{"id":"0d533521-b46d-4a89-8a91-44f43fc5ad13","arxiv_id":"2505.20283","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"An unsupervised method that represents deformable 3D objects as sparse, editable blobs and re-poses them across categories without manual rigging.","lead":"The paper introduces CANOR, a method that learns a rigging representation for 3D deformable objects from animated 3D sequences, without category-specific knowledge or manual rigging. A generalist might read it because it could make animating new types of 3D objects as easy as dragging a few control points.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central claim of intuitive blob-based editing is not tested: training only decodes regressed target blobs from the same identity, so pose/identity disentanglement and arbitrary-edit generalization remain unvalidated.","rationale":"The reader conditioned acceptance on releasing code/data and adding error bars; those are reasonable but do not target the central claim's weakest point. The paper's main contribution is an editable rig, yet every quantitative experiment uses the same reconstruction-style protocol: encoder predicts blobs from a source and from a target of the same identity, decoder reconstructs the target. That protocol cannot distinguish a true pose/identity disentanglement from a shape autoencoder that memorizes per-identity appearance. It also never presents a user-edited blob configuration to the decoder at evaluation time. Because the training objective (Eq. 11) contains reconstruction and voxel occupancy losses only, there is no reason to believe the latent decomposition is semantically disentangled. The reader's weakest assumption about blob count and category scope is relevant, but it is secondary: even within one category, the editing claim is not evidenced. I therefore keep the CONDITIONAL verdict but shift the condition: before accepting the paper's headline, the authors should demonstrate cross-identity pose transfer and manual-edit plausibility, not just report re-posing numbers on the same identity.","tokens_in":13003,"tokens_out":7820,"duration_ms":85496,"concrete_test":"On DeformingThings4D, train with a held-out set of full sequences. For a held-out source identity A and a different held-out target identity B, decode with BI(A) and BP(B); compute Chamfer distance from the output to A's own meshes and to B's meshes, and compare against same-identity re-posing (BI(A) plus BP(A in the target pose)). If the cross-identity output is closer to B than to A, or if its distance to A's meshes is substantially worse than the reported 0.937 IoU / 0.017 CD1, then BP carries identity information and the disentanglement claim fails. Separately, apply a manual edit (translate or rotate one blob by a fixed offset) on a held-out shape; if the decoded mesh is not a plausible instance of A or degrades sharply relative to regressed-pose reconstruction, the intuitive-editing claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (Abstract) is that users can manipulate pose by editing blob parameters while preserving instance identity. The training protocol (Sec. 3.4, Eq. 10) only forms tuples (PM, {EI(PM), EP(PM')}, M') with the same identity M, and BP is regressed from an actual target frame M'. There is no explicit disentanglement loss between BP and BI; the separation is only architectural (Sec. 3.2). Two requirements of the claim are therefore never verified. (1) Identity leakage: nothing prevents BP from carrying identity information, since BP is always predicted from a mesh of the same identity during training; if identity leaks into BP, editing blobs will alter identity, contradicting 'preserving rich instance-specific information.' (2) Arbitrary-edit generalization: the decoder is trained to map BP values that come from real target frames to occupancy; user-edited blob positions and rotations that lie off the training manifold may decode to implausible shapes. Section 4.2's evaluation regresses BP from the target shape of the same identity, mirroring training, so it cannot detect either failure. The clay-monster demo (Sec. 4.3) is qualitative and not quantified. Without these checks, the headline capability is an assumption rather than a demonstrated result.