acceptodds
Under review as a conference paper at ICLR 2027

KriDrag: Spatially Autocorrelated Conflict-Free Drag-Based Image Editing

Abstract

Drag-based image editing provides an intuitive interface for spatial image editing through a small number of user-specified point movements. Existing methods primarily rely on interpolation to construct the displacement field, paying insufficient attention to the role of explicit spatial modeling. Moreover, the displacement fields induced by multiple drags are directly aggregated, often overlooking their mutual conflicts. In this work, we introduce KriDrag, a spatially autocorrelated and conflict-free framework for drag-based image editing. KriDrag constructs a Covariance-Guided Displacement Field, where each displacement influences its surrounding positions through spatial autocorrelation. For multiple drags, KriDrag constructs a continuous and smooth displacement field that preserves coherent motion across neighboring regions while suppressing conflicting displacements between drags. Experiments on DragBench and ReD-Bench demonstrate the efficiency and effectiveness of our method for drag-based image editing, particularly on multi-drag benchmarks.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Drag-based Image Editing, Training-Free Image Editing, Multi-handle Control

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_ti5YOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes KriDrag, a DBIE framework using spatially correlated displacement fields. It identifies the problem of conflicting movements in multi-point dragging and introduces Gaussian covariance to model the spatial influence of each drag point. It combines covariance-based weights with inverse-distance weighting to construct a smooth displacement field for latent warping. Experiments on DragBench and ReD-Bench, including the constructed multi-drag settings, demonstrate improved editing quality and efficiency.

Critical strengths
  1. The paper has a clear motivation and is also well written. The method is easy to follow, and the experiments are comprehensive, including ablations, and analyses of different covariance kernels.

  2. The use of Gaussian covariance to model the spatial influence of each drag point is reasonable. Combining it with inverse-distance weighting helps produce smoother displacement fields and reduce interference between multiple drags. Experiments prove its effectiveness.

Critical weaknesses

The proposed method mainly introduces Gaussian-based spatial weighting into the displacement field, while reusing inverse-distance weighting and the existing latent warping pipeline. The technical contribution seems incremental, and the use of Gaussian Process modeling does not introduce a fundamentally different approach.

Questions for authors
  1. How does KriDrag perform in extreme cases, such as when two nearby drag points have large displacements in opposite directions?

  2. Why does the method model each drag point independently using a Gaussian Process, rather than jointly modeling the spatial relationships between multiple drag points?

2 comments

Rebuttal by Authors [1/2]

We thank you for recognizing the clear motivation, smooth displacement modeling, and comprehensive evaluation of our work. Below we address your comments with exact references to the revised manuscript and Appendix.

[W1] Technical contribution

[W1-1] Controllable per-handle spatial response

We agree that Gaussian kernels, inverse-distance weighting (IDW), and the latent relocation pipeline build upon established work. Our specific contribution is constructing a controllable per-handle spatial response within this pipeline, as stated in the manuscript:

  • Lines 310–312 (Section 4.1) state: "Our implementation follows the latent-space relocation and reconstruction pipeline of FastDrag (Zhao et al., 2024). The only difference lies in the construction of the dense displacement field."
  • Lines 170–173 (Section 3.2) state: "Unlike FastDrag, which uses a geometric construction based on the enclosing mask circle (Zhao et al., 2024), we construct a Covariance-Guided Displacement Field (CGDF) centered at each handle, with a length scale controlling its spatial extent."
  • Eqs. (2)–(4) (Lines 178, 193, 204) & Algorithm 1 (Lines 248–250) show how αi(p)\alpha_i(p) and wi(p)w_i(p) interact: while normalized IDW weights always sum to one (∑iwi(p)=1\sum_i w_i(p)=1) even far from all handles, scaling wi(p)w_i(p) by αi(p)\alpha_i(p) yields v(p)=s(p)∑iβi(p)div(p)=s(p)\sum_i \beta_i(p)d_i, where "The coefficient s(p)s(p) controls the attenuation of the local displacement mixture and decreases when all handles have weak influence" (Lines 195–196).

[W1-2] Controlled empirical evidence

Three controlled comparisons substantiate this spatial profile beyond uniform motion dampening:

  1. Iso-fidelity ablation (Section 4.3, Table 6, Lines 450–457 & 466–470): At matched average fidelity (IF≈0.894\text{IF} \approx 0.894) on the 97 DragBench multi-handle cases, Gaussian attenuation (l=12l=12) achieves lower MD (25.76) and DAI10\text{DAI}_{10} (0.1437) than both Constant scaling (αi=1,g≈0.733\alpha_i=1, g \approx 0.733: MD=26.74,DAI10=0.2209\text{MD}=26.74, \text{DAI}_{10}=0.2209) and FastDrag’s geometric response (g≈1.073g \approx 1.073: MD=29.33,DAI10=0.1641\text{MD}=29.33, \text{DAI}_{10}=0.1641).
  2. Local vs. Global aggregation (Appendix C.1, Table S1): Local aggregation reduces MD from 31.0331.03 to 24.5024.50 on DragBench and from 28.7628.76 to 26.2426.24 on ReD-Bench compared to global aggregation.
  3. Independent vs. Joint GP control (Appendix C.4, Lines 948–958 & Table S5, Lines 1088–1097): Independent responses outperform a jointly conditioned one-hot GP control under identical IDW composition.

