acceptodds
Under review as a conference paper at ICLR 2027

Where to Look, Where to Edit: Enhancing Editability in Inversion-Free Editing

Abstract

Inversion-free methods based on rectified-flow models enable efficient image and video editing by avoiding costly inversion. However, their limited editing strength often leads to poor alignment with target prompts. We find that this limitation results from insufficient attention to target-prompt semantics within editing regions. Selectively increasing this attention can improve target-prompt guidance and enable stronger edits while maintaining structural fidelity. This motivates us to propose Spatially Adaptive Attention Transfer Edit (SAATEdit), a training-free and inversion-free framework for more effective image and video editing. SAATEdit extracts spatial priors from source-prompt-guided attention to localize editing regions and uses these priors to compute bounded spatial gains for targeted attention enhancement. Benefiting from spatial priors and targeted attention enhancement, SAATEdit achieves stronger edits and better alignment with target prompts while maintaining structural fidelity. Experiments on PIE-Bench and FiVE-Bench demonstrate that SAATEdit achieves superior image and video editing performance with minimal inference overhead.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
image editing, video editing

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_YyQiOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes SAATEdit, a training-free and inversion-free framework for image and video editing. Built on FlowEdit, it transfers source-object attention logits to target editing tokens through Attention Reference Transfer (ART). Spatial Adaptive Enhancement (SAE) further scales and clips the transferred logits to strengthen localized editing. Experiments on PIE-Bench and FiVE-Bench demonstrate improved editing performance over FlowEdit and Wan-Edit.

Critical strengths
  1. The motivation is clear and reasonable. Insufficient editing strength and inaccurate localization are relevant limitations of inversion-free editing.
  2. The method is evaluated on both images and videos, showing some applicability across modalities.
Critical weaknesses
  1. The contribution relative to existing attention-based editing methods is insufficiently established. FlowDirector [1], FlowAnchor [2], and SAM-Flow [3] already explore attention-guided localization, spatial attention modulation, or localized velocity updates. ReFlex [4] also combines source-attention injection and adaptation to improve editing. Although their implementations differ, these works are closely related to the proposed localization and enhancement mechanisms. Direct comparisons with the relevant inversion-free methods, and at least a detailed discussion of ReFlex, are needed to clarify the novelty. The related work about attention manipulation, which only covers Prompt-to-Prompt, is also insufficient.

  2. The method is not fully specified for basic editing scenarios. Appendix C allows a null source anchor for object addition and a null target anchor for deletion. However, Eq. 6 requires a non-empty source token set, while Eq. 7 requires target tokens to receive the transferred reference. Neither the equations nor the algorithm explain how these cases are handled. The source-derived reference may also constrain edits involving substantial changes in object shape or spatial extent, as acknowledged in the limitations. These issues leave the claimed applicability to general editing tasks unclear.

  3. The experimental comparisons are incomplete, and preservation of unedited content remains a concern. Video experiments omit FlowDirector and FlowAnchor, while image experiments omit representative inversion-free methods such as FlowAlign [5]. Although aggregate preservation metrics improve over FlowEdit, visible unintended changes remain. For example, adding a collar in Figure 6 also changes the dog’s facial expression. Stronger baseline comparisons at matched editing strength would help establish whether the method offers a better editing–preservation trade-off.

  4. Generalization across image backbones is not demonstrated. Image experiments are limited to FLUX. Results on other flow-based backbones, such as SD3, would clarify whether the source-attention reference remains reliable and whether the same ART/SAE settings retain their benefits.

  5. The video evaluation provides limited evidence for addition and deletion. These categories contain only 10 samples each, compared with 100 for each replacement or attribute category. Deletion achieves only 5% FiVE-Acc, while the reported addition score is based on a very small sample. The overall improvement therefore does not establish effectiveness across editing types, particularly those affected by the missing-anchor issue.

[1] FlowDirector: Training-Free Flow Steering for Precise Text-to-Video Editing. CVPR 2026.

[2] FlowAnchor: Stabilizing the Editing Signal for Inversion-Free Video Editing. ACM MM 2026.

[3] SAM-Flow: Source-Anchored Masked Flow for Training-Free Image Editing. arXiv 2026.