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CANOR, an unsupervised method that learns a blob-based rigging representation for deformable object categories from 3D sequences. An encoder maps an input point cloud to a sparse set of anisotropic blobs, each parameterized by position, rotation, scale, opacity, and a feature vector; the parameters are split into pose-related (position, rotation) and identity-related (scale, opacity, feature). A transformer decoder voxelizes edited blobs into a feature volume, conditions on point-wise features from the encoder, and predicts an occupancy field from which a mesh is extracted. Training uses two frames of the same identity: pose blobs are regressed from the target frame, identity blobs from the source frame, and the decoder must reconstruct the target. The method is evaluated by re-posing the first frame of held-out sequences to later frames, comparing IoU and Chamfer distances against SkeRig, KeypointDeformer, and an adapted Neural Deformation Graph. Ablations on DeformingThings4D and a qualitative clay-monster editing demonstration are also presented.","tokens_in":13322,"tokens_out":8431,"duration_ms":89319,"significance":"CANOR targets a genuine gap: automatic, category-agnostic rigging without manual rigs, correspondences, or category-specific priors. The representation is interpretable, and the self-supervised training protocol in Eq. (10) is a reasonable way to obtain training signal. If the editing capability holds, the method could be practically valuable, especially since the clay-monster experiment shows the pipeline works with casual iPhone scans. Strengths include the diversity of evaluated categories, the clear ablations in Table 2 showing the benefit of identity conditioning and anisotropic blobs, and the consistent quantitative improvements over the baselines on most datasets. However, the central claim of intuitive blob editing preserving identity is not directly validated, and the quantitative comparison rests on small test sets without error bars. The contribution is promising but not yet fully established.","major_comments":[{"comment":"The training objective never requires the model to disentangle pose from identity. Pose parameters BP are always regressed from a target frame M′ that shares the identity of M, and the reconstruction loss can be minimized even if BP encodes identity information. The decoder's additional conditioning on the source point features F (Sec. 3.3) may hide this leakage in the same-identity setting, but then the claimed disentanglement is not actually verified by the reported experiments. I recommend adding a cross-identity re-posing experiment (identity blobs from instance A, pose blobs from instance B) and reporting whether the decoded shape preserves A's identity while taking B's pose; in addition, probe the information content of BP, for example by testing whether BP alone can be used to predict instance identity.","section":"Sec. 3.4 (Eq. 10); Sec. 4.2"},{"comment":"The evaluation only exercises the same operation used in training: pose parameters are produced by the encoder from a real target frame of the same identity. The clay-monster demonstration (Sec. 4.3, Fig. 5) is purely qualitative. Because the decoder is trained exclusively on BP values that lie on the image of EP for real target meshes, arbitrary user edits (e.g., dragging a blob far from its training distribution) may decode to implausible shapes, and the current metrics cannot detect this failure. Add a quantitative editing study: apply random and structured translations/rotations to blob parameters, measure reconstruction quality and identity preservation, and if possible include interpolation experiments between two real poses of the same identity with comparison to ground-truth intermediate frames, or a small user study, or a comparison against baselines under identical edits.","section":"Sec. 4.2; Sec. 4.3"},{"comment":"Results are reported as means over only the five longest held-out sequences per dataset, without error bars, standard deviations, or multiple training seeds. With such a small test set, the claimed large margins over baselines are not statistically supported; for example, on Eyeglasses the IoU of Ours (0.770) is below Neural Deformation Graph (0.791), so the 'large margin' claim does not hold uniformly. Report per-sequence results or standard deviations and, ideally, results across at least three seeds.","section":"Table 1; Sec. 4.2"}],"minor_comments":[{"comment":"There is an extra bracket in 'Q[i] · F[j]]T'; the expression should read Q[i] · F[j]^T.","section":"Eq. (3)"},{"comment":"The exponent is written as (gi − xj/sj)(gi − xj/sj)^T, which is matrix-valued; presumably the squared norm ||(gi − xj)/sj||^2 is intended.","section":"Eq. (7)"},{"comment":"In the sentence about the final MLPs, 'the aggregated feature vector sets FP and FP' should read 'FP and FI'.","section":"Sec. 3.2"},{"comment":"There are several