acceptodds
Under review as a conference paper at ICLR 2027

FlowResolve: Constraining Motion and Allocating Guidance in Video Editing

Abstract

Text-driven editing modifies image or video content from natural language instructions. Inversion-free methods do it without mapping the image or video back to a noise latent, steering the sampling trajectory directly with an editing signal. On images this has shown impressive efficiency and structure preservation. Extending it to video remains challenging, and typically fails in three respects. The subject drifts away from the source motion, the background is perturbed, and the edit is not strong enough on the subject. We trace the motion drift to the backbone’s prior, which constrains motion but cannot guarantee consistency with the source motion. The other two failures follow from an editing signal that is misallocated. Based on this analysis we propose FlowResolve, a training-free framework with three components. Multi-scale temporal self-similarity guidance (MTS) constrains motion explicitly. Edit-guided localisation (EGL) confines the editing signal to the region the prompt describes, and reduces the drift at the same time. Target attention focusing (TAF) instead controls the editing strength within that region. Together these mechanisms stabilise the editing signal and add an explicit motion constraint, guiding the flow-based evolution toward the intended target distribution. Extensive experiments on FiVE-Bench show that FlowResolve improves background preservation, motion fidelity and instruction following.

Primary area:
generative models
Keywords:
video editing, text-guided video editing, diffusion models, video generation, motion fidelity

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

FlowResolve augments Wan-Edit with three training-free interventions: temporal self-similarity guidance (MTS), localisation of the excess target-guidance term using an edit-differential map (EGL), and mask-conditioned target-token attention amplification (TAF). MTS matches multi-stride relations in clean-sample predictions using stop-gradient latent updates. On 419 FiVE-Bench pairs, the system improves source preservation and local semantic alignment. Evaluation includes all eight component combinations, guidance and stride sweeps, and a source-reference ablation.

Critical strengths
  1. EGL offers an interpretable intervention: it localises the guidance-gap term while retaining the other two velocity components (Eqs. 8-9). TAF changes attention logits while preserving normalisation. The system improves over Wan-Edit in M.PSNR (26.00$\to$27.68 dB), EPE (1.804$\to$1.434), and local CLIP (24.81$\to$25.47). Appendix F also matches Wan-Edit's local CLIP at γ=1.5\gamma=1.5 while retaining better fidelity, so the improvement is not explained simply by a weaker edit.

  2. Table 2 contains the complete three-module factorial evaluation, supplemented by stride, guidance-strength, and source-reference studies. MTS reduces EPE in every tested component context, including 1.804$\to$1.671 when added alone. The non-adjacent-stride results and qualitative comparisons provide relevant support for its motivation. The paper appropriately separates temporal self-consistency from agreement with source motion and reports both optical-flow and tracklet-based measures.

Critical weaknesses
  1. The evidence does not isolate the specific guidance decomposition from simpler spatial control. The EGL ablation compares the complete rule with no EGL, not the same map applied to the entire velocity difference. This matters because FlowAnchor [1] and FlowDirector already localise or modulate editing signals, although their mechanisms differ. TAF additionally uses SAM 2 masks, without a Wan-Edit control receiving identical masks. A controlled spatial-gating comparison would establish whether selectively weighting T3 contributes beyond spatial information alone. FlowAnchor and DFVEdit [2] are cited but not evaluated; FlowAnchor's unreleased code [3] is a legitimate limitation, not by itself grounds for rejection.

  2. The added value of MTS in the complete system remains uncertain. Adding it to EGL+TAF reduces EPE by 2.6% (1.473$\to$1.434), but lowers local CLIP (25.58$\to$25.47), background PSNR (28.32$\to$27.68), and MFS (0.9092$\to$0.9082). The existing full-model sweep partly addresses edit strength: at γ=4.5\gamma=4.5, local CLIP is nearly matched (25.61), but EPE is 1.479 and MFS 0.9069. Other settings trade these metrics rather than clearly improving the frontier. These results do not establish a clear improvement of the motion-editing trade-off; the full-system 20.5% EPE gain does not isolate MTS. Full-frame motion metrics and selected examples do not resolve whether object trajectories improve reliably at comparable editing strength, and the twenty additional target forwards lack measured runtime.

