approximation with their RegionE counterparts, one at a time, on Step1X-Edit with GEdit-Bench (Table 3, Section 5.3). All variants share the same evaluation schedule and latency (2.68× speedup).

Region Part.Velocity Approx.SC ↑PQ ↑DF ↓PSNR ↑SSIM ↑LPIPS ↓
REVERegionE8.2068.0250.07031.730.9420.056
RegionEREVE8.1888.0260.04731.810.9460.051
REVEREVE8.2388.0500.04631.790.9470.049

The two components improve complementary aspects of editing quality. Our temporally adaptive extrapolation mainly improves detail fidelity, reducing DF from 0.070 to 0.046 compared with RegionE's velocity rescaling. Our velocity-aware partitioning mainly improves instruction following and perceptual quality, raising SC from 8.188 to 8.238 and PQ from 8.026 to 8.050, in line with its higher localization precision (App. K). Combining both achieves the best results on five of the six metrics at identical latency, showing that both components contribute to the improvement of REVE over RegionE.

2. Independent human evaluation (W2, Q2). We conduct a human evaluation on Step1X-Edit with 66 GEdit-Bench samples, 6 from each of the 11 tasks (Table 14, App. L). Given the reference image and the instruction, 24 annotators rate the anonymized outputs of all six acceleration methods, shown in random order, on a five-point scale. The criteria are fine detail within the edit region (Detail), preservation of the non-edit region (Preservation), and overall editing quality (Quality).

MethodDetail ↑Preservation ↑Quality ↑
TaylorSeer3.6823.5213.240
TeaCache3.6723.6393.326
RAS3.2323.5933.003
SpotEdit3.4573.6043.124
RegionE3.8193.6653.368
REVE (Ours)3.8813.6983.386

REVE receives the highest mean rating on all three criteria, including fine-detail fidelity and non-edit preservation. Since these ratings rely on neither the unaccelerated outputs nor a VLM evaluator, they provide independent evidence that the improvements of REVE persist under human judgment.

3. Robustness of the VLM-based improvements (W2). To further assess the SC and PQ improvements, we compute paired bootstrap 95% confidence intervals of the per-sample differences between REVE and RegionE (Table 7, App. E). The mean differences are positive in all 12 model–benchmark–metric settings, and the PQ improvement is statistically significant in 4 of 6 settings. This consistent direction across models and benchmarks indicates that the perceptual improvements are systematic rather than incidental.

Thanks for providing the additional results and analysis. I don't have any further requests or questions for the authors. I recommend weak acceptance.

AuthorsOpenReview ↗

Thank You for Your Follow-up

Thank you for your follow-up and for your positive evaluation of our work. Your comments on component isolation and human evaluation have helped us improve the paper. We sincerely appreciate your time and effort in reviewing our work.

Reviewer_g5tPOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes REVE, a training-free acceleration method for flow-matching image editors. Building on RegionE’s spatial computation and KV caching framework, it identifies edit regions through velocity agreement, reuses cached velocities in non-edit regions, and extrapolates historical velocities in edit regions with a time-dependent weight. Experiments on three models and two benchmarks report 2.54–2.73× speedups over unaccelerated inference, with improved output fidelity relative to the compared acceleration methods.

Critical strengths
  1. Targeted analysis of velocity approximation. Equal-latency comparisons support treating edit and non-edit regions differently. Direction and magnitude perturbations provide further evidence about the source of the extrapolation benefit.

  2. Consistent empirical results. Evaluation across three editors and two benchmarks reports improved fidelity at comparable or lower latency than the tested baselines, with additional results using adaptive timestep skipping.

Critical weaknesses
  1. Limited and insufficiently isolated advances over RegionE. REVE retains RegionE’s spatial computation, KV caching, and full-image evaluation structure, with its main changes concentrated in the region criterion and velocity approximation. The ablations do not isolate these changes through controlled replacements of RegionE’s components under matched evaluation schedules, leaving their individual contributions unclear.

  2. Incomplete characterization of the quality–cost tradeoff. The paper reports overall latency, but does not separately quantify region-partitioning and extrapolation overhead or the memory cost of retaining velocity history. Without component-level measurements and quality–latency curves, it is difficult to establish how much of the improvement reflects the proposed approximation rather than differences in token coverage or evaluation schedules.

