{"id":"c3bef5ec-208c-4f8f-8afa-81573122155e","arxiv_id":"2506.07713","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"FlowV2V performs video editing by editing the first frame and propagating it with shape-aligned optical flow, reporting improved consistency on DAVIS-EDIT.","lead":"This paper describes a video editing technique, FlowV2V, that edits the first frame of a video and then uses estimated optical flow to propagate that edit through the remaining frames. The authors report better temporal consistency and output quality than prior methods on a 50-video benchmark, especially for non-rigid motion such as face rotation.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (3) replaces per-pixel flow inside the edited mask with a single averaged vector, so rotation/articulation inside the object is discarded; the non-rigid motion claim is unsupported unless SCFC is shown to recover that structure.","rationale":"The reader's weakest assumption was that mask propagation and inherited average motion would fail under shape change, occlusion, or non-rigid deformation. My concern is sharper: even with perfect source flow and perfect masks, Eq. (3) replaces the intra-object flow field with a per-region average, discarding exactly the rotational and articulated structure that motivates the paper's headline examples. This is an internal property of the construction, not merely an evaluation gap, so it is the most load-bearing threat to the central claim. The supplementary's RAFT-error analysis does not cover this failure mode, and SCFC is trained on corrupted ground-truth flow, not on constant-flow inputs, so it cannot be assumed to recover the missing structure. The proposed test uses synthetic data with known flow, isolating the mechanism and avoiding confounds from RAFT error or editing artifacts. Since the paper currently provides no code, no quantitative ablation of IMP/SCFC, and no error bars, the conditional verdict remains appropriate until this test is run; a positive result would confirm the mechanism, while a negative result would require the central non-rigid claim to be revised. My read agrees partially with the reader: we both identify the pseudo-flow assumption as critical, but I locate the failure point earlier, at the constant-flow averaging in Eq. (3), rather than at mask warping alone.","tokens_in":12348,"tokens_out":4798,"duration_ms":67015,"concrete_test":"Create a synthetic clip of a disk rotating at constant angular velocity with a known mask, and perform an appearance-only edit so the shape is unchanged. First, run IMP and compare the pseudo-flow inside the mask with the ground-truth disk flow: compute mean endpoint error and the average curl of the flow field. If the pseudo-flow curl is near zero while the source curl is positive, the control signal cannot represent rotation. Second, run the full FlowV2V pipeline on this clip and recover the rotation angle of the edited output by template tracking; if the tracked angle advance is near zero or systematically smaller than the source, the rotation claim is falsified. This isolates the Eq. (3) averaging from RAFT error, since flow is known exactly in synthetic data.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Sec. 3.2, Eq. (3) sets the pseudo-flow inside the edited first-frame mask to a constant value F1 equal to the average source flow over the source mask M1; outside the mask it keeps source flow. Eq. (4) then propagates masks by source flow, and the same averaging is repeated per frame. The advertised non-rigid cases (face rotation, multi-object articulation, Fig. 1, Fig. 4) require spatially varying flow inside the object: rotation has nonzero curl, articulation has shear and divergence. A per-region average removes this structure, so the flow condition cannot transmit the very deformation the paper claims to handle. SCFC (Sec. 3.3) is trained on random-mask corruption of ground-truth flow (supp. A), so it may inpaint plausible structure, but the paper never demonstrates that it can recover rotation or articulation from an input whose interior is constant; the information is absent from the input. The supplementary's RAFT-error discussion (Sec. D) concerns textureless regions, occlusion, and lighting, not this structural information loss, so it does not address the gap. The central non-rigid editing claim is therefore not supported by the mechanism as described.