{"id":"a0ae037f-007a-48e0-8014-18f91794a148","arxiv_id":"2507.08285","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":10,"one_line_summary":"FlowDrag combines 3D mesh deformation with diffusion-based drag editing, using the resulting 2D vector flow to steer the denoising process, and adds a ground-truth benchmark built from video frames.","lead":"FlowDrag edits images by building a 3D mesh of the object, deforming it with a rigidity-preserving algorithm, and using the projected 2D motion vectors to guide a diffusion model. It also introduces VFD-Bench, a video-derived benchmark with ground-truth edited frames, and reports improved geometric consistency over prior drag-editing methods.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Benchmark superiority may reflect selection bias: per-image mesh choice and hyperparameters are tuned on the same author-created VFD-Bench with no fixed protocol, so the central outperformance claim is not yet supported.","rationale":"In good faith, the paper's idea is coherent: use a 3D mesh deformation to derive a dense 2D flow that preserves local rigidity, then inject that structure into motion-supervised diffusion editing. The progressive SR-ARAP machinery is standard, and VFD-Bench is a sensible attempt to supply ground truth that DragBench lacks. The qualitative comparisons and user study give some support for the claim that FlowDrag produces plausible rigid edits. The load-bearing weakness is in the evaluation and the specification of the mechanism, not in the deformation math. The per-image DiffMesh/DepthMesh rule in Section 6.1 is uncontrolled; the ablation and vector-count choices in Section 6.4 and Table 3 are made on the same author-created benchmark; and no error bars, seeds, or code are provided. In addition, the paper's own Section 7 concedes that the method works only for moderate rigid edits, so the scope of the superiority claim is narrower than the abstract's blanket statement. I therefore do not think the current paper establishes that mesh-guided vector flow is the cause of the reported gains; a fixed-protocol, split-honest evaluation with an ablation that removes the flow from the loss would settle it. This matches the reader's conditional assessment, with the additional mechanistic under-specification noted.","tokens_in":17363,"tokens_out":6237,"duration_ms":79863,"concrete_test":"Release code and data, split VFD-Bench into a 100-pair development set and a 150-pair test set, and pre-register all choices: a fixed mesh-type rule (or always DiffMesh), alpha, beta, vector count, and sampling strategy. Evaluate once on the test set and report paired per-sample differences with standard errors against GoodDrag. In the same run, include an ablation that replaces the sampled flow set with uniformly sampled positions inside the mask while keeping layout injection; if PSNR and 1-LPIPS do not drop, the flow field is not the active ingredient.","verdict_should_be":"UNCHANGED","load_bearing_attack":"To attribute FlowDrag's reported advantage to mesh-guided vector flow, the evaluation must use a fixed protocol. It does not. Section 6.1 chooses the mesh type per image: DiffMesh is used 'primarily,' but DepthMesh replaces it whenever DiffMesh 'exhibits significant artifacts or deviates substantially from the original image,' with no objective threshold and no validation split. Section 6.4 and Fig. 8 select beta and the number of vectors by their effect on VFD-Bench metrics, and Table 3 selects magnitude-based sampling on the same benchmark. Since VFD-Bench is an author-created benchmark and the numbers are used as the test set, the pipeline has effectively been tuned on the evaluation itself. The reported margins are small enough for this to matter: DragBench MD is 22.88 vs 22.96 for GoodDrag, and the VFD-Bench advantages are 0.41 PSNR, 0.03 1-LPIPS, and 7.08 MD, with no error bars. A second, mechanistic gap is that Section 4.3 says only that the summation in Eq. 3 is 'restricted' to q in the sampled flow set; it never states that the flow displacement replaces the user direction delta_i. If the flow field merely selects which patches are supervised while the direction remains the user's delta_i, the geometry-aware deformation vectors are not actually entering the motion-supervision loss, and the title mechanism is untested. Together these gaps make the central outperformance claim unfalsifiable in its current form.