{"id":"9440ee10-910b-4c98-adf0-012f88dfc211","arxiv_id":"2412.02306","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"PaNDaS is a neural deformation model that combines per-triangle local features with a global code, enabling masked partial deformations and interpolations of 3D meshes without inference-time optimization.","lead":"PaNDaS learns per-triangle deformation features from a neutral 3D mesh and a global target code, allowing users to deform or interpolate only selected parts of a hand, body, or face. Because the model controls deformations locally, it enables partial shape mixing, motion transfer, and statistics without per-pose optimization.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The masking mechanism assumes G((ft,0)) = source for arbitrary source features, but no loss enforces it and no quantitative partial-deformation metric is reported; this unvalidated identity is load-bearing for all local-control claims.","rationale":"The reader's weakest assumption identifies essentially the same load-bearing gap: masked decoding assumes G((ft,0)) = X without an explicit training signal. My stress test sharpens this by noting that the assumption must hold for arbitrary source feature fields, including non-neutral sources, and that no quantitative partial-deformation metric is reported anywhere in Sec. 4.4 or Sec. 4.5. This is the single most load-bearing concern because the paper's core novelty is localized control via masking; if zeroing the global code does not return the source in the masked region, the partial deformation, pose mixing, and partial statistics demonstrations are not validated. The concern is not a claim of internal contradiction or fraud: the reconstruction and full-mesh interpolation numbers in Tables 1-3 are plausible and the architecture is coherent. But they only demonstrate that the network can deform a neutral template toward a target when the full global code is active. They do not establish that the same network, given only per-triangle source features with zero code, outputs the source on arbitrary meshes. The paper itself flags related limitations, including boundary artifacts in Sec. 5 and Poisson coupling in Sec. 3.5, which makes the absence of a direct identity test conspicuous. A single experiment measuring the zero-code identity error would settle whether the central mechanism works as assumed. Because the reader already assigned CONDITIONAL largely on this basis, my read does not change the verdict; it reinforces the condition that the masking identity and outside-mask leakage be demonstrated quantitatively before the localized-control claims are accepted.","tokens_in":13268,"tokens_out":5167,"duration_ms":69195,"concrete_test":"Measure the zero-code identity error on held-out non-neutral source meshes. For each test source X (e.g., MANO poses other than the neutral template), compute the full-mask output X_id = G((F(X),0)) and report E_id = ||X_id - X||_2 / ||X||_2. If E_id is at the same order as the full reconstruction error in Table 1, then masked-zero decoding does not preserve the source and partial-deformation results are biased. Separately, for a concrete partial mask M (e.g., the left arm), report the outside-mask residual E_out = || (1 - V_M)(G((F(X), M⊙z)) - X) ||_2 / ||X||_2, where V_M is the vertex-to-masked-triangle indicator, to quantify leakage outside the selected region. Repeat across identities and compare with a baseline that keeps outside vertices fixed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central partial-deformation operation in Sec. 3.5 decodes the feature field (ft, Mt ⊙ z), and when Mt = 0 it relies on the identity behavior G((ft,0)) = X for the source mesh X. This is never enforced. The training loss in Sec. 3.4, Eq. (3) with Lrec in Eq. (4) and Ln in Eq. (5), is evaluated only against target meshes Y in pairs (X,Y) where X is the dataset's neutral pose; there is no identity term pushing G((ft,0)) toward X. Moreover, because F(X) is trained only on neutral source meshes, there is no reason to expect the zero-code behavior to transfer to non-neutral source meshes, which are exactly the inputs used in the advertised motion-transfer and partial-statistics applications. Zeroing z in a masked region also means decoding (F(X),0) outside the training distribution: z is defined as zY - zX, so for a non-neutral X the feature field is not one seen during training. The Poisson solve couples all vertices, so even per-triangle identity Jacobians do not guarantee that anchored regions remain fixed; Sec. 3.5 explicitly concedes there is no strict confinement guarantee, and Sec. 5 concedes boundary artifacts. Figure 6 is qualitative and measures only decay of the deformation norm, not whether the outside-mask region actually equals X. The ablation in Table 5 tests reconstruction components but never tests masked decoding. Thus the method's unique contribution, localized control via masking, rests on an unverified and unquantified identity property.