[4] ReFlex: Text-Guided Editing of Real Images in Rectified Flow via Mid-Step Feature Extraction and Attention Adaptation. ICCV 2025.

[5] FlowAlign: Trajectory-Regularized, Inversion-Free Flow-based Image Editing. ICLR 2026.

Questions for authors
  1. How are ART and SAE implemented when either anchor is null?
  2. Does SAATEdit retain its benefits on SD3 using the same ART window and SAE gains?
4 comments

Q1. Contribution relative to FlowDirector, FlowAnchor, SAM-Flow, and ReFlex

We thank the reviewer for pointing out these related works. In the revision, we have expanded the Related Work section to discuss attention manipulation methods beyond P2P, added FlowDirector [1], FlowAnchor [2], SAM-Flow [3], and ReFlex [4] to our comparison tables, and added the corresponding qualitative comparisons in Appendix J. Below we clarify our mechanism differences and empirical advantages over each method.

Comparison with ReFlex [4]. As detailed in our response to Reviewer 2, ReFlex is an inversion-based method (requiring noised inversion, mid-step residual injection, top-kk I2I-SA replacement, and mask blending). In I2T-CA, ReFlex mainly replaces unrelated background tokens to preserve image structure and scales target tokens with a fixed weight (α=4\alpha=4). In contrast, SAATEdit is strictly inversion-free and does not process background tokens: it extracts the structural prior from source object tokens, transfers this prior to target editing tokens, and applies a per-patch spatially adaptive gain. On PIE-Bench (FLUX) (Table R1), SAATEdit matches ReFlex on CLIP-E, but achieves much better structure preservation (Struct 27.57 vs. 42.60) and is 1.55×1.55\times faster (14.10 s vs. 21.80 s). Fig. 14 (Appendix J) further corroborates this global-structure deficiency: ReFlex also modifies regions beyond the target—e.g., it changes the clothing color in the first row, the global saturation in the third row, and the overall image style in the eighth row—consistent with its much larger Structure Distance.

Table R1. ReFlex vs. SAATEdit on PIE-Bench.

MethodStruct ↓PSNR ↑MSE ↓CLIP-W ↑CLIP-E ↑Time (s) ↓
ReFlex [4]42.6022.578.9626.3823.1421.80
SAATEdit (ours)27.5722.838.2726.5123.1414.10

Comparison with FlowDirector [1] and FlowAnchor [2]. FlowDirector builds a soft spatial mask M~\tilde{M} from cross-attention maps to gate the editing velocity (V^edit=V~edit⊙M~\hat{V}_{\mathrm{edit}} = \tilde{V}_{\mathrm{edit}} \odot \tilde{M}). In contrast, SAATEdit does not build a mask; it directly modulates the I2T-CA attention scores with a spatially adaptive gain. On FiVE-Bench (Table R2), our method outperforms FlowDirector on all five metrics and is 3.1×3.1\times faster (4.70 min vs. 14.40 min). Meanwhile, FlowAnchor requires an external mask (GT, hand-drawn, or bounding box) to separate edit and non-edit regions, which increases user cost and fails when no source mask applies. Fig. 13 (Appendix J, top-right and bottom-right examples) also corroborates this on the Add category. While this mask stabilization gives FlowAnchor better structure scores, our method achieves a much higher overall FiVE-Acc (59.04 vs. 37.35, Table R2). As shown in Table R3, FlowAnchor fails on Add (0.00 vs. 83.30) due to its mask dependency, and our method also leads across all replacement, color, and material categories.

Table R2. Video comparison on FiVE-Bench.

MethodStruct ↓PSNR ↑LPIPS ↓CLIP-W ↑FiVE-Acc ↑Time (min) ↓
FlowDirector [1]22.7922.15111.5926.9958.7414.40
FlowAnchor [2]10.6027.8474.5925.7137.351.59
Wan-Edit + Ours12.0625.3888.6227.4159.044.70

Table R3. Per-category FiVE-Acc of FlowAnchor and our method.