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"FlowV2V is presented as a video-editing pipeline that casts the task as flow-driven image-to-video generation. The method first edits the first frame (FFE), then iteratively propagates the source video's optical flow inside the source object mask (IMP) to obtain a pseudo flow sequence aligned with the edited object, refines this sequence with a flow-calibration network (SCFC), and finally conditions a Stable Video Diffusion-based I2V model on the calibrated flow (FD-I2V). The method is evaluated on the DAVIS-EDIT benchmark against several existing video-editing methods, with the abstract reporting improvements of 13.67% on DOVER and 50.66% on warping error. The paper also includes ablations on the first-frame editor and the type of control signal.","tokens_in":12572,"tokens_out":6104,"duration_ms":68516,"significance":"If the proposed method truly handles non-rigid motion as claimed, it would address a known limitation of depth-based video editing (StableV2V) by using optical flow as a richer control signal. The pipeline is pragmatic and built from off-the-shelf components, and the paper includes a reasonably broad comparison and ablations. However, the central mechanism for non-rigid motion transfer is not convincingly supported, and the quantitative claims suffer from missing uncertainty measures and a potential metric-coupling issue. The idea of using flow with shape alignment is interesting, but the current evidence is not sufficient to substantiate the headline claims.","major_comments":[{"comment":"The pseudo flow inside the edited object mask is set to a single constant vector per frame—the average flow over the source mask. This removes all spatial variation of the flow field within the object, including the rotational and shear components that define non-rigid motion. Consequently, the described mechanism cannot by itself transfer non-rigid motions such as the face rotation and multi-object articulation highlighted in Figures 1 and 4. The paper relies on SCFC (Sec. 3.3) to refine the pseudo flow, but SCFC is trained on ground-truth flow corrupted by random mask dropping (supp. A), where missing regions are zeroed out rather than filled with a constant average. It is not demonstrated that SCFC can recover a spatially varying flow field from a constant-valued input, since the deformation information is not present in the input. Supplementary Section D discusses RAFT estimation errors on textureless regions, occlusion, and lighting, but does not address this structural information loss. Without additional evidence, the central claim of non-rigid motion editing is unsupported.","section":"Sec. 3.2, Eq. (3)"},{"comment":"The headline improvement figures in the abstract (13.67% on DOVER, 50.66% on warping error) are not directly traceable to the numbers in Table 1, and all results are point estimates over 50 videos with no confidence intervals or significance tests. For example, the DOVER improvement over StableV2V in the text-based setting is about 13.9% on DAVIS-EDIT-S but only about 8.7% on DAVIS-EDIT-C, so the reported single percentage is not representative of the results as presented. More importantly, the warping error (WE) metric may be computed using the same RAFT optical flow that FlowV2V uses as control; the paper does not specify how WE is defined. If WE is evaluated by warping generated frames with the source video's RAFT flow, then FlowV2V is partially self-scoring because its output is explicitly conditioned on that flow. Please clarify the exact definition of WE and provide per-set results, error bars, or statistical tests to support the comparative claims.","section":"Table 1"},{"comment":"The mask propagation step in Eq. (4) warps binary masks with optical flow. Standard bilinear warping will produce fractional values, which conflicts with the statement that the masks are binary-valued. The paper does not specify whether nearest-neighbor sampling, thresholding, or some other discretization is used. If nearest-neighbor is used, masks may be inaccurate at object boundaries; if bilinear, the masks become soft and the subsequent average computation in Eq. (3) and the iterative propagation are affected. This technical detail is important for reproducibility and should be clarified in the main text or supplementary.","section":"Eqs. (3)-(4)"}],"minor_comments":[{"comment":"There are several typos and formatting inconsistencies: 'Hardamard' should be 'Hadamard' in Sec. 3.2; 'psuedo' should be 'pseudo'; the dataset name appears as 'DA VIS-EDIT' in most places but 'DAVIS-EDIT' in the abstract and sometimes 'DA VIS-Edit'; 'I2Edit' in Sec. 5.1 should be 'I2VEdit'; 'WAN' in Sec. 2 should be 'Wan'; and 'Step-Video3' is likely 'Step-Video-Ti2V'.","section":"Throughout"},{"comment":"The table includes results under both 512×512 and the original aspect ratio (marked with '*'), yet the abstract reports only a single set of improvement numbers without specifying which setting they refer to. The differences are substantial (e.g., DOVER 77.20 vs. 71.77 on DAVIS-EDIT-S for FlowV2V), so the abstract should clarify the setting to avoid misleading readers.","section":"Table 1 and Sec. 4"},{"comment":"The sentence 'we adopt a pre-trained image segmentation model (i.e., SAM) to obtain a binary mask sequence M = {M_i} and the first frame mask \\hat{M}_1 from the input video V and the edited first frame \\hat{I}_1' is ambiguous: it should be clarified that SAM is applied per frame to obtain the source masks, and only the edited first frame mask is obtained by segmenting \\hat{I}_1, with subsequent edited masks propagated via flow.","section":"Sec. 3.2"},{"comment":"The ablation studies are qualitative only. To support the claim that the flow-based condition is superior to depth or edge, quantitative results on a subset of DAVIS-EDIT would be more convincing. Similarly, the effect of the first-frame editor is described qualitatively without quantitative metrics.","section":"Sec. 5.2"},{"comment":"A user study is announced in Sec. 4 ('we conduct a user study to measure the human preferences'), but no user-study results appear anywhere in the paper or the supplementary. Please either add the results or remove the mention.","section":"Sec. 5.1"}],"recommendation":"major_revision","confidential_remarks":"The paper's core idea is potentially interesting, but the current review leaves me uneasy about the mismatch between the abstract's numbers and the table, the lack of evidence for the non-rigid motion mechanism, and the absence of code or detailed hyperparameters. Even after revision, the authors should probably temper the claims in the abstract or provide substantially stronger experimental support. The paper may fit the journal's scope if the mechanism concern is addressed, but as it stands, the support is not there."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nFlowV2V recombines known pieces—first-frame editing, RAFT flow, mask warping, a ProPainter-style calibration network, and a flow adapter on SVD—into a pipeline that has not appeared before. That is the honest novelty: not a new component but a new combination, and the paper is upfront about borrowing from AnyV2V, StableV2V, MOFA-Video, and ProPainter. The writing is clear, the supplement is honest about RAFT's failures, and the qualitative results in the non-rigid cases look better than the depth-based baseline.\n\nThe soft spots are real. The biggest is Eq. (3): inside the edited mask, the pseudo-flow is set to a single averaged vector. That throws away rotation, shear, and articulation—exactly the structure needed for the face-rotation and multi-object cases the paper leads with. A constant vector can only translate an object; it cannot turn a head. SCFC may inpaint some structure, but it is trained to restore randomly-masked ground-truth flow, not to reconstruct detail from an input whose interior is constant. The paper never shows that this recovery happens. The multi-object case has the same problem: one average over the whole mask would merge the motions of different objects, and no per-object handling is described.\n\nThe evaluation also needs work. Fifty instances, no error bars, no significance tests. The abstract's 13.67% and 50.66% improvements do not match any single cell in Table 1. The warping error is computed with RAFT, and since RAFT also generates the control signal, that metric is at least partially self-consistent with the method. The ablations are qualitative only. No code or weights are provided.\n\nNone of this kills the idea. Flow-based control is a sensible direction, and the paper is good enough that a serious referee should see it. What it needs is a hard look at whether SCFC can actually recover non-rigid motion from constant pseudo-flow, a per-instance evaluation with confidence intervals, and a reconciliation of the reported numbers.\n\nI'd send it to review, but ask for code and a direct experiment on the rotation case. If that checks out, it could be a useful contribution to the I2V editing literature.