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PaNDaS, a learning-based method for non-rigid deformation of triangle meshes. For a source mesh X (typically the neutral pose) and a target mesh Y, the method extracts per-triangle local features f_t with DiffusionNet and a global deformation code z = EncD(Y) - EncD(X), where EncD aggregates Laplacian-frequency projections of per-triangle features. The deformation generator G predicts per-triangle Jacobians from the concatenated feature field (f_t, z), and a Poisson solve yields vertex displacements. At inference, masking the global-code part of the feature field (f_t, M_t ⊙ z) is proposed as a mechanism for partial/localized deformation, and the paper also demonstrates feature-field interpolation, pose mixing, motion transfer, and local statistics. Experiments on MANO, DFAUST, and COMA report reconstruction errors in Table 1, interpolation errors in Tables 2 and 3, and qualitative partial-deformation results in Figures 5, 6, and 9, with an ablation study in Table 5.","tokens_in":13610,"tokens_out":6751,"duration_ms":74072,"significance":"If the localized-control mechanism works as claimed, PaNDaS would provide a useful and flexible tool for partial non-rigid mesh deformation without per-instance optimization, on top of a clearly described architecture that builds on Neural Jacobian Fields and DiffusionNet but introduces a novel combination of local feature fields, global deformation codes, and Laplacian-based aggregation. The main reconstruction and interpolation evaluations are not circular, and the paper includes ablations and comparisons on three benchmark datasets. However, the central localized-control claim rests on an identity property that is neither enforced during training nor quantitatively validated, and the abstract's unqualified 'state-of-the-art' statement is contradicted by the COMA row of Table 1. These issues are load-bearing for the paper's main contribution, so the manuscript requires substantial additional evidence before it can be accepted.","major_comments":[{"comment":"The masked decoding operation (f_t, M_t ⊙ z) is never trained or quantitatively validated. When M_t = 0 the generator receives (f_t, 0), but no term in L in Eqs. (3)-(5) enforces G((f_t,0)) = X, and the training pairs use only the neutral pose as X, so the identity behavior is not guaranteed for non-neutral source meshes, which are exactly the inputs used for motion transfer and partial statistics. The Poisson solve couples all vertices, so even per-triangle identity Jacobians would not confine the deformation to the mask, and Fig. 6 reports only the norm of the deformation, not whether the outside-mask region equals X. The ablation in Table 5 does not test masked decoding. Please add an explicit identity constraint (e.g., a loss term on G((F(X),0)) = X) or provide a quantitative evaluation of leakage, such as outside-mask error against the source on unseen identities and non-neutral sources.","section":"Sec. 3.5 and Sec. 3.4, Eqs. (2)-(5)"},{"comment":"The abstract claims 'state-of-the-art accuracy' and Sec. 1 states that PaNDaS 'set a new state-of-the-art,' but Table 1 shows that ARAPReg outperforms PaNDaS on COMA on all three reported metrics (MSE 0.128 vs 0.160, HD 0.0070 vs 0.0086, CD 0.474 vs 0.839). The authors themselves describe the COMA result as 'competitive' in Sec. 4.2. This unqualified SOTA claim should be corrected to reflect that PaNDaS is best or tied on MANO and DFAUST but not on COMA, or the authors should explain why the COMA setting is deprioritized.","section":"Abstract and Sec. 1 vs. Table 1"},{"comment":"Partial deformations are evaluated only qualitatively; there is no metric measuring how well the masked region matches the target or how well the anchored region is preserved. Since localized control is the paper's central novelty, the authors should report region-wise errors (e.g., MSE inside and outside the mask) and compare leakage against a baseline that enforces fixed vertices. This is closely related to the first major comment but concerns the experimental evaluation rather than the training objective.","section":"Sec. 4.4 and Fig. 5"}],"minor_comments":[{"comment":"The number of Laplacian frequencies s=4 is stated as sufficient, but no sensitivity analysis is provided; please report how reconstruction and interpolation quality vary with s.","section":"Sec. 3.2"},{"comment":"The caption refers to a 'per-triangle displacement field (v_i)' although v_i is defined as a per-vertex displacement in Sec. 3.1; please reword for consistency.","section":"Fig. 3 caption"},{"comment":"Some individual sequences are better with ARAPReg (e.g., DFAUST 'one leg jumping' and 'one leg landing', COMA 'high smile' and 'lips up'); the text should acknowledge these cases explicitly rather than reporting only the mean values.","section":"Tables 2 and 3"},{"comment":"The limitations section mentions boundary artifacts from masking