[1] Chen et al. (2026). FlowAnchor: Stabilizing the Editing Signal for Inversion-Free Video Editing.

[2] Cai et al. (2025). DFVEdit: Conditional Delta Flow Vector for Zero-shot Video Editing.

[3] FlowAnchor authors (2026). Official FlowAnchor repository and release status.

Questions for authors
  1. How does selective T3 modulation compare with applying the identical map to the entire update, and with a Wan-Edit baseline using the same SAM 2 masks?

  2. Using the existing strength sweep, what evidence establishes MTS's incremental object-motion benefit over EGL+TAF? Please provide paired uncertainty, foreground-motion results, and actual runtime for the two variants.

  3. Please specify MTS learning rate, loss weight, pooling and neighbourhood sizes, and clarify which phrase/detector initializes masks for replacement or addition. Explain the comparison boundary for FlowAnchor and DFVEdit.

5 comments
AuthorsOpenReview ↗

Response to Reviewer GkFJ : W1 and Q1

W1 & Q1: Spatial Gating and Masks.

We agree that the original submission did not disentangle the effect of guidance decomposition from spatial gating. We therefore added the two controls requested in Q1 (Table R1; App. I). EGL gates only T3 because of the structure of the update. Regrouping Eq. 8 gives

T1+T2=V~(Zttar,t,P∗,ωsrc)−V~(Ztsrc,t,P,ωsrc),T3=Δω (Vttar,c−Vttar,u).T_1+T_2=\tilde V(Z_t^{tar},t,P^*,\omega_{src})-\tilde V(Z_t^{src},t,P,\omega_{src}),\qquad T_3=\Delta\omega\,(V_t^{tar,c}-V_t^{tar,u}).

T1+T2 is the source-referenced FlowEdit difference that carries the edit and vanishes where the target and source branches agree. T3 has no source counterpart and may remain nonzero in unchanged regions. Thus, gating the entire update can suppress the edit, whereas gating T3 selectively attenuates this unwanted guidance contribution.

The controls in Table R1 confirm this distinction. With EGL's soft map, gating the entire update (row c) improves background preservation but substantially reduces edit success compared with gating only T3 (row b). With the exact object mask, the two variants (rows d and e) are closer: full-update gating preserves slightly more background, while T3-only gating retains slightly more edit success. This highlights why selective gating is particularly important for EGL's soft, mask-free map.

Row e also shows that mask-based full-update gating provides strong background preservation but reduces edit success relative to ungated Wan-Edit. EGL achieves substantial background gains without benchmark masks, using a map derived from the model's own edit signal. In FlowResolve, only TAF uses the benchmark mask to strengthen editing inside the object, explaining the higher edit success of the full model. We now identify mask-dependent methods in Table 1 and acknowledge this limitation in the conclusion. FlowAnchor and DFVEdit are evaluated in our response to Q3.

Table R1: Spatial-gating controls on all 419 pairs, without MTS and TAF (App. I, Table 7). w: EGL map; m: benchmark object mask used by TAF.

UpdateL.CLIP-T↑FiVE-Acc (%)↑M.PSNR↑EPE↓MFS↑
aT1+T2+T3 (Wan-Edit)24.8140.626.001.8040.8988
bT1+T2+w⊙T3 (EGL)24.3736.028.051.3030.9098
cw⊙(T1+T2+T3)23.9023.630.701.0680.9183
dT1+T2+m⊙T324.6637.229.931.2080.9139
em⊙(T1+T2+T3) (Wan-Edit mask)24.6035.031.531.0470.9159
AuthorsOpenReview ↗

Response to Reviewer GkFJ : W2 and Q2

W2 & Q2: Added Value of MTS.