CategoryFlowAnchor [2]Wan-Edit + Ours
Rigid replacement25.0054.60
Non-rigid replacement14.5055.60
Color78.5081.30
Material38.0048.00
Add object0.0083.30
Delete object5.005.00

Comparison with SAM-Flow [3]. Similar to FlowDirector, SAM-Flow extracts a soft mask from cross-attention to restrict differential velocity updates without modifying attention, whereas SAATEdit directly modifies the I2T-CA score sub-block at the attention level. On PIE-Bench (SD3) (Table R4 and the Pareto plot in Fig. 12), the two methods sit at different ends of the structure–edit trade-off: SAM-Flow leans toward structure preservation, while SAATEdit achieves stronger editing quality and is 12×12\times faster (4.39 s vs. 52.72 s). Table R4. SAM-Flow vs. Ours on PIE-Bench(SD3) .

MethodStruct ↓PSNR ↑MSE ↓CLIP-W ↑CLIP-E ↑Time (s) ↓
SAM-Flow [3]14.9026.873.7126.9323.4752.72
FlowEdit + Ours33.6021.0910.7327.6824.234.39

In summary, these four methods are related to ours, but differ clearly in core mechanism, editing scope, and efficiency. All the above discussions have been added to the revised manuscript.

Q2. Applicability to basic editing types

We thank the reviewer for pointing this out. We have updated the Sec. 4.3 and Appendix D to explicitly specify the null-anchor cases: when Tsrc=∅\mathcal{T}_{src}=\emptyset (addition) or Ttar=∅\mathcal{T}_{tar}=\emptyset (deletion), ART (Eqs. 6–7) is skipped and only SAE is applied. Moreover, our per-category comparisons below show that the method improves editing capability across all basic edit types, including cases where the source-derived reference is least direct.

Video editing (FiVE-Bench). As reported in Table 10 and Fig. 8 of the revision, our method improves FiVE-Acc over Wan-Edit across all six basic categories (Table R5)—including Add object, Delete object, and Non-rigid replacement (which involves shape changes)—and also improves LPIPS across these categories (e.g., Rigid replacement 91.90→89.4091.90\to 89.40 and Add object 101.60→93.40101.60\to 93.40). Delete remains low for both methods due to the difficulty of this category rather than a failure of the enhancement path.

Table R5. Per-category FiVE-Acc on FiVE-Bench (higher is better).

CategoryWan-EditWan-Edit + Ours
Rigid replacement53.0354.60
Non-rigid replacement48.9955.60
Color66.6781.30
Material31.8248.00
Add object77.7883.30
Delete object0.005.00

Image editing (PIE-Bench). Similarly, in Table 12 and Fig. 9 of the revision, adding our method to FlowEdit improves CLIP-Whole across all four basic groups: Add, Delete, Change, and Style (Table R6). For edits with substantial shape or spatial changes, the reference is less informative, which is why Delete shows a larger structure cost. Overall, these results confirm that our method is applicable to general editing scenarios and consistently improves editing capability across them, while the structure cost for large shape changes is acknowledged.

Table R6. Per-category CLIP-Whole on PIE-Bench (higher is better).

GroupFlowEditFlowEdit + Ours
Add27.3727.96
Delete25.1725.69
Change26.2326.30
Style26.1028.41

Q3. Experimental comparisons and preservation of unedited content

We thank the reviewer for these suggestions. We have addressed the missing baselines in the revision: FlowDirector and FlowAnchor are now included in the video comparisons (Table 1 and Table 10; summarized in Tables R2 and R3), and FlowAlign is compared on both FLUX and SD3 (Table 2). We also strengthen the matched-strength comparison on SD3 with strong inversion-based and inversion-free baselines, including UniEdit, SNR-Edit, and SplitFlow.

FlowAlign. Table R7 summarizes the comparison. On FLUX, FlowAlign has better structure (Struct 25.39 vs. 27.57) but lower CLIP alignment (CLIP-W 25.28 vs. 26.51; CLIP-E 22.29 vs. 23.14). On SD3, the same pattern holds (Struct 15.17 vs. 33.60; CLIP-W 25.94 vs. 27.68; CLIP-E 22.33 vs. 24.23). FlowAlign therefore occupies the structure-preserving side of the trade-off, while our method occupies the editing side.