\n\nBest,\n[Your name]","headline":"Plausible flow-based editing pipeline whose headline non-rigid claim is undercut by the constant-flow averaging in Eq. (3); evaluation also thinner than reported, but worth a serious referee.","tokens_in":13133,"tokens_out":4108,"would_cite":false,"duration_ms":46075,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Optical flow, not depth, drives shape-changing video edits","keywords":["video editing","image-to-video generation","optical flow","diffusion models","temporal consistency","non-rigid motion","shape alignment","flow-conditioned generation"],"falsifier":"A controlled synthetic test would settle the load-bearing assumption: render a video of an object moving with known optical flow, edit the first frame into a differently shaped object, then compare FlowV2V's pseudo flow to the true motion of the edited object. If the pseudo flow diverges as soon as the edited shape rotates or moves differently from the source object, the average-flow premise is falsified; the paper's own admitted failures on occluded objects and lighting changes are natural real-world approximations of this test.","tokens_in":12142,"feed_emoji":"🎬","tokens_out":8513,"duration_ms":89898,"temperature":0.7,"pith_summary":"FlowV2V tries to establish that video editing can be treated as a flow-driven image-to-video generation problem: edit the first frame with any convenient image editor, then let a generative model propagate the edit to all remaining frames under the control of optical flow that has been reshaped to match the edited object. The point is that depth maps, which previous first-frame methods use as the motion condition, cannot represent non-rigid motion such as face rotation, expression changes, or multi-object deformation. On the DAVIS-EDIT benchmark the reported numbers support the method: a 13.67% improvement in DOVER and a 50.66% reduction in warping error over existing state-of-the-art, together with better handling of portrait and multi-object cases. The paper also identifies the boundary of the claim: the final video quality is capped by the generative capability of the image-to-video backbone.","feed_headline":"Optical flow, not depth, drives shape-changing video edits","feed_subtitle":"A first-frame-to-video pipeline with shape-aligned flow lifts DOVER by 13.67 percent and cuts warping error by half on DAVIS-EDIT.","key_machinery":"The mechanism that carries the argument is the iterative motion propagation (IMP) routine plus the shape-consistent flow calibration network. IMP builds the pseudo optical flow sequence by assuming the average flow over the edited object region equals the average flow over the original object region, then uses the source video's optical flow to warp the edited-region mask forward frame by frame. The calibration network, modeled on ProPainter, removes flow regions that would mislead the generation, and a flow-driven adapter injects the calibrated flow into Stable Video Diffusion so the edited first frame propagates along the desired motions.","core_discovery":"The central discovery is that a shape-aligned pseudo optical flow sequence, not depth, is the right condition for propagating an edited first frame through time. The pipeline first edits the first frame, then computes RAFT optical flow on the source video, averages the flow inside the source object mask, and paints that average flow into the edited object's mask. Warping the edited mask forward with the source flow lets the same operation be repeated for every frame, producing a full pseudo flow sequence; a shape-consistent flow calibration network removes redundant regions, and a flow-conditioned Stable Video Diffusion model generates the edited video. The net effect, as measured on DAVIS-EDIT, is a substantial jump in temporal consistency and sample quality, with the largest gains on the subset where the edited object changes shape significantly.","pith_inferences":["A testable extension the paper does not pursue: weight the flow transfer by per-pixel confidence or rigidity, which could repair the admitted failures in textureless regions, occlusions, and lighting changes.","If the flow-transfer assumption is the bottleneck, evaluating the method on videos with object-independent motion (an object that rotates while translating, or masks that lose track under occlusion) should show the pseudo flow drifting exactly where the assumption breaks.","The reported inference time of about 2.5 minutes at 512x512 suggests the same paradigm could be applied to interactive draft editing, where the first frame is edited by the user and the flow-driven propagation is rerun on demand."],"forward_implications":["Non-rigid video edits such as face rotation, expression changes, and multi-object replacement become tractable in a first-frame-based pipeline without per-video tuning.","The choice of first-frame editor becomes modular: text, instruction, and reference-image prompts all plug into the same flow-driven propagation, so the method inherits each editor's strengths.","If the reported DAVIS-EDIT numbers hold, depth-conditioned propagation may be superseded