We agree that the full-system EPE reduction does not isolate MTS: most of the gain comes from EGL, while MTS provides only a small additional improvement on the full frame. We have therefore narrowed our claims in Sec. 4.4. To address Q2, App. J adds foreground-motion metrics (EPE-FG and MFS-FG), paired significance tests, comparisons at matched TAF strengths, and measured runtime.

The results show that MTS consistently and significantly reduces foreground EPE across all tested TAF strengths (Table R2; App. J, Table 11), indicating improved motion of the edited object. In contrast, in the signed-rank test it does not significantly improve full-frame EPE or MFS, and it introduces small but significant decreases in L.CLIP-T and M.PSNR. Comparisons at matched TAF strengths confirm this trade-off across all three γ values (Table R3; Fig. 11); the background-quality cost occurs only when TAF is enabled (Table 2, rows 1→2 and 7→5).

MTS adds 60 network passes to the 140 used by Wan-Edit (200 in total), without backpropagation through the backbone. We correct the undercount in the submitted Sec. 4.1 and report measured runtime in Table R4.

Table R2: Full model vs. EGL+TAF at γ = 3.5, paired over the 419 pairs: mean difference with 95% bootstrap CI and two-sided Wilcoxon signed-rank p (App. J, Table 10).

MetricEGL+TAFFull modelMean Δ95% CIp
EPE-FG↓5.1784.726−0.452[−0.573, −0.347]< 1e-39
EPE↓1.4731.434−0.039[−0.068, −0.011]0.20
MFS-FG↑0.89590.8957−0.0002[−0.0018, 0.0013]0.44
MFS↑0.90920.9082−0.0010[−0.0023, 0.0004]0.12
L.CLIP-T↑25.5825.47−0.110[−0.157, −0.063]< 1e-6
M.PSNR↑28.3227.68−0.645[−0.721, −0.571]< 1e-46

Table R3: Full model minus EGL+TAF at the same γ (App. J, Table 9).

γL.CLIP-T (full / EGL+TAF)ΔEPE-FGΔEPEΔMFSΔM.PSNR (dB)
2.525.24 / 25.34−9.4%−2.8%−0.0019−0.70
3.025.38 / 25.49−9.0%−3.0%−0.0022−0.68
3.525.47 / 25.58−8.7%−2.6%−0.0010−0.65

Table R4: Time per edit over all 419 pairs on one RTX 3090, excluding model loading and video I/O (App. J, Table 12).

Network passesMean ± std (s)Median (s)95th percentile (s)
EGL+TAF140726 ± 189779949
Full model200976 ± 26810511291
AuthorsOpenReview ↗

Response to Reviewer GkFJ : Q3

Q3: Hyperparameters, Masks, and the Comparison with FlowAnchor and DFVEdit.

We thank the reviewer for pointing out these missing details. All MTS hyperparameters are provided in Table R5 and App. N. App. C further evaluates four pooling and neighbourhood configurations, showing little sensitivity to these choices (Table R6).

All FiVE-Bench experiments use the benchmark-provided per-video object masks for TAF. For videos without masks, App. D describes our automatic pipeline: Grounding DINO is prompted with the noun phrase of an object in the source video: the replaced object for replacement and the host object for addition (e.g., "dog" for "a dog wearing sunglasses"), and likewise the edited or removed object for the other edit types. SAM 2 then propagates the resulting box through the video. The target tokens in Eq. 10 are the sub-tokens of the benchmark-annotated target-object phrase (e.g., "sunglasses"). Since FiVE computes L.CLIP-T inside and M.PSNR outside the object masks, we mark FlowResolve as mask-based in Table 1 and provide the same masks to Wan-Edit (App. I) and FlowAnchor. Importantly, only TAF consumes the masks; EGL and MTS do not. We also acknowledge this mask dependency in the revised paper.

For a fair comparison, FlowAnchor and DFVEdit are evaluated using their official code and default settings, with the same Wan2.1-T2V-1.3B backbone and all 419 pairs. FlowAnchor receives the same benchmark masks as TAF, following its FiVE-Bench evaluation protocol, whereas DFVEdit uses no masks. FlowResolve outperforms both methods on all six metrics (Table R7), with consistent advantages also on the Wan2.1-T2V-14B backbone (App. M).

