acceptodds
Under review as a conference paper at ICLR 2027

PersistEdit: Steering Persistent Noise for Inversion-Free Image Editing

Abstract

Recent inversion-free flow-based methods enable text-guided image editing without explicitly inverting source images into noise. They typically exploit averaging across timesteps through independent Gaussian noise sampling to approximate the noise-marginalized editing direction. However, finite sampling leaves realization-specific deviations that can favor different ways of realizing target semantics, shaping path-specific editing tendencies. Particularly, independent resampling repeatedly shifts these tendencies, disrupting the coherent accumulation of their effects over successive editing steps and leading to incomplete edits with residual source content or visible artifacts. To address this issue, we introduce PersistEdit, a training-free and inversion-free method to organize and selectively exploit realization-specific effects through noise control, which involves two key components, i.e., Noise Persistence and Semantic Steering. Specifically, Noise Persistence models noise evolution as a Gaussian-Markov process that combines inherited memory with newly sampled innovation, promoting coherent editing tendencies across updates. Semantic Steering selects target-favorable innovations using semantic cues within the estimated editing region while retaining inherited memory, allowing selected semantic preferences to persist across updates. Together, these designs produce a temporally coherent and semantically directed editing trajectory, improving target-semantic alignment while maintaining source-image fidelity. Extensive experiments on PIE-Bench show that PersistEdit outperforms existing methods in overall performance. Compared with FlowEdit, it improves background PSNR by 4.37 dB and edited-region CLIP similarity by 2.32%.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
image editing, training-free, inversion-free

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_j2AMOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces PersistEdit, a training-free and inversion-free image editing method built upon FlowEdit. The paper argues that independently resampling noise at each step leads to inconsistent realization-specific editing tendencies. To address this, PersistEdit introduces noise persistence to correlate noise across steps within the editing region, and semantic steering to select target-favorable noise innovations based on attention responses. It additionally applies a source-latent preservation constraint outside the editing mask. Experiments on part of the PIE-Bench show promising editing and background-preservation trade-off.

Critical strengths
  1. The paper is easy to follow.

  2. The paper provides a clear motivation for studying cross-step noise dependence in inversion-free flow editing and proposes a targeted solution. The analysis of persistent noise offers an interesting perspective on the realization-specific variations in FlowEdit.

Critical weaknesses
  1. The mask estimation pipeline and its computational cost are unclear. According to Algorithm 1, the complete editing mask must be estimated before the editing process. However, TAM aggregates attention responses across steps 5–28 and multiple layers, which appears to require additional multi-step model forward. The paper does not explain how this pass is implemented or whether its cost is included in the reported inference time. Figure 9 only compares early-layer and full-layer candidate scoring for steering, rather than providing an end-to-end comparison with FlowEdit. The paper should compare complete runtime, NFE with baselines, including both TAM and candidate selection.

  2. The contributions of the individual designs are not sufficiently isolated. Noise Persistence is motivated as a way to strengthen semantic editing by allowing target changes to develop more completely. However, the “+Persistence” configuration also produces a large improvement in background preservation, which appears to be partly caused by the separate source-latent preservation constraint. The current ablation does not isolate persistence from this constraint. Moreover, the mask itself is a central component because it determines where persistence and steering are applied and where the source latent is copied. The paper should separately evaluate persistence-only, preservation-only, steering-only, and their combinations, with estimated masks and no masks. In particular, applying persistence or steering globally would help clarify whether their effectiveness depends mainly on the localized mask.

  3. The experimental evaluation is not sufficiently comprehensive. The experiment misses SNR-Edit, a closely related and open-source method that directly improves FlowEdit through noise rectification. Other recent flow-based editing methods, such as FIA-Edit [1] and DRFS [2], should also be compared or discussed more carefully. In addition, the claimed transferability to SD3-Medium is supported only by qualitative examples. Quantitative comparisons against methods that support the SD family, such as SNR-Edit, DNAEdit, and FlowEdit, are necessary. Besides, completely removing the style-transfer category is questionable. FlowEdit is designed as a general editing method, and prior methods such as DNAEdit and FlowAlign evaluate on the complete PIE-Bench. Style transfer would also provide a useful setting for examining persistence and steering when the editing mask covers most or all of the image. Results on the excluded category should be provided. Finally, evaluation on long-prompt and complex editing such as DNA-Bench should also be compared and discussed.

  4. The metrics in Table 1 are incomplete. Whole-image CLIP similarity and Structure Distance, both commonly reported on PIE-Bench, should be included.

[1] FIA-Edit: Frequency-Interactive Attention for Efficient and High-Fidelity Inversion-Free Text-Guided Image Editing. AAAI 2026.

[2] Delta Rectified Flow Sampling for Text-to-Image Editing. CVPR 2026.

Questions for authors
  1. How is the number of K selected? Any ablation?

  2. How well does the attention score corresponds to the final editing quality? A quantitative comparison among highest-score, random, and lowest-score candidate selection would be more convincing than the current qualitative examples.

7 comments

W1. End-to-End Runtime.