for shape-changing edits, and warping error can be cut by roughly half.","The upper bound on output quality is set by the image-to-video backbone, so the method improves automatically as stronger generation models appear."],"supporting_citations":[{"why":"Supplies the DAVIS-EDIT benchmark and the depth-aligned StableV2V method that FlowV2V must beat; the quantitative comparison is defined against it.","marker":"[14]"},{"why":"Provides the RAFT optical flow extractor that produces the source flow sequence used in iterative motion propagation.","marker":"[29]"},{"why":"Provides the SAM segmentation model that yields the binary masks of the original and edited objects, which drive the shape-aware flow deformation.","marker":"[11]"},{"why":"The ProPainter video inpainting architecture is the basis for the shape-consistent flow calibration network that refines the pseudo flow.","marker":"[42]"},{"why":"Supplies the Stable Video Diffusion backbone that the flow-driven adapter conditions to generate the final video.","marker":"[2]"},{"why":"Supplies the flow-based adapter design for injecting motion control into a frozen image-to-video diffusion model.","marker":"[20]"}],"fun_headline_variants":["Flow beats depth for shape-changing video edits","Shape-aligned flow lifts video edit DOVER by 13.67%","Warping error cut 50% with flow-conditional video generation","Video editing as flow-driven image-to-video generation","Non-rigid video edits stay consistent with optical flow"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the edited object inherits the average motion of the original object region, and that the edited mask can be tracked by warping with the source video's flow; this premise fails when the edited object rotates, occludes, or moves differently from the source.","fun_headline_variants_meta":{"raw":{"variants":["Flow beats depth for shape-changing video edits","Shape-aligned flow lifts video edit DOVER by 13.67%","Warping error cut 50% with flow-conditional video generation","Video editing as flow-driven image-to-video generation","Non-rigid video edits stay consistent with optical flow"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000294,"raw_usage":{"total_tokens":1717,"prompt_tokens":959,"completion_tokens":758,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":575,"completion_tokens_details":{"reasoning_tokens":674}},"tokens_in":575,"tokens_out":758,"duration_ms":8593,"temperature":1.0,"reasoning_tokens":674,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T05:27:48.043813+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A controlled synthetic test would settle the load-bearing assumption: render a video of an object moving with known optical flow, edit the first frame into a differently shaped object, then compare FlowV2V's pseudo flow to the true motion of the edited object. If the pseudo flow diverges as soon as the edited shape rotates or moves differently from the source object, the average-flow premise is falsified; the paper's own admitted failures on occluded objects and lighting changes are natural real-world approximations of this test.","supporting_citations":[{"cited_title":"StableV2V: Stablizing Shape Consistency in Video-to-Video Editing","cited_arxiv_id":null,"evidence_quote":"Supplies the DAVIS-EDIT benchmark and the depth-aligned StableV2V method that FlowV2V must beat; the quantitative comparison is defined against it."},{"cited_title":"RAFT: Recurrent All-Pairs Field Transforms for Optical Flow","cited_arxiv_id":null,"evidence_quote":"Provides the RAFT optical flow extractor that produces the source flow sequence used in iterative motion propagation."},{"cited_title":"Berg, Wan-Yen Lo, Piotr Dollár, and Ross B","cited_arxiv_id":null,"evidence_quote":"Provides the SAM segmentation model that yields the binary masks of the original and edited objects, which drive the shape-aware flow deformation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The ProPainter video inpainting architecture is the basis for the shape-consistent flow calibration network that refines the pseudo flow."},{"cited_title":"Stable Video Diffusion: Scaling Latent Video Diffusion Models to Large Datasets","cited_arxiv_id":null,"evidence_quote":"Supplies the Stable Video Diffusion backbone that the flow-driven adapter conditions to generate the final video."},{"cited_title":"Mofa-video: Controllable image animation via generative motion field adaptions in frozen image-to-video diffusion model","cited_arxiv_id":null,"evidence_quote":"Supplies the flow-based adapter design for injecting motion control into a frozen image-to-video diffusion model."}],"review_version":1}