Table R5: MTS settings (App. N).

SettingValue
OptimiserAdam, η = 0.003, (β1, β2, ε) = (0.9, 0.999, 1e-8), re-initialised at every guided step
Loss weightλ_MS = 20
Guided steps, inner updatesG = 10, J = 2
Pooling, neighbourhoodρ = 2, κ = 3 (descriptor dimension D = Cκ² = 16 × 9 = 144)
Strides, weightsS = {1, 2, 3}, λ = (1, 2, 4)
Source referencesource-branch prediction at the same noise and timestep

Table R6: Descriptor size, with everything else at the default (App. C, Table 3).

κ, ρL.CLIP-T↑EPE↓MFS↑
1, 125.431.4090.9087
1, 225.441.4200.9092
3, 125.441.4210.9083
3, 2 (default)25.471.4340.9082

Table R7: FlowAnchor and DFVEdit on Wan2.1-T2V-1.3B, all 419 pairs (from Table 1).

L.CLIP-T↑FiVE-Acc (%)↑M.PSNR↑BG.LPIPS↓EPE↓MFS↑
DFVEdit24.4430.524.660.09682.1440.8958
FlowAnchor (same masks)24.1033.826.470.07831.9720.8986
FlowResolve25.4749.327.680.04961.4340.9082
Reviewer_yadNOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes FlowResolve to improve source-motion preservation and instruction following without inversion. Three complementary components are proposed: Multi-scale temporal self-similarity guidance (MTS), Edit-guided localisation (EGL), and Target attention focusing (TAF). Extensive experiments on FiVE-Bench demonstrate that FlowResolve achieves the best background preservation and motion fidelity.

Critical strengths
  1. Text-driven video editing without explicit inversion is an important and challenging problem.

  2. The experiments provide detailed comparisons against representative baselines, along with systematic full ablations (Table 2).

Critical weaknesses
  1. The paper is poorly written. Many claims are made without sufficient theoretical explanation or empirical evidence, and the underlying motivation is not clearly justified.

  2. Figure 1 is difficult to follow. An effective pipeline figure should allow readers to easily understand how the method works under the hood. However, this figure is overloaded with too much content and notation, making it hard to identify the key focus.

  3. Limited novelty. The method appears to be an incremental combination of three standard heuristic techniques: MTS is basic temporal feature alignment, EGL is difference-based spatial masking/attenuation, and TAF is a common logit-boosting operation on cross-attention.

  4. Table 2 indicates that the benefit of the TAF component is very marginal, and it even introduces trade-offs in background preservation and motion fidelity.

Questions for authors

See Weaknesses.

2 comments
AuthorsOpenReview ↗

Response to Reviewer yadN : W1–W3

W1: Writing, Motivation and Evidence.

We thank the reviewer for this feedback. The paper follows one line of argument: three observed failures (motion drift, background perturbation, and weak edits) → two underlying causes (a motion prior that is not guaranteed to follow the source motion, and a misallocated editing signal) → three corresponding components. The abstract and introduction follow this structure, and the revision adds quantitative evidence for each causal claim. Specifically, gating only T3 improves background and motion, whereas also gating T1+T2 suppresses much of the edit (Table 2; App. I), supporting EGL's decomposition. At matched TAF strengths, MTS reduces motion error on the edited object (App. J), supporting the need for source-motion preservation. TAF consistently improves edit success across all settings in Table 2, addressing weak edits.

W2: Figure 1.

We agree that Fig. 1 is overloaded. A single editing step proceeds as follows: MTS first corrects the target latent by matching multi-stride temporal self-similarities of its clean-sample prediction to those of the source. TAF then biases the cross-attention logits of the target tokens within the object region, only in the target conditional pass. Finally, EGL rescales only T3, the guidance-gap term of the velocity difference, using a mask-free map derived from the model's own edit signal. We retain Fig. 1 as an overview and clarify these operations step by step in Sec. 3.

