{"id":"a48cb9d1-a3de-440c-80bd-352416952016","arxiv_id":"2411.18197","paper_version":3,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A data-driven framework automatically rigs and skins diverse 3D humanoid characters, including mesh and Gaussian splat representations, in under one second.","lead":"This paper presents a fast automatic system that turns any 3D humanoid character into an animatable one, adding bones, skin weights, and a neutral pose in about a second. It works on both mesh models and 3D Gaussian splats, and handles unusual shapes and poses that earlier tools struggle with.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Teacher-forcing mismatch in coarse-to-fine and autoregressive modules: training uses ground-truth bones, inference uses predicted bones, and no experiment isolates this, so reported accuracy may not hold for arbitrary shapes/poses.","rationale":"The reader identifies the frozen ShapeNet encoder as the weakest assumption, and I agree that generalization to arbitrary shapes and poses is the core risk to the central claim. However, the most immediately verifiable weakness is the teacher-forcing gap: the fine stage and the structure-aware transformer are trained with ground-truth bone information and deployed without it, yet no experiment separates the contribution of oracle information from genuine generalization. This is a load-bearing concern because the headline promises 'regardless of shapes and poses' and the coarse-to-fine representation is precisely what supposedly enables unusual poses and fine-grained hands. If the reported numbers depend on oracle canonicalization or oracle parent attributes, the actual system on out-of-distribution inputs is likely worse than the quantitative evaluation suggests. The frozen-encoder concern is plausible but less decisive: ShapeNet-derived features may transfer low-level geometry, and the paper's failures could equally arise from exposure bias or from the limited 95-character Mixamo training set. The proposed check isolates the teacher-forcing effect directly. Since the paper already acknowledges failure cases and the method may still be strong for Mixamo-like characters, the CONDITIONAL verdict remains appropriate; the condition should include quantifying this mismatch and releasing the pipeline for verification.","tokens_in":19611,"tokens_out":7575,"duration_ms":75760,"concrete_test":"On the Mixamo test split, run the released model fully autoregressively (predicted coarse joints drive canonicalization/hierarchical sampling, predicted parent bones drive the structure-aware transformer). Then rerun the same checkpoints with teacher forcing: (i) replace coarse-predicted joints with ground-truth joints only for canonicalization and hand sampling, and (ii) replace predicted parent-bone attributes with ground-truth attributes in the causal transformer. Compare joint/pose/weight errors. If either ablation improves scores substantially (e.g., joint error below 1.5% vs 2.11%), the deployed pipeline's accuracy is not reliably characterized, and claims about out-of-distribution shapes and poses need re-scoping.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.3 and S2.1 describe a coarse-to-fine representation in which the fine stage canonicalizes the input and performs hierarchical hand sampling using bone positions. S2.1 states explicitly that in training these positions are ground-truth: 'we apply canonical transformation and hierarchical sampling to the shapes in advance based on the ground-truth bone positions.' At inference, the coarse stage's predicted bones are used instead. Similarly, Sec. 3.4's structure-aware transformer is trained with ground-truth parent-bone latents (Fig. 3: 'encoded from the autoregressive outputs (in inference) or the ground-truth values (in training)') and at inference feeds its own predictions. Both are teacher-forcing mismatches. If coarse localization is imperfect—especially on the exaggerated shapes and unusual poses the paper claims to support—the canonical frame is rotated/scaled incorrectly and hand samples can miss the actual hands, and errors in parent bones propagate to finger and toe predictions. Table 3 reports full-pipeline numbers on the Mixamo split, but no ablation replaces predicted coarse bones with GT bones or feeds GT parent attributes during decoding, so the reader cannot tell how much of the reported 2.11% joint / 13.6% pose error depends on oracle information. The acknowledged failure cases (Fig. S14(a)-(c)) are exactly regimes where the coarse stage and autoregressive decoding are least reliable.