acceptodds
Under review as a conference paper at ICLR 2027

AniRig: Recovering Animation-Ready Rigged Assets from Arbitrary-Topology 3D Characters

Abstract

High-quality character meshes are increasingly produced by generation, reconstruction, scanning, and artist workflows. Skeletal animation requires a rest mesh, a skeleton, and skinning weights that work together. Recovering this rigged asset from a single observed mesh remains a bottleneck, especially when the input is already posed. Existing rigging and skinning methods often assume rest-pose inputs or evaluate skeletons and weights separately. Component-level accuracy alone does not establish whether these predictions support coherent rest recovery and subsequent animation. We introduce AniRig, a semantic-hierarchy-conditioned framework for recovering an animation-ready rigged asset from a single observed character mesh with arbitrary topology. The hierarchy specifies joint names and parent-child connectivity, and can be user-provided or inferred from rendered mesh views by a Vision-Language Model. AniRig jointly predicts observed and rest skeletons in a single diffusion model under this hierarchy, and estimates structured sparse skinning weights by jointly evaluating a compact set of candidate joints. A differentiable pose-to-rest construction composes these predictions with the observed mesh to recover the canonical rest mesh and enables coupled fine-tuning with asset-level surface supervision. On Articulation-XL2.0, AniRig improves skeleton and skinning weight accuracy. On an independent real-motion posed-mesh benchmark, asset-level fine-tuning improves canonical rest recovery and held-out re-posing consistency. Qualitative results show that the recovered assets support skeletal animation reuse across matched hierarchies.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D computer vision, character rigging, skeletal animation, skinning, 3D character animation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_GoiWOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

AniRig recovers an animation-ready rigged asset from a single observed 3D character mesh, conditioned on a user-provided or VLM-inferred skeletal hierarchy. The method jointly predicts observed- and rest-pose skeletons, estimates sparse skinning weights, and reconstructs a canonical rest mesh through differentiable inverse articulation. Asset-level fine-tuning uses rest-recovery and re-posing losses to jointly refine the predictions. Experiments evaluate skeleton and skinning accuracy, as well as complete-asset recovery on an independent real-motion benchmark.

Critical strengths
  1. Clear presentation and well-motivated problem. The paper clearly motivates the need to recover a complete, reusable rigged asset from a posed mesh, rather than evaluating skeletons and skinning weights independently. The method and evaluation are organized around this practical and meaningful problem.

  2. Coherent method with supporting evaluation. Paired skeleton prediction, structured sparse skinning, and differentiable pose-to-rest reconstruction form a coherent framework. Asset-level supervision jointly refines skeleton and skinning predictions, while component comparisons, ablations, and an independent real-motion benchmark provide evidence for the proposed design.

Critical weaknesses
  1. Insufficient validation of hierarchy and semantic conditioning. Although the appendix provides a VLM hierarchy example, it depicts a character in a rest-like pose. Since the paper targets rest recovery from posed meshes, more challenging articulated cases—with self-occlusion, dense joints, and complex appendages—are needed to assess hierarchy reliability. These cases should show the six-view inputs, inferred and reference hierarchies, and resulting rigs. Furthermore, it is unclear whether frozen CLIP embeddings meaningfully distinguish fine-grained joint names such as tail_01 and tail_02. The effects of varying skeletal topologies and naming conventions across animal categories on training and generalization are also insufficiently explored.

  2. Limited comparative evidence for downstream usability. Component-level comparisons do not fully establish the practical benefits of the recovered assets. Side-by-side examples comparing AniRig and suitable baseline pipelines under the same re-posing or rig-editing operations would better substantiate the claims of animation readiness and editability. These comparisons should highlight deformation artifacts and any required manual corrections.

  3. Incomplete discussion of relevant work. The related-work discussion should address Auto-Connect: Connectivity-Preserving RigFormer with Direct Preference Optimization, which studies connectivity-preserving skeleton generation and geometry-aware sparse bone selection, and Stroke3D: Lifting 2D Strokes into Rigged 3D Model via Latent Diffusion Models, which generates controllable rigged assets using latent diffusion. While their task settings differ, discussing these connections would clarify AniRig’s contributions relative to prior work.