[Q1] Performance in extreme cases (nearby handles with large opposing displacements)

[Q1-1] Direct answer on close-range opposing pulls and failure boundaries (Appendix C.7, Figure S5)

When two nearby handles request large opposing displacements, their local Gaussian responses overlap and partially cancel out, so KriDrag can under-edit. This boundary—along with texture stretching and rigid-object rotation—is explicitly analyzed in Appendix C.7:

  • Lines 1172–1176 (Ultra-close opposing pulls): "Row 2 (bouquet) represents an extreme opposing configuration where two handles are only 29.2 pixels apart but request large displacements (117.8 and 129.5 pixels in the opposite direction). Because the handles are closely spaced, their local influence fields still overlap and partially cancel out, resulting in under-expansion albeit avoiding the severe structural tearing of GeoDrag."
  • Lines 1169–1172 (Fine-detail degradation): "In Row 1 (penguins), outward local adjustments blur fine details and leg structures, though avoiding the complete object erasure seen in DragDiffusion."
  • Lines 1176–1180 (Absence of rigidity priors): "Row 3 (pencil) highlights the absence of explicit rigidity priors: although the active drags approximate a rigid rotation, local displacement modeling bends and fragments the pencil shaft, whereas baselines preserve shaft continuity only by failing to move the pencil toward the targets."

[Q1-2] Evaluation across directional regimes when handles are separated (Section 4.4, Table 7 & Figure 7)

When handles have sufficient spatial separation, spatial attenuation effectively mitigates cross-drag interference across differing directions (Section 4.4, Lines 474–485 & Table 7, Lines 486–498):

  • Lines 474–480 state: "By resolving spatial influence locally before combining displacements, KriDrag mitigates cross-drag interference and reduces MD by 13.7% over the strongest baseline for angles >45∘>45^\circ [24.64 vs. 28.54]. In the near-opposing subset (>135∘>135^\circ [43 cases]), KriDrag achieves a low mean MD comparable to GeoDrag (24.31 vs. 24.39, while outperforming all other baselines at 26.78–37.89) and maintains higher image fidelity (IF=0.896\text{IF}=0.896 vs. 0.855), whereas GeoDrag achieves lower DAI at the cost of structural degradation and DragNeXt under-edits."

Rebuttal by Authors [2/2]

[Q2] Why model each drag point independently rather than jointly via a GP?

[Q2-1] Independent response construction with coupled composition

While each handle’s spatial response is evaluated independently, multi-handle spatial relationships are coupled when composing the final displacement field:

  • Eq. (2) (Line 178) & Appendix B.2 (Eqs. S3–S5, Lines 850–859) show that independent evaluation yields 0<αi(p)≤11+ϵ≤10 < \alpha_i(p) \le \frac{1}{1+\epsilon} \le 1 and ∂αi(p)∂si≤0\frac{\partial \alpha_i(p)}{\partial s_i} \le 0, so "the spatial influence decreases monotonically with the distance from the corresponding handle point" (Lines 858–859).
  • Eqs. (3)–(4) (Lines 191, 205) couple all handles via normalized IDW weights wi(p)≥0,∑iwi(p)=1w_i(p) \ge 0, \sum_i w_i(p)=1, which allows nearby compatible drags to contribute jointly to the same region and bounds the composed field magnitude by ∥v(p)∥=∥∑iwi(p)αi(p)di∥≤max⁡i∥di∥\|v(p)\| = \|\sum_i w_i(p)\alpha_i(p)d_i\| \le \max_i \|d_i\|.
  • Lines 951–955 (Appendix C.4) describe the joint control: "Unlike our approach, which evaluates each handle's field independently, the joint variant solves for all Gaussian fields simultaneously to enforce a response of roughly 1 at each target handle and 0 at all others. All other configurations—including IDW composition, length scale (l=12l=12), inputs, and generation settings—remain identical." Enforcing zero response at other handles can suppress useful overlap between nearby handles.

[Q2-2] Paired experimental verification (Appendix C.4, Table S5, Lines 1088–1097)

Across all 97 DragBench and 91 ReD-Bench multi-handle cases (Table S5):

Dataset (Table S5, Lines 1088–1097)MethodMD ↓IF ↑DAI1 ↓DAI10 ↓DAI20 ↓
DragBench (97)Ours (Independent)24.680.89400.15800.14380.1339
DragBench (97)Joint GP + IDW31.550.89780.20640.17600.1540
ReD-Bench (91)Ours (Independent)24.860.86020.19700.16600.1621
ReD-Bench (91)Joint GP + IDW35.330.87670.23650.20180.2079

As stated in Lines 955–957, "independent responses consistently achieve better editing accuracy (lower MD and DAI) on both benchmarks, whereas the joint variant yields only a bit higher image fidelity (IF)." Table S5 supports our design under this one-hot conditioned control and shared IDW composition, without ruling out other joint GP formulations.