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Make-It-Animatable, a data-driven framework that takes a 3D humanoid character (mesh or 3D Gaussian splat) and, in a single forward pass, predicts blend weights, a 52-bone skeleton, and pose-to-rest transformations, thereby making the character ready for animation. The architecture combines a frozen ShapeNet-pretrained point-based shape encoder, a coarse-to-fine representation that canonicalizes the input and hierarchically samples hand regions, and a structure-aware transformer that autoregressively predicts bone attributes with causal attention along the kinematic tree. The model is trained on 95 Mixamo characters (with over 40 million effective frame samples) and fine-tuned on 35 VRoid characters with extra bones. Experiments report a large improvement over RigNet on a public MoModelsResource dataset for skeleton prediction (IoU 82.5% vs 53.5%), internal ablations on Mixamo, and qualitative comparisons with commercial and research baselines. The abstract claims that any humanoid, regardless of shape or pose, can be made animatable in less than one second.","tokens_in":19915,"tokens_out":7549,"duration_ms":67752,"significance":"If the central claim holds, this is a practically valuable contribution to automatic rigging and skinning: it unifies bones, blend weights, and pose-to-rest estimation in one fast feed-forward network, supports both meshes and 3D Gaussian splats, and handles arbitrary input poses without a standard T/A-pose requirement. The quantitative skeleton-prediction improvement over RigNet on a public test set is substantial, and the release of code and demos would facilitate reproducibility. The main weakness is that the full-pipeline quantitative evidence is limited to a small Mixamo-derived validation split, while the headline 'any humanoid, regardless of shape and pose' claim is supported only by selective qualitative results that include acknowledged failures.","major_comments":[{"comment":"There is a teacher-forcing mismatch between training and inference in the two core modules. S2.1 states that the fine stage applies canonical transformation and hierarchical sampling 'based on the ground-truth bone positions'; at inference, the coarse stage's predicted bones are used instead. Likewise, Fig. 3 shows that the structure-aware transformer encodes parent-bone latents from ground-truth attributes during training but from its own autoregressive outputs during inference. Table 3 reports no ablation that replaces predicted coarse bones with ground-truth bones or that feeds ground-truth parent attributes at decode time. Consequently, the reported joint error (2.11%) and pose error (13.6%) do not isolate the error contribution of this mismatch, and the mismatch is likely to be most severe precisely in the arbitrary poses and out-of-distribution shapes that the paper claims to support (Fig. S14). I request explicit ablations that quantify this gap.","section":"S2.1 and Fig. 3"},{"comment":"The quantitative evaluation of the full pipeline (blend weights and pose transformations) is conducted only on the Mixamo validation split, which consists of roughly 4–5 characters out of the 95-character training collection. The ModelsResource-RigNetv1 test set (Table 2) is used only for skeleton-prediction metrics (IoU, precision, recall, Chamfer distances), not for blend-weight or pose errors. As a result, the paper's central generalization claim for arbitrary shapes and poses is supported by qualitative examples, which include acknowledged failure cases in Fig. S14(b)-(c). The abstract's 'any humanoid ... regardless of its shapes and poses' is stronger than the evidence presented.","section":"Sec. 4.1, Tables 2 and 3"},{"comment":"The 'percentage error' metrics for weights, joints, and poses are not defined in the main text or the supplement. Without the normalization and computation formulas, these numbers cannot be reproduced or compared across methods. Please provide explicit metric definitions, including how the percentage is computed for each asset type.","section":"Table 3"},{"comment":"The abstract and Table 1 claim processing in less than one second, but Fig. S10(d) states that the Wukong model with over one million triangular faces takes 'within 3 seconds'. Since the vertex/face count strongly affects inference time, the speed claim should be conditioned on a representative input range (e.g., the 8k-vertex mesh used in Table 1) and the unconditional 'less than one second' wording should be revised.","section":"Sec. 1 and Fig. S10(d)"},{"comment":"The paper asserts that the ShapeNet-pretrained encoder 'already has sufficient capacity for low-level geometry perception' and therefore freezes its parameters. No experiment compares freezing with fine-tuning or otherwise validates that ShapeNet features transfer to exaggerated cartoon proportions and non-human accessories. The failure cases in Fig. S14(b)-(c) are consistent with a lack of feature adaptation. Given that the encoder cannot be updated during training, the claim that the model generalizes to 'any 3D humanoid model' needs quantitative support from an ablation or a feature-level analysis.","section":"Sec. 3.2 and Fig. S14"}],"minor_comments":[{"comment":"Please report the exact number of characters in the Mixamo validation split; with 95 characters and a 95/5 split, the validation set is only about 4–5 characters, which is small for reliable error bars.","section":"Sec. 4.1"},{"comment":"The paper says the geometry-aware attention is 'compatible with inputs like 3DGS in inference' but does not explain how per-point normal values are obtained for Gaussian splat inputs. Clarify this point.","section":"Sec. 3.2"},{"comment":"There is a typo: 'the remaining 5% for for validation' should read 'for