Matched editing strength. The FLUX Pareto sweep in Figure 7 compares the methods across the editing-strength sweep neditn_{\text{edit}}, the SD3 Pareto plot in Figure 12 (Appendix I) provides the same comparison on SD3, and the SD3 block of Table 2 includes UniEdit [7], SNR-Edit [8], SplitFlow [9], FlowEdit, and FlowAlign. For all compared methods, the hyperparameters follow the official configurations. This provides a stronger matched-strength comparison than an aggregate number alone.

The collar example in Figure 6. We agree with the reviewer's observation: in that example FlowEdit changes the dog's facial expression when adding the collar, and our method shows the same change. Our method improves editing strength and prevents the enhancement from spreading, but it does not substantially improve the base editor's structural fidelity; when FlowEdit exhibits a structural failure mode, our method often inherits it. We do not claim to be able to fix such local failure modes. The improvement in the aggregate structure-preservation metrics mainly comes from the spatial constraint of the enhancement operation, not from correcting every unintended change. We have made this explicit in the revised limitation.

Table R7. FlowAlign vs. SAATEdit on FLUX and SD3.

BackboneMethodStruct ↓PSNR ↑MSE ↓CLIP-W ↑CLIP-E ↑
FLUXFlowAlign [5]25.3922.568.4025.2822.29
FLUXOurs27.5722.838.2726.5123.14
SD3FlowAlign [5]15.1728.032.4225.9422.33
SD3Ours33.6021.0910.7327.6824.23

Q4. Generalization across image backbones

We thank the reviewer for this suggestion. We have added an SD3 block to Table 2, covering FlowEdit, FlowAlign, UniEdit, SNR-Edit, SplitFlow, and Sam-Flow. SD3 is also a joint-attention MMDiT, so the same image-to-text sub-block is available. We apply ART/SAE with a light SD3-specific setting (nart=2n_{art}=2, αbg=0.5\alpha_{bg}=0.5, αedit=0.85\alpha_{edit}=0.85); no architectural change or retraining is required.

Table R8 reports the full SD3 comparison. On SD3, our method improves CLIP-Whole from 27.41 to 27.68 and CLIP-Edit from 23.92 to 24.23, at a structure cost (Struct 30.26→33.60; PSNR 21.64→21.09; MSE 9.87→10.73). Among all SD3 methods in Table 2, ours attains the highest CLIP-Whole and CLIP-Edit, confirming that the source-attention reference remains reliable on a different backbone and that ART/SAE retain their editing benefits with a light setting adjustment.

Figure 12 (Appendix I) reports the corresponding Pareto sweep on SD3: our method consistently improves over FlowEdit across the editing-strength sweep and occupies the high-editability side of the trade-off. It does not dominate the strongest SD3 methods, which is consistent with the trade-off discussed in Q1 and Q3.

Table R8. Full SD3 comparison on PIE-Bench.

Method (SD3)Struct ↓PSNR ↑MSE ↓CLIP-W ↑CLIP-E ↑Time (s) ↓
FlowAlign15.1728.032.4225.9422.332.98
UniEdit25.8324.335.5626.4623.081.07
Sam-Flow14.9026.873.7126.9323.4752.79
SNR-Edit26.8522.128.8027.3023.775.03
SplitFlow25.5622.538.1027.4023.957.50
FlowEdit [3]30.2621.649.8727.4123.924.11
FlowEdit+Ours33.6021.0910.7327.6824.234.39

Q5. Limited samples for addition and deletion on FiVE-Bench

We thank the reviewer for this concern. FiVE-Bench [6] is a widely used benchmark for training-free video editing; its core focus is fine-grained video editing, with a particular emphasis on object-level local edits. It therefore provides sufficient discriminative power for the training-free methods compared in this paper. FlowAnchor [2] itself also adopts FiVE-Bench as its evaluation benchmark, which further supports its relevance for this comparison. We report the complete per-category results in Table 10 and the corresponding radar plots in Figure 8, rather than relying only on the aggregate number.