We construct the TAM mask before editing in a separate source-conditioned pass. To reduce this cost, we aggregate attention over eight sampled timesteps, which yields mask IoU close to that of 24-step TAM. The binary-mask IoU results on PIE-Bench are shown in Table 1 below and have also been added to Appendix A.2. We have updated Appendix D (Table 5) with end-to-end runtime comparisons on FLUX.1-dev. As shown in Table 2 below, PersistEdit takes 22.19 s per image, including 2.32 s for TAM construction, all Steering candidate evaluations, and decoding. With an existing edit mask, the runtime is 19.87 s per image.

Table 1. Effect of the number of TAM aggregation timesteps on mask IoU on PIE-Bench (FLUX.1-dev).

TAM timesteps14824
Timestep indices{13}{5, 13, 20, 28}{5, 8, 12, 15, 18, 21, 25, 28}5–28
Pixel-level IoU ↑52.21%53.74%56.36%56.95%
Token-level IoU ↑58.67%60.21%64.04%64.13%

Table 2: Per-image runtime (s) on FLUX.1-dev.

ModelRF-InvFlowEditFlowAlignKV-EditOurs (excluding TAM)Ours (including TAM)
FLUX.1-dev17.3515.6115.6123.2719.8722.19

W2. Component Contributions and Spatial Localization

We have updated Section 4.3 to explicitly separate the contributions of Noise Persistence, Semantic Steering, and Background Constraint (BC), including Persistence-only, Steering-only, and BC-only controls. All variants reported in Section 4.3 use TAM-estimated masks rather than ground-truth masks. Persistence and Steering individually improve semantic alignment, and their combination further improves CLIP-Edited and ImageReward to 22.61 and 78.96, respectively. These results support their complementary roles in promoting more complete target-semantic changes. BC accounts for the main background-preservation gain, increasing PSNR from 22.57 to 26.98 dB.

To further examine the benefits of Noise Persistence, we tested whether it improves semantic alignment at comparable accumulated-update variance. We compared Noise Persistence with an i.i.d. control that samples noise independently across timesteps and approximately matches the variance of accumulated updates within the editing region. Semantic Steering and Background Constraint were disabled in both settings. As shown in Table 3 below, Noise Persistence achieves higher CLIP-Edited (23.25 vs. 22.97) and ImageReward (59.04 vs. 52.72) at comparable Structure Distance and background PSNR. These gains support our proposed mechanism that cross-step noise dependence helps editing tendencies accumulate coherently. Experiment details are provided in Appendix E.4.

We additionally compare global and localized Persistence, with Semantic Steering and BC disabled (Table 4 below). Global Persistence achieves higher semantic alignment scores, but reduces source fidelity: PSNR decreases from 22.56 to 21.24 dB, while Structure Distance increases from 28.07 to 35.04. This demonstrates that semantic gains remain achievable without masks, while localization helps preserve source content. Semantic Steering likewise selects candidate innovations using target-token attention within the editing region; a global scoring objective would remove this spatial constraint. Figure 13 further suggests that early Steering helps establish the target object’s placement and spatial extent. Together, these observations motivate region-aware control, for which TAM provides automatic spatial support.

Table 3. Comparison with a variance-matched i.i.d. control on PIE-Bench.

MethodIn-mask accumulated-update varianceStructure Distance ×10³ ↓PSNR ↑CLIP-Whole ↑CLIP-Edited ↑IR ×10² ↑
FlowEdit0.10521.5724.4425.2022.6646.70
Noise Persistence0.17026.5423.8425.6423.2559.04
Variance-matched i.i.d.0.17326.3123.8525.3622.9752.72

Table 4: Effect of TAM-based localization on Noise Persistence. Semantic Steering and Background Constraint are disabled.

Persistence ScopeStructure Distance ×10³ ↓HPS ×10² ↑AS ↑PSNR ↑LPIPS ×10³ ↓MSE ×10⁴ ↓CLIP-Edited ↑IR ×10² ↑
Global (no mask)35.0429.606.6621.24146.25114.3222.7789.82
Localized (TAM)28.0728.816.6122.5698.0883.5222.5672.02

W3 & W4. Expanded Evaluation and Metrics (1/2)

Additional Metrics and SNR-Edit. Thank you for these suggestions. Table 5 below adds CLIP-Whole and Structure Distance, along with SNR-Edit using its official FLUX.1-dev implementation. SNR-Edit’s structure-aware noise rectification helps limit trajectory drift, consistent with its slightly lower Structure Distance. PersistEdit improves background PSNR by 2.37 dB and CLIP-Edited by 0.99 over SNR-Edit.

FIA-Edit on SD3.5 and DRFS on SD3. We added quantitative comparisons with FlowEdit, FlowAlign, and DRFS on SD3-Medium in Appendix G (Table 12), alongside qualitative examples in Fig. 15. DRFS shifts noisy latents toward the target trajectory and achieves strong semantic alignment. PersistEdit comes within 0.14 of DRFS on CLIP-Edited while improving background PSNR by 3.96 dB and achieving lower Structure Distance. On SD3.5-Medium (Table 6 below), PersistEdit and FIA-Edit both improve background PSNR over FlowEdit (+2.85 and +4.20 dB, respectively). PersistEdit also improves CLIP-Edited by 0.18 over FlowEdit, compared with a change of −0.83 for FIA-Edit.

Table 5. Quantitative Comparison on PIE-Bench; * indicates a newly added baseline.