Reviewer_KY32OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents KriDrag, a training-free method for drag-based image editing that focuses on composing multiple user-specified motions. Given an image, handle–target pairs, and an editing mask, the method constructs a Gaussian spatial response around each handle and combines the resulting displacement fields using inverse-distance weights. The combined field is integrated into an existing latent relocation and diffusion reconstruction pipeline. Experiments on DragBench and ReD-Bench, including their multi-handle subsets, report lower MD and DAI scores than the evaluated baselines, together with low reported inference cost.

Critical strengths
  • The proposed displacement field has a compact formulation and can be integrated into an existing editing pipeline without additional training or iterative latent optimization. The range of each handle’s influence is controlled by an interpretable Gaussian length scale. The reported results in Table 1 suggest a useful balance between MD/DAI performance, runtime, and memory consumption.

  • Tables 1 and 2 report better MD and DAI scores than the baselines evaluated on each benchmark and its multi-handle subset. The method also improves image fidelity relative to FastDrag and GeoDrag, making the gains more informative than improvements in localization alone. The selected qualitative examples provide supporting evidence across different image contents and editing operations.

Critical weaknesses
  • The paper motivates interference between drags, but it does not specify a criterion for deciding when an edit is conflict-free. The Limitations paragraph at the end of Section 5 appropriately acknowledges that conflicts are alleviated rather than fully resolved, which is weaker than the “conflict-free” wording in the title, the abstract, and the opening sentence of Section 5. Figure 3 and Table S1 provide relevant evidence for the field construction and local aggregation, but the evaluation does not separately quantify performance as drag directions, handle separation, or overlap between influence regions vary. Narrowing the “conflict-free” wording and, if feasible, adding controlled cases with well-defined intended edits would clarify the method’s scope.

  • Table 3 shows that the wiw_i-only variant has better MD and DAI but lower IF. The authors describe its Figure 4 results as moving the entire selected object, suggesting that better metrics do not necessarily mean the intended edit was achieved. Table 6 provides useful support for the Gaussian response at similar average IF, but these metrics still do not directly measure whether the result matches the user’s intent. A small evaluation with clearly defined editing goals would strengthen this aspect of the paper.

  • Section 4.1 provides the general experimental settings, but the paper does not explain how the configurations with comparable average IF in Table 6 were obtained. It is unclear whether the different Gaussian MD relative to the main multi-handle results reflects changed settings or evaluation details. In Appendix C.2, the hard-partition control combines nearest-handle assignment with regional length-scale recalibration. Explaining how Table 6 was constructed, including any changes from Section 4.1, would improve reproducibility; a fixed-response comparison could additionally clarify the assignment rule’s contribution.

  • KriDrag’s spatially decaying weights can be viewed as a way to define each drag’s spatial influence and reduce interference between multiple drags. This motivation is closely related to DragNeXt, which discusses ambiguity in the regions and types of drag operations and uses explicitly specified regions and geometric transformations to guide latent optimization. Although the authors introduce DragNeXt in the manuscript, but a direct experimental comparison is missing. Could the authors include DragNeXt in Tables 1–2 and compare editing results on the examples in Figures 1, 3, S1, and S2. Such comparisons would clarify the advantages and trade-offs of spatially decaying influence weights relative to explicit region-guided geometric transformations, particularly in reducing interference, preserving editing quality, and limiting user effort.

    [a] Dragnext: Rethinking drag-based image editing, AAAI 2026.

Questions for authors
  • How do the authors define and measure a conflict between drag instructions, and which conditions are covered by the “conflict-free” description? Could they provide a small controlled analysis varying handle separation and motion direction for cases with clearly specified intended deformations? Representative successful and unsuccessful cases under these conditions would help establish the limits of the claimed conflict reduction.

  • Could the authors include DragNeXt in Tables 1–2 and compare editing results on the examples in Figures 1, 3, S1, and S2, and more clearly highlight the advantage of the proposed approach over this relevant prior work.

  • Compared with DragNeXt’s explicit specification of the region affected by each drag operation, what distinct advantages does KriDrag’s soft spatial decay offer? Please discuss this comparison in more detail in the paper to clarify the strengths and distinguishing features of the proposed approac

4 comments

Rebuttal by Authors [1/2]

Thank you for your positive assessment and constructive suggestions. Below we address overlapping questions and weaknesses with exact references to the revised paper.