Questions for authors
  1. Under matched evaluation schedules, how much does replacing RegionE’s region criterion or velocity approximation individually improve quality and latency?

  2. What runtime and peak-memory overheads do region partitioning and historical-velocity extrapolation introduce?

  3. Does the time-dependent extrapolation weight outperform a tuned constant weight at the same computational budget?

1 comment
AuthorsOpenReview ↗

Response to Reviewer g5tP

We thank the reviewer for the positive assessment and the constructive questions. All new results are included in the revised paper, with revisions highlighted in blue.

1. Contribution of each component under matched schedules (W1, Q1). We replace our region partitioning and velocity approximation with their RegionE counterparts, one at a time, on Step1X-Edit with GEdit-Bench (Table 3, Section 5.3). All variants share the same evaluation schedule and therefore the same latency.

Region Part.Velocity Approx.SC ↑PQ ↑DF ↓PSNR ↑SSIM ↑LPIPS ↓Latency (s)
REVERegionE8.2068.0250.07031.730.9420.05629.47
RegionEREVE8.1888.0260.04731.810.9460.05129.47
REVEREVE8.2388.0500.04631.790.9470.04929.47

The two components improve complementary aspects of editing quality. Our velocity extrapolation mainly improves detail fidelity, reducing DF from 0.070 to 0.046 compared with RegionE's velocity rescaling. This is consistent with our perturbation ablation (Table 2), which shows that extrapolation-derived directions, left unchanged by rescaling, drive much of the detail improvement. Our velocity-agreement criterion mainly improves SC and PQ, in line with its higher localization precision (App. K). Combining both achieves the best results on five of the six metrics at identical latency, so both components contribute to the gains over RegionE.

2. Runtime and memory overhead (W2, Q2). We measure the cost of each component on Step1X-Edit (Table 8, App. F).

ComponentRuntimePeak Memory
Region partitioning2.87 ms (0.010%)1.0 MiB (0.003%)
Velocity extrapolation7.86 ms (0.027%)1.0 MiB (0.003%)
REVE (total)29.47 s36.47 GiB

Both components operate on latent tensors and require no additional model evaluations, so their combined runtime is about 0.04% of the total latency. Retaining the velocity history requires only one additional velocity field from the previous evaluation, which accounts for less than 0.01% of the peak memory.

3. Quality–latency tradeoff (W2). We compare REVE and RegionE across speedups on a 330-sample subset of GEdit-Bench, varying the number of edit-region evaluations for REVE and the skipping threshold for RegionE (Figure 5, Section 5.2). REVE achieves lower DF and LPIPS than RegionE at all evaluated speedups, with comparable or higher SC and PQ. As the speedup increases, RegionE degrades sharply, whereas REVE degrades gradually, so its advantage widens. Regarding token coverage and schedules, rows 1 and 3 of Table 3 share identical masks and schedules, so their difference isolates the effect of our velocity approximation.

4. Time-dependent vs. tuned constant weight (Q3). We compare our time-dependent weight with constant weights at the same 2.68× speedup on a 330-sample subset of GEdit-Bench (Table 11, App. H).

α(t)SC ↑PQ ↑DF ↓PSNR ↑SSIM ↑LPIPS ↓
0.008.2037.9670.07231.570.9390.061
0.258.2067.9550.05931.610.9420.056
0.508.2337.9420.04831.480.9430.055
0.758.1827.9270.04131.180.9420.055
1.008.2067.9670.04130.720.9390.058
(1−t)0.5(1−t)^{0.5} (ours)8.2457.9790.04231.660.9450.052

The time-dependent weight outperforms every constant weight on five of the six metrics. Constant weights trade detail fidelity against similarity to the unaccelerated output: increasing α lowers DF but also lowers PSNR. By attenuating extrapolation during structure formation and strengthening it during detail refinement, the time-dependent weight mitigates this trade-off. Its average value is about 0.55, yet it outperforms α = 0.5 on all six metrics, which indicates that the gain comes from its temporal variation rather than its average magnitude.

Comments on the paper

1 comment
Area_Chair_YsNgOpenReview ↗

Reviewer-author discussion period is ending. Please feedback.

Dear Reviewers,

Thanks for your time and effort in reviewing ICLR2027 submissions.

As the reviewer-author discussion period will end very soon, please take time to read the author’s rebuttal and give your feedback. Your continuous engagement is highly appreciated.

Best regards,

Your AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.