MethodStructure Distance ×10³ ↓HPS ×10² ↑AS ↑Background PSNR (dB) ↑Background LPIPS ×10³ ↓Background MSE ×10⁴ ↓CLIP-Whole ↑CLIP-Edited ↑IR ×10² ↑
P2P71.9325.406.2517.83208.30220.6925.6222.08-4.47
P2P (PnP-Inv)11.1626.076.3327.2552.6434.1325.5621.7622.67
PnP25.3126.596.3822.41109.8981.1925.8922.3340.89
MasaCtrl23.9224.936.1322.8381.1479.1825.1321.145.44
RF-Inv51.9627.506.4619.34211.26167.7425.4121.9340.86
KV-Edit24.4027.226.4227.3256.4943.4525.4622.1258.76
DNAEdit17.6927.686.5125.2585.8447.4525.0621.8555.02
FlowEdit24.8128.326.5922.5797.8183.3625.7122.0156.48
FlowAlign13.8727.816.5125.8953.3942.7525.3021.5638.49
SNR-Edit*16.8127.666.4924.5777.1454.3525.3021.5338.21
Ours17.4128.416.5426.9449.8036.8625.8122.5271.03

Table 6. Quantitative Comparison on SD3.5-Medium

MethodStructure Distance ×10³ ↓Background PSNR (dB) ↑CLIP-Whole ↑CLIP-Edited ↑IR ×10² ↑HPS ×10² ↑AS ↑
FlowEdit23.9423.2027.0523.19102.9228.476.48
FIA-Edit9.9427.4026.1922.3675.2527.336.41
Ours19.3126.0527.2023.37112.1828.796.44

W3 & W4. Expanded Evaluation and Metrics (2/2)

Style Transfer on PIE-Bench. We initially excluded style transfer from the aggregate because its masks leave little unedited area for assessing background preservation, and methods generally score higher on CLIP-Edited for this category. Table 7 reports all ten categories. PersistEdit leads on style transfer (26.24) and object replacement, addition, and removal. Non-rigid pose changes remain challenging.

Long-Prompt Editing on DNA-Bench. DNA-Bench uses the same source images and tasks as PIE-Bench with longer, more detailed prompts. We include change_style. PersistEdit gains 1.36 dB in background PSNR over DNAEdit and leads on CLIP-Edited (23.45) and ImageReward (82.71) (Table 8). With the same eight TAM timesteps, pixel/token IoU is 61.24%/68.35% on DNA-Bench vs. 56.36%/64.04% on PIE-Bench. These results demonstrate PersistEdit’s effectiveness for long-prompt editing.

Table7. Category-wise CLIP-Edited comparison on PIE-Bench.

Method#0:random#1:change object#2:add object#3:delete object#4:change attribute content#5:change attribute pose#6:change attribute color#7:change attribute material#8:change background#9:change style
RF-Inversion22.5420.1924.4217.6822.9721.8721.3423.3723.4726.01
KV-Edit22.9221.2225.1717.8022.1122.0420.1323.2923.3826.12
DNA-Edit22.5720.7324.6018.2322.0222.0120.5123.4622.2925.67
FlowEdit22.8320.5424.5918.3722.5422.2320.8023.6222.5125.98
FlowAlign22.4120.5723.9717.7522.4021.8220.2423.6221.5324.61
Ours23.2521.4725.4718.4722.7821.2221.3223.8423.7726.24

Table8. Quantitative Comparison on DNA-Bench.

MethodStructure Distance ×10³ ↓Background PSNR (dB) ↑CLIP-Whole ↑CLIP-Edited ↑IR ×10² ↑HPS ×10² ↑AS ↑
FlowEdit29.0721.6128.1323.2774.9828.676.67
DNAEdit18.7325.3928.3422.9970.9328.136.48
Ours20.4326.7528.5423.4582.7128.836.56

Q1. Candidate Pool Size

We ablated K∈{1,2,4,8}K\in\{1,2,4,8\} on PIE-Bench (Appendix F.2). CLIP-Edited rises from 22.37 at K=1K=1 to 22.54 at K=4K=4, with only 0.02 more at K=8K=8. We use K=4K=4 to balance semantic alignment and scoring cost.

Q2. Attention Score and Editing Quality

From lowest-score to random to highest-score selection, CLIP-Edited rises from 21.82 to 22.01 to 22.18, and ImageReward from 51.37 to 56.48 to 64.11 (Table 9 below). Highest-score selection improves alignment over random at similar Structure Distance and background PSNR. With our qualitative examples, this supports attention scores as a proxy for target alignment.

Table 9. Effect of Attention-Based Candidate Selection on Editing Quality

Candidate SelectionStructure Distance ×10³ ↓PSNR (dB) ↑CLIP-Whole ↑CLIP-Edited ↑IR ×10² ↑HPS ×10² ↑AS ↑
Lowest-score24.8622.4025.6521.8251.3728.296.60
Random24.8122.5725.7122.0156.4828.326.59
Highest-score24.8322.5625.7822.1864.1128.536.61
Reviewer_ofptOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents PersistEdit, a FlowEdit-based inversion-free editing method evaluated primarily with FLUX.1-dev. It argues that independent noise resampling disrupts the consistent accumulation of target content. PersistEdit combines three components: temporally correlated noise, attention-guided noise selection, and temporal attention masking with a background-preservation constraint. On PIE-Bench excluding the style category, PersistEdit reports the best LPIPS, edited-region CLIP, and ImageReward scores compared with nine baselines, improving background PSNR by 4.37 dB and CLIP by 0.51 over FlowEdit. The evaluation includes ablations, hyperparameter analyses, a 25-participant user study, and qualitative results with SD3-Medium.