Table R9 isolates the Add and Delete categories and compares our method with the base editor Wan-Edit [6]. On Add, FiVE-Acc improves from 77.78 to 83.30 and LPIPS from 101.60 to 93.40; on Delete, FiVE-Acc improves from 0.00 to 5.00 and LPIPS from 79.40 to 75.00. The corresponding radar plots in Fig. 8 (Appendix) show the same trend across all categories.

The small sample sizes of the Add and Delete subsets are defined by FiVE-Bench itself [6], not reduced by us. The Delete category is intrinsically difficult: the object must disappear while the background remains consistent. All methods obtain low FiVE-Acc (0.00–15.00), and our method improves over Wan-Edit. We therefore do not claim a large absolute gain on Delete; instead, we report it separately so that the reader can see the per-category behavior.

Table R9. Add and Delete categories on FiVE-Bench (base editor vs. ours).

MethodAdd FiVE-Acc ↑Add LPIPS ↓Delete FiVE-Acc ↑Delete LPIPS ↓
Wan-Edit [6]77.78101.600.0079.40
Wan-Edit + Ours83.3093.405.0075.00

[1] FlowDirector: Training-Free Flow Steering for Precise Text-to-Video Editing

[2] FlowAnchor: Stabilizing the Editing Signal for Inversion-Free Video Editing

[3] SAM-Flow: Source-Anchored Masked Flow for Training-Free Image Editing

[4] ReFlex: Text-Guided Editing of Real Images in Rectified Flow via Mid-Step Feature Extraction and Attention Adaptation

[5] FlowAlign: Trajectory-Regularized, Inversion-Free Flow-based Image Editing

[6] FiVE-Bench: A Fine-Grained Video Editing Benchmark for Evaluating Emerging Diffusion and Rectified Flow Models

[7] UniEdit-Flow: Unleashing Inversion and Editing in the Era of Flow Models

[8] SNR-Edit: Structure-Aware Noise Rectification for Inversion-Free Flow-Based Editing

[9] SplitFlow: Flow Decomposition for Inversion-Free Text-to-Image Editing

I thank the authors for their substantial effort in adding the missing baselines, SD3 experiments, and further explanations. These additions strengthen the empirical evaluation.

However, my concern about the incremental methodological novelty remains. More importantly, the undefined handling of addition and deletion was a fundamental omission in the original formulation, independently identified by all three reviewers. The revised explanation partially addresses this issue, but does not fully restore my confidence in the method’s completeness and soundness. It is still unclear how “all entity tokens” are extracted. Besides, the reliance on per-sample LLM processing limits the method’s practical applicability. Some editing gains also come with reduced preservation.

In recognition of the additional experiments and clarifications, I will raise my score from 1 to 2 (Weak rejection), while maintaining my rejection recommendation.

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

This paper proposes SAATEdit, a training-free and inversion-free framework for image and video-editing. The paper posits that previous inversion-free method like FlowEdits often fail at editing due to strong source-image prior causing week differential velocity between source and target branches. To address this issue, the method proposes two components, Attention Reference Transfer (ART) and Spatial Adaptive Enhancement (SAE). ART transfers the spatial attention of the source entity to relevant target tokens, while SAE is an adaptive scheme that modulates the strength. The method is validated on one image and one video editing benchmark.

Critical strengths
  1. The paper is well written and easy to follow.
Critical weaknesses
  1. Novelty and positioning against prior work
    The method novelty appears limited relative to prior attention-manipulation method in DiT editing. A particularly close prior work is ReFlex, which already performs attention adaptation for DiT image editing, including source-to-target attention replacement and attention amplification of selected target tokens, despite not being inversion-free. Moreover, these attention-based operators are fairly standard ever since P2P. Therefore, the main novelty seems to be identifying a particular problem of one branch (inversion-free), adapting widely-established operators to solve it, and its specific target-token selection rule. This distinction is potentially meaningful, but may be slightly incremental, and the paper does not discuss or compare against the particularly close prior work ReFlex.
    Kim, J., Park, J., Song, Y., Kwak, N., & Rhee, W. (2025). ReFlex: Text-Guided Editing of Real Images in Rectified Flow via Mid-Step Feature Extraction and Attention Adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 15939–15948.