Questions for authors

As weakness.

3 comments
AuthorsOpenReview ↗

Response to Reviewer GoiW

Thank you for suggesting more challenging examples and a closer analysis of semantic conditioning.

W1. Hierarchy reliability and semantic conditioning

Dense skeleton with multiple branches. To assess hierarchy inference for a dense, branched character, we add a scorpion example and compare the inferred hierarchy with the original asset's skeleton. Most inferred topology agrees with this reference and needs only small local additions, supporting reliable inference of the main skeletal structure. The original output has 59 joints; we manually add only two missing mouthpart joints and their connections, retaining all others. The augmented hierarchy and reference each contain 61 joints with matching topology. Figure 10 shows the six-view inputs, augmented hierarchy, reference skeleton, and resulting rig (Appendix D.3, Figure 10, p. 31, L1620-1623 and L1648-1653).

Severe self-intersection. To examine recovery under severe interpenetration, we construct a humanoid pose from a standard T-pose, with both arms passing through the torso to opposite sides and the legs crossed. Its unedited VLM hierarchy matches the connectivity of the original asset's reference skeleton, supporting the reliability of VLM hierarchy inference for this severely intersecting input. AniRig places the arm chains near the waist, indicating a joint-localization failure. Comparable poses are absent from training. The output remains a connected humanoid skeleton (Figure 11, p. 33, L1759-1762; discussion in Appendix E, pp. 32-33, L1726-1731).

Fine-grained joint names. For fine-grained names such as tail_01 and tail_02, we observe in practice that frozen CLIP produces distinguishable joint-name embeddings. These joint-name embeddings, together with token-index and parent-index positional encodings, jointly condition the model, enabling it to distinguish corresponding joints within a skeletal chain. The humanoid cat in Figure 5's fifth row and the scorpion in Figure 10 show recovered tail chains using such names (Figure 5, p. 15; Figure 10, p. 31; Appendix B.1, p. 16, L822-825).

Cross-category skeletal topology. We train on character-specific hierarchies with shared model parameters; at test time, no category- or topology-specific fine-tuning is required. The cross-category asset recovery and animation results in Figure 5, together with the dense, branched scorpion in Figure 10, illustrate the same model's ability to adapt to different bone chains and branching structures (Figure 5, p. 15; Figure 10, p. 31; Appendix C.2, p. 24, L1260-1262).

Cross-category name substitution. To isolate naming, we exchange humanoid and quadruped names in both directions, e.g., hand_left and front_hoof_left, while fixing the mesh, joint order, parent-child connections, and inference sampling. After the exchange, the main bone chains retain their overall spatial layout and remain consistent with each character's geometry. These examples illustrate AniRig's generalization across skeletal naming conventions (Appendix C.7, p. 27, L1418-1422; Figure 8, p. 27).

W2. Downstream re-posing comparison

We compare AniRig with Make-It-Animatable (MIA) on the 40 humanoids from our real-motion benchmark. AniRig uses VLM-inferred hierarchies; MIA uses its pipeline's hierarchies. In Figure 7, both methods recover assets from the same posed mesh, then receive the same target motion through retargeting, with GT shown for comparison. All assets shown in this comparison were obtained without manual correction.

Case (a): limb deformation. MIA produces flattened, stretched arms and twisting, compression, and folding around the hips and legs, which persist after re-posing. AniRig recovers extended arms and separated legs and maintains more intact arm, torso, and leg contours under the target motion.

Case (b): head deformation. MIA's recovered head contour is skewed, and this deformation persists after re-posing. AniRig better preserves the head shape in both stages, with results closer to GT.

These differences persist under the same target motion, showing that AniRig's recovery improvements translate into cleaner downstream deformations without manual asset correction.

Rest-recovery results on the same 40 humanoids (Table 6) show lower CD and higher F-Score for AniRig:

MethodRest CD ↓Rest F-Score ↑
MIA0.1050.490
AniRig0.08110.621

Table 5 and Figure 5 provide further evidence of animation use: Table 5 reports held-out re-posing under the benchmark hierarchy, and Figure 5 shows assets recovered with unedited VLM hierarchies driven by compatible animations.