typos: 'curcial' in Sec. 4.1, 'Addtional' in Sec. 3.3, 'deformes' in Sec. 4.2, and 'state of the arts' in the Table 1 caption.","section":"Sec. 4.1; Sec. 3.3; Sec. 4.2; Table 1 caption"},{"comment":"The number of blobs nb is set per category (8 to 24), so the 'category-agnostic' claim is qualified by this per-category manual choice; Table 2 reports sensitivity only on DeformingThings4D, and a brief per-category sensitivity discussion would be useful.","section":"Sec. 4.1"}],"recommendation":"major_revision","confidential_remarks":"The core idea and training pipeline are sound, and the paper addresses an important problem. The main risk is that the headline editing capability is asserted without direct validation; the requested cross-identity and editing experiments are necessary before the central claim can be accepted. I would encourage the editor to request these experiments rather than reject, as they are well within the scope of a revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe thing to know about this paper is that the core representation is promising and the re-posing results are strong, but the headline claim that users can edit blobs to manipulate poses while preserving identity is not actually tested by the experiments. The training and evaluation protocol only ever uses pose parameters regressed from real target frames of the same identity, so identity leakage into the pose blobs and off-manifold generalization of arbitrary edits both go unvalidated. That is a real gap, not a nitpick.\n\nWhat's actually new: the specific combination of spatial blobs plus an instance-aware feature volume for 3D deformable objects, with a two-frame training scheme that encourages pose/identity separation, and a learned decoder that voxelizes edited blobs into an occupancy field. I think the paper is honest about the components building on BlobGAN and KeypointDeformer, and the extension to 3D dynamic categories is non-trivial. The ablations in Table 2 show the identity conditioning and anisotropic blobs each matter, and the blob count sensitivity is sensible.\n\nThe soft spots, in order of importance. First, the editing claim. There is no explicit disentanglement loss between pose and identity parameters; the separation is only architectural. Since BP is always predicted from a mesh of the same identity during training, nothing prevents identity information from leaking into the pose blobs. And since the decoder only sees BP values from actual target frames, arbitrary user edits may decode to implausible shapes. Section 4.2's evaluation regresses BP from the target shape of the same identity, mirroring training, so it cannot detect either problem. The clay-monster demo is qualitative only. The fix is not hard: test identity preservation by editing blobs across different identities, and test off-manifold edits by interpolating or perturbing blob parameters and measuring shape plausibility. Without that, the central use case is assumed, not demonstrated.\n\nSecond, the evaluation has no error bars or multiple seeds, and some baseline setups are unusual (e.g., SkeRig with correspondences from NDP, NDG adapted to rigging). Those comparisons could be fair, but they need more detail to be sure. Third, no code or data release, so reproducing Table 1 is not possible.\n\nWho is this for: anyone working on automatic rigging, 4D shape representation, or interpretable 3D generative models. It deserves a serious referee: the representation is novel enough and the experiments broad enough to warrant the time, but the revision needs to address the editing validation and release artifacts. I'd accept it for peer review with heavy revision expected.\n\nBest.","headline":"A promising blob-based rigging representation whose headline 'intuitive editing' claim is not actually tested by the experiments.","tokens_in":13840,"tokens_out":3580,"would_cite":true,"duration_ms":31511,"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":"The paper claims that the pose of a deformable 3D object can be captured by a sparse set of editable blobs learned from animated shape sequences, with no category-specific rigging knowledge.","keywords":["neural rigging","4D deformable objects","blob representation","pose-identity disentanglement","category-agnostic learning","unsupervised shape representation","shape re-posing","occupancy fields"],"falsifier":"Train the method on a category with many independently articulated parts, such as a hand with twenty-odd joints, using the maximum 24 blobs, and measure re-posing IoU on held-out sequences; if users cannot express finger poses