{"id":"6bd31181-24e1-486e-b2c3-1096fee66baa","arxiv_id":"2501.08676","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"FlexiClip proposes residual temporal corrections and a neural ODE for clipart animation, reporting modest gains on most metrics but worse geometric deviation than AniClipart.","lead":"FlexiClip adds a corrective 'temporal Jacobian' to the local deformations used in clipart animation, and fits it with a learned differential equation and a video-diffusion score loss. Its reported gains over AniClipart are small, and its own geometric-deviation metric gets worse, so the 'locality-preserving' claim is not supported.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed pfODE reformulation in Eq. 12-13 is not a reformulation of Eq. 10: no score function appears, no explicit state dependence is defined, and no convergence mechanism is derived, so the continuous-time temporal-modeling contribution is unsubstantiated.","rationale":"The reader's weakest assumption identifies the same load-bearing concern I find: the paper's central theoretical claims rest on Eq. 12-13 being a valid reformulation of the pfODE in Eq. 10, but no derivation connects them. I checked Sec. 3.3 and Appendix C: the appendix derives only the standard Gaussian-kernel pfODE and stops at GG^T = C_dot(t), with no mention of temporal Jacobians. Equation 12 contains no score term, and the attention features C^P_W and C^R_{W-1} are not shown to play the roles of C(t) and A(t). The convergence of the temporal correction to zero is asserted, but the only actual mechanism forcing it small is the L2 penalty in Eq. 18. This means the two headline innovations, continuous-time pfODE modeling and GFlowNet-inspired flow matching, are not established. I give credit for a clearly described pipeline, a user study, and consistent improvements on most reported metrics; however, the empirical evaluation is thin (single runs, no error bars, no released code), so it cannot independently carry the weight of the theoretical claims. The paper's own geometric-deviation metric also worsens versus AniClipart, which further strains the 'locality-preserving' claim. For these reasons I find no basis to change the reader's REJECT verdict.","tokens_in":14839,"tokens_out":6146,"duration_ms":66638,"concrete_test":"Re-derive Eq. 12-13 from Eq. 10 by positing a Gaussian-smoothed density over the total Jacobian field J_t = JP_t + JR_t and writing the probability-flow ODE for JR_t; verify that a score function grad log p(J_t) and a covariance schedule appear in the resulting equation. If no such derivation exists, compute the density p_t(J_t) implied by Eq. 13 and check whether it satisfies the Fokker-Planck equation corresponding to Eq. 10; if it does not, Eq. 12-13 is not a pfODE.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Equation 12 defines dJR_t/dt = f_R(JP_0, C^P_W, C^R_{W-1}, t; theta_R), and Eq. 13 obtains JR_t by integration. The right-hand side depends only on the base Jacobian, windowed attention features, and time; it contains no explicit dependence on the current state JR_t and no score function. In Eq. 10, the pfODE is driven by the score function grad_x log p_t(x) together with explicit covariance-rate and rescaling schedules C(t) and A(t). None of these objects appear in Eq. 12. The text equates C^P_W with C(t) and C^R_{W-1} with A(t), but these are attention-encoded features, not a defined noise covariance or scale schedule, and no probability density over Jacobian fields is specified. Appendix C derives only the standard Gaussian-kernel pfODE and ends at GG^T = C_dot(t); it never connects this derivation to temporal Jacobians or to Eq. 12. Thus the claim that Eq. 12-13 is a reformulation of Eq. 10 is unsupported. Moreover, the assertion that the correction 'reduced to zero' is not a property of Eq. 13; the only mechanism enforcing smallness of JR_t is the ad-hoc L2 penalty added in Eq. 18. If Eq. 12-13 is just an integrated MLP with an L2 regularizer, the paper's central pfODE and GFlowNet-inspired flow-matching claims lose their stated grounding, even if the implementation still produces plausible animations.