Revised manuscript: Appendix C.3, Table 6, p. 24, L1278-1294; Figure 7, p. 25, L1341-1345; Table 5, p. 9, L461-471; Figure 5, p. 15.

W3. Related work

Thank you for recommending these two related works. We have added both to Related Work. Auto-Connect's connectivity-preserving skeleton generation and geodesic-aware top-kk bone selection relate to our skeleton prediction and sparse influence selection, respectively. Stroke3D uses latent diffusion to generate rigged assets from 2D strokes and text. This shares AniRig's goal of producing animation-ready assets, with Stroke3D addressing asset generation and AniRig addressing pose-to-rest recovery from an observed mesh.

Revised manuscript: Auto-Connect in Section 2.2, p. 3, L140-141; Stroke3D in Section 2.3, p. 3, L159-161.

Thanks for your response, I will keep my score.

Reviewer_yv1mOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary
  1. This paper studies the recovery of an animation-ready rigged asset from a single 3D character mesh. Given an observed mesh and a semantic skeletal hierarchy, the method predicts an observed skeleton, a rest skeleton, vertex-level skinning weights, and a canonical rest mesh.

  2. The proposed AniRig combines paired skeleton prediction, structured sparse skinning, and differentiable pose-to-rest construction in one asset-recovery framework. It further uses rest-surface and re-posing supervision for asset-level fine-tuning.

  3. The paper presents a relatively comprehensive evaluation across skeleton prediction, skinning weight prediction, rest recovery, and held-out re-posing. It also includes ablations on candidate-set size, module removal, and asset-level fine-tuning.

Critical strengths
  1. The task has high practical value. Recovering an editable and re-poseable rigged asset from a single posed mesh is important for graphics and 3D content production.
  2. The proposed method is well motivated. Paired skeleton prediction, structured sparse skinning, and differentiable pose-to-rest construction are naturally coupled at the task level rather than treated as independent predictions.
  3. The method shows strong experimental results. The paper evaluates AniRig on multiple benchmarks and task settings, and reports competitive results for skeleton, skinning, and asset-level metrics. The candidate-size study, module ablations, and asset-level fine-tuning experiments also provide support for the main design choices.
Critical weaknesses
  1. The training and inference conditions for asset-level fine-tuning may not be fully consistent. Appendix B.4 states that ground-truth observed skeletons are used for orientation guidance during fine-tuning, while inference relies on the predicted observed skeleton. The authors should clarify whether this guidance provides information unavailable at inference time and report the rest-recovery and re-posing results without ground-truth orientation guidance.

  2. The fairness of the paired-skeleton ablation needs clarification. In Table 3, the variant without rest-skeleton prediction uses the ground-truth rest skeleton during asset-level fine-tuning. This may provide extra information to that variant. The authors should clearly specify the inputs, supervision, and use of ground-truth skeletons for each ablation during both training and inference.

  3. The paper reports rest recovery with an unedited VLM-inferred hierarchy, but does not report the corresponding held-out re-posing result. The authors should clarify whether this is because the VLM hierarchy cannot be directly matched to the ground-truth motion, or because of another evaluation limitation. If re-posing cannot be evaluated under this setting, the authors should discuss how this limits the scope of the complete-asset evaluation.

Questions for authors
  1. For asset-level fine-tuning, what ground-truth information from the observed skeleton is used for orientation guidance? Is this information unavailable during inference? What are the rest-recovery and held-out re-posing results when this guidance is removed?

  2. Do all variants in Table 3 use the same inputs, supervision, and training procedure? Why is the ground-truth rest skeleton available to the variant without rest-skeleton prediction? Would the conclusion remain unchanged under a unified setting without this extra information?

  3. Why is held-out re-posing not reported for the unedited VLM-inferred hierarchy? Is the issue the lack of direct joint correspondence with the ground-truth motion, or another protocol limitation? How does this affect the interpretation of the complete-asset results?

1 comment
AuthorsOpenReview ↗

Response to Reviewer yv1m

Thank you for asking us to clarify orientation guidance during training and inference.