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"FlowDrag is a drag-based image editing method that constructs a 3D mesh from an input image (either from a monocular depth estimator, DepthMesh, or from an image-to-3D diffusion model, DiffMesh), deforms this mesh using a progressive SR-ARAP energy, projects the per-vertex displacements into a 2D vector flow field, and injects this flow plus the deformed mesh's projected layout into the UNet denoising process of a motion-based drag-editing pipeline. The paper also introduces VFD-Bench, a new benchmark of 250 input/ground-truth pairs derived from consecutive video frames, with handle/target annotations. The central claim is that FlowDrag outperforms existing drag-editing methods (DiffEditor, DragDiffusion, DragNoise, FreeDrag, GoodDrag) on both VFD-Bench and DragBench while better preserving geometric consistency.","tokens_in":17687,"tokens_out":4523,"duration_ms":50531,"significance":"The high-level idea of using explicit 3D mesh deformation to produce a dense, geometry-aware displacement field for drag editing is promising and well motivated: existing motion-supervision methods only match user-specified points and ignore broader structure. The proposed VFD-Bench is also a constructive contribution, because DragBench has no ground truth and conventional fidelity metrics can penalize successful geometric edits. If the central outperformance claim were supported by a fixed, non-circular evaluation protocol, the paper would be a solid contribution to the drag-editing literature. However, as presented, the quantitative evidence is not currently reliable: the authors select several key hyperparameters and even the mesh-generation mode on VFD-Bench itself and then report final numbers on that same benchmark, with no error bars or significance tests. The mechanism by which the flow field enters the motion-supervision loss is also underspecified. These issues are load-bearing for the abstract's blanket claim of superiority, though they are potentially fixable with a clearer protocol and additional experiments.","major_comments":[{"comment":"The paper selects hyperparameters by ablating on VFD-Bench and then reports final performance on VFD-Bench. Specifically, the Inter-Step Smoothness parameter β is chosen from Table 4, the vector count (10) is chosen from Fig. 8, and the magnitude-based sampling strategy is chosen from Table 3, all using VFD-Bench metrics. The final comparison in Table 2 is then reported on the same VFD-Bench set. This is test-set tuning: the reported gains (e.g., MD 28.23 vs. 35.31 for GoodDrag) may reflect choices fitted to this specific benchmark rather than genuine method superiority. To support the claimed outperformance, the authors must either pre-register all hyperparameters, use a disjoint validation split for ablations, or report VFD-Bench results under a fixed protocol, and they should include per-method error bars and significance tests.","section":"Section 6.4, Table 3, Fig. 8"},{"comment":"The selection between DiffMesh and DepthMesh is made per image without a fixed, objective criterion: \"When DiffMesh exhibits significant artifacts or deviates substantially from the original image, we employ DepthMesh instead.\" No threshold is defined, no ablation quantifies when each mesh is preferred, and no statistics are reported on how often each mesh type is used or on the sensitivity of the final numbers to this choice. As written, this is a post hoc selection rule that makes the evaluation unfalsifiable. The authors should specify a deterministic rule (for example, a quantitative fidelity threshold between the rendered DiffMesh and the input image) and demonstrate that the reported results are robust to this choice, or report results separately for DepthMesh and DiffMesh.","section":"Section 6.1, Mesh Generation"},{"comment":"The description of vector-flow sampling does not specify whether the user direction δ_i in the motion-supervision loss is replaced by the flow displacement at the sampled point. The text says \"We then restrict the summation in Eq.3 to q ∈ ˆΦ,\" but the mesh-derived vector (∆x_i, ∆y_i) from Eq. (10) never explicitly appears in Eq. (3). If the flow field only changes which patch positions q are summed over while the direction remains the user's normalized δ_i, then the deformation geometry does not actually control the movement direction, and the central mechanism claimed by the paper is not exercised. The authors must state clearly whether δ_i is replaced by the flow vector, provide the revised loss function if so, and include an ablation that isolates the effect of using flow directions versus user directions.","section":"Section 4.3, Eq. (3)"},{"comment":"The abstract's statement that \"FlowDrag outperforms existing drag-based editing methods on both VFD Bench and DragBench\" is not supported by Table 1. On DragBench, FlowDrag is second on 1-LPIPS (0.82 vs. DiffEditor's 0.89), and the MD advantage over GoodDrag (22.88 vs. 22.96) is within the range where noise or a single case could change the ranking. The authors acknowledge the 1-LPIPS caveat in Section 6.3, but the claim should be qualified to say \"best MD and competitive fidelity on DragBench, and best overall on VFD-Bench,\" and the MD difference should be accompanied by a significance test.","section":"Abstract, Table 1, Section 6.3"}],"minor_comments":[{"comment":"The text says \"β = 0.8 yields the best results, achieving maximal mean edge length ratio and mean ARAP error,\" but a lower ARAP error is better and the table indeed shows the lowest error at β=0.8. This should read \"minimal mean ARAP error.