[Q1/W1] Definition of interference and controlled analysis

[Q1-1/W1-1] Scope and evaluation criterion

We agree “conflict-free” was too strong; we removed it across the Title, Abstract (Lines 036–038), Intro (Lines 066–078), and Conclusion (Lines 537–539), and defined interference:

  • Lines 067–073 (Intro): "Combining all drag directions into a single overall motion can cause spatially separated drags with different directions to interfere with each other—specifically, the motion propagation from one drag unintentionally pulls or distorts content targeted by another drag or surrounding regions that should remain static, even though each drag should primarily affect its own local region. The angle between drag directions, the separation between handles, and the overlap of their influence regions characterize the configurations in which such interference may arise."
  • Lines 073–076 (Intro): "To address this, the influence of each drag should first be determined according to its spatial relationship with the query location, after which the relevant displacements can be combined locally. This mitigates cross-region interference so that spatially separated drags remain largely independent, while nearby compatible drags can contribute jointly."
  • Given an editing goal, we evaluate both target arrival and untargeted/static region preservation (Lines 371–377, 968–970, 1108–1115).

[Q1-2/W1-2] Directional regimes (Sec. 4.4, Table 7 & Fig. 7)

Sec. 4.4 & Table 7 (Lines 472–498) evaluate angle-conditioned subsets across all 188 multi-handle cases (97 DragBench, 91 ReD-Bench):

  • Lines 474–480: "By resolving spatial influence locally before combining displacements, KriDrag mitigates cross-drag interference and reduces MD by 13.7% over the strongest baseline for angles >45∘>45^\circ [24.64 vs. 28.54, with lowest DAI10=0.179\text{DAI}_{10}=0.179]. In the near-opposing subset (>135∘>135^\circ [43 cases]), KriDrag achieves a low mean MD comparable to GeoDrag (24.31 vs. 24.39, while outperforming all other baselines at 26.78–37.89) and maintains higher image fidelity (IF=0.896\text{IF}=0.896 vs. 0.855), whereas GeoDrag achieves lower DAI at the cost of structural degradation and DragNeXt under-edits."

[Q1-3/W1-3] Close-handle pulls & failure boundaries (App. C.7, Fig. S5)

As angle thresholds alone do not restrict separation or rigidity, App. C.7 details three failure modes:

  • Lines 1169–1172 (Fine-detail degradation): "In Row 1 (penguins), outward local adjustments blur fine details and leg structures, though avoiding the complete object erasure seen in DragDiffusion."
  • Lines 1173–1176 (Ultra-close opposing pulls): "Row 2 (bouquet) represents an extreme opposing configuration where two handles are only 29.2 pixels apart but request large displacements (117.8 and 129.5 pixels in the opposite direction). Because the handles are closely spaced, their local influence fields still overlap and partially cancel out, resulting in under-expansion albeit avoiding the severe structural tearing of GeoDrag."
  • Lines 1176–1180 (Absence of rigidity priors): "Row 3 (pencil) highlights the absence of explicit rigidity priors: although the active drags approximate a rigid rotation, local displacement modeling bends and fragments the pencil shaft, whereas baselines preserve shaft continuity only by failing to move the pencil toward the targets."

[W2] Evaluation with defined goals (App. C.6, Fig. S4)

In Sec. 4.3 (Lines 371–377) and App. C.6, we state:

  • Lines 969–970 & 1098–1110: "While broad regional translation is appropriate when moving an entire object, MD and DAI only evaluate target localization and target-patch agreement without penalizing unintended motion spillover into untargeted static structures within the mask."
  • Fig. S4 compares input, w-only, and KriDrag across five cases with explicit goals:
Col.Case & TypeExplicit Goal & w-only Spillover
Col. 1I119 (Carafe neck)Shorten neck, preserve lower body/glass; w-only "distorts the lower vessel and nearby glass".
Col. 2I106 (Plastic bags)Move bags right, retain hand; w-only "erases the foreground hand".
Col. 3WINDOW (Window)Widen upper opening, preserve lower frame; w-only "also stretches the lower opening".
Col. 4DB155 (Sofa)Raise left backrest, preserve right backrest; w-only "simultaneously lifts the right backrest".
Col. 5INTERIOR (Room)Move desk right & bin left, keep shelf; w-only "distort[s] the desk and bin geometry".

Rebuttal by Authors [2/2]

[W3] Table 6 construction and Soft vs. Hard assignment ablation

[W3-1 & W3-2] Matching average IF in Table 6 and relation to Table 2

  • Lines 466–470 (Section 4.3) state: "To compare spatial responses at matched average fidelity (IF≈0.894\text{IF} \approx 0.894) on the 97 multi-handle DragBench cases (Table 6), we sweep and linearly interpolate a global gain gg using average IF alone (g≈0.733g \approx 0.733 for Constant; g≈1.073g \approx 1.073 for FastDrag). All three variants are evaluated under the same protocol and sample order in an independent 97-case run, where KriDrag achieves the lowest MD and DAI across all scales."
  • Why Gaussian MD in Table 6 (25.7625.76) differs from Table 2 (24.5024.50): Table 6 (Lines 450–457) was evaluated in a standalone run over the 97 multi-handle cases (Line 470), whereas Table 2 (Lines 341–354) reports the 97 multi-handle subset extracted from the 205-case full-benchmark run (Lines 292–294). Model settings are identical, but because seed_everything(42) is initialized once before traversing each manifest, traversing 97 cases directly vs. 205 cases produces different random-number sequences during DIFT feature extraction.