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FlexiClip, a text-driven clipart animation method that builds on AniClipart's cubic Bézier keypoint trajectories and spatial Jacobian field deformation. Its claimed contributions are (i) temporal Jacobians that incrementally correct spatial Jacobians, (ii) a continuous-time reformulation via probability-flow ODEs (pfODEs), and (iii) a GFlowNet-inspired flow matching loss to reduce temporal noise. The system is trained with video Score Distillation Sampling, and experiments compare FlexiClip against AniClipart and several T2V/I2V models on bitmap metrics, animation metrics, and a user study.","tokens_in":15249,"tokens_out":5843,"duration_ms":52248,"significance":"The idea of adding a learned temporal correction to spatial Jacobians is plausible, and the qualitative results suggest the implemented system can produce appealing animations. However, the paper's central theoretical contributions are not established: the asserted pfODE reformulation in Eqs. (12)-(13) lacks any derivation or connection to Eq. (10), and the 'flow matching loss' is not shown to correspond to GFlowNet detailed balance or to any standard flow matching objective. More seriously, Table 2 shows that FlexiClip has a higher (worse) Geometric Deviation than AniClipart, directly contradicting the 'Locality-Preserving' claim in the title and the abstract's promise of geometric integrity. The quantitative evaluation reports single numbers without variance or significance tests. If the theoretical framing were removed, the remaining technical contribution is an incremental residual correction with L2 regularization.","major_comments":[{"comment":"The claim that Eq. (12)-(13) is a reformulation of the pfODE in Eq. (10) is not supported. The pfODE in Eq. (10) is driven by the score function ∇_x log p_t(x) with explicit schedules C(t) and A(t); none of these appear in Eq. (12), whose right-hand side depends only on the base Jacobian, attention features, and time, with no dependence on the current state J^R_t. The text identifies C^P_W with C(t) and C^R_{W-1} with A(t), but these are attention-encoded features rather than a covariance or scale schedule, and no probability density over Jacobian fields is specified. Appendix C derives the Gaussian-kernel pfODE and ends at G G^T = Ċ(t), but never connects this derivation to temporal Jacobians or to Eq. (12). Thus the central 'continuous-time pfODE' contribution is unsubstantiated; Eq. (12)-(13) is essentially an integrated MLP. The statement that the correction 'reduced to zero' is also not a property of Eq. (13); the only mechanism enforcing smallness is the ad-hoc penalty added in Eq. (18).","section":"Sec. 3.3 (Eqs. 12-13)"},{"comment":"The 'flow matching loss' is claimed to implement a detailed-balance objective from GFlowNets, but no mapping to GFlowNet states, transitions, or forward/backward policies is provided. Eq. (17) is an L2 difference of two score-function evaluations, and Eq. (18) appends a penalty on the temporal Jacobian magnitude; there is no derivation showing that this corresponds to detailed balance, to flow matching in the vector-field sense, or to any GFlowNet training objective. The terminology appears to be a post-hoc analogy rather than a formal connection. The authors should either supply the formal derivation or abandon the pfODE/GFlowNet framing.","section":"Sec. 3.4 (Eqs. 17-18)"},{"comment":"The title and abstract claim 'locality-preserving' deformation, yet Table 2 reports Geometric Deviation (GD, lower is better) of 52.34 for FlexiClip versus 50.98 for AniClipart. Sec. 4.5 admits this is because FlexiClip lacks ARAP deformation, which 'inherently minimizes shape distortion.' The sentence in Sec. 4.3 that 'GD has increased but DS decreased showing smoother animations' does not resolve the contradiction; a higher GD means the deformation is less rigid and less local, directly undercutting the central claim. This is a load-bearing inconsistency, not a presentation issue.","section":"Table 2 and Secs. 4.3, 4.5"},{"comment":"All quantitative results in Tables 1-4 are reported as single numbers with no variance, number of independent runs, or significance tests. Given the small margins on several metrics (e.g., TC 8.14 vs 8.51, X-CLIP 0.2102 vs 0.2075) and the worse GD, the claimed improvements over AniClipart are not established. The user study in Table 3 reports only mean percentages without confidence intervals or statistical tests, so the claim of significant outperformance is unsupported.","section":"Secs. 4.1-4.3 and Table 3"}],"minor_comments":[{"comment":"The GFlowNet preliminaries are duplicated nearly verbatim in two consecutive paragraphs; please condense to a single