Critical strengths

S1. A simple modification with a useful empirical insight. Noise persistence offers a training-free way to mitigate under-editing without additional network evaluations. Comparisons with noise averaging and velocity smoothing support the value of preserving temporal noise correlation and motivate placing memory in the noise.

S2. Strong empirical performance with thorough ablations. The complete pipeline achieves competitive alignment and preservation on PIE-Bench, supported by consistent user preferences and extensive component and design ablations. Qualitative transfer to SD3-Medium provides additional evidence of applicability beyond the primary backbone.

Critical weaknesses

W1. Unclear attribution of preservation gains. Table 3 bundles persistence with masked source-latent resetting, although the ablations suggest that resetting drives much of the preservation gain. Steering barely changes PSNR (22.57 → 22.56), and persistence without the reset slightly reduces it (22.57 → 22.31 for ρ = 0 → 0.9), while varying only the reset window shifts PSNR by 5 dB (Table 7). Without a FlowEdit baseline using the same TAM mask and reset schedule, the contribution beyond latent blending remains unclear. Unequal mask access (only KV-Edit receives the TAM mask in Table 1) and the missing SNR-Edit comparison further weaken the evaluation. As a result, the headline +4.37 dB over FlowEdit cannot be credited to the noise mechanism.

W2. The mechanism is not distinguished from increased editing strength. Higher persistence improves CLIP but worsens Structure Distance (Fig. 7), while more within-step noise averaging lowers CLIP (Table 4). This is consistent with less noise averaging producing stronger edits, i.e., movement along FlowEdit’s existing fidelity–editability trade-off. Frontier comparisons against FlowEdit swept over its own strength controls (target CFG, n_max, n_avg) are needed to demonstrate an improvement. Figure 6 cannot isolate the mechanism: persistence is applied only inside the mask, so high in-mask residual similarity holds by construction. Steering adds only 0.17 CLIP at 27% extra runtime, and its argmax selection breaks the N(0, I) marginals established in Eqs. (5)–(8). If ρ acts only as a strength knob, the central claim of coherent accumulation is not established.

W3. Insufficient evaluation of mask reliability and failure modes. Mask evaluation is limited to layer-wise soft mIoU on 100 COCO images, with scores of 0.10 to 0.26 (Fig. 8). Final binary-mask accuracy on PIE-Bench, robustness to mask errors, and edits extending beyond the source silhouette remain unexamined. Excluding style edits without a limitations or failure-case analysis further leaves the method’s scope unclear.

W4. Incomplete comparative and cross-backbone evaluation. Running FlowAlign on FLUX with default settings may understate its performance, given that its original quantitative evaluation uses SD3. Missing runtime comparisons with FlowEdit and mask-estimation costs leave efficiency unclear. Cross-backbone transferability is supported only by qualitative SD3 results.

Questions for authors

Q1. Attribution of gains. Please evaluate FlowEdit with the same TAM mask and reset schedule, and PersistEdit without resetting against unmasked baselines. Do persistence and steering still improve CLIP/ImageReward at matched or better LPIPS and Structure Distance?

Q2. Fidelity and editability. Please compare performance frontiers by sweeping FlowEdit’s target CFG, n_max, and n_avg and PersistEdit’s ρ, including ρ = 1. Does PersistEdit improve the frontier or merely shift the operating point?

Q3. Mechanism validation. Can temporal coherence be distinguished from reduced averaging? Useful controls include AR(1) versus block-constant noise with matched accumulated-update variance, and global versus masked persistence.

Q4. Mask definition and accuracy. How are TAM masks and steering scores defined for addition or removal edits when the relevant source or target concept is absent? Please also report final binary-mask IoU against PIE-Bench ground truth.

5 comments

W1. Component Contributions and Evaluation Fairness

Attribution of Preservation Gains. Thank you for the constructive comment. we revised Sec. 4.3 and Table 3 of the paper to isolate Noise Persistence, Semantic Steering, and Background Constraint (BC), including a BC-only control. All variants use TAM-estimated masks.

As shown in Table 1, Persistence and Steering individually improve semantic alignment, and combining them further improves CLIP-Edited and ImageReward to 22.61 and 78.96, respectively, supporting the complementary roles of temporal coherence and semantic selection. BC accounts for the main background-preservation gain. With BC retained, adding Persistence and Steering increases CLIP-Edited from 21.64 to 22.52 and ImageReward from 39.49 to 71.03, while maintaining comparable PSNR (26.98 vs. 26.94 dB).

Evaluation Details. KV-Edit requires an explicit edit mask, so we generate its masks with the same TAM procedure as for PersistEdit. We evaluate SNR-Edit with its official FLUX.1-dev implementation, default editing hyperparameters, and 28 sampling steps. SNR-Edit incorporates segmentation-derived structural priors into noise rectification to improve source preservation. As shown in Table 2, it improves background preservation over FlowEdit, accompanied by lower semantic alignment scores in our evaluation. PersistEdit achieves stronger background preservation together with better semantic alignment.