W3: Novelty.

We agree that the individual operators are established. Our contribution lies in their specific roles, the motivation for this design, and the controls that validate it.

EGL. Unlike difference-based masking of an entire update, EGL leaves T1+T2 unchanged and gates only the source-free T3 using a map derived from the model's own edit signal, without an external mask. T1+T2 is the equal-guidance FlowEdit difference, referenced to the source and vanishing where the source and target branches agree; T3 has no such source reference. Gating the entire update therefore suppresses the edit, as confirmed by our control (App. I). With β = 1, EGL's rescaling is disabled, recovering the original Wan-Edit guidance update. Unlike DiffEdit [1], which estimates a mask from noise predictions and blends latents, EGL directly localises the guidance gap without latent blending.

MTS. MTS does not align appearance features, which would penalise the intended appearance change. Instead, it matches second-order temporal relations: cosine self-similarities across frames at strides 1–3, computed on temporally mean-removed descriptors of clean-sample predictions, with no backpropagation through the DiT. Related motion-transfer methods match pairwise frame differences of space-time features (DMT [2]) or temporal-mean-removed motion features (MOFT [3]). MTS instead preserves source motion during appearance editing. Non-adjacent strides are important: using stride 1 alone barely changes EPE (App. C).

TAF. Attention re-weighting is an established operator (Prompt-to-Prompt [4]), and we do not claim it as novel. Its role is complementary to EGL: since EGL can only attenuate (w ≤ 1), the edit must be strengthened inside the object without globally increasing ω_tgt. TAF achieves this by biasing target-token logits only in the target conditional pass. The resulting additional guidance enters T2 and T3 and also shapes the map used by EGL.

[1] G. Couairon et al., DiffEdit: Diffusion-based Semantic Image Editing with Mask Guidance. ICLR 2023.

[2] D. Yatim et al., Space-Time Diffusion Features for Zero-Shot Text-Driven Motion Transfer. CVPR 2024.

[3] Z. Xiao et al., Video Diffusion Models are Training-free Motion Interpreter and Controller. NeurIPS 2024.

[4] A. Hertz et al., Prompt-to-Prompt Image Editing with Cross Attention Control. ICLR 2023.

AuthorsOpenReview ↗

Response to Reviewer yadN: W4

W4: Benefit and Cost of TAF.

TAF is the only component that improves edit success and semantic alignment in every setting of Table 2 (Table R11 below). FiVE-Acc, the benchmark's edit-success rate, is now reported for all methods and ablations, and it makes this benefit easier to see than L.CLIP-T alone.

TAF does have a cost, mainly in motion: EPE rises in all four settings. Its background cost appears only together with MTS; without MTS, TAF barely changes M.PSNR. Within the full model, raising the TAF strength γ improves L.CLIP-T at a small cost in M.PSNR (App. F, Table 5). At γ = 1.5, FlowResolve matches Wan-Edit's L.CLIP-T while keeping clearly better background and motion fidelity, so its fidelity gains do not come from a weaker edit.

Table R11: Adding TAF in the four settings of Table 2.

Setting (rows of Table 2)FiVE-Acc (%) without → with TAFΔL.CLIP-TΔM.PSNR (dB)ΔEPE
alone (1 → 3)40.6 → 53.6+0.91+0.11+11.4%
with MTS (2 → 4)33.7 → 51.3+1.09−0.36+10.2%
with EGL (7 → 6)36.0 → 50.4+1.21+0.27+13.0%
with MTS + EGL (5 → 8)31.7 → 49.3+0.89−1.18+10.1%
Reviewer_FHPTOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes FlowResolve, a training-free video editing method built on Wan2.1-T2V-1.3B. It combines temporal constraints to preserve source motion, localized guidance to reduce background changes, and attention enhancement within an object mask to strengthen the edit. The method is evaluated on FiVE-Bench against six training-free editing methods.

Critical strengths

The results show better background preservation and motion fidelity than Wan-Edit, with an improvement in local text alignment. The ablations cover all combinations of the three components and help explain the trade-off between editing strength and preservation.