description.","section":"Sec. 2.3"},{"comment":"The displayed Jacobian matrix for a triangular face has an inconsistent size and layout; since J0,f is said to be in R^{2×2}, the partial-derivative table should be a 2×2 matrix and should be written unambiguously.","section":"Eq. (6)"},{"comment":"The overall loss is described as a weighted sum of 'Eq.14, 16, 17', but Eq. (16) is a definition of the denoising objective, not a loss term; the actual loss terms are Eq. (17) (with the penalty from Eq. (18)) and Eq. (14). Please clarify which equations enter the total loss and define λ consistently with its earlier use in Eq. (2).","section":"Eq. (19) and Sec. 3.4"},{"comment":"The text refers to 'Fig.6' and 'Fig.7' when the corresponding items are captioned as Table 6 and Table 7; the cross-references should be corrected.","section":"Appendix A"},{"comment":"The pfODE in Eqs. (7)-(9) is attributed to 'Song et al. (Lim et al., 2023)', but the cited work is by Lim et al.; the probability flow ODE was introduced by Song et al. (2021). The citation should be corrected, and the derivation in Appendix C should include the rescaling term A(t) that appears in Eq. (10).","section":"Sec. 2.2 and References"}],"recommendation":"reject","confidential_remarks":"The paper's theoretical framing appears to be rhetorical rather than substantive, and the main quantitative result (higher GD) contradicts the title's promise of locality preservation. The authors might reposition the work as an empirical study of residual Jacobian corrections for clipart animation, but as it stands the central claims are not defensible in a top-tier venue. The missing derivations and the absence of variance reporting would require a major rewrite and a re-evaluation of the claimed contributions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe useful piece in this paper is the concrete recipe: learn a small residual (temporal Jacobian) on top of the spatial Jacobian field, integrate it over time with a neural ODE, and regularize it to stay small. That combination appears to be new, and the reported numbers (CLIP 0.956 vs 0.940, DS 13.76 vs 18.49, user study) suggest it helps for clipart animation. I believe the effect is real: the residual correction plus L2 penalty is a plausible smoothness regularizer.\n\nEverything else in the framing is trouble. The paper claims a pfODE reformulation, but Eq. (12) and (13) are just an MLP whose input is the initial Jacobian, attention features, and time; there is no dependence on the current state, no score function, and no derivation linking them to Eq. (10). Appendix C ends at the standard Gaussian kernel and GG^T = C_dot; it never touches temporal Jacobians. So the pfODE and GFlowNet parts are naming, not theory. The 'flow matching loss' Eq. (17)-(18) is an L2 difference between two score estimates plus an L2 penalty on the correction; calling it GFlowNet-inspired doesn't make it one.\n\nThe biggest single problem is the paper's own Table 2: Geometric Deviation is worse than AniClipart (52.34 vs 50.98), while the title and abstract claim 'locality-preserving.' The text admits this and spins it as 'smoother animations' because DS improved, but GD is literally a measure of geometric distortion. That is a direct contradiction with the paper's central claim, not a minor caveat.\n\nAlso, all quantitative results are single runs with no error bars, no code or data released, and the user study numbers (94-95% vs 83% for AniClipart) are implausibly clean for 30 raters and 55 images.\n\nWho is this for? A reader working specifically on clipart animation might extract a usable trick: add a learned residual Jacobian with a small L2 penalty to smooth motion. That is worth a brief look. But the paper does not merit citation as a theory or as a state-of-the-art claim; it needs major revision to strip the overclaiming and provide basic statistical evidence.