W1/Q1. GT orientation guidance

During asset-level fine-tuning, we align predicted observed bone directions to GT observed bone directions, then apply the inverse of each joint's global rest-to-observed rotation from pose augmentation to construct the pose-to-rest rotation. Translations are anchored to the predicted observed/rest joint positions. At inference, GT directions and augmentation rotations are unavailable; rotations are constructed from the two predicted skeletons.

Without orientation guidance, fine-tuning uses the same transform construction as inference:

\qquad t_j=\hat j_j^r-R_j\hat j_j^o.$$ $\hat S^o,\hat S^r$ are predicted observed/rest skeletons sharing hierarchy $\mathcal T$. $\operatorname{HierarchicalAlign}$ aligns corresponding bone directions along this hierarchy to obtain joint rotation $R_j$; $\hat j_j^o,\hat j_j^r$ are predicted joint positions. This control replaces the pose-to-rest transform construction; subsequent re-posing, losses, and other supervision retain their original settings. All three settings use the benchmark hierarchy; re-posing is evaluated on held-out frames: | Variant | Rest CD ↓ | Rest F-Score ↑ | Re-posing CD ↓ | Re-posing F-Score ↑ | |---|---:|---:|---:|---:| | Ours (benchmark hier., w/o fine-tuning) | 0.070 | 0.655 | 0.103 | 0.553 | | Ours (benchmark hier., w/o orientation guidance) | 0.066 | 0.672 | 0.095 | 0.598 | | **Ours (benchmark hier.)** | **0.064** | **0.692** | **0.089** | **0.621** | Fine-tuning improves all four metrics even when the pose-to-rest transforms are constructed from predicted skeletons, showing that its benefit persists without GT orientation guidance. Retaining guidance further improves rest recovery and held-out re-posing. *Revised manuscript: Appendix B.4, p. 20, L1075-1079 and p. 21, L1080-1089; Appendix C.5, Table 8, p. 26, L1366-1382.* ### **W2/Q2. Paired-skeleton ablation** The Table 3 variant without rest-skeleton prediction **needs the GT rest skeleton for asset-level fine-tuning**, because constructing the rest/re-posed surfaces and computing their losses requires a rest skeleton. We therefore use the GT rest skeleton for this variant during training; the full model uses its predicted rest skeleton. All variants use the observed mesh and semantic hierarchy as network conditioning inputs during training and inference. Training uses GT observed-skeleton supervision for every variant; the rest-skeleton branch additionally receives direct GT rest-skeleton supervision. `Ours w/o asset-level fine-tuning` is the pretrained paired checkpoint. Table 3 evaluates observed-skeleton prediction errors at inference. To isolate paired prediction, Table 7 compares the paired model with the variant without rest-skeleton prediction immediately after component pretraining, before asset-level fine-tuning. Pretraining data, steps, optimizer, and learning rate are matched: | Before asset-level fine-tuning | J2J ↓ | J2B ↓ | B2B ↓ | |---|---:|---:|---:| | Ours w/o rest-skeleton prediction | 1.736 | 1.304 | 1.139 | | **Ours w/o asset-level fine-tuning** | **1.660** | **1.208** | **1.020** | All three errors improve with paired prediction before the fine-tuning-time GT rest reference is introduced. *Revised manuscript: Appendix C.4, Table 7, p. 26, L1350-1364; Table 3 in Section 4.3, p. 8, L405-420.* ### **W3/Q3. Held-out re-posing with unedited VLM hierarchies** Held-out re-posing directly applies GT motion's per-joint local rotations. VLM-inferred hierarchies can use different joint definitions from the motion skeletons, so applying these rotations requires establishing correspondence. We therefore use the benchmark hierarchy for quantitative held-out re-posing. For unedited VLM hierarchies, Table 5 evaluates rest recovery quantitatively. The current re-posing evidence is qualitative: Figure 5 shows compatible animations driving the recovered assets, and Figure 7 shows retargeted examples. Accordingly, the complete-asset results quantify rest recovery for unedited VLM hierarchies, while the held-out re-posing metrics assess deformation under direct correspondence with GT motion. We have added this correspondence requirement and its implications for complete-asset evaluation to Limitations. *Revised manuscript: Appendix C.1, p. 23, L1220-1223; Appendix E, p. 32, L1719-1724; Table 5, p. 9, L461-471; Figure 5, p. 15.*
Reviewer_BZWROpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the problem of recovering animation-ready 3D assets from an observed input mesh. To achieve this, the rest pose of the input mesh should be recovered so that the rigged mesh can be animation reusable. The proposed approach predicts corresponding observed and rest skeletons, estimates skinning weights, and recovers a rest mesh through a differentiable approximate pose-to-rest construction. The method is conditioned on a semantic hierarchy specifying joint names and parent-child connectivity, which can be provided by user-specified or inferred by a VLM from multi-view renderings of the input mesh.