Table 1: Ablation study. BC denotes Background Constraint.

ComponentHPS ×10² ↑AS ↑PSNR ↑LPIPS ×10³ ↓MSE ×10⁴ ↓CLIP ↑IR ×10² ↑
Base28.326.5922.5797.8183.3622.0156.48
+Steering28.536.6122.5697.8983.4722.1864.11
+Persistence28.816.6122.5698.0883.5222.5672.02
+Persistence+Steering29.006.6522.5199.4283.5722.6178.96
+Background Constraint27.696.5026.9848.6333.8621.6439.49
+Persistence+Steering+BC28.416.5426.9449.8036.8622.5271.03

Table 2: Quantitative comparison on PIE-Bench using FLUX.1-dev.

MethodHPS ×10² ↑AS ↑PSNR ↑LPIPS ×10³ ↓MSE ×10⁴ ↓CLIP-Edited ↑IR ×10² ↑
FlowEdit28.326.5922.5797.8183.3622.0156.48
SNR-Edit27.666.4924.5777.1454.3521.5338.21
PersistEdit28.416.5426.9449.8036.8622.5271.03

W2. Editing Strength and Noise Control

Editing Strength and Noise Control. We expanded Figure 7 to compare a ρ\rho sweep at fixed target CFG with FlowEdit’s target-CFG sweep, with Semantic Steering and BC disabled. Within the evaluated range, increasing CFG yields diminishing gains in CLIP-Edited while Structure Distance continues to increase. Noise Persistence achieves higher CLIP-Edited at comparable Structure Distance, yielding a more favorable semantic–structural trade-off. These results support benefits beyond simply increasing guidance strength.

Residual coherence. Figure 6 illustrates how temporal noise correlation is reflected in editing-velocity residuals. These residuals arise from nonlinear, state-dependent model evaluations, and their cosine similarity is not directly imposed by the noise construction. See our response to Q3 for how we distinguish temporal coherence from reduced averaging.

Semantic Steering. The costs and benefits of Semantic Steering cannot be assessed in isolation. It complements Noise Persistence, which carries noise information across steps so that editing tendencies can accumulate. Steering uses target-token attention to select continuations that favor the intended edit. With BC disabled, ImageReward rises from 56.48 for Base to 72.02 with Persistence alone and 78.96 with both components. Figure 13 also suggests that early Steering helps establish the target object's placement and spatial extent. As stated in Sec. 3.3, the Gaussian marginals in Eqs. (5)–(8) apply with Steering disabled, and the attention-based argmax can shift the selected innovations away from N(0,I)\mathcal{N}(0,I). Prior work shows that small, structured perturbations to Gaussian initialization [1] and score-based noise selection during denoising [2] can improve generation. Our ablation likewise shows that attention-based noise selection can improve editing outcomes.

[1] A Noise is Worth Diffusion Guidance. ICLR 2026.

[2] Test-Time Scaling of Diffusion Models via Noise Trajectory Search. NeurIPS 2025.

W3. Mask Reliability and Editing Limitations

Mask reliability. We have updated Appendix A.2 (Table 4) to report the accuracy of the final binary TAM masks on PIE-Bench. The 24-step TAM used in our main experiments achieves 56.95% pixel-level IoU and 64.13% token-level IoU against the ground-truth masks. CLIP-Edited is nearly identical with TAM and ground-truth masks (22.52 vs. 22.54), suggesting that semantic alignment is robust to the observed mask errors.

Editing Limitations. Table 3 shows that our method handles substantial changes to foreground content, achieving the highest CLIP-Edited scores for object replacement, addition, and removal. We agree that edits extending beyond the source silhouette remain challenging. The change attribute content category includes many such edits, such as object enlargement, and our method scores 22.78, slightly below RF-Inversion’s 22.97. When the edit mask misses areas that the enlarged object needs to occupy, it can restrict the edit. This is a common limitation of mask-based methods such as KV-Edit. Large non-rigid pose changes also remain challenging, with a CLIP-Edited score of 21.22 versus 22.23 for FlowEdit.

Table3. Category-wise CLIP-Edited comparison on PIE-Bench.

Method#0:random#1:change object#2:add object#3:delete object#4:change attribute content#5:change attribute pose#6:change attribute color#7:change attribute material#8:change background#9:change style
RF-Inversion22.5420.1924.4217.6822.9721.8721.3423.3723.4726.01
KV-Edit22.9221.2225.1717.8022.1122.0420.1323.2923.3826.12
DNA-Edit22.5720.7324.6018.2322.0222.0120.5123.4622.2925.67
FlowEdit22.8320.5424.5918.3722.5422.2320.8023.6222.5125.98
FlowAlign22.4120.5723.9717.7522.4021.8220.2423.6221.5324.61
Ours*23.2521.4725.4718.4722.7821.2221.3223.8423.7726.24

W4. Cross-Backbone Comparison and Runtime

Thank you for the constructive feedback. To make the comparison with FlowAlign fairer, we added a comparison on SD3-Medium in Appendix G, using the backbone from FlowAlign’s original quantitative evaluation. Appendix G also presents qualitative results for our method on SD3-Medium.