  2. Value and scope of the method
    The motivation in Fig. 2 naturally supports the hypothesized failure mode for the specific type of editing illustrated there: a localized attribute change on an existing object. Here, the argument that a diffused attention not spatially focused on the intended target object causing a weak ΔV\Delta V makes sense.
    However, whether that diagnosis extends to other edit categories, is much less clear.
    For example, a style change is already intended to affect many spatial locations rather than a localized region. Another example will be object addition, where the newly introduced object token may already receive concentrated attention in the target region. Therefore, it is not convincing enough that insufficient target-region attention and weak $\Delta V$ is a general failure mode of inversion-free editing across edit categories.
    The per-category results in Table 9 do not answer this question, because they report only the final performance of SAATEdit rather than how much SAATEdit improves each category over FlowEdit.

  3. Method under-specified on one-side editing categories
    One-sided edits such as object addition and deletion will lead to null source or target anchors, whose handling is unclear by the current formulation presented in the method section.

Questions for authors
  1. Novelty: It will benefit the author to clarify their novelty and include relevant baselines such as ReFlex.
  2. Scope: Does the motivating failure hypothesis actually holds across different edit categories? It would also benefit to show the evaluation against baselines per-category, be clear about whether it improves over all categories in the evaluation, and if not, discuss it as limitations.
  3. Method Clarity: It may be valuable to actually clarify what the method does with one-sided edits such as addition or deletion.
3 comments

Q1. Novelty and positioning against prior work (ReFlex)

We thank the reviewer for pointing out ReFlex [1], and we have added a discussion in Sec. 3.3 and included it as a baseline in Table 2. While both methods are based on attention manipulation (specifically I2T-CA), their mechanisms differ significantly. ReFlex mainly replaces unrelated background tokens to preserve image structure, and simply scales target editing tokens with a fixed weight (α=4\alpha=4). In contrast, to improve inversion-free editing strength, our method does not process background tokens. Instead, it extracts the structural prior from source object tokens, transfers this prior to target editing tokens, and applies a per-patch spatially adaptive enhancement to amplify the edit region while suppressing the background.

Therefore, compared to prior attention-manipulation methods, the novelty of SAATEdit lies in: (i) adapting attention operations to inversion-free image and video editing (without ReFlex's noised inversion, latent caching, I2I-SA replacement, or mask blending); (ii) using an LLM to automatically extract source-object and target-edit words instead of manual specification [2] (while supporting an SAE-only path for one-sided edits); and (iii) proposing a spatially adaptive enhancement that outperforms ReFlex's fixed-weight amplification. Our ablation (Table 13, Appendix H) verifies this: replacing our spatially adaptive gain with a constant gain of 2.0 degrades all five metrics, showing that spatial adaptivity is key to balancing strong edits and structure preservation.

MethodStruct ↓PSNR ↑MSE ↓CLIP-W ↑CLIP-E ↑
ART + SAE (ours)27.5722.838.2726.5123.14
ART + Fixed weight29.2722.058.9626.1522.85

We also compare with ReFlex under the same FLUX.1-dev / PIE-Bench [3] setting (full results in Table 2):

MethodStruct ↓PSNR ↑MSE ↓CLIP-W ↑CLIP-E ↑NFETime (s) ↓
ReFlex [1]42.6022.578.9626.3823.144921.80
SAATEdit (ours)27.5722.838.2726.5123.144814.10

SAATEdit matches ReFlex on CLIP-E, but performs better on PSNR, MSE, CLIP-W, and especially Structure Distance (27.57 vs. 42.60). Under a similar NFE count—48 NFE for SAATEdit (24 source + 24 target) vs. 49 NFE for ReFlex (28 inversion + 21 generation)—our method is 1.55×1.55\times faster (14.10 s vs. 21.80 s) because it does not process the large I2I-SA blocks at each step. Thus, SAATEdit achieves a better structure–edit trade-off without inversion or high runtime cost. We will expand this discussion in the revision.

Q2. Value and scope across edit categories

We thank the reviewer for this comment. First, our diagnosis in Fig. 2 does extend to other edit categories: we have added new attention visualizations in the revision (Appendix K) showing that weak target-token attention also occurs in one-sided edits (such as Add and Style, where ART is skipped and only SAE is applied). Our visualizations confirm that applying SAE to focus and amplify target attention directly improves editing capability, even when no source object defines the target region.