Critical weaknesses
  • My main concern is the limited qualitative evidence of effective edits beyond recoloring.

  • Most qualitative examples focus on recoloring. Can the authors provide more examples of object replacement, shape changes, and lighting edits? For recoloring, a simple baseline using the same tracked masks would help establish the method’s added value.

  • How is the editing mask selected from the instruction? For adding sunglasses to a dog, is TAF applied to the whole dog, its face, or the eye region? Appendix D describes mask generation and propagation but does not explain how the detector phrase and mask granularity are chosen for such additions.

  • FlowResolve is evaluated only on Wan2.1-T2V-1.3B. Do the improvements persist on a larger or different backbone? Such results would help establish whether the benefits extend beyond the current model.

  • I would suggest including video visualizations for better comparisons. Frame samples alone are insufficient to assess flickering, temporal artifacts, and motion preservation. Quantitative metrics alone may not fully capture the overall editing quality.

Questions for authors

Most of my questions are included in the weaknesses part.

4 comments
AuthorsOpenReview ↗

Response to Reviewer FHPT: W1 and W2

W1: Edits Beyond Recolouring.

We agree that the qualitative examples in the submission overrepresented colour edits. However, 319 of the 419 FiVE-Bench pairs are non-colour edits (replacements, material edits, additions, or removals). The revision therefore reports FiVE-Acc, the benchmark's VLM-based edit-success rate, for all methods and by edit type (Table R8; App. K). FlowResolve outperforms Wan-Edit in five of six categories and ties on removal. Among the four categories with 100 pairs, the largest gain is on material edits. Overall, FlowResolve ranks second to FlowDirector in edit success, while FlowDirector has substantially lower background and motion fidelity (Table 1).

App. H (Figs. 8–10) adds nine selected (not randomly sampled) non-colour examples, each compared with FlowAnchor using the same masks: three rigid replacements (musician → robot, motorcyclist → Spiderman, dog → cat), one non-rigid replacement (man → panda), three material edits (metallic man, wooden musician, plush dog), and two additions (a bright yellow fedora, a cap). We omit removal examples because there are only 10 such pairs and no method edits them reliably.

FiVE-Bench has no lighting category, and FlowResolve is designed for object-level edits. Global edits such as relighting have no object region to localise; with TAF disabled and β = 1, FlowResolve reduces to Wan-Edit with MTS. We therefore make no claim of an advantage for such edits.

Table R8: FiVE-Acc (%) by edit type, with the number of pairs per type (App. K, Table 14).

MethodRigid repl. (100)Non-rigid repl. (100)Colour (100)Material (100)Addition (9)Removal (10)All (419)
RAVE47.020.549.024.05.610.033.9
TokenFlow34.032.039.517.550.00.030.4
UniEdit-Flow58.050.545.047.050.05.049.0
DNAEdit56.550.052.035.066.75.047.7
FlowDirector65.049.074.032.088.95.054.5
Wan-Edit49.038.553.524.538.910.040.6
FlowResolve52.540.562.044.566.710.049.3

W2: Same-Mask Baseline for Recolouring.

We added the requested baseline: Wan-Edit with the same benchmark masks used by TAF, applied to the entire update (Table R9; App. I). It achieves stronger background preservation (M.PSNR: 31.53 vs. 26.00) but lower edit success (35.0% vs. 40.6%), illustrating the trade-off introduced by mask-based gating. In FlowResolve, only TAF uses the masks to strengthen edits inside the object, while EGL achieves background preservation without masks.

Table R9: Same-mask baseline on all 419 pairs (App. I, Table 7; Table 1).

L.CLIP-T↑FiVE-Acc (%)↑M.PSNR↑EPE↓
Wan-Edit24.8140.626.001.804
Wan-Edit + same masks24.6035.031.531.047
FlowResolve25.4749.327.681.434
AuthorsOpenReview ↗

Response to Reviewer FHPT: W3–W5

W3: Mask Selection.