by editing blobs or the IoU falls well below the roughly 0.86 level reported elsewhere, the claim that a small fixed blob set captures generic deformable pose spaces is falsified.","tokens_in":12802,"feed_emoji":"🕹️","tokens_out":5326,"duration_ms":48344,"temperature":0.7,"pith_summary":"The paper tries to establish that the pose of a deformable 3D object can be captured by a sparse set of spatially grounded blobs, learned automatically from animated 3D shape sequences of a category, with no hand-crafted skeleton or category-specific rigging expertise. Each object is encoded as a set of anisotropic blobs that carry pose parameters (position and rotation) plus identity parameters (scale, opacity, feature), and a decoder turns edited blobs back into a full mesh. If this holds, a user can intuitively re-pose objects simply by dragging blobs, and a new category can be rigged from a handful of scans. The authors report re-posing accuracy above 0.86 IoU on four of five tested categories, with the fifth at 0.770, outperforming three rigging baselines.","feed_headline":"Sparse 3D blobs re-pose objects without a rig","feed_subtitle":"A category-agnostic neural method separates pose from identity, letting users drag blobs to animate quadrupeds, faces, fish, and more.","key_machinery":"The load-bearing object is the blob: an anisotropic sphere parameterized as $b = (x, r, s, o, f)$, where $x$ is the center, $r$ a rotation quaternion, $s$ the radius, $o$ the opacity, and $f$ a local feature vector. The encoder uses a learnable codebook of 8 to 24 query tokens that cross-attend over point cloud features, aggregating positional encodings into pose features and fused point features into identity features; the decoder voxelizes blobs into an $8 \\times 8 \\times 8$ feature volume that is refined by self-attention and conditioned on the original point features, then queried to predict occupancy. This machinery makes pose and identity separable: editing $x$ and $r$ changes pose, while $s$, $o$, and $f$ carry the instance.","core_discovery":"The central claim is that the low-dimensional structure of deformable motion, traditionally encoded by artist-designed skeletons and skinning weights, can be discovered purely from data: an encoder maps a point cloud to a sparse set of feature-embedded anisotropic blobs, deliberately split into pose-related parameters (center, rotation) and identity-related parameters (radius, opacity, feature), and a transformer decoder voxelizes edited blobs into an occupancy field to reconstruct the re-posed surface. Because the codebook of blobs is shared across instances and learned end-to-end, the representation is category-agnostic: the same pipeline is applied to quadrupeds, faces, fish, refrigerators, and eyeglasses without changing the model or adding priors. The paper demonstrates that dragging blob positions and rotations produces plausible pose edits while identity conditioning, via cross-attention between the decoded feature volume and encoder point features, preserves instance-specific surface details.","pith_inferences":["A natural stress test is whether blob indices stay semantically consistent across instances; if they do, the representation yields dense correspondences for free, a property the paper does not explicitly evaluate.","The per-category training leaves open a stronger cross-category model: one codebook that could rig unseen categories by fine-tuning only the blob count, which is not what the paper demonstrates.","Because pose and identity parameters are separated, swapping identity blobs between two instances of a category should transfer one object's appearance onto another's pose, a testable extension the paper does not report.","The fixed upper bound of 24 blobs suggests categories with many independent parts or topological changes may need hierarchical or variable-count blobs to stay expressible."],"forward_implications":["Users can re-pose a trained object category by dragging blob positions and rotations, with no 3D modeling expertise.","A new deformable category can be rigged from a few scanned sequences; the paper demonstrates this with 12 iPhone-scanned clay figures.","The same unsupervised pipeline handles both rigid articulated motion (refrigerators, eyeglasses) and non-rigid deformation (fish, faces, quadrupeds).","Identity conditioning preserves fine-grained surface detail during re-posing; ablations show removing it drops IoU from 0.937 to 0.853 on DeformingThings4D.","Reported re-posing accuracy beats all three baselines on nearly every metric across the five datasets."],"supporting_citations":[{"why":"Defines