\n\nMy recommendation: send it to peer review—the empirical recipe is concrete enough to deserve referee time—but expect a reject or very major revision. In its current form, the theoretical grounding is not defensible.","headline":"A useful residual-Jacobian trick buried under unsupported pfODE and GFlowNet labels, with a headline claim contradicted by its own geometry metric.","tokens_in":15742,"tokens_out":3549,"would_cite":false,"duration_ms":33703,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"FlexiClip claims that clipart animation artifacts come from temporal noise in per-frame deformations, and it removes that noise by adding a learned temporal Jacobian field integrated over time with a probability-flow ODE and trained with…","keywords":["clipart animation","temporal Jacobian","probability flow ODE","flow matching loss","score distillation sampling","Bézier trajectories","video diffusion priors","GFlowNet"],"falsifier":"Check whether the learned temporal Jacobian actually converges toward zero for a stationary first frame and only grows where motion occurs, by recording $\\|J_t^R\\|$ over training and over the 24 frames; and run an ablation that replaces the ODE integrator with an ordinary residual MLP of the same capacity. If the ODE variant does not beat the residual variant on DS and TC, or if the corrections do not shrink as claimed, the pfODE reformulation is not load-bearing.","tokens_in":14624,"feed_emoji":"🎬","tokens_out":6508,"duration_ms":58133,"temperature":0.7,"pith_summary":"FlexiClip aims to show that the jerky, distorted motion in clipart animation is not an inevitable limit of Bézier-based deformation but a temporal-noise problem that can be corrected. It proposes adding a learned temporal Jacobian to each frame's spatial Jacobian, integrating that correction over time with a probability-flow ODE, and training the whole pipeline with a GFlowNet-inspired flow matching loss that forces corrections to shrink as motion stabilizes. On the AniClipart benchmark it reports higher visual identity (CLIP 0.9563 vs 0.9401), better text-video alignment (X-CLIP 0.2102 vs 0.2075), and smoother deformations (DS 13.76 vs 18.49), while acknowledging slightly higher geometric deviation (GD 52.34 vs 50.98) because it abandons ARAP rigidity. If it holds, the recipe makes text-driven clipart animation both smoother and more faithful to the prompt without requiring natural-video training data.","feed_headline":"Temporal corrections make clipart animation smoother and truer to text","feed_subtitle":"FlexiClip's Jacobian corrections lift CLIP score to 0.9563 and cut deformation smoothness error to 13.76.","key_machinery":"The load-bearing object is the temporal Jacobian $J_t^R$, a per-triangle corrective matrix added to the spatial Jacobian from Bézier trajectories. It is paired with two mechanisms: the pfODE reformulation in Eqs. (12)–(13), which integrates the correction over continuous time using attention over past windows in place of a score function, and the flow matching loss $L_{\\text{flow}} = \\mathbb{E}_{t',t}\\|\\nabla_X \\log p_{t'}(X,J_t) - \\nabla_X \\log p_{t'}(X,J_t^P)\\|^2 + \\|J_t - J_t^P\\|^2$, which pulls the score of fully corrected frames toward the score of spatial-only frames so that the correction term is driven to zero once temporal noise is removed.","core_discovery":"The paper's claim is that temporal coherence in clipart animation can be separated from spatial deformation: each frame's total Jacobian is $J_t = J_t^P + J_t^R$, where $J_t^P$ comes from Bézier keypoint trajectories and $J_t^R$ is a learned correction that starts at zero and evolves under an ODE $\\mathrm{d}J_t^R/\\mathrm{d}t = f_R(J_0^P, C_W^P, C_{W-1}^R, t; \\theta_R)$, integrated with Euler's method. The ODE is presented as a reformulation of the probability-flow ODE in which the usual score function is replaced by attention-encoded windows of spatial and temporal Jacobians, and a flow matching loss enforces that $J_t^R$ cancels only temporal noise while leaving the spatial pose intact. FlexiClip then demonstrates that this design beats AniClipart on identity, alignment, and smoothness metrics and wins a 30-participant user study on those same criteria.","pith_inferences":["Editorial inference: the same spatial/temporal Jacobian split could apply to natural-image or 3D shape animation, wherever a per-frame deformation predictor produces temporal jitter; the flow matching loss would then act as a universal smoothness regularizer.","Editorial inference: if Eqs. (12)–(13) are not a genuine pfODE (no score function appears), then the paper's theoretical framing reduces to adding a residual correction net with a flow-matching regularizer; a strong ablation comparing ODE vs residual integration would settle this.","Editorial inference: a testable extension is using a lightweight denoiser instead of the pretrained video diffusion model in Eq. (17), which would reveal how much of the quality gain comes from