but does not discuss how these artifacts affect the mixing and statistics applications; a brief note would help readers understand the practical impact.","section":"Sec. 5"},{"comment":"The name 'PaNDAS' is written in all caps in Fig. 1 while the rest of the paper uses 'PaNDaS'; please unify the capitalization.","section":"Fig. 1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is likely within the journal's scope, and the missing validation of the localization identity is obtainable with additional experiments rather than being an unfixable flaw. I recommend major revision: the authors should add a quantitative evaluation of masked decoding, consider enforcing the identity behavior in training, and qualify the state-of-the-art claim in light of the COMA results."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. The architecture is genuinely new: per-triangle features from DiffusionNet, concatenated with a global code computed as a difference of spectral projections, feeding a Jacobian-field decoder. That combination buys something real — localized control at the feature level, no optimization at inference, and competitive full-deformation and interpolation numbers on MANO and DFAUST, with COMA close behind ARAPReg. The ablation is honest and the remeshing robustness is a plus.\n\nThe soft spot is the one the stress-test flags, and it lands. The partial-deformation mechanism sets the global code to zero on masked triangles and relies on the decoder treating (ft,0) as “output the source.” Nothing in the training loss enforces that. The encoder only sees neutral X during training, so for a non-neutral source the zero-code feature field is out-of-distribution. And because the Poisson solve couples all vertices, even per-triangle identity Jacobians do not pin down the region outside the mask. The paper acknowledges boundary artifacts but never measures how much deformation bleeds outside the mask. Figure 6 shows norm decay, not whether the outside region actually equals X. That missing quantitative test is the difference between “we demonstrate localized control” and “we have a plausible mechanism that we show qualitatively.” The reader’s note about COMA is fair but minor; the authors themselves say the setting is easier, and being competitive on faces is still a reasonable outcome.\n\nFor full deformation and interpolation, I trust the numbers. They are supervised, evaluated against baselines, and the normal loss and ablation make sense. The lack of error bars is annoying but common in this subfield.\n\nWho should read this: anyone building learned deformation models for meshes, especially for animation or editing. It deserves a serious referee. A revision should add a quantitative partial-deformation metric (e.g., mean displacement outside the mask relative to inside) and either an identity loss or an explicit test of the zero-code behavior on non-neutral sources.\n\nRecommendation: send to peer review with expectations of revision. The core is solid; the headline claim needs measurement.","headline":"A genuinely useful architecture for local mesh deformation control, but the flagship partial-deformation claim is plausible rather than proven: the masking identity is never enforced and never measured.","tokens_in":14134,"tokens_out":2145,"would_cite":true,"duration_ms":25793,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"PaNDaS represents a mesh deformation as a per-triangle feature field with a maskable global pose code, so selected regions can be deformed, poses mixed, and localized statistics computed with no optimization at inference.","keywords":["non-rigid shape deformation","partial shape deformation","3D mesh interpolation","Jacobian fields","localized deformation control","pose mixing","shape statistics","learning-based deformation"],"falsifier":"Take a trained PaNDaS model, pick a non-neutral source mesh $X$ (for example a hand with bent fingers), set the global code to $z=0$, and decode; if the output differs from $X$ by more than the model's usual reconstruction error, then the zero-code identity that masking relies on fails for non-neutral sources. A complementary check is to apply a mask to a small patch on such a source and measure the displacement of vertices far outside the mask: large motion there would contradict the claimed locality.","tokens_in":13074,"feed_emoji":"🧍","tokens_out":9298,"duration_ms":94343,"temperature":0.7,"pith_summary":"PaNDaS (Partial Non-rigid Deformations and interpolations of Surfaces) is a learning-based model for non-rigid deformation of triangular surface meshes that is designed to deform only selected parts of a shape. Its central idea is to represent a deformation not by one global latent vector but by a per-triangle feature field obtained by concatenating local features of the source mesh with a global code that encodes the difference between target and source poses. Because the global code is replicated across triangles, the user can multiply