acceptodds
Withdrawn from ICLR 2027

VestEdit: Inversion-Free Video Editing via Spatio-Temporal Transport Stabilization

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Inversion-free flow editing modifies videos directly through a pretrained backbone's velocity field, avoiding costly noise inversion. However, naive frame-wise velocity construction can entangle the spatial support of an edit with motion across video frames. The resulting residual may leak into non-edited regions or follow directions that are inconsistent with source-frame correspondences, causing flicker, boundary deformation, and background drift. We introduce VestEdit, a training-free correction layer inserted between velocity construction and ODE integration. A semantic-topological spatial projection suppresses residuals outside the editable support, while a norm-preserving spectral projection redirects the spatially corrected field toward an attention-induced low-frequency subspace. On the six FiVE tasks, VestEdit achieves a FiVE-Acc of 50.13 and obtains the best reported six-task average text-alignment scores, with CLIP-S of 27.71 and CLIP-E of 27.87, without additional backbone forward evaluations. These results indicate that explicit spatial and spectral conditioning of the editing field can improve the preservation–alignment trade-off of inversion-free video editing. Code will be released.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Text-guided Video Editing, Inversion-free Editing, Training-free Editing, Spatio-temporal Consistency, Velocity Field Correction

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_JsSvOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes VestEdit for inversion-free text-guided video editing. The proposed training-free correction layer is placed between velocity construction and ODE solver. There are two components in this correction layer: First, the spatial projection suppresses residuals outside the estimated editable support. Then, this spatially corrected field is redirected via a norm-preserving spectral projection called kinetic-energy-preserving projection (KE-OPM), to an attention-induced low-frequency subspace. VestEdit is analyzed on FiVE benchmark.

Critical strengths

The motivation is convincing that existing inversion-free methods only consider source-target difference and ignore cross-frame dynamics. Then the method naturally connects with this motivation by proposing spatial correction to suppress residuals and spectral correction to redirect inconsistent high-frequency motion to a low-frequency motion manifold in the temporal direction. The method looks sound.

Critical weaknesses
  1. A main critical weakness is that the submission does not provide any video results even though it is a video editing paper. This is especially important since one of the main advantages of the proposed method is improving temporal aspects of the edited video (e.g,. flickering, motion). It is not possible to judge whether the temporal/motion side has improved without any video results.

  2. The experiment section is weak: (i) The proposed ideas are only validated on a single backbone (Wan 2.1). How does it work on other backbones like LTX or MiniMax H3? (ii) The metrics are not defined or explained clearly. How is motion fidelity measured? What is Dist? (iii) NIQE is reported to assess visual quality (l367), however, the proposed method performs worst in terms of NIQE. (iv) Ablation study (Table 4) shows that not using the proposed spatial projection is better than using it across all metrics. Also, not using the proposed KE-OPM is better in terms of Dist. If these metrics are not reliable, visual comparisons could be provided. Otherwise, there is no empirical evidence supporting the proposed components.

[Not critical but worth noting] The overall presentation of the paper should be improved. For example, related work in the main paper doesn't give enough details. I understand that the authors are limited due to the page limit but some details from the related work in the appendix should be brought forward. I would significantly reduce the preliminaries section which is almost taking a page and bring in more details from Appendix B and C.

[Minor] Fig 5 (a) caption: "retain-support (red)" should be "retain-support (gray)" ?

Questions for authors

Does the proposed method also work on other backbones? Please provide empirical results.

Clarification: How is motion fidelity measured? What is Dist?

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

This paper presents VestEdit, a training-free correction module for inversion-free text-guided video editing. The method combines a semantic-topological spatial constraint with a norm-preserving graph-spectral projection to reduce editing leakage and improve temporal consistency. However, the current ablations do not establish the benefits of the proposed components, and several methodological and evaluation details remain insufficiently specified.