[W3-3] Fixed-response Soft vs. Hard assignment (Appendix C.2, Table S2)

  • Lines 905–913 & Table S2 (Lines 918–924) isolate the assignment rule with fixed Gaussian responses (l=12l=12, without length-scale recalibration) alongside the recalibrated control in Table S3 (Lines 926–934):
Strategy (Fixed l=12l=12, Table S2)MD ↓IF ↑DAI1 ↓DAI10 ↓DAI20 ↓
Soft assignment (KriDrag)24.390.89400.15810.14370.1338
Hard assignment (w/o recalibration)22.860.89150.20350.18280.1546

As stated in Lines 908–912, "hard assignment yields a slightly lower mean MD (22.86 vs. 24.39, where the paired confidence interval for the MD difference includes zero), but lowers IF (0.8915 vs. 0.8940) and substantially degrades target-patch consistency across all scales (DAI1,DAI10\text{DAI}_1, \text{DAI}_{10}, and DAI20\text{DAI}_{20} worsen by 28.7%, 27.2%, and 15.6%, respectively)."

[Q2/Q3/W4] Comparison with DragNeXt

[Q2-1/W4-1] Added comparisons in Tables 1, 2, 7 & Figures 1, 7, S1–S3, S5

Benchmark (Tables 1 & 2, Lines 324–354)MethodMD ↓DAI1 ↓DAI10 ↓DAI20 ↓IF ↑Time (s)Mem (GB)
DragBench Full (205, Table 1)DragNeXt42.260.3010.2870.2640.93375.6018.43
DragBench Full (205, Table 1)KriDrag23.230.1390.1140.1060.8772.052.59
ReD-Bench Full (120, Table 1)DragNeXt42.820.3520.3140.2810.92570.0417.22
ReD-Bench Full (120, Table 1)KriDrag26.980.1750.1490.1470.8712.232.81
DragBench Multi (97, Table 2)DragNeXt38.900.2990.2880.2530.948——
DragBench Multi (97, Table 2)KriDrag24.500.1580.1440.1340.894——
ReD-Bench Multi (91, Table 2)DragNeXt42.330.3620.3180.2820.924——
ReD-Bench Multi (91, Table 2)KriDrag26.240.1970.1660.1620.860——

[Q2-2/Q3/W4-2/W4-3] Adaptation scope & trade-offs

  • Adaptation protocol (Figure S5 Caption, Lines 1180–1185): Because standard point benchmarks only provide point pairs and a shared mask, our evaluation adapts DragNeXt using "a shared-mask translation adapter" with benchmark text prompts (and boundary repairs on 6 ReD-Bench cases). These results characterize performance under the shared-mask point benchmark protocol rather than DragNeXt’s full capabilities with manually tailored per-operation regions and transformation modes (Section 4.2, Lines 359–362).
  • Trade-offs (Section 2, Lines 126–127 & Appendix A, Lines 825–830): "While explicit region masks and transformation modes better enforce coherent rigid-body movements when manually specified, KriDrag's soft spatial decay avoids manual region partitioning and iterative optimization, reducing inference time by over 90% and GPU memory by over 80% under the standard point-based interface."

Thanks for the authors' response. My main concerns have been addressed and I will keep my positive score.

Thank you for your careful consideration of our response and for maintaining your positive assessment. We are glad that the rebutal have addressed your main concerns.

Thank you for your suggestions for improving our paper. We will incorporate these changes into the final proof.

Reviewer_qtnAOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces KriDrag, a spatially autocorrelated, conflict-free drag editing framework. The core motivation is to construct dense displacement fields by modeling spatial dependencies via covariance-controlled spatial influence. Experiments indicate that KriDrag can perform precise object-level spatial transformations, demonstrating the potential importance of explicitly modeling spatial correlations in drag-based image editing.

Critical strengths
  1. KriDrag is a simple and efficient framework that integrates a lightweight displacement-field construction, and it doesn't require per-image iterative optimization.

  2. The experimental results show consistent improvements over the evaluated baselines across both DragBench and ReD-Bench.

Critical weaknesses
  1. The novelty is limited. The proposed field essentially combines Gaussian distance attenuation with FastDrag's inverse-distance weighting, and the overall pipeline is the same.