For skeleton generation, a skeleton predictor module is introduced that adapts a diffusion process to jointly generate the observed and rest skeleton. The bone tokens encode joints in both observed and rest configurations, in order to preserve correspondence between the two skeletons and allow the predicted components to be combined for rest recovery. The skeleton predictor is conditioned on joint-name embeddings that provide semantic information, along with joint and parent indices that specify the skeleton’s hierarchy. Additionally, mesh features acquired by a forward pass of a TripoSG encoder, supply the geometric conditioning needed for the joints’ accurate placement.

For skinning weights prediction, the proposed skinning module first scores all mesh surface query points and joint pairs, and selects a set of candidate joints that highly influence the query point. Structured interaction compares the selected joints, while separate support and weight heads predict influence membership and magnitude. Coverage supervision encourages the selector to retain relevant joints before detailed weight prediction. Surface and voxel heat diffusion, together with accessibility cues, provide geometric priors beyond spatial proximity.

The skeleton and skinning modules are pretrained separately and then fine-tuned together using losses on recovered rest and re-posed surfaces. Evaluation on Articulation-XL2.0 demonstrates improved skeleton placement and skinning accuracy. An independent benchmark of 100 characters with real animation clips evaluates rest recovery and held-out re-posing, showing improvements from asset-level fine-tuning. Qualitative examples further illustrate animation reuse of the generated rigged 3D assets.

Critical strengths
  1. The joint observed- and rest-pose skeleton modeling aid the skeleton generation of the observed mesh, as substantiated by the relevant ablation (Table 3). Predicting corresponding observed and rest skeletons also allow skinning weights to be refined through the input posed mesh and its rest-pose counterpart. Finally, it improves the rest recovery and re-posing of input mesh, as seen by the fine-tuning ablation, where the rest Chamfer and held-out re-posing Chamfer distances improve from 0.070 to 0.064 and from 0.103 to 0.089, respectively.

  2. The structured sparse skinning formulation is supported by clear component comparisons and ablations. Coverage supervision addresses the inability of the weight scorer to recover influences discarded during candidate selection, while candidate interaction helps distinguish competing joints. Removing these components increases L1 error from 0.233 to 0.349 and 0.324, respectively. The full model also improves skinning recall over SkinTokens while maintaining similar precision. These results substantiate the proposed design and show that its benefit extends beyond reducing the size of the candidate set.

Critical weaknesses
  1. The quantitative validation of complete-asset recovery is limited to the observed input and the method's own fine-tuning ablation. Table 5 establishes that fine-tuning improves the proposed system, but does not independently establish its advantage over an existing pose-to-rest approach. The authors explain the difficulty of comparing character-specific hierarchies with Make-It-Animatable, although a comparison on compatible humanoid configurations would still be informative. Re-posing is also evaluated only with the benchmark hierarchy, leaving the deformation quality of the fully inferred pipeline supported primarily by qualitative examples.