\"","section":"Section 6.4, Table 4 caption"},{"comment":"VFD-Bench pairs are consecutive video frames, which may include lighting changes, camera motion, or non-rigid background motion even within the user mask. The paper should describe the frame-filtering criteria (e.g., how much background change was tolerated, whether frames with occlusion or blur were removed) and report any manual cleaning protocol.","section":"Section 5, Appendix C"},{"comment":"The algorithm's input declaration lists \"D: Depth map\" but Step 1 computes the depth map from the input image I. The input/output signature should be made consistent (e.g., input is I and τd, τb; D is an internal variable).","section":"Appendix B, Algorithm 1"},{"comment":"Equation (11) is referenced as \"Eq. 11\" but the equation itself is not numbered in the text; please add the number or change the reference.","section":"Appendix A.1, Eq. (11)"},{"comment":"The user study reports average rankings from 25 volunteers on 50 images, but no inter-rater agreement, per-method standard error, or statistical test is provided. A paired test (e.g., Wilcoxon) across images would strengthen the claim that FlowDrag is preferred.","section":"Section 6.3, User Study"},{"comment":"The paper acknowledges in the limitations that FlowDrag \"primarily supports rigid edits\" and struggles with content creation/removal, but this scope restriction is not carried into the abstract or the main claims. Please state the rigid-edit scope prominently when summarizing the method's capabilities.","section":"Section 7, Limitations"}],"recommendation":"major_revision","confidential_remarks":"The core idea and benchmark are genuinely interesting, but the evaluation protocol is not yet at the standard required for publication. The test-set tuning on VFD-Bench and the per-image mesh-choice rule are the two load-bearing issues; both are addressable with a fixed protocol, validation splits, and error bars. The ambiguity in Eq. (3) about whether the flow direction actually replaces δ_i is more fundamental: if the vector field only changes patch selection, the paper's title mechanism is untested. I recommend major revision rather than rejection because these issues can be resolved within the paper's scope, and the proposed benchmark is a useful resource. Please also confirm whether VFD-Bench and the code will be released, as this is essential for a benchmark contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know up front. VFD-Bench is a genuinely good idea: drag-editing evaluation has been stuck with no-reference metrics that penalize successful rigid edits, and building input-GT pairs from video frames directly addresses that. The second thing is that the paper's central outperformance claim is not yet supported by its own tables. On DragBench, FlowDrag lands at 0.82 1-LPIPS versus 0.89 for DiffEditor and 0.86 for GoodDrag, and the MD advantage over GoodDrag is 0.08—noise range without error bars. The VFD-Bench margins are larger, but they come from an author-created benchmark where beta, vector count, and the sampling strategy were all tuned on that same benchmark, with no fixed rule for choosing between DiffMesh and DepthMesh per image.\n\nWhat is actually new: connecting SR-ARAP mesh deformation to the diffusion motion-supervision loop is a reasonable integration, and the qualitative figures do show better structural coherence on rigid edits like the statue and the face rotation. The idea of projecting a deformed mesh into a 2D vector flow and injecting those vectors into latent optimization is worth exploring. The ablation and sensitivity studies are more thorough than most drag-editing papers; the authors also say plainly that the method targets moderate rigid edits, not large structural changes.\n\nThe soft spots, in proportion. First, the mechanism is under-specified. Section 4.3 says the summation in Eq. 3 is restricted to the sampled flow vectors, but it never says the flow displacement replaces the user's delta_i. If the flow field only selects which patches get supervised, the geometry-aware direction is not actually entering the loss and the title mechanism is untested. That needs a clear sentence or an updated equation. Second, the per-image mesh choice is uncontrolled—\"when DiffMesh exhibits significant artifacts\" is a subjective criterion, and because the paper compares methods on the same images, this can only bias the results in FlowDrag's favor. Third, the evaluation uses VFD-Bench for both model selection and final reporting. Fig. 8 peaks at 10 vectors, and Table 3 picks magnitude-based sampling; those choices need a validation split before the numbers mean anything. There are no error bars or significance tests anywhere, and no code or data release.