We updated Appendix D with end-to-end runtime comparisons on FLUX.1-dev, as shown in Table 4 below. PersistEdit takes 22.19 s per image, including 2.32 s for TAM construction using eight sampled timesteps, compared with 15.61 s for FlowEdit. When an edit mask is already available, for example when supplied by the user as in KV-Edit, PersistEdit can skip TAM construction and run in 19.87 s per image.

Table 4: Per-image runtime (s) on FLUX.1-dev.

ModelRF-InvFlowEditFlowAlignKV-EditOurs (excluding TAM)Ours (including TAM)
FLUX.1-dev17.3515.6115.6123.2719.8722.19

Q1. Expanded Ablation Study

We added two controls in Sec. 4.3: FlowEdit with the same TAM mask and reset schedule and PersistEdit without Background Constraint. The constraint lowers FlowEdit’s CLIP-Edited from 22.01 to 21.64, consistent with FlowAlign’s observation that unintended edits can raise CLIP. With the same constraint, PersistEdit achieves higher CLIP-Edited (22.52 vs. 21.64) and ImageReward (71.03 vs. 39.49) at comparable PSNR (26.94 vs. 26.98 dB).

Q2. Fidelity–Editability Frontier

We updated Sec. 4.4 and Figure 7 to compare a ρ\rho sweep at fixed target CFG with FlowEdit’s target-CFG sweep, with Semantic Steering and Background Constraint disabled. Higher CLIP-Edited at comparable Structure Distance shows an improved frontier over the tested range.

Q3. Mechanism Validation

Variance-Matched Comparison. Thank you for suggesting these controls. Table 5 below compares AR(1), block-constant noise, and a variance-matched i.i.d. control on 140 PIE-Bench random examples using ground-truth masks, with Steering and Background Constraint disabled. Block-constant noise reuses the previous in-mask noise with probability p=0.6p=0.6 and resamples otherwise, matching AR(1)’s lag-one correlation at ρ=0.6\rho=0.6. At each step, the i.i.d. control computes two velocities from independent noise draws at the same latent and applies v~t=vt(1)+λM⊙(vt(1)−vt(2))\tilde v_t=v_t^{(1)}+\lambda M\odot(v_t^{(1)}-v_t^{(2)}). We tune λ\lambda per example to approximately match AR(1)’s in-mask accumulated-update variance, estimated from 20 seeds per example.

The three variants achieve similar in-mask accumulated-update variance, Structure Distance, and background PSNR. AR(1) and block-constant noise yield similar semantic alignment and both outperform FlowEdit on CLIP-Edited and ImageReward. These results suggest that similar semantic gains can be achieved with different temporally correlated samplers. AR(1) provides a simple implementation that retains noise history while introducing a fresh innovation at each step. The variance-matched i.i.d. control improves over FlowEdit, suggesting that larger stochastic updates explain part of the gain. At comparable fidelity and with slightly lower accumulated-update variance, AR(1) still achieves higher CLIP-Edited (23.25 vs. 22.97) and ImageReward (59.04 vs. 52.72). This supports a contribution from cross-step noise dependence beyond variance amplification alone.

Global vs. Masked Persistence. We also compared global and localized Persistence with Steering and Background Constraint disabled (Table 6), restricting Persistence to the TAM region raises background PSNR by 1.32 dB and reduces Structure Distance (35.04 to 28.07), with a 0.21 drop in CLIP-Edited. Global Persistence’s higher semantic scores may partly reflect off-mask edits, consistent with FlowAlign’s observation. This supports masked Persistence for source preservation.

Table 5. Noise samplers and variance-matched controls

MethodIn-mask accumulated-update varianceStructure Distance ×10³ ↓PSNR ↑CLIP-Whole ↑CLIP-Edited ↑IR ×10² ↑
FlowEdit0.10521.5724.4425.2022.6646.70
AR(1)0.17026.5423.8425.6423.2559.04
Block-constant0.16926.1923.8725.6523.2160.03
Variance-matched i.i.d.0.17326.3123.8525.3622.9752.72

Table 6. Global versus localized Persistence

Persistence scopeStructure Distance ×10³ ↓HPS ×10² ↑AS ↑PSNR ↑LPIPS ×10³ ↓MSE ×10⁴ ↓CLIP-Edited ↑IR ×10² ↑
Global (no mask)35.0429.606.6621.24146.25114.3222.7789.82
Localized (TAM)28.0728.816.6122.5698.0883.5222.5672.02

Q4. Mask definition and accuracy

For addition, TAM uses a source concept indicating placement (e.g., “rabbit” when adding a hat), and Steering uses the added object (“hat”). For removal, TAM uses the object to remove, and Steering uses “without” if present in the target prompt or a retained concept otherwise (e.g., “table” when removing water).

We updated Appendix A.2 with final binary-mask IoU on PIE-Bench (Table 7 below).

Table 7. Effect of the number of TAM aggregation timesteps on mask IoU on PIE-Bench (FLUX.1-dev).

TAM timesteps14824
Timestep indices{13}{5, 13, 20, 28}{5, 8, 12, 15, 18, 21, 25, 28}5–28
Pixel-level IoU ↑52.21%53.74%56.36%56.95%
Token-level IoU ↑58.67%60.21%64.04%64.13%
Reviewer_CTESOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

PersistEdit targets a plausible limitation of inversion-free editing: independent noise resampling can interrupt consistent semantic changes. Correlated noise and attention-based innovation selection are reasonable responses, and quantitative results plus a human study support their usefulness. The remaining concerns are incomplete separation from background preservation, unmatched inference computation, and limited evidence that the proposed mechanism generalizes beyond the evaluated semantic-editing setting.