  2. The structured sparse skinning architecture in Appendix B.2 is somewhat difficult to follow, as several important operations are introduced descriptively without sufficiently explicit formulations. This is particularly noticeable in the Candidate selection and Candidate interaction paragraphs. Although Appendix B.3 describes the geometric priors, the learned combination of these priors, geometric proximity, and surface–joint compatibility into the selector’s pair scores is not explicitly specified. Similarly, the bone-relative features, mutual query–joint updates, and geometry-dependent attention biases and gated messages described in lines 859–865 would benefit from mathematical definitions. Figure 6 presents these operations at a high level, but does not sufficiently clarify their internal structure or connections. Expanding the candidate-interaction subfigure and providing equations for the selection and interaction mechanisms would make the architecture easier to assess and reproduce.

Questions for authors

For the skeleton and skinning comparisons on the Articulation-XL2.0 test split, were all competing methods trained or fine-tuned on the same training split as AniRig? If released checkpoints were used, please specify the training datasets and splits for each method, as differences in training data could affect the interpretation of the reported improvements. Could you additionally report these comparisons on the ModelsResource test split [1]? This would help establish whether the improvements generalize beyond the training dataset.

[1] Zhan Xu, Yang Zhou, Evangelos Kalogerakis, Chris Landreth, and Karan Singh. Rignet: Neural rigging for articulated characters. ACM TOG, 39(4), 2020

2 comments
AuthorsOpenReview ↗

Response to Reviewer BZWR

Thank you for suggesting a complete-asset baseline comparison and for highlighting the evaluation of the fully inferred pipeline.

W1. Asset comparison and fully inferred evaluation

We compare with Make-It-Animatable (MIA) on the same 40 humanoids from the real-motion benchmark. AniRig uses VLM-inferred hierarchies; MIA uses its pipeline's hierarchies. Table 6:

MethodRest CD ↓Rest F-Score ↑
MIA0.1050.490
AniRig0.08110.621

These results provide an independent pose-to-rest baseline comparison on compatible humanoids, complementing the original fine-tuning ablation.

Figure 7 uses the same posed meshes and retargets the same motion to each rig, without manual asset correction. In (a), MIA's stretched, flattened arms and twisting/folding around the hips and legs persist after re-posing. AniRig recovers extended arms and separated legs and retains more intact limb contours under the target motion. In (b), MIA's head skew persists after re-posing; AniRig's head shape is closer to GT.

Held-out re-posing directly applies GT motion's per-joint local rotations and therefore needs joint correspondence. VLM-inferred hierarchies can use different joint definitions from the motion skeletons, so our quantitative held-out re-posing evaluation uses the benchmark hierarchy. For unedited VLM hierarchies, Table 5 reports quantitative rest recovery; the current re-posing evidence is qualitative, from compatible animations in Figure 5 and retargeted examples in Figure 7. We have clarified this evaluation scope in Limitations.

Revised manuscript: Appendix C.3, Table 6, p. 24, L1278-1294; Figure 7, p. 25, L1341-1345; Appendix C.1, p. 23, L1220-1223; Appendix E, p. 32, L1719-1724.

W2. Skinning formulation

Appendix B.2 now defines candidate selection in Eq. (6), bone-relative pair features in Eq. (7), candidate interaction in Eq. (8), and candidate attention in Eq. (9). Figure 6 expands the candidate scorer.

Candidate selection (qq: surface query; jj: joint):