\n\nThe benchmark alone justifies a serious referee. The method is plausible, the limitations are stated honestly, and the evaluation framework is a step forward for the field. I would send this to review, but with a request for major revision: fixed mesh-selection protocol, held-out hyperparameter tuning, error bars, and a precise statement of how the flow field enters the loss. I'd engage with the work and probably cite VFD-Bench once it's released; I wouldn't take the performance tables as they stand.","headline":"VFD-Bench is a genuinely useful benchmark idea, but the paper overclaims on DragBench and its VFD-Bench numbers are weakened by tuning on the same benchmark and an uncontrolled mesh-selection rule.","tokens_in":18310,"tokens_out":3859,"would_cite":true,"duration_ms":44028,"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":"FlowDrag reports that injecting a mesh-derived 2D vector flow field into diffusion denoising improves accuracy and geometric coherence of drag edits, outperforming existing methods on both VFD-Bench and DragBench.","keywords":["drag-based image editing","diffusion models","3D mesh deformation","as-rigid-as-possible","vector flow field","ground-truth benchmark","video frame pairs","geometric consistency"],"falsifier":"On VFD-Bench, replace FlowDrag's mesh-derived vector field with a smooth interpolation of each handle-to-target displacement spread over the edit mask, keeping everything else identical; if the interpolated field matches FlowDrag's PSNR and Mean Distance within noise, then the SR-ARAP mesh deformation is not responsible for the reported gains, whereas if the mesh-derived field wins, the mechanism is confirmed.","tokens_in":17087,"feed_emoji":"🖱️","tokens_out":7073,"duration_ms":70877,"temperature":0.7,"pith_summary":"Drag-based image editing lets a user drag handle points to targets, but existing methods match only those points and ignore the rest of the object, so rigid parts such as hats, torches, or limbs get distorted. FlowDrag's proposal is to give the edit an explicit geometric backbone: reconstruct a 3D mesh of the object, deform it with a progressive smoothed-rotation as-rigid-as-possible energy that moves the handle vertex toward the target while keeping local rotations consistent, project the per-vertex displacements back into a 2D vector flow field, and feed that field into the UNet denoising loop in place of point-only motion supervision. The paper also introduces VFD-Bench, a 250-pair benchmark built from consecutive video frames, so that for the first time drag edits have real ground-truth edited images rather than proxy fidelity scores. On that benchmark, and on DragBench, the paper reports that FlowDrag achieves better handle-to-target alignment and geometric fidelity than existing drag editors, with the largest advantages on edits that should preserve rigidity.","feed_headline":"Mesh-guided drag editing beats point-only methods on two benchmarks","feed_subtitle":"A deformed 3D mesh becomes a 2D vector field that keeps rigid parts from distorting during drag edits.","key_machinery":"The load-bearing object is the 2D vector flow field $\\Phi = \\{(\\Delta x_i, \\Delta y_i)\\}$, the projection of per-vertex displacements between the original mesh $\\mathcal{M}$ and the deformed mesh $\\hat{\\mathcal{M}}$ onto the image plane. It is produced by progressive SR-ARAP deformation, an extension of the as-rigid-as-possible energy that moves the handle vertex toward its target over $K$ steps, smooths rotations between adjacent vertices, and penalizes large position changes between steps. The field supplies the motion-supervision loss with a sparse set of sampled vectors, replacing the single handle-to-target direction with a shape-aware displacement map, while the deformed mesh projection is separately injected into the UNet's attention features at an early denoising timestep to provide global layout context.","core_discovery":"FlowDrag's central claim is that a continuous 2D vector flow field derived from a 3D mesh deformation is sufficient to make diffusion-based drag editing geometrically coherent. The method maps the user's handle point, target point, and edit mask onto a mesh, deforms the mesh with progressive SR-ARAP, and defines the flow field as the projected per-vertex displacements between original and deformed meshes. That field, rather than the single handle-to-target direction, is what drives motion supervision and point tracking, and the projected deformed mesh is additionally injected as an early-timestep layout feature. The paper's experiments report that this geometry-aware guidance improves alignment