validation'.","section":"Supplementary S3.1"},{"comment":"The sentence 'Mixamo fails to fail to distinguish the left and right sides' contains a duplicated 'fail to'.","section":"Supplementary Fig. S7"}],"recommendation":"major_revision","confidential_remarks":"The paper is on a solid track: the method is well-designed, the skeleton-prediction comparison on a public dataset is convincing, and the code release is a plus. The main barrier is that the headline claims ('any humanoid', 'in less than one second') exceed the current evidence. The teacher-forcing mismatch and the lack of quantitative generalization tests are the most important items to address. If the authors can provide the requested ablations and metric definitions, and appropriately qualify the speed and generalization claims, the paper could become acceptable. I do not see an unfixable technical flaw in the core approach."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a genuinely useful auto-rigging paper. The unified mesh/3DGS handling via a particle-based autoencoder is the real contribution — it gives you one pipeline for both representations, and the speed story holds up. The coarse-to-fine canonicalization and the structure-aware transformer are sensible engineering around that core. The quantitative results against RigNet on the ModelsResource bipedal subset (IoU 82.5 vs 53.5) are strong, and the Mixamo ablations are consistent and tell a coherent story. The comparisons with commercial tools are soft, as those usually are, but the paper does not oversell them beyond what the figures show.\n\nNow the soft spots. The headline claim is broader than the experiment: fixed 52-bone Mixamo skeleton, frozen ShapeNet encoder, 95 training characters, and acknowledged failures on out-of-distribution shapes and hard poses. That is a mismatch between the abstract's \"any humanoid... regardless of shapes and poses\" and what the method can actually do. It does not sink the paper — it just means the honest scope is \"many typical bipedal characters in near-rest poses.\"\n\nThe more technical concern is the teacher-forcing gap that the stress test identified. In training the fine stage uses ground-truth bone positions for canonicalization and hierarchical sampling, and the structure-aware transformer consumes ground-truth parent latents. At inference both use predicted values. The paper does not report an ablation that replaces predicted bones or latents with ground truth. Without that, I cannot tell how much of the reported 2.11% joint error and 13.6% pose error depends on oracle information. The failure cases in Fig. S14 are exactly where that error would bite, which makes the missing diagnostic more conspicuous. This is fixable: one oracle-conditioned run would tell you whether the gap is small or load-bearing.\n\nMinor note: the low-rank dynamics framing in Sec. 3.1 is a dressed-up description of linear blend skinning. It is not wrong, just heavier than necessary. Also, the paper says \"code available\" on the project page; I would want the code and the Mixamo preprocessing before fully trusting the numbers, but the method is reproducible enough to referee.\n\nBottom line: this deserves a serious referee. The teacher-forcing question should be a major comment, not a reject reason. I would take it to reading group and would likely cite it when discussing fast auto-rigging or mesh/GS-agnostic animation.","headline":"Good practical auto-rigging paper with real speed and input-flexibility gains; the abstract overclaims generality, and the teacher-forcing mismatch in training needs a diagnostic the authors did not run.","tokens_in":20454,"tokens_out":2476,"would_cite":true,"duration_ms":22391,"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":"A single feed-forward framework rigs, skins, and re-poses any 3D humanoid character in under a second.","keywords":["3D character animation","automatic rigging","skinning weights","pose-to-rest transformation","particle-based shape autoencoder","neural field","3D Gaussian splatting","structure-aware transformer"],"falsifier":"Run the framework on a curated set of bipedal characters far outside both the Mixamo training distribution and ShapeNet-like geometry — for example chibi figures whose head is half their total height, or characters with detached floating accessories — with artist-made ground-truth skeletons and weights, and measure joint IoU and Chamfer distances on those cases; if the predictions degrade to the level of the failure cases the paper itself shows in Fig. S14(b)-(c), the 'regardless of shapes and poses' claim is bounded.","tokens_in":19414,"feed_emoji":"🦴","tokens_out":14647,"duration_ms":114702,"temperature":0.7,"pith_summary":"Make-It-Animatable sets out to prove that rigging, skinning, and pose normalization of a 3D humanoid character can be collapsed into one feed-forward prediction that finishes in under a second. The method takes meshes or 3D Gaussian splats — a particle-based representation used by modern 3D generators — in arbitrary poses and shapes, and outputs the three animation assets the industry needs: blend weights (per-vertex influence of each bone), a skeleton, and a pose-to-rest transformation that