Second, we now report per-category comparisons for both video and image editing. For video editing (Table 10 and Fig. 8 in the revision), our method improves both FiVE-Acc and LPIPS over Wan-Edit [1] across all six categories (Table R1). The gains are largest for Color, Material, Non-rigid and Add object, while Delete remains near zero for all methods due to the difficulty of this category rather than the enhancement path.

Table R1. Per-category comparison on FiVE-Bench.

CategoryWan-Edit (Acc ↑)+ Ours (Acc ↑)Wan-Edit (LPIPS ↓)+ Ours (LPIPS ↓)
Rigid replacement53.0354.6091.9089.40
Non-rigid replacement48.9955.6096.4094.50
Color66.6781.3089.6085.50
Material31.8248.0088.9086.00
Add object77.7883.30101.6093.40
Delete object0.005.0079.4075.00

For image editing (Table 12 and Fig. 9 in the revision), adding our method to FlowEdit [3] also improves CLIP-Whole across all four groups (Table R2), with clear gains in Style (+2.31) and Add (+0.59). Together, these visual and quantitative results confirm that our method consistently improves editing capability across diverse edit categories.

Table R2. Per-category CLIP-Whole on PIE-Bench (higher is better).

GroupFlowEditFlowEdit + Ours
Add27.3727.96
Delete25.1725.69
Change26.2326.30
Style26.1028.41

Q3. Method clarity for one-sided edits (addition and deletion)

We thank the reviewer for this question. This point was raised by all three reviewers, and we have made the procedure explicit in Sec. 4.3 and Appendix D.

ART requires both a non-empty source token set Tsrc\mathcal{T}_{src} and a non-empty target token set Ttar\mathcal{T}_{tar}, because it transfers a spatial prior from the source object tokens to the target editing tokens. For addition, Tsrc=∅\mathcal{T}_{src}=\emptyset (the new content is absent from the source prompt). For deletion, Ttar=∅\mathcal{T}_{tar}=\emptyset (the target prompt does not name the removed object). In both cases ART is skipped by setting M^tar=Mtar\hat{M}^{tar}=M^{tar}, and only SAE is applied. The reference is read from the target branch itself,Rj=max⁡i∈TrefMj,itarR_j = \max_{i \in \mathcal{T}_{\mathrm{ref}}} M^{tar}_{j,i}with Tref=Ttar\mathcal{T}_{\mathrm{ref}}=\mathcal{T}_{tar} for addition and Tref=Ttarent\mathcal{T}_{\mathrm{ref}}=\mathcal{T}^{\mathrm{ent}}_{tar} (all entity tokens of the target prompt) for deletion. The same set replaces Ttar\mathcal{T}_{tar} as the set of modulated columns in Eq. (10).

We have added Appendix D ("Additional Discussion on Different Edit Types"), which discusses these two cases and lists the extracted SRC_WORD / TAR_WORD for representative examples in Table 6. The per-category results in Tables 10 and 12 report these cases separately.

[1] ReFlex: Text-Guided Editing of Real Images in Rectified Flow via Mid-Step Feature Extraction and Attention Adaptation

[2] Prompt-to-Prompt Image Editing with Cross-Attention Control

[3] PnP Inversion: Boosting Diffusion-based Editing with 3 Lines of Code

I genuinely appreciate the authors' efforts and the quality of their work. The additional results have addressed my previous doubts regarding the utility and empirical merits of the method across different editing categories.

My remaining concern, however, is still with the mechanistic novelty of the method. That being said, I consider this a genuinely borderline case. The key question, in my view, is whether the strong empirical results and practical utility sufficiently outweigh the relatively incremental methodological contributions to warrant acceptance.

My assessment lies close to the boundary between Weak Reject and Weak Accept. At this point, I am inclined to maintain my current score, while emphasizing that I view the paper as borderline rather than a rejection. I would be open to reconsidering my recommendation in light of the broader reviewer discussion, as I see reasonable arguments on both sides.

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

SAATEdit modifies attention in existing inversion-free image and video editors. ART transfers a source-object attention pattern to the target editing tokens, and SAE adjusts its strength across the image. The method requires no retraining.