u_{qj}&=B_{qj}+\lambda_h\kappa_jr^{\mathrm{heat}}_{qj} +\lambda_c\zeta g_{qj} \langle\widehat P^{(0)}_q,\widehat H^{(0)}_j\rangle,\\\\ \mathcal C_q&=\operatorname{TopK}_{j\\ \mathrm{valid}} (u_{qj},\min(K_c,K)). \end{aligned}$$ $B_{qj}$ is learned from joint/segment distances, distance falloff, and accessibility. The learned $r^{\mathrm{heat}}_{qj}$ combines surface/voxel heat and their agreement; $\kappa_j$ is the valid heat-field fraction. $P^{(0)},H^{(0)}$ are initial query/joint features; hats denote normalized learned projections. Compatibility uses learned scale $\zeta$ and gate $g_{qj}$ from distance falloff and accessibility. Fixed branch weights: $\lambda_h=1.35,\lambda_c=0.75$. $K,K_c$ denote joint count and candidate budget. **Bone-relative pair features and candidate interaction.** Pair features embed signed query offsets from the bone midpoint along the bone and two orthogonal transverse axes, joint/segment distances, distance falloff, surface-normal alignment toward the segment, and priors. These offsets distinguish similarly distant bones. Each block updates a query from $\mathcal C_q$, then the shared joint from all selecting queries $\mathcal Q_j=\\{q:j\in\mathcal C_q\\}$ using updated queries, then the pair from the updated query/joint states and its previous pair state. Weights normalize over $\mathcal C_q$ and $\mathcal Q_j$, respectively, using current features and geometry-enriched pair states. **Candidate attention.** The cue vector $r_{qj}$ contains distance, projection, accessibility, radial, and selector-score information. $\Delta r_{qjk}=r_{qj}-r_{qk}$ compares $j,k\in\mathcal C_q$. $E^{(2)}$ is the pair state after two blocks, and $\bar E$ its layer-normalized features: $$\begin{aligned} \pi^h_{qjk}&=\operatorname{softmax}_{k\in\mathcal C_q} \left(\frac{Q_h(\bar E_{qj})^\top K_h(\bar E_{qk})}{\sqrt{d_h}} +[\phi_{\mathrm{bias}}(\Delta r_{qjk})]_h\right),\\\\ m^{\mathrm{rel}}_{qj}&=\operatorname{Avg}_{k} \left[\sigma(\phi_{\mathrm{gate}}(\Delta r_{qjk})) (E^{(2)}_{qj}-E^{(2)}_{qk})\right],\\\\ E^{\mathrm{res}}_{qj}&=E^{(2)}_{qj}+m^{\mathrm{attn}}_{qj}+m^{\mathrm{rel}}_{qj}. \end{aligned}$$ Attention has four heads ($h$: index; $d_h$: width); $Q_h,K_h$ are learned projections. $\phi_{\mathrm{bias}}$ learns head-wise geometric logit biases. $\pi^h_{qjk}$ weights projected pair features; merging heads gives $m^{\mathrm{attn}}$. $\sigma(\phi_{\mathrm{gate}}(\Delta r_{qjk}))$ gates pair-feature differences in $m^{\mathrm{rel}}$; Avg divides zero-masked messages by the padded candidate count. Their sum with $E^{(2)}$ gives $E^{\mathrm{res}}$, then a residual feed-forward layer updates the pair features before bilinear fusion and the support and weight heads. *Revised manuscript: Appendix B.2, Eqs. (6)-(7), p. 18, L918-947; Eqs. (8)-(9), pp. 18-19, L949-1001; Figure 6, p. 17, L895-900.* ### **Q1. Training data and ModelsResource** Thank you for asking about training provenance and cross-dataset evaluation. We clarify the training data and report the additional results below. RigNet and Anymate use AniRig's Articulation-XL2.0 training split. UniRig and Puppeteer use official checkpoints trained on the same data and splits. SkinTokens uses an official checkpoint trained on 70% Articulation-XL2.0, 20% VRoid Hub, and 10% ModelsResource. VHD is non-learning. ModelsResource test results: |Method|J2J ↓|J2B ↓|B2B ↓| |---|---:|---:|---:| |RigNet|7.223|5.987|4.329| |Anymate|4.668|3.891|3.234| |UniRig|3.964|3.021|2.570| |Puppeteer|3.841|2.881|2.475| |SkinTokens|2.893|2.012|1.547| |**AniRig**|**2.079**|**1.470**|**1.242**| |Method|L1 Distance ↓|Recall ↑|Precision ↑| |---|---:|---:|---:| |RigNet|0.707|80.2%|65.7%| |VHD|0.514|88.3%|69.0%| |UniRig|0.531|80.6%|73.7%| |Puppeteer|0.443|81.6%|79.7%| |SkinTokens|0.327|87.3%|**82.7%**| |**AniRig**|**0.307**|**89.6%**|81.4%| AniRig has the lowest three skeleton errors and skinning L1 Distance, and highest Recall, consistent with Articulation-XL2.0 and supporting generalization to ModelsResource. *Revised manuscript: Appendix C.2, p. 24, L1253-1257; Appendix C.6, Tables 9-10, pp. 26-27, L1384-1414.*

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of Revisions and Responses to Reviewers

Dear Area Chair and Reviewers,

Thank you for your careful reviews and constructive suggestions. We have revised the manuscript and prepared detailed replies to each reviewer. Additions and substantive revisions are highlighted in blue.