We apologize for not specifying this in the submission; Sec. 3.4 and App. D now clarify the procedure. On FiVE-Bench, TAF uses the benchmark-provided per-video object mask. For "a dog wearing sunglasses", the mask covers the entire dog, not the sunglasses. For videos without masks, App. D describes an automatic pipeline: Grounding DINO is prompted with the relevant source-object phrase (replaced, edited, removed, or host object, depending on the edit type), and SAM 2 propagates its box through the video.

TAF does not specify where the sunglasses appear. It adds log γ to the logits of the target-object tokens annotated by FiVE-Bench (here, "sunglasses") for queries inside the dog mask; the model's cross-attention determines their spatial placement. TAF is the only mask-dependent module. With TAF disabled, the remaining modules use no masks and still improve background and motion over Wan-Edit (Table 2, row 5).

W4: Other Backbones.

We evaluated Wan-Edit and FlowResolve on three additional backbones, Wan2.2-TI2V-5B, Wan2.1-T2V-14B, and Wan2.2-T2V-A14B, using all 419 pairs and FlowResolve's default Wan2.1-T2V-1.3B settings without retuning (Table R10; App. M). FlowResolve improves over Wan-Edit on all six metrics across all three backbones. On Wan2.1-T2V-14B, it also outperforms DFVEdit and FlowAnchor on every metric in Table R10.

Table R10: Other backbones, all 419 pairs (App. M, Table 15).

BackboneMethodL.CLIP-T↑FiVE-Acc (%)↑M.PSNR↑BG.LPIPS↓EPE↓MFS↑
Wan2.2-TI2V-5BWan-Edit24.2232.926.410.06151.6560.9074
FlowResolve25.0943.426.940.04961.4600.9092
Wan2.1-T2V-14BWan-Edit24.5436.826.540.06501.4790.9075
DFVEdit24.0930.826.030.07491.5010.9074
FlowAnchor24.0032.527.160.06471.8200.9044
FlowResolve24.9742.827.580.05011.2730.9096
Wan2.2-T2V-A14BWan-Edit25.0447.123.760.09881.9470.8986
FlowResolve25.6552.725.300.07011.5970.9031

W5: Video Results.

We agree that individual frames cannot reveal flicker or trajectory errors. The supplementary material now includes side-by-side videos (source | FlowResolve | FlowAnchor) for 12 examples in App. H, covering rigid and non-rigid replacement, colour, material, and addition edits, using the same prompts. We include FlowAnchor as a strong temporal-consistency baseline. Other baselines and ablations are quantitatively compared using the temporal metrics in Tables 1 and 4.

Thanks for adding the experiments and the side-by-side videos, including the material editing results and the tests on larger Wan models. One question about the videos: the FlowAnchor outputs look much blurrier and have much more artifacts than the examples on its official project page. This visual results raise the concern about fair comparisions.

Thank you for raising this concern. We followed the default configuration and official implementation provided by the FlowAnchor authors, and used the masks provided by Five-Bench to generate the results. No modifications were made to the official code or its default settings. The observed visual artifacts may be attributable to certain issues in the official implementation. We have strictly followed the officially recommended setup to ensure a fair comparison.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response: summary of reviews and revisions

We thank all reviewers for their time and constructive feedback. We have uploaded a revised PDF, with new and revised text in blue. The main text remains within the 10-page limit, and the new material is in Appendices I–N.

Strengths noted by the reviewers

  1. An interpretable modification to the update. Reviewer GkFJ highlighted that EGL gates only the guidance-gap term, leaving the other two velocity terms intact, while TAF biases attention logits without changing attention normalisation.
  2. Improved fidelity without weakening text alignment. Reviewers GkFJ and FHPT noted improvements over Wan-Edit in background preservation, motion fidelity, and local text alignment. GkFJ further noted that at γ = 1.5, FlowResolve matches Wan-Edit in L.CLIP-T while retaining better fidelity.
  3. Comprehensive ablations and comparisons. Reviewers GkFJ, FHPT, and yadN noted the full ablation over all combinations of the three components (Table 2). GkFJ also highlighted the stride, guidance-strength, and source-reference studies, as well as the separate evaluation of temporal self-consistency and agreement with source motion. Reviewer yadN noted the detailed comparisons with representative baselines.