  2. Some claims are made without sufficient experimental support. For example, while the authors claim the framework is conflict-free, the experiments do not directly evaluate whether conflicting motions are effectively resolved. Furthermore, multiple claims made in the introduction lack experiments for support.

  3. In the ablation study (Table 3), removing the Gaussian responses actually improves performance. Although the authors attribute this improvement to undesirable region translation, the reported metrics do not systematically substantiate this interpretation.

Questions for authors

See weaknesses.

3 comments

Rebuttal by Authors [1/2]

We thank you for your constructive feedback. In the revised manuscript, we removed “conflict-free” phrasing across the Title, Abstract, Introduction, and Conclusion, narrowing our claim to mitigating cross-region interference, and added quantitative and qualitative analyses to substantiate our spatial responses.

[W1] Contribution beyond FastDrag

[W1-1] Controllable per-handle spatial influence

We agree that Gaussian kernels and inverse-distance weighting (IDW) are established tools. Our contribution is constructing a controllable spatial response for each drag within the existing latent relocation pipeline. This distinction is explicitly stated in the paper:

  • Lines 310–314 (Section 4.1) state: "Our implementation follows the latent-space relocation and reconstruction pipeline of FastDrag (Zhao et al., 2024). The only difference lies in the construction of the dense displacement field."
  • Lines 170–173 (Section 3.2) state: "Unlike FastDrag, which uses a geometric construction based on the enclosing mask circle (Zhao et al., 2024), we construct a Covariance-Guided Displacement Field (CGDF) centered at each handle, with a length scale controlling its spatial extent."
  • Eq. (2) (Line 178) & Lines 182–185 define the per-handle Gaussian response αi(p)=kl(p,hi)1+ϵ\alpha_i(p) = \frac{k_l(p, h_i)}{1+\epsilon}, where "A smaller ll concentrates motion near the handle, whereas a larger ll spreads it farther. The radial response is orientation-agnostic and decays smoothly without a hard cutoff, independently of the enclosing mask circle."
  • Eqs. (3)–(4) (Lines 193, 204), Lines 194–207, & Algorithm 1 (Lines 248–250) make the composition clear: while normalized IDW weights always sum to one (∑iwi(p)=1\sum_i w_i(p)=1) even far from all handles, modulating wi(p)w_i(p) by αi(p)\alpha_i(p) as ri(p)=wi(p)αi(p)r_i(p)=w_i(p)\alpha_i(p) yields v(p)=s(p)∑iβi(p)div(p)=s(p)\sum_i \beta_i(p)d_i, where "The coefficient s(p)s(p) controls the attenuation of the local displacement mixture and decreases when all handles have weak influence" (Lines 195–196), while "βi(p)\beta_i(p) determines the relative contribution of each handle" (Line 206).

[W1-2] Controlled empirical evidence

Three analyses in the manuscript substantiate this spatial profile beyond uniform motion scaling:

  1. 3D field visualization (Figure 3, Lines 216–239): Shows that "our approach generates smoother spatial variations without abrupt changes around the drag points" (Lines 233–235), whereas FastDrag "exhibits a pronounced peak around each user-provided drag point" (Lines 278–280).
  2. Iso-fidelity ablation (Section 4.3, Table 6, Lines 450–457 & 466–470): Compares Gaussian attenuation (l=12l=12) against Constant scaling (αi=1,g≈0.733\alpha_i=1, g \approx 0.733) and FastDrag’s geometric response (g≈1.073g \approx 1.073) at matched average fidelity (IF≈0.894\text{IF} \approx 0.894) on the 97 multi-handle DragBench cases:
Spatial Response (Table 6)IF ↑MD ↓DAI1 ↓DAI10 ↓DAI20 ↓
Gaussian (KriDrag, l=12l=12)0.8939825.760.15810.14370.1338
Constant (αi=1,g≈0.733\alpha_i=1, g \approx 0.733)0.8943426.740.25610.22090.1887
FastDrag Geometric (g≈1.073g \approx 1.073)0.8942529.330.18560.16410.1496
  1. Independent vs. Joint GP ablation (Appendix C.4, Lines 950–959 & Table S5, Lines 1088–1097): Under identical IDW composition, independent responses reduce MD from 31.5531.55 to 24.6824.68 on DragBench and from 35.3335.33 to 24.8624.86 on ReD-Bench compared to a jointly conditioned one-hot GP control.