brings the character to a standard T-pose. This would matter because rigging and skinning are among the most labor-intensive steps in 3D content production, and existing automatic tools either need manual joint placement, only accept T- or A-poses, or are bound to realistic human body templates. The paper argues its framework handles exaggerated cartoon proportions, asymmetric designs, fine finger bones, and extra structures like ears and tails, and completes in roughly half a second where learned baselines take tens of seconds to minutes.","feed_headline":"Rig, skin, and re-pose any 3D character in under a second","feed_subtitle":"Automates rigging and skinning for meshes and Gaussian splats, with no template and no pose restrictions.","key_machinery":"The load-bearing machinery is the particle-based shape autoencoder paired with the structure-aware transformer. The autoencoder converts any point-samplable representation into a compact neural field: 32,768 surface points are downsampled by farthest point sampling and cross-attention builds 512 latent vectors that can be queried at arbitrary coordinates, which is the step that unifies meshes and 3D Gaussian splats and makes inference cost nearly independent of input particle count. A geometry-aware attention branch injects mesh normals adaptively, so semantically distant but spatially close points do not corrupt blend weights. The second pillar is structure-aware modeling: a next-child-bone prediction transformer treats each bone as a token, adds the parent bone's decoded attribute latent to the child's shape embedding, and applies masked causal attention over the kinematic tree, so deep bones are predicted conditioned on their ancestors instead of independently. The pose-to-rest output is regressed as dual quaternions, which the paper finds converges more reliably than a vanilla 6-DoF prediction, and three body-prior losses (bone connectivity, left-right symmetry, limb parallelism) keep rest poses well-formed. Conceptually, the paper frames character motion as a low-rank decomposition $D \\approx B_t W_s$, so making a character animatable reduces to recovering the spatial weight matrix (the blend weights) and the temporal basis (the pose-to-rest transformations).","core_discovery":"The central claim is that the complete animation asset set of a character — K bone head-and-tail positions, a pose-to-rest transformation per bone, and per-vertex blend weights — can be read directly off a neural field encoding of the input geometry, with no template mesh, no manual annotation, and no assumption about the input pose. The framework samples the input character into a point cloud, encodes it with a frozen ShapeNet-pretrained shape autoencoder into 512 latent vectors, decodes blend weights at arbitrary surface coordinates, and decodes bone attributes from learned per-bone queries. A structure-aware transformer organizes bone decoding as next-child-bone prediction: each bone's embedding is summed with its parent's decoded latent and fused through causal masked attention down the kinematic tree, so finger and toe poses inherit the correct arm and leg context. A coarse stage localizes the skeleton from uniform samples, and a fine stage re-aligns the shape into a canonical hip-centered frame and re-samples half of its points near the hands, which is what makes fine finger bones and arbitrary global poses converge. Trained end-to-end on 95 Mixamo characters with 2,453 motion sequences, the framework reports a skeleton-matching IoU of 82.5 percent against 53.5 percent for RigNet on a shared test set, sub-second inference that barely depends on vertex count, and identical handling of mesh and Gaussian-splat inputs.","pith_inferences":["A natural extension the paper leaves implicit: pipeline the framework directly onto the output of text-to-3D or image-to-3D generators, so every generated bipedal character ships animation-ready and 'animatable' becomes a default property of generation rather than a post-process.","The next-child-bone decoding scheme is a general recipe for any prediction problem with a fixed tree-structured dependency; porting it to non-bipedal skeletons (quadrupeds, winged characters) is the paper's own stated future work and would test whether the mechanism itself, rather than the human-shaped training set, carries the accuracy.","The frozen ShapeNet encoder is the likeliest generalization ceiling; replacing it with an encoder pretrained on stylized or cartoon character meshes is a direct, measurable upgrade that the paper's own failure cases point to."],"forward_implications":["Characters in arbitrary poses — running, seated, asymmetric — are automatically transformed into a canonical T-pose, so existing motion libraries become immediately applicable to inputs that commercial auto-riggers reject.","Because the input is just a point-samplable shape, meshes and 3D Gaussian splats flow through the same pipeline, so static assets from text-to-3D generators become animatable without being bound to an SMPL-like template.","Fine-grained finger and hand bones are produced that commercial tools like Meshy and Tripo omit, enabling expressive hand articulation.","New skeleton