1. Main concerns raised across reviewers

  • C1. Whether the gains come from the guidance decomposition or spatial information, and what role the object masks play (GkFJ, FHPT).
  • C2. The added value of MTS in the full system, its statistical significance, and its runtime (GkFJ).
  • C3. Missing baselines (FlowAnchor, DFVEdit) and evaluations on other backbones (GkFJ, FHPT).
  • C4. Evidence of effective edits beyond recolouring and the benefit of TAF (FHPT, yadN).
  • C5. Missing implementation details, including MTS hyperparameters and mask and detector-phrase selection (GkFJ, FHPT).
  • C6. Writing, Figure 1, and novelty (yadN).

2. Our responses and revisions

C1. Spatial gating and masks. We clarify that TAF uses FiVE-Bench's per-video object masks in all benchmark experiments (Sec. 3.4 and App. D); Table 1 also identifies the mask-based methods. EGL and MTS use no masks, and FlowResolve still improves background and motion fidelity over Wan-Edit when TAF is disabled. New controls (App. I) show that applying the EGL map to the entire update, rather than only to T3, substantially weakens edits. Wan-Edit with the same masks applied to the entire update preserves more background but has lower edit success than standard Wan-Edit.

(See W1 & Q1 for Reviewer GkFJ; W2 and W3 for Reviewer FHPT.)

C2. Added value of MTS. We compare EGL+TAF with the full model at matched TAF strengths (App. J). In the full system, MTS significantly reduces motion error on edited objects; its full-frame EPE gain is small and not significant in the signed-rank test, and it incurs modest costs in edit strength and background fidelity. We have narrowed the MTS claims in Sec. 4.4 and clarified its computational cost: MTS adds 60 network passes to Wan-Edit's 140. On an RTX 3090, an edit takes 976 ± 268 s, vs. 726 ± 189 s for EGL+TAF.

(See W2 & Q2 for Reviewer GkFJ.)

C3. Baselines and backbones. App. L adds FlowAnchor and DFVEdit. FlowResolve outperforms both on every metric in Table 1 except WarpSSIM, for which FlowAnchor performs best. On Wan2.2-TI2V-5B, Wan2.1-T2V-14B, and Wan2.2-T2V-A14B, FlowResolve improves on Wan-Edit across all reported metrics, using the Wan2.1-T2V-1.3B settings without retuning (App. M).

(See Q3 for Reviewer GkFJ; W4 for Reviewer FHPT.)

C4. Edits beyond recolouring and the benefit of TAF. FiVE-Acc, the benchmark's VLM-based edit-success metric, is now reported for all methods and ablations. FlowResolve raises it from 40.6% (Wan-Edit) to 49.3%, second only to FlowDirector (54.5%), which has substantially lower background and motion fidelity. FlowResolve improves over Wan-Edit in five of the six edit types and ties on removal. TAF increases edit success in every setting of Table 2. App. H adds nine selected non-colour examples next to FlowAnchor, and the supplementary material includes side-by-side videos for 12 examples.

(See W1, W2, and W5 for Reviewer FHPT; W4 for Reviewer yadN.)

C5. Missing details. App. N now lists all MTS settings, including the Adam learning rate, loss weight, pooling size, and neighbourhood size. App. D specifies Grounding DINO and the noun phrase used to initialise the mask for each edit type; SAM 2 then propagates it through the video.

(See Q3 for Reviewer GkFJ; W3 for Reviewer FHPT.)

C6. Writing, Figure 1, and novelty. In this response, we link each causal claim to supporting measurements in the revised paper, provide an explanation of an editing step, and clarify how each component differs from the prior operators it builds on. Figure 1 itself is unchanged. We have also corrected several errors in the text.

(See W1–W3 for Reviewer yadN.)

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.