it by a binary mask and thereby restrict the deformation to a chosen region, interpolate between poses, mix several pose codes on different regions, and compute localized shape statistics. The paper reports that this field-based representation outperforms global-latent baselines in reconstruction accuracy on hand and body datasets, is competitive on face expressions, and produces smoother interpolations for strongly nonlinear motions, all without any optimization at inference time.","feed_headline":"Mask a region, and PaNDaS deforms only that part of the mesh","feed_subtitle":"It deforms only the masked region, mixes poses from different parts, and runs without any per-query optimization.","key_machinery":"The load-bearing object is a feature field over the source mesh whose per-triangle vectors combine a local geometric descriptor with a global deformation code; the field is what makes masking and mixing possible. The global code is produced by a deformation encoder that projects per-triangle target features onto the first $s=4$ eigenvectors of the cotangent Laplacian, $p_k = \\frac{1}{\\mathrm{Area}(Y)}\\sum_t \\mathrm{Area}(t)\\, g_t e^k_t$, then passes $(p_1,p_2,p_3,p_4)$ through a linear layer and takes the difference $z = z_Y - z_X$, so identical source and target give $z=0$. The deformation generator is a neural Jacobian field: it predicts per-triangle $3\\times 3$ Jacobians from the feature field and recovers a consistent displacement field by solving the Poisson equation $\\nabla_X v = \\nabla^T M J$. This pipeline lets the user manipulate the feature field directly (mask, interpolate, sum codes) while the Poisson solve couples the result into a smooth mesh.","core_discovery":"The paper's central claim is that localized control of non-rigid deformation can be achieved by making the deformation generator depend on a per-face feature field $\\tilde f_t = (f_t, z)\\in\\mathbb{R}^{l+r}$, where $f_t$ is a local geometric feature on triangle $t$ of the source mesh and $z = \\mathrm{Enc}_D(Y) - \\mathrm{Enc}_D(X)$ is a global deformation code. The generator $G$ maps this field to a per-triangle Jacobian field $(J_t)_t$, which is integrated into vertex displacements through a Poisson solve, $\\nabla_X v = \\nabla^T M J$. Masking the code, $\\tilde f_t^{\\mathrm{partial}} = (f_t, M_t\\odot z)$, yields deformations that the paper shows are concentrated near the masked region, with the displacement norm decaying rapidly away from the mask boundary. From this construction the paper derives partial interpolation, pose mixing by summing masked codes, partial motion transfer, and a statistical model of deformations in which Euclidean operations on feature fields correspond to natural shape operations. The paper supports the claim with experiments on MANO hand meshes, DFAUST bodies, and COMA faces, reporting state-of-the-art reconstruction error on MANO and DFAUST and improved interpolation on strongly nonlinear sequences.","pith_inferences":["Because the zero-code identity $G((f_t,0))=X$ is only learned implicitly from neutral source meshes, the method's partial-deformation guarantees are likely to degrade when the source mesh is already far from neutral; a direct test would be to decode $z=0$ from a strongly posed source and measure the residual displacement.","The hard binary mask creates boundary artifacts, as the paper itself notes; a weighted or learned mask would likely smooth the transition region and could be trained with a small auxiliary loss on the deformation outside the mask.","The global code discards all Laplacian frequency components above $s=4$, so very fine high-frequency pose details are filtered out; keeping per-part or per-frequency codes could improve fidelity for detailed expressions while preserving locality.","Since the final displacements come from a Poisson solve, the deformed region is never strictly compactly supported; measuring the exact support size on a dense sequence of masks would quantify how 'local' the control really is."],"forward_implications":["A user can deform only the triangles covered by a binary mask, and the resulting displacement decays quickly away from the mask boundary, so partial edits of a mesh need no skeleton, texture, or per-query optimization.","New poses can be generated by summing several masked pose codes on different regions, giving a mix-and-match strategy for combining, say, one pose's arm motion with another pose's leg motion.","Partial motion transfer is available: because the encoder is trained to encode pose, one can restrict the transferred motion to a chosen part of the target identity.","Euclidean statistics on feature fields translate back into shape statistics, so one can compute mean shapes and principal deformation components that respect locality and work on unregistered scans.","On the reported benchmarks, the method beats global-latent deformation generators in reconstruction MSE on MANO and DFAUST and stays competitive on