Critical strengths
  1. The direct comparison with FlowEdit is well controlled. The editing-strength sweep is useful, and the gains come with little extra runtime.

  2. The video editing improvement is encouraging. The component ablations also support the use of spatial modulation to preserve the background.

Critical weaknesses
  1. I could not follow the procedure for object addition and deletion. Appendix C can return a null source or target anchor. Eq. (6) then has no source tokens to obtain the reference from, or Eq. (7) has no target column to modify. There is no fallback in Algorithm 1. How were these cases handled in the reported experiments?
Questions for authors

How are null anchors handled? For these cases, is ART still applied and does it improve over the base editor?

1 comment

Q1. How are null anchors handled? Is ART still applied?

We thank the reviewer for pointing out this gap in the presentation. ART requires both a non-empty source token set Tsrc\mathcal{T}_{src} and a non-empty target token set Ttar\mathcal{T}_{tar}, because it transfers a spatial prior from the source object tokens to the target editing tokens. For addition, Tsrc=∅\mathcal{T}_{src}=\emptyset (the new content is absent from the source prompt); for deletion, Ttar=∅\mathcal{T}_{tar}=\emptyset (the target prompt does not name the removed object). In both cases ART is skipped by setting M^tar=Mtar\hat{M}^{tar}=M^{tar}, and only SAE is applied. The reference is read from the target branch itself,Rj=max⁡i∈TrefMj,itarR_j = \max_{i \in \mathcal{T}_{\mathrm{ref}}} M^{tar}_{j,i}with Tref=Ttar\mathcal{T}_{\mathrm{ref}}=\mathcal{T}_{tar} for addition and Tref=Ttarent\mathcal{T}_{\mathrm{ref}}=\mathcal{T}^{\mathrm{ent}}_{tar} (all entity tokens of the target prompt) for deletion. We have made this explicit in Sec. 4.3 and added Appendix D ("Additional Discussion on Different Edit Types"), which discusses the two one-sided cases and lists the extracted SRC_WORD / TAR_WORD for representative examples in Table 6.

Q2. Does the SAE-only path improve over the base editor?

Video (FiVE-Bench [1]). Tables R1 and R2 report LPIPS and FiVE-Acc on the Add and Delete categories. On Add, the SAE-only path improves both structure and editing: LPIPS decreases from 101.60 (Wan-Edit [1]) to 93.40, and FiVE-Acc increases from 77.78 to 83.30. On Delete, both Wan-Edit and our method remain near zero (0.00 vs. 5.00 FiVE-Acc), reflecting the intrinsic difficulty of deletion on FiVE-Bench. The radar plots in Figure 8 and the full per-category results in Table 10 show the same trend.

Table R1. LPIPS on the Add and Delete categories of FiVE-Bench (lower is better).

CategoryWan-EditWan-Edit + Ours
Add object101.6093.40
Delete object79.4075.00

Table R2. FiVE-Acc on the Add and Delete categories of FiVE-Bench (higher is better).

CategoryWan-EditWan-Edit + Ours
Add object77.7883.30
Delete object0.005.00

Image (PIE-Bench [2]). Tables R3 and R4 report SSIM and CLIP-Whole on the Add and Delete categories, comparing FlowEdit [3] and our method. The SAE-only path improves CLIP-Whole for both categories, while SSIM is lower; Delete shows the larger structure cost. The complete per-category results are reported in Table 12.

Table R3. SSIM on the Add and Delete categories of PIE-Bench (higher is better).

CategoryFlowEditFlowEdit + Ours
Add0.8640.847
Delete0.8100.764

Table R4. CLIP-Whole on the Add and Delete categories of PIE-Bench (higher is better).

CategoryFlowEditFlowEdit + Ours
Add27.3727.96
Delete25.1725.69

[1] FiVE-Bench: A Fine-Grained Video Editing Benchmark for Evaluating Emerging Diffusion and Rectified Flow Models

[2] PnP Inversion: Boosting Diffusion-based Editing with 3 Lines of Code

[3] FlowEdit: Inversion-Free Text-Based Editing Using Pre-Trained Flow Models

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.