- 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
-
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$25.47). Appendix F also matches Wan-Edit's local CLIP at while retaining better fidelity, so the improvement is not explained simply by a weaker edit.
-
Table 2 contains the complete three-module factorial evaluation, supplemented by stride, guidance-strength, and source-reference studies. MTS reduces EPE in every tested component context, including 1.804$\to$1.671 when added alone. The non-adjacent-stride results and qualitative comparisons provide relevant support for its motivation. The paper appropriately separates temporal self-consistency from agreement with source motion and reports both optical-flow and tracklet-based measures.
Critical weaknesses
-
The evidence does not isolate the specific guidance decomposition from simpler spatial control. The EGL ablation compares the complete rule with no EGL, not the same map applied to the entire velocity difference. This matters because FlowAnchor [1] and FlowDirector already localise or modulate editing signals, although their mechanisms differ. TAF additionally uses SAM 2 masks, without a Wan-Edit control receiving identical masks. A controlled spatial-gating comparison would establish whether selectively weighting T3 contributes beyond spatial information alone. FlowAnchor and DFVEdit [2] are cited but not evaluated; FlowAnchor's unreleased code [3] is a legitimate limitation, not by itself grounds for rejection.
-
The added value of MTS in the complete system remains uncertain. Adding it to EGL+TAF reduces EPE by 2.6% (1.473$\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 , 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
-
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?
-
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.
-
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
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).
| Metric | EGL+TAF | Full model | Mean Δ | 95% CI | p |
|---|---|---|---|---|---|
| EPE-FG↓ | 5.178 | 4.726 | −0.452 | [−0.573, −0.347] | < 1e-39 |
| EPE↓ | 1.473 | 1.434 | −0.039 | [−0.068, −0.011] | 0.20 |
| MFS-FG↑ | 0.8959 | 0.8957 | −0.0002 | [−0.0018, 0.0013] | 0.44 |
| MFS↑ | 0.9092 | 0.9082 | −0.0010 | [−0.0023, 0.0004] | 0.12 |
| L.CLIP-T↑ | 25.58 | 25.47 | −0.110 | [−0.157, −0.063] | < 1e-6 |
| M.PSNR↑ | 28.32 | 27.68 | −0.645 | [−0.721, −0.571] | < 1e-46 |
Table R3: Full model minus EGL+TAF at the same γ (App. J, Table 9).
| γ | L.CLIP-T (full / EGL+TAF) | ΔEPE-FG | ΔEPE | ΔMFS | ΔM.PSNR (dB) |
|---|---|---|---|---|---|
| 2.5 | 25.24 / 25.34 | −9.4% | −2.8% | −0.0019 | −0.70 |
| 3.0 | 25.38 / 25.49 | −9.0% | −3.0% | −0.0022 | −0.68 |
| 3.5 | 25.47 / 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 passes | Mean ± std (s) | Median (s) | 95th percentile (s) | |
|---|---|---|---|---|
| EGL+TAF | 140 | 726 ± 189 | 779 | 949 |
| Full model | 200 | 976 ± 268 | 1051 | 1291 |
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).
| Setting | Value |
|---|---|
| Optimiser | Adam, η = 0.003, (β1, β2, ε) = (0.9, 0.999, 1e-8), re-initialised at every guided step |
| Loss weight | λ_MS = 20 |
| Guided steps, inner updates | G = 10, J = 2 |
| Pooling, neighbourhood | ρ = 2, κ = 3 (descriptor dimension D = Cκ² = 16 × 9 = 144) |
| Strides, weights | S = {1, 2, 3}, λ = (1, 2, 4) |
| Source reference | source-branch prediction at the same noise and timestep |
Table R6: Descriptor size, with everything else at the default (App. C, Table 3).
| κ, ρ | L.CLIP-T↑ | EPE↓ | MFS↑ |
|---|---|---|---|
| 1, 1 | 25.43 | 1.409 | 0.9087 |
| 1, 2 | 25.44 | 1.420 | 0.9092 |
| 3, 1 | 25.44 | 1.421 | 0.9083 |
| 3, 2 (default) | 25.47 | 1.434 | 0.9082 |
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↑ | |
|---|---|---|---|---|---|---|
| DFVEdit | 24.44 | 30.5 | 24.66 | 0.0968 | 2.144 | 0.8958 |
| FlowAnchor (same masks) | 24.10 | 33.8 | 26.47 | 0.0783 | 1.972 | 0.8986 |
| FlowResolve | 25.47 | 49.3 | 27.68 | 0.0496 | 1.434 | 0.9082 |
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).
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.