Critical strengths
  1. The motivation is clear and supported by targeted analysis. Section 4.4 measures residual coherence with steering and background preservation disabled. Appendices D.1-D.2 also compare persistence with noise averaging and velocity smoothing.

  2. The design is technically reasonable and explicitly qualified. The Gaussian-Markov process provides controllable correlation, and Section 3.3 correctly states that its Gaussian identities describe the process before semantic selection. This avoids claiming that selected innovations retain the original distribution.

  3. The evaluation includes independent perceptual evidence. Table 2 reports a randomized, anonymized human study with 25 participants. The method uses attention rather than CLIP for selection, so the CLIP evaluation is not directly optimized by the same scoring function.

Critical weaknesses
  1. The main preservation gain is not isolated from background replacement. Table 3's “+Persistence” variant includes a background consistency constraint, whereas the base has PSNR 22.57 and the variant reaches 26.96. Section 4.4 appropriately isolates persistence, but uses ground-truth masks and does not establish how much of Table 1's gain remains under estimated masks with background preservation held constant. A background-constraint-only control is needed to attribute the headline preservation improvement.

  2. Equal sampling steps do not ensure equal computation. Semantic steering evaluates four candidates at early steps, and Appendix A.3 processes candidates through the first 19 Transformer blocks. TAM also aggregates attention across timesteps. Table 1 reports step counts without complete runtime or function-evaluation budgets. Figure 9 compares two versions of PersistEdit, rather than an equal-budget baseline. End-to-end cost, including mask construction, should accompany comparisons with FlowEdit and attention-based alternatives.

  3. The reliability and scope of the mechanism need stronger validation. Correlated residuals need not be beneficial residuals, and high target-token attention is only a proxy for successful editing. The paper acknowledges over-editing at high persistence, but gives limited category-wise or seed-wise failure analysis. The main quantitative evaluation uses one benchmark with style transfer excluded; SD3 transfer in Appendix F is qualitative. User-specified concept tokens and mask errors further limit applicability to edits with ambiguous localization.

Questions for authors
  1. Ablation: In Table 3, is background replacement enabled only for rows containing Persistence? Is an existing result available with the same estimated mask and preservation window but persistence disabled?

  2. Cost: Does the 19.87 s timing include TAM construction, all candidate evaluations, and decoding? What are the corresponding baseline runtimes under the same hardware and settings?

  3. Evaluation protocol: Were persistence, preservation windows, and steering duration chosen on data separate from the final evaluation? Please clarify the number of distinct images and repeated judgments in the human study and whether existing seed-wise variability was measured.

5 comments

W1. Attribution of Preservation Gains

Thank you for the constructive comment. Following your recommendation, we have revised Section 4.3 to explicitly separate Noise Persistence , Semantic Steering, and Background Constraint. All variants use TAM-estimated masks rather than ground-truth masks. As shown in Table 1 below, the complementary roles of Noise Persistence and Semantic Steering in enhancing target-semantic editing remain evident. Background Constraint substantially improves background preservation but leads to a decrease in CLIP similarity. This suggests that the higher text alignment achieved by the unconstrained baseline may partly arise from unintended changes to the source content, consistent with the observations in FlowAlign.

Table 1: Ablation study. BC denotes Background Constraint.

ComponentHPS ×10² ↑AS ↑PSNR ↑LPIPS ×10³ ↓MSE ×10⁴ ↓CLIP ↑IR ×10² ↑
Base28.326.5922.5797.8183.3622.0156.48
+Steering28.536.6122.5697.8983.4722.1864.11
+Persistence28.816.6122.5698.0883.5222.5672.02
+Persistence+Steering29.006.6522.5199.4283.5722.6178.96
+Background Constraint27.696.5026.9848.6333.8621.6439.49
+Persistence+Steering+BC28.416.5426.9449.8036.8622.5271.03

W2. End-to-End Computational Cost

To quantify the computational cost of our method, we added end-to-end runtime comparisons on FLUX.1-dev using the same NVIDIA RTX A6000 GPU (Appendix D, Table 5). To reduce mask construction cost, we aggregate attention over eight representative timesteps, achieving mask IoU close to that of 24-step TAM, as shown in Table 2 below. We have added this comparison to Appendix A.2. As shown in Table 3 below, PersistEdit takes 22.19 s per image, including TAM construction, all Semantic Steering candidate evaluations, and decoding. With a pre-existing edit mask, the runtime is 19.87 s per image.

Table 2. Effect of the number of TAM aggregation timesteps on mask IoU on PIE-Bench (FLUX.1-dev).

TAM timesteps14824
Timestep indices{13}{5, 13, 20, 28}{5, 8, 12, 15, 18, 21, 25, 28}5–28
Pixel-level IoU ↑52.21%53.74%56.36%56.95%
Token-level IoU ↑58.67%60.21%64.04%64.13%

Table 3: Per-image runtime (s) on FLUX.1-dev.