the skeleton-and-skinning rigging paradigm that the blob representation is designed to replace.","marker":"[31]"},{"why":"Introduces blobs as a mid-level neural representation for 2D scenes, the direct inspiration for spatially grounded blobs.","marker":"[11]"},{"why":"Demonstrates blob-like latent particles for 2D face manipulation, motivating editable blob parameters.","marker":"[9]"},{"why":"Supplies the quadruped animation dataset used for training and the primary ablation study.","marker":"[24]"},{"why":"Supplies the facial expression dataset used to test non-rigid re-posing.","marker":"[6]"},{"why":"Supplies the fish dataset used as a non-rigid category.","marker":"[38]"},{"why":"Supplies the articulated object categories (refrigerator, eyeglasses) used to test rigid motion.","marker":"[43]"},{"why":"Serves as the keypoint-based baseline whose cage deformation is compared against the learned decoder.","marker":"[17]"},{"why":"Serves as the neural deformation graph baseline for implicit rigging.","marker":"[7]"},{"why":"Serves as the skeleton-based rigging baseline requiring dense correspondences.","marker":"[20]"}],"fun_headline_variants":["Blobs replace rigs for category-agnostic animation","Drag blobs to re-pose any object","Neural blobs unlock rig-free object animation","Category-agnostic reposing via learned blobs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The pipeline assumes that a fixed small set of 8 to 24 blobs per category can faithfully represent the pose space and identity of any deformable object without any hierarchical structure, an assumption tested on only five categories.","fun_headline_variants_meta":{"raw":{"variants":["Blobs replace rigs for category-agnostic animation","Drag blobs to re-pose any object","Neural blobs unlock rig-free object animation","Category-agnostic reposing via learned blobs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000141,"raw_usage":{"total_tokens":1143,"prompt_tokens":902,"completion_tokens":241,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":518,"completion_tokens_details":{"reasoning_tokens":179}},"tokens_in":518,"tokens_out":241,"duration_ms":6328,"temperature":1.0,"reasoning_tokens":179,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:55:41.923489+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the method on a category with many independently articulated parts, such as a hand with twenty-odd joints, using the maximum 24 blobs, and measure re-posing IoU on held-out sequences; if users cannot express finger poses by editing blobs or the IoU falls well below the roughly 0.86 level reported elsewhere, the claim that a small fixed blob set captures generic deformable pose spaces is falsified.","supporting_citations":[{"cited_title":"Joint-dependent local deformations for hand an- imation and object grasping","cited_arxiv_id":null,"evidence_quote":"Defines the skeleton-and-skinning rigging paradigm that the blob representation is designed to replace."},{"cited_title":"Blobgan: Spatially disentangled scene representations","cited_arxiv_id":null,"evidence_quote":"Introduces blobs as a mid-level neural representation for 2D scenes, the direct inspiration for spatially grounded blobs."},{"cited_title":"4dcomplete: Non-rigid motion esti- mation beyond the observable surface","cited_arxiv_id":null,"evidence_quote":"Supplies the quadruped animation dataset used for training and the primary ablation study."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the facial expression dataset used to test non-rigid re-posing."},{"cited_title":"Sketchfab - the best 3d viewer on the web","cited_arxiv_id":null,"evidence_quote":"Supplies the fish dataset used as a non-rigid category."},{"cited_title":"Shape2motion: Joint analysis of motion parts and attributes from 3d shapes","cited_arxiv_id":null,"evidence_quote":"Supplies the articulated object categories (refrigerator, eyeglasses) used to test rigid motion."},{"cited_title":"Keypointdeformer: Unsupervised 3d keypoint discovery for shape control","cited_arxiv_id":null,"evidence_quote":"Serves as the keypoint-based baseline whose cage deformation is compared against the learned decoder."},{"cited_title":"Palafox, Michael Zollh ¨ofer, Justus Thies, Angela Dai, and Matthias Nießner","cited_arxiv_id":null,"evidence_quote":"Serves as the neural deformation graph baseline for implicit rigging."},{"cited_title":"Robust and accurate skele- tal rigging from mesh sequences.ACM Trans","cited_arxiv_id":null,"evidence_quote":"Serves as the skeleton-based rigging baseline requiring dense correspondences."}],"review_version":1}