temporal correction versus SDS guidance."],"forward_implications":["Text-to-clipart animation can be made smoother by learning per-frame Jacobian corrections rather than predicting each frame's pose independently.","The flow matching loss, by comparing score estimates of corrected versus spatial-only frames, can serve as a temporal denoising regularizer for any Jacobian-based deformation model.","The pipeline supports rotation, multiple text conditions, multiple interacting objects, and layered animations in a single framework, so the correction mechanism generalizes beyond simple keypoint motions.","Dropping ARAP rigidity in favor of learned temporal corrections lowers geometric deviation quality slightly (GD 52.34 vs 50.98) but improves deformation smoothness and motion energy, indicating a deliberate trade-off between rigidity and expressiveness.","Because the method needs only one input image and a text prompt, it extends SDS-based animation to clipart without retraining the video diffusion model."],"supporting_citations":[{"why":"AniClipart baseline that FlexiClip extends; provides the Bézier trajectory framework, metrics, comparison numbers, and the 30 test images.","marker":"Wu et al., 2024"},{"why":"Source of the pfODE formulation the paper reuses, cited in-text via 'Song et al. (Lim et al., 2023)'.","marker":"Lim et al., 2023"},{"why":"Supplies the dimension-preserving pfODE with rescaling (Eq. 10) that FlexiClip adapts to temporal Jacobians.","marker":"de Albuquerque & Pearson, 2024"},{"why":"GFlowNet foundations; the detailed balance objective is adapted into the flow matching loss.","marker":"Bengio et al., 2023"},{"why":"DreamFusion SDS; supplies the gradient approximation used in the video SDS loss.","marker":"Poole et al., 2022"},{"why":"ModelScope text-to-video model, the pretrained video diffusion prior used for SDS and as the denoiser in the flow matching loss.","marker":"Wang et al., 2023"},{"why":"DiffVG differentiable renderer that rasterizes deformed meshes into frames for the SDS loss.","marker":"Li et al., 2020"}],"fun_headline_variants":["Jacobian ODEs keep clipart motion smooth frame by frame","FlexiClip's ODE corrections end clipart animation jitter","Temporal Jacobians make clipart move naturally without distortion","Flow matching loss fixes clipart animation's temporal glitches"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper assumes that the neural ODE in Eqs. (12)–(13) is a valid probability-flow ODE reformulation of Eq. (10), so that integrating the learned MLP over time removes temporal noise and drives the correction to zero; if that equivalence fails, the temporal Jacobian is just a residual network and the claimed theoretical grounding collapses.","fun_headline_variants_meta":{"raw":{"variants":["Jacobian ODEs keep clipart motion smooth frame by frame","FlexiClip's ODE corrections end clipart animation jitter","Temporal Jacobians make clipart move naturally without distortion","Flow matching loss fixes clipart animation's temporal glitches"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000473,"raw_usage":{"total_tokens":2392,"prompt_tokens":1032,"completion_tokens":1360,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":648,"completion_tokens_details":{"reasoning_tokens":1289}},"tokens_in":648,"tokens_out":1360,"duration_ms":12306,"temperature":1.0,"reasoning_tokens":1289,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:20:43.930454+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Check whether the learned temporal Jacobian actually converges toward zero for a stationary first frame and only grows where motion occurs, by recording $\\|J_t^R\\|$ over training and over the 24 frames; and run an ablation that replaces the ODE integrator with an ordinary residual MLP of the same capacity. If the ODE variant does not beat the residual variant on DS and TC, or if the corrections do not shrink as claimed, the pfODE reformulation is not load-bearing.","supporting_citations":[{"cited_title":"Score-based generative modeling through stochastic evolution equations in hilbert spaces","cited_arxiv_id":null,"evidence_quote":"Source of the pfODE formulation the paper reuses, cited in-text via 'Song et al. (Lim et al., 2023)'."},{"cited_title":"and Pearson, J","cited_arxiv_id":null,"evidence_quote":"Supplies the dimension-preserving pfODE with rescaling (Eq. 10) that FlexiClip adapts to temporal Jacobians."}],"review_version":1}