COMA; interpolation errors on nonlinear DFAUST motions are the lowest among compared methods."],"supporting_citations":[{"why":"Supplies the per-triangle Jacobian prediction and Poisson integration that the deformation generator is built on, and is the primary baseline for full deformations.","marker":"[1]"},{"why":"Provides the heat-diffusion face-wise feature extractor, robust to remeshing, that produces the local features and target features used in the field.","marker":"[34]"},{"why":"The registered human-body 4D dataset used for training and evaluating body reconstruction and interpolation.","marker":"[7]"},{"why":"The face dataset used to evaluate expression deformations and interpolation sequences, where the method is competitive.","marker":"[31]"},{"why":"The hand mesh dataset used to evaluate reconstruction and interpolation, where the method reports its best MSE.","marker":"[32]"},{"why":"A regularized latent-space baseline for deformation generators, the strongest competitor on COMA and a main interpolation baseline.","marker":"[19]"},{"why":"A graph-convolution latent baseline used for full and partial deformation comparisons, including a partial interpolation failure case.","marker":"[46]"},{"why":"Supplies the localized As-Rigid-As-Possible vertex displacement implementation used as the baseline for partial deformations.","marker":"[28]"}],"fun_headline_variants":["Deform only the masked region without per-query optimization","PaNDaS: localized deformation by masking, no optimization","Point-level control for partial non-rigid deformation","Mix poses and deform only the masked part of the mesh","Mask a region, deform locally, no inference-time solving"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The masking trick works only if setting the global deformation code to zero makes the generator output the source mesh for any source shape, but this identity is not enforced by a loss term and is only implicitly learned from training pairs in which the source is a neutral pose.","fun_headline_variants_meta":{"raw":{"variants":["Deform only the masked region without per-query optimization","PaNDaS: localized deformation by masking, no optimization","Point-level control for partial non-rigid deformation","Mix poses and deform only the masked part of the mesh","Mask a region, deform locally, no inference-time solving"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000665,"raw_usage":{"total_tokens":3049,"prompt_tokens":970,"completion_tokens":2079,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":586,"completion_tokens_details":{"reasoning_tokens":1999}},"tokens_in":586,"tokens_out":2079,"duration_ms":16537,"temperature":1.0,"reasoning_tokens":1999,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T23:37:11.803719+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a trained PaNDaS model, pick a non-neutral source mesh $X$ (for example a hand with bent fingers), set the global code to $z=0$, and decode; if the output differs from $X$ by more than the model's usual reconstruction error, then the zero-code identity that masking relies on fails for non-neutral sources. A complementary check is to apply a mask to a small patch on such a source and measure the displacement of vertices far outside the mask: large motion there would contradict the claimed locality.","supporting_citations":[{"cited_title":"Kim, Siddhartha Chaudhuri, Jun Saito, and Thibault Groueix","cited_arxiv_id":null,"evidence_quote":"Supplies the per-triangle Jacobian prediction and Poisson integration that the deformation generator is built on, and is the primary baseline for full deformations."},{"cited_title":"Diffusionnet: Discretization agnostic learning on surfaces","cited_arxiv_id":null,"evidence_quote":"Provides the heat-diffusion face-wise feature extractor, robust to remeshing, that produces the local features and target features used in the field."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The registered human-body 4D dataset used for training and evaluating body reconstruction and interpolation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The hand mesh dataset used to evaluate reconstruction and interpolation, where the method reports its best MSE."},{"cited_title":"Huang, X","cited_arxiv_id":null,"evidence_quote":"A regularized latent-space baseline for deformation generators, the strongest competitor on COMA and a main interpolation baseline."},{"cited_title":"Fully convolutional mesh autoencoder using efficient spatially varying kernels","cited_arxiv_id":null,"evidence_quote":"A graph-convolution latent baseline used for full and partial deformation comparisons, including a partial interpolation failure case."},{"cited_title":"Smooth non-rigid shape matching via effective dirichlet energy optimization","cited_arxiv_id":null,"evidence_quote":"Supplies the localized As-Rigid-As-Possible vertex displacement implementation used as the baseline for partial deformations."}],"review_version":1}