topologies cost only a fine-tune: adding extra bones such as ears and tails requires updating the final weight-decoder layer and adding learnable queries, not retraining the network.","Inference stays sub-second on typical meshes and around three seconds on a one-million-triangle model, so production-grade assets can be rigged at interactive speed."],"supporting_citations":[{"why":"Supplies the downsampled-point querying neural-field architecture (3DShape2VecSet) that the paper's shape autoencoder is built on.","marker":"[58]"},{"why":"Mixamo; the source of the 95-character, 2,453-motion training set with artist-made skeletons, blend weights, and poses.","marker":"[1]"},{"why":"RigNet; the learned auto-rigging baseline that the framework is quantitatively compared against on skeleton prediction.","marker":"[57]"},{"why":"TARig; the template-aware rigging baseline used for quality and inference-speed comparison.","marker":"[38]"},{"why":"ShapeNet; the pretraining corpus for the frozen shape encoder whose feature transfer the whole method depends on.","marker":"[11]"},{"why":"The ModelsResource-RigNetv1 dataset; the shared test set on which the quantitative skeleton-prediction numbers are reported.","marker":"[56]"},{"why":"SMPL; the fixed template behind the TADA and HumanGaussian baselines that the paper contrasts with its template-free approach.","marker":"[37]"},{"why":"VRoid Studio; source of the 35 anime characters with ears and tails used to fine-tune and validate extra-bone predictions.","marker":"[45]"}],"fun_headline_variants":["Auto-rig any 3D humanoid in under a second, no template needed","Neural field predicts full rig for any humanoid mesh or Gaussian splat","One-shot animation prep: bones, skins, and poses from any 3D humanoid","Sub-second auto-rigging that works for any pose and shape without manual work"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a shape encoder pretrained on ShapeNet's everyday objects already has enough low-level geometry understanding for any character shape — exaggerated cartoon proportions, accessories, stylized non-human bodies — and because this encoder is frozen during training, the promise of 'any shape' holds only as far as those pretrained features transfer.","fun_headline_variants_meta":{"raw":{"variants":["Auto-rig any 3D humanoid in under a second, no template needed","Neural field predicts full rig for any humanoid mesh or Gaussian splat","One-shot animation prep: bones, skins, and poses from any 3D humanoid","Sub-second auto-rigging that works for any pose and shape without manual work"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000449,"raw_usage":{"total_tokens":2332,"prompt_tokens":1079,"completion_tokens":1253,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":695,"completion_tokens_details":{"reasoning_tokens":1164}},"tokens_in":695,"tokens_out":1253,"duration_ms":10854,"temperature":1.0,"reasoning_tokens":1164,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:25:26.291430+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the framework on a curated set of bipedal characters far outside both the Mixamo training distribution and ShapeNet-like geometry — for example chibi figures whose head is half their total height, or characters with detached floating accessories — with artist-made ground-truth skeletons and weights, and measure joint IoU and Chamfer distances on those cases; if the predictions degrade to the level of the failure cases the paper itself shows in Fig. S14(b)-(c), the 'regardless of shapes and poses' claim is bounded.","supporting_citations":[{"cited_title":"3DShape2VecSet: A 3D shape representation for neural fields and generative diffusion models.ACM TOG, 42 (4):92:1–92:16, 2023","cited_arxiv_id":null,"evidence_quote":"Supplies the downsampled-point querying neural-field architecture (3DShape2VecSet) that the paper's shape autoencoder is built on."},{"cited_title":"Mixamo, 2024","cited_arxiv_id":null,"evidence_quote":"Mixamo; the source of the 95-character, 2,453-motion training set with artist-made skeletons, blend weights, and poses."},{"cited_title":"RigNet: Neural rigging for articu- lated characters","cited_arxiv_id":null,"evidence_quote":"RigNet; the learned auto-rigging baseline that the framework is quantitatively compared against on skeleton prediction."},{"cited_title":"TARig: Adaptive template- aware neural rigging for humanoid characters","cited_arxiv_id":null,"evidence_quote":"TARig; the template-aware rigging baseline used for quality and inference-speed comparison."},{"cited_title":"Predicting animation skeletons for 3D articulated models via volumetric nets","cited_arxiv_id":null,"evidence_quote":"The ModelsResource-RigNetv1 dataset; the shared test set on which the quantitative skeleton-prediction numbers are reported."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"SMPL; the fixed template behind the TADA and HumanGaussian baselines that the paper contrasts with its template-free approach."},{"cited_title":"VRoid Studio, 2024","cited_arxiv_id":null,"evidence_quote":"VRoid Studio; source of the 35 anime characters with ears and tails used to fine-tune and validate extra-bone predictions."}],"review_version":1}