Critical strengths
  1. Clear motivation and coherent design. The paper identifies spatial leakage and cross-frame inconsistency as important failure modes and proposes complementary mechanisms to address them. The overall presentation makes the intended roles of the two components easy to understand.
  2. Practical inference-time formulation. VestEdit operates on a frozen backbone without additional training or backbone forward evaluations. This makes the approach potentially useful as an extension to existing inversion-free editing pipelines.
Critical weaknesses
  1. The ablations do not substantiate the central methodological contribution. In Tab. 4, removing the spatial module improves both Structure Distortion, from 19.65 to 13.47, and Motion Fidelity, from 86.84 to 89.74. Temporal averaging also outperforms the full method on both reported quality metrics. The authors appropriately acknowledge that these metrics may favor conservative edits, but no per-variant editing-success or alignment results are provided to support that explanation. Consequently, the evidence does not establish why the full method should be preferred over its simpler alternatives.
  2. The interaction between spatial anchoring and spectral correction is insufficiently justified. Eq. 9 anchors the field to the source-conditioned velocity outside the editable support. However, Eq. 11 subsequently projects and rescales the entire spatially corrected field, potentially modifying those background velocities again. Although this limitation is acknowledged in the appendix, its practical extent is not quantified, leaving a gap between the intended localization mechanism and the final update.
  3. The stability argument remains heuristic. Smoothness on a graph of video tokens does not directly imply smoothness with respect to ODE integration time or a reduction in the field Jacobian. The paper acknowledges this distinction, but the local truncation error discussion therefore offers limited support for the claimed stabilization mechanism. Likewise, interpreting a fixed low-frequency attention subspace as a source-consistent motion manifold requires more direct empirical validation.
  4. The timing results also require clarification: the full method takes 5.21 s/frame in Tab. 1 but 3.18 s/frame in Tab. 4. Without matched hardware, sampling, and timing settings, the practical overhead and efficiency advantages cannot be assessed reliably.
  5. The evaluation does not directly compare against the closely related inversion-free video editors discussed in the related-work section, limiting the assessment of the method’s incremental contribution.
Questions for authors
  1. Component effectiveness: Are editing-success and text-alignment scores available for the Tab. 4 variants? In particular, does the full method outperform temporal averaging and the variant without spatial correction on these objectives sufficiently to justify its worse reported structure and motion scores?
  2. Actual projection implementation: Is Eq. 11 applied to the complete spatially corrected velocity, as written, or only to the editable residual? Is any masking or background anchoring reapplied afterward? This distinction materially affects the interpretation of the proposed spatial constraint.
  3. Graph construction: How is the temporal attention affinity constructed and symmetrized to obtain the orthogonal eigendecomposition assumed in Eq. 3? Please specify the graph dimensions, attention aggregation, and whether the spectral basis is recomputed at every integration step.
  4. Runtime comparability: What explains the 5.21 versus 3.18 s/frame results for the full method? Under identical hardware, resolution, frame count, solver steps, and timing procedures, what is VestEdit’s total runtime relative to Wan-Edit, including attention extraction and spectral decomposition?
  5. Reliability of the main gain: What are the evaluation sample counts and aggregation procedure underlying the reported FiVE-Acc improvement? Are paired uncertainty estimates or repeated-run results available to determine whether the 1.09-point gain is robust?
Reviewer_14KSOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a training-free correction for inversion-free video editing with rectified-flow models. Following FlowEdit/Wan-Edit, the raw editing velocity is built from source- and target-conditioned predictions. VestEdit modifies this velocity in two stages before each ODE step:

  1. Semantic-topological spatial projection: a soft mask is formed from the positive cross-attention difference between target and source prompts, multiplied elementwise with source self-attention. The mask confines the editing residual to the estimated edit region and keeps the rest on the source velocity.
  2. Kinetic-energy-preserving orthogonal projection (KE-OPM): the masked field is projected onto the lowest 15% eigenvectors of a normalized graph Laplacian built from the backbone's temporal self-attention. The result is then rescaled to the pre-projection L2 norm.
Critical strengths
  1. VestEdit is training-free, leaves the backbone and the editing rule unchanged, and needs no additional backbone forward evaluations. In principle it can be added to any text-to-video rectified flow models.
  2. Consistent background-preservation gains over the direct baseline. VestEdit has lower LPIPS and higher SSIM than Wan-Edit on every one of the six tasks.
Critical weaknesses
  1. Unreliable experimental results

The headline numbers are internally inconsistent across tables. This makes it hard to trust the main claims.