ModelRF-InvFlowEditFlowAlignKV-EditOurs (excluding TAM)Ours (including TAM)
FLUX.1-dev17.3515.6115.6123.2719.8722.19

W3. Mechanism Reliability and Evaluation Scope

Residual Correlation and the Attention Proxy. Thank you for raising these points. Residuals can reflect different spatial configurations, extents, or appearances under the same target prompt, shaping path-specific editing tendencies. However, the editing updates are already guided by the source and target conditions. Noise Persistence reduces repeated switching between these tendencies, allowing target content to develop more coherently. Our ablation demonstrates its effectiveness: adding Persistence alone improves CLIP-Edited from 22.01 to 22.56 and ImageReward from 56.48 to 72.02. To further examine the benefits of correlated residuals, we compared Noise Persistence with a variance-matched i.i.d. control (Appendix E.4, Table 13). At comparable accumulated-update variance and source fidelity, Noise Persistence achieves higher CLIP-Edited (23.25 vs. 22.97) and ImageReward (59.04 vs. 52.72), supporting a contribution from temporal coherence to semantic alignment. To reduce unintended edits associated with less target-aligned residual effects, Semantic Steering uses target-token attention as a selection proxy. The ablation supports its empirical utility: Steering alone improves CLIP-Edited and ImageReward to 22.18 and 64.11, respectively, while combining it with Persistence further improves these scores to 22.61 and 78.96. These results support both the effectiveness of the proxy and the complementary roles of temporal coherence and semantic selection.

Over-Editing and the CFG Trade-off. To address the concern about over-editing, we updated Section 4.4 and Figure 7 to compare a sweep over ρ\rho at fixed target CFG with a target-CFG sweep for FlowEdit. Within the tested range, Noise Persistence achieves higher CLIP-Edited at comparable Structure Distance, offering a better trade-off between editability and structural fidelity than increasing editing strength through target CFG alone.

Category Results, Seed Robustness, and SD3 Transfer. Using the same settings as the main experiments, we additionally report category-wise CLIP-Edited results, including style transfer, as shown in Table 4. Compared with the FlowEdit baseline, our method achieves higher scores in nine of ten categories, particularly object replacement and addition, consistent with more complete target-semantic expression. Pose changes remain challenging, highlighting a limitation in non-rigid editing. We also analyze seed robustness in Appendix E.3: our method exhibits lower mean per-image variance in Structure Distance, PSNR, and ImageReward, with comparable CLIP-Edited variance (+0.7%) relative to FlowEdit. Furthermore, we provide additional quantitative results on SD3-Medium in Appendix G to further validate the cross-backbone generalizability of our method across flow-based models.

Table4. Category-wise CLIP-Edited comparison on PIE-Bench.

Method#0:random#1:change object#2:add object#3:delete object#4:change attribute content#5:change attribute pose#6:change attribute color#7:change attribute material#8:change background#9:change style
RF-Inversion22.5420.1924.4217.6822.9721.8721.3423.3723.4726.01
KV-Edit22.9221.2225.1717.8022.1122.0420.1323.2923.3826.12
DNA-Edit22.5720.7324.6018.2322.0222.0120.5123.4622.2925.67
FlowEdit22.8320.5424.5918.3722.5422.2320.8023.6222.5125.98
FlowAlign22.4120.5723.9717.7522.4021.8220.2423.6221.5324.61
Ours23.2521.4725.4718.4722.7821.2221.3223.8423.7726.24

Q1. Extended Ablation Study

In the original ablation study, the background consistency constraint was enabled only in configurations that included Persistence. We have updated Section 4.3 and Table 3 to disentangle the contributions of Noise Persistence, Semantic Steering, and Background Constraint. All variants use TAM-estimated masks.

Q2. End-to-end runtime comparison

The 19.87 s per image reported in Figure 9 includes all candidate evaluations and decoding but excludes TAM construction. We have updated Appendix D (Table 5) to report end-to-end runtimes on the same hardware, including 22.19 s per image for PersistEdit with 8-step TAM and runtimes for the baselines.

Q3. Evaluation Protocol Details

TAM’s high-confidence layers are selected using 100 examples from COCO. The other hyperparameter choices are informed by empirical analyses on PIE-Bench, with systematic sensitivity analyses reported for persistence strength (Sec. 4.4), the steering window (Appendix F.1), candidate pool size KK (Appendix F.2), and the preservation window (Appendix F.3). The main comparisons use fixed settings across evaluation images.

The human study involves 25 participants. Each participant completes eight comparisons against each of five baselines, totaling 40 comparisons, using examples randomly sampled from the random category of PIE-Bench, which contains 140 examples. This yields 200 comparisons per baseline. Examples may overlap across participants.

Appendix E.3 additionally reports seed-wise variability on 140 images with 20 seeds per input. Compared with FlowEdit, our method exhibits lower mean per-image variance in Structure Distance, PSNR, and ImageReward, with comparable CLIP-Edited variance (+0.7%).

Comments on the paper

1 comment
Area_Chair_3RDzOpenReview ↗

Discussion

Dear reviewers,

We have mixed signals from the review ratings

  • ofpt: 3: Weak acceptance
  • CTES: 2: Weak rejection
  • j2AM: 2: Weak rejection

Could you take a look at each other's reviews and the author responses and initiate a discussion to reach consensus? Were your concerns resolved and questions answered? How do you weigh the strengths and weaknesses?

Please note that ICLR 2027 has Policy on LLM use.

Best, 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.