Rebuttal by Authors [2/2]

[W2] Interference definition and supporting evidence

[W2-1] Scope and definition

We agree “conflict-free” overstates the guarantee. We removed “conflict-free” phrasing throughout the manuscript and explicitly defined spatial interference and our scope:

  • Lines 037–039 (Abstract) : "we introduce KriDrag, a spatially autocorrelated framework for drag-based image editing that can tame spatial interference in multi-handle image editing."
  • Lines 067–073 (Introduction) define interference: "Combining all drag directions into a single overall motion can cause spatially separated drags with different directions to interfere with each other—specifically, the motion propagation from one drag unintentionally pulls or distorts content targeted by another drag or surrounding regions that should remain static, even though each drag should primarily affect its own local region. The angle between drag directions, the separation between handles, and the overlap of their influence regions characterize the configurations in which such interference may arise."
  • Lines 073–076 (Introduction) our mechanism: "the influence of each drag should first be determined according to its spatial relationship with the query location, after which the relevant displacements can be combined locally. This mitigates cross-region interference so that spatially separated drags remain largely independent, while nearby compatible drags can contribute jointly."
  • Lines 537–539 (Section 5) & Lines 1108-1120 (Appendix C.7) explicitly state the boundary: "while local spatial responses mitigate cross-region interference, they do not guarantee fine-detail preservation under ultra-close opposing pulls or enforce global rigid-body constraints."

[W2-2] Evaluation across directional regimes (Section 4.4, Table 7 & Figure 7)

Lines 474–485 and Table 7 (Lines 486–498) evaluate all 188 multi-handle cases across DragBench (97) and ReD-Bench (91) grouped by maximum pairwise drag angle:

Method (Table 7)>45°: MD↓ / IF↑>45°: DAI 1/10/20↓>135°: MD↓ / IF↑>135°: DAI 1/10/20↓
DragDiffusion35.44 / 0.8990.274 / 0.260 / 0.23037.89 / 0.8960.300 / 0.273 / 0.247
DragNoise31.12 / 0.8940.260 / 0.227 / 0.19831.93 / 0.8930.283 / 0.239 / 0.215
FreeDrag28.54‾\underline{28.54} / 0.917‾\underline{0.917}0.300 / 0.262 / 0.22426.78 / 0.915‾\underline{0.915}0.314 / 0.271 / 0.234
FastDrag31.35 / 0.8870.266 / 0.227 / 0.20732.59 / 0.8920.322 / 0.248 / 0.231
GeoDrag28.74 / 0.8600.194 / 0.180‾\underline{0.180} / 0.16224.39‾\underline{24.39} / 0.8550.199 / 0.172 / 0.164
DragNeXt36.70 / 0.9460.326 / 0.298 / 0.26435.00 / 0.9470.338 / 0.299 / 0.274
KriDrag (Ours)24.64 / 0.8950.209‾\underline{0.209} / 0.179 / 0.169‾\underline{0.169}24.31 / 0.8960.245‾\underline{0.245} / 0.206‾\underline{0.206} / 0.194‾\underline{0.194}

As stated in Lines 475–480: "KriDrag mitigates cross-drag interference and reduces MD by 13.7% over the strongest baseline for angles >45∘>45^\circ. In the near-opposing subset (>135∘>135^\circ), KriDrag achieves a low mean MD comparable to GeoDrag (24.31 vs. 24.39, while outperforming all other baselines at 26.78–37.89) and maintains higher image fidelity (IF=0.896\text{IF}=0.896 vs. 0.855), whereas GeoDrag achieves lower DAI at the cost of structural degradation and DragNeXt under-edits."

[W3] Gaussian ablation and regional translation

For a single handle (K=1K=1), Eq. (3) (Line 191) gives w1(p)≡1w_1(p) \equiv 1, so removing α1\alpha_1 reduces v(p)v(p) to a constant shift v(p)=t1−h1v(p)=t_1-h_1 across Ω\Omega:

  • Lines 371–377 (Section 4.3) : "Although removing the spatial responses (wiw_i only) yields lower MD and DAI in Table 3, Figure 4 shows that it shifts the entire masked region together rather than localizing the deformation. While uniform regional translation is a valid editing operation when moving an entire object, for localized edits it causes unintended motion spillover into untargeted structures within the mask. Because MD measures target-point localization and DAI measures target-patch agreement without penalizing spillover into surrounding static regions, broad shifts can lower target error while degrading source preservation (IF)."
  • Lines 1137–1160 (Appendix C.6, Figure S4) verify this across five cases with clear preservation goals: "Without Gaussian spatial attenuation, w-only exhibits severe motion spillover across all five cases: shortening the carafe neck distorts the lower vessel and nearby glass (Col. 1); shifting the bags rightward erases the foreground hand (Col. 2); widening the upper window also stretches the lower opening (Col. 3); raising the left sofa backrest simultaneously lifts the right backrest (Col. 4); and opposing drags in the interior scene distort the desk and bin geometry (Col. 5)."