(1) Table 1 results are not convincing. VestEdit's CLIP-S is never higher than Wan-Edit's on any of the six tasks: in Tables 5–10. Yet Table 1 reports 27.71 (VestEdit) vs. 27.59 (Wan-Edit). Moreover, Wan-Edit's Table 1 entries equal the simple mean of its task-wise values, whereas VestEdit's do not in Dist, LPIPS, CLIP-E and Fid. Also the reproduced Wan-Edit baseline is substantially weaker than the results in FiVE (Li et al., 2025b, Table 2). The paper's main empirical claim, improved background preservation over Wan-Edit, holds only relative to the authors' reproduction. Could the authors (i) report the full Wan-Edit and VestEdit configurations, (ii) explain the gap to the published Wan-Edit numbers, and (iii) evaluate VestEdit under the official FiVE protocol with the official Wan-Edit settings, so that it can be compared directly with the published results?

(2) Table 2 FiVE-Acc does not match its sub-criteria. For DMT, AnyV2V and Pyramid-Edit, the reported FiVE-Acc equals the mean of FiVE-YN, FiVE-MC, FiVE-∪ and FiVE-∩ exactly. But it doesn’t match in VestEdit, Wan-Edit and TokenFlow.

(3) Runtime is unexplained. VestEdit runs Wan-Edit's computation plus an extra correction, yet it is reported as faster (5.21 vs. 8.20 s/frame). The same full model is reported at 3.18 s/frame in the ablation (Table 4). The paper should explain how they evaluated runtime under a single protocol.

(4) Table 3 seems to compare Edit3 results of VACE, Wan-Edit, VestEdit against overall averages of TokenFlow, DMT and AnyV2V. So the claim that VestEdit leads the six compared methods on Edit3 is not supported. In Tables 3 and 5–10, TokenFlow, DMT and AnyV2V have exactly the same value for every metric on all six tasks, and these values equal the Table 1 averages.

  1. Heuristic method design

(1) The role of source self-attention in the spatial mask is not justified.

  • How the self-attention map is aggregated and mapped to match the cross-attention map is unclear. Also, the effect of each term is not validated in ablation study.
  • The paper claims that self-attention identifies coherent regions. However, the output value is closer to how much attention a token receives. A high saliency score does not imply that the token belongs to a coherent structure.

(2) A large Dirichlet energy along the source attention graph does not necessarily indicate flickering. KE-OPM treats any velocity component that differs across tokens linked by source temporal attention as undesirable. This assumption fails in several common editing scenarios such as (i) temporally varying edits (e.g. lighting/time-of-day changes), (ii) object addition/removal (iii) erroneous correspondences such as repetitive textures or multiple similar instances (e.g., several identical ducks). Although the authors acknowledge these issues, the paper neither measures whether this energy correlates with perceived flickering nor reports how much of the edit is removed by the projection.

Questions for authors
  1. Ablation results doesn't support VestEdit.

In the ablation study, the full model does not outperform Structure Distortion and Motion Fidelity. Removing either module, or replacing KE-OPM with naive temporal averaging, improves Distortion. The authors attribute this to these metrics rewarding conservative outputs. But the spatial mask is designed to make the output more conservative (to anchor the background), so why does removing it reduce Distortion from 19.65 to 13.47? Could the authors report per-variant FiVE-Acc, CLIP-S/E and locality metrics and add variants isolating the self-attention term?

  1. Missing baseline comparison on recent training-free video editing methods.

FlowDirector, DFVEdit, IF-V2V and ContextFlow are recent training-free, inversion-free video editing methods that the paper does not compare against. Can the authors include them, ideally on the same backbone? TokenFlow, DMT and AnyV2V are built on much older backbones, so their gap to VestEdit likely reflects backbone strength more than the editing method.

  1. Missing direct evaluation of the core claim.

The paper's motivation is flicker, boundary tearing and background drift, but none of the reported metrics measures temporal consistency directly. Could the authors report evidence by metrics such as CLIP/DINO frame-to-frame consistency, VBench temporal-flickering/subject-consistency scores?

  1. Question on VACE evaluation setting.

VACE is built on Wan2.1, a rectified-flow model, but Table 1 and App. A list it as a diffusion-model (DM) method. Which VACE model and evaluation setting were used for the reproduced numbers?

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

Related papers

Discussion (0)

Sign in to comment.