of handle points to targets and preserves object structure on both the new VFD-Bench, which has ground-truth edited frames, and the existing DragBench.","pith_inferences":["Our inference: if the vector flow field is the real source of the gains, feeding the same field into gradient-guidance drag editors should also improve their geometric coherence; this is not tested in the paper.","Our inference: VFD-Bench's paired frames could measure the paper's mesh-faithfulness assumption directly, by correlating per-sample edit failures with reconstruction error between the generated mesh and the actual frame-to-frame object change.","Our inference: the optimal-vector-count result suggests the flow field helps most where displacement is concentrated, so a natural follow-up is to test whether the winning sampled vectors sit on occlusion boundaries or high-curvature regions."],"forward_implications":["Edits that rotate, relocate, or change the pose of a rigid object become more stable, because the flow field carries shape context beyond the two clicked points.","VFD-Bench gives drag editing a ground-truth evaluation protocol, so future methods can be compared on actual edit accuracy rather than proximity to the unedited input.","The method is deliberately bounded: it suits moderate drags and rigid edits, and is not designed for large content creation or removal.","Sampling around ten flow vectors with the largest magnitudes gives the best fidelity, so the vector-selection step is as important as the mesh deformation itself."],"supporting_citations":[{"why":"Defines the ARAP deformation energy that FlowDrag's mesh deformation is built on.","marker":"Sorkine & Alexa, 2007"},{"why":"Adds the smoothed-rotation term (SR-ARAP) that FlowDrag extends with inter-step smoothness.","marker":"Levi & Gotsman, 2014"},{"why":"Supplies Marigold, the single-image depth estimator used to build DepthMesh.","marker":"Ke et al., 2024"},{"why":"Supplies Hunyuan3D 2.0, the image-to-3D diffusion model used to build DiffMesh.","marker":"Zhao et al., 2025"},{"why":"Provides GoodDrag, the motion-based baseline whose pipeline FlowDrag adopts and modifies, and whose DDIM-inversion settings are reused.","marker":"Zhang et al., 2024"},{"why":"Defines DragDiffusion, the motion-supervision baseline and DragBench source that FlowDrag compares against.","marker":"Shi et al., 2024"},{"why":"Defines DiffEditor, the gradient-guidance baseline compared on both benchmarks.","marker":"Mou et al., 2024"},{"why":"Defines DDIM, whose inversion is used to obtain the latent that FlowDrag optimizes.","marker":"Song et al., 2020a"},{"why":"Provides the Stable Diffusion backbone and VQ-VAE encoder-decoder used in all experiments.","marker":"Rombach et al., 2022"},{"why":"Defines DIFT features, used by the Mean Distance metric on both benchmarks.","marker":"Tang et al., 2023"}],"fun_headline_variants":["3D mesh flow field stabilizes drag edits in diffusion models","Geometry-aware drag: FlowDrag beats point-only editing","Drag editing gets a 3D mesh backbone for stable deformations","FlowDrag: Mesh-derived vector fields for precise drag edits","Vector flow from 3D mesh improves drag-based image editing"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole method rests on the assumption that the single-image depth map or image-to-3D reconstruction yields a mesh that is geometrically faithful to the object, because the vector flow field inherits every error in that mesh; the paper's uncontrolled rule of switching between DiffMesh and DepthMesh when one looks bad is never quantitatively validated.","fun_headline_variants_meta":{"raw":{"variants":["3D mesh flow field stabilizes drag edits in diffusion models","Geometry-aware drag: FlowDrag beats point-only editing","Drag editing gets a 3D mesh backbone for stable deformations","FlowDrag: Mesh-derived vector fields for precise drag edits","Vector flow from 3D mesh improves drag-based image editing"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000238,"raw_usage":{"total_tokens":1481,"prompt_tokens":884,"completion_tokens":597,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":500,"completion_tokens_details":{"reasoning_tokens":513}},"tokens_in":500,"tokens_out":597,"duration_ms":6313,"temperature":1.0,"reasoning_tokens":513,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T18:22:39.142434+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On VFD-Bench, replace FlowDrag's mesh-derived vector field with a smooth interpolation of each handle-to-target displacement spread over the edit mask, keeping everything else identical; if the interpolated field matches FlowDrag's PSNR and Mean Distance within noise, then the SR-ARAP mesh deformation is not responsible for the reported gains, whereas if the mesh-derived field wins, the mechanism is confirmed.","supporting_citations":[],"review_version":1}