Dear Reviewer qtnA,

Thank you for your time and effort in reviewing our submission. We have carefully considered your comments and provided detailed responses. We look forward to your feedback.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Global Response

We thank the AC and all reviewers (KY32, qtnA, and ti5Y) for their constructive feedback. We respectfully highlight the post-rebuttal discussion outcomes:

  • We recalibrated the our claims and evaluated directional conflicts and failure boundaries (raised by qtnA [W2], KY32 [W1, Q1], and ti5Y [Q1]) by removing conflict-free in the title and across the manuscript, explicitly defining spatial interference (Lines 067–076), and adding angle-stratified evaluations on all 188 multi-handle cases across DragBench and ReD-Bench (Section 4.4, Table 7, Figure 7). For drag angles >45°, KriDrag achieves the lowest MD of 24.64, outperforming the strongest baseline (28.54) by 13.7%, alongside the best DAI-10 (0.179). In the near-opposing subset (>135°, 43 cases), KriDrag achieves a low mean MD comparable to GeoDrag (24.31 vs. 24.39, outperforming all other baselines at 26.78–37.89) while maintaining substantially higher image fidelity (IF = 0.896 vs. 0.855), whereas GeoDrag achieves lower DAI at the cost of object erasure and structural artifacts. We also added explicit failure boundary analyses in Appendix C.7 (Figure S5) covering ultra-close opposing pulls (29.2 px handle separation with 117.8 and 129.5 px opposing displacements), fine-detail degradation on complex textures, and rigid-body rotations lacking explicit rigidity priors.

  • We substantiated controllable per-handle spatial response and independent GP modeling (raised by qtnA [W1], ti5Y [W1, Q2], and KY32 [W3]) by clarifying how independent Gaussian responses (Section 3.2, Eq. 2) allow total motion strength to decay at weakly supported locations while coupling handles through normalized IDW weights (Section 3.3, Eqs. 3–4), and by adding paired ablations within the fixed relocation pipeline. In Section 4.3 (Table 6), we specified the gain interpolation protocol matching dataset-average fidelity (IF ≈ 0.894) on the 97 multi-handle DragBench cases: Gaussian responses (l = 12) achieve 25.76 MD and 0.1437 DAI-10, outperforming both Constant uniform scaling (g ≈ 0.733, 26.74 MD, 0.2209 DAI-10) and FastDrag’s geometric response (g ≈ 1.073, 29.33 MD, 0.1641 DAI-10). In Appendix C.4 (Table S5), comparing independent responses against a jointly conditioned one-hot GP control under identical IDW composition across all 188 multi-handle cases shows that independent responses reduce MD from 31.55 to 24.68 (21.8%) on DragBench and from 35.33 to 24.86 (29.6%) on ReD-Bench. In Appendix C.2 (Table S2), isolating the assignment rule with fixed Gaussian responses (l = 12, without length-scale recalibration) shows that while nearest-handle hard assignment slightly lowers mean MD (22.86 vs. 24.39, paired CI includes zero), it lowers IF (0.8915 vs. 0.8940) and degrades patch consistency across all scales (DAI-1, DAI-10, and DAI-20 worsen by 28.7%, 27.2%, and 15.6%). We also clarified that the MD difference between the standalone 97-case run in Table 6 (25.76) and the 97-case subset from the 205-case run in Table 2 (24.50) stems from different random-number sequences in DIFT evaluation after a single initial seed_everything(42) call.

  • We clarified w-only ablation trade-off and verified structural preservation intent (raised by qtnA [W3] and KY32 [W2]) by showing analytically that for a single handle, normalized IDW yields w_1(p) = 1 everywhere, reducing w-only to a constant whole-mask translation v(p) = t_1 - h_1 (Section 4.3, Lines 371–377). While uniform translation is valid when moving an entire object, in localized edits it shifts untargeted static structures within the mask—an artifact not penalized by target-local metrics (MD and DAI) but reflected by w-only dropping to the lowest IF in Table 3 (0.826 on Full and 0.872 on Multi-handle DragBench, vs. 0.877 and 0.894 for KriDrag). Side-by-side comparisons across five cases with explicit editing and preservation goals in Appendix C.6 (Figure S4) confirm that w-only distorts untargeted regions (e.g., lifting the untargeted right sofa backrest or erasing a foreground hand), whereas KriDrag confines deformations to the intended local parts.

  • We expanded the evaluation to include DragNeXt (raised by KY32 [W4, Q2, Q3]) by adding quantitative and qualitative comparisons in Tables 1, 2, 7 and Figures 1, 7, S1–S3, S5, and discussing the trade-offs between explicit region specification and soft point-based decay in Sections 2, 4.2, and Appendix A. Evaluated under the shared-mask point-benchmark adaptation (with boundary repairs on 6 ReD-Bench cases), DragNeXt retains high source similarity (IF 0.924–0.948) due to conservative under-editing (MD 38.90–42.82), whereas KriDrag reduces MD to 23.23–26.98 and DAI-10 to 0.114–0.166 while reducing inference time by ~97% (2.05–2.23 s vs. 70.04–75.60 s) and GPU memory by 84%–86% (2.59–2.81 GB vs. 17.22–18.43 GB).

We are grateful to reviewers KY32, qtnA, and ti5Y for their comments, which helped us improve our paper, and we thank the Area Chairs for their consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.