The main updates are:

  1. Complete-asset comparison. We compare with Make-It-Animatable (MIA) on the same 40 humanoids, obtaining lower rest-recovery CD and higher F-Score (Appendix C.3, Table 6). AniRig uses VLM-inferred hierarchies; MIA uses its pipeline's hierarchies. Figure 7 also compares rest recovery and re-posing after retargeting the same motion to each recovered rig. All assets in this comparison are shown without manual correction.

  2. Controlled ablations. We evaluate paired skeleton prediction at the same pretraining stage, before asset-level fine-tuning (Appendix C.4, Table 7). We also remove GT orientation guidance during fine-tuning while retaining the other supervision; this setting improves all four asset metrics over the model without fine-tuning (Appendix C.5, Table 8).

  3. Method and training details. We add mathematical definitions for candidate selection and interaction, expand the candidate scorer in Figure 6 (Appendix B.2), disclose baseline training data and checkpoints (Appendix C.2), and report ModelsResource comparisons (Appendix C.6, Tables 9-10).

  4. Hierarchy and semantic conditioning. For the dense scorpion with multiple branches, most VLM-inferred topology matches the reference and only two mouthpart joints need adding, supporting reliable inference of its main skeletal structure (Appendix D.3). In the severe self-intersection case, the unedited VLM hierarchy matches the connectivity of the original asset's reference skeleton, supporting reliable hierarchy inference for this input (Appendix E). Cross-category name substitution preserves plausible skeletal layouts (Appendix C.7).

  5. Related work. We add discussions of Auto-Connect and Stroke3D in Section 2.

Shared clarification on fully inferred re-posing (BZWR and yv1m). VLM-inferred hierarchies can use different joint definitions from the GT motion skeletons. Applying the GT per-joint rotations requires establishing correspondence, so quantitative held-out re-posing uses the benchmark hierarchy. For unedited VLM hierarchies, the current re-posing evidence is qualitative, from compatible animations (Figure 5) and retargeted examples (Figure 7). Rest recovery is evaluated quantitatively in Tables 5 and 6. We have stated this evaluation scope in Limitations (Appendix E).

Our individual replies provide the experimental results, technical explanations, and precise locations in the revised manuscript. All page and line references below refer to that version.

AuthorsOpenReview ↗

What the New Experiments Show

Dear Area Chair and Reviewers,

Thank you for the discussion. We have responded to each review and revised the manuscript accordingly. The new experiments clarify where paired skeleton prediction and asset-level fine-tuning help, and how the resulting assets perform when animated.

Paired prediction and training supervision. At the end of pretraining, predicting the observed and rest skeletons together reduces all three observed-skeleton errors compared with predicting the observed skeleton alone (Table 7). This advantage is already present before the observed-only variant uses a ground-truth rest skeleton for asset-level fine-tuning. In a separate control using the benchmark hierarchy, fine-tuning without ground-truth orientation guidance still improves all four rest-recovery and held-out re-posing metrics over the pretrained model (Table 8).

Quality of the recovered assets. On the same 40 humanoids, AniRig recovers the rest mesh more accurately than Make-It-Animatable (Chamfer distance: 0.0811 vs. 0.105; F-Score: 0.621 vs. 0.490; Table 6). In the two examples in Figure 7, we retargeted the same motion to both methods’ rigs without manual asset correction. AniRig better preserves the arm contours in one case and the head shape in the other, showing how differences in rest recovery carry through to the animated characters.

AniRig’s central contribution is to recover a reusable rigged character from a single posed mesh and a supplied or inferred skeletal hierarchy. The revised evidence connects the learning approach to that practical goal: the ablations identify gains at each training stage, the ModelsResource comparisons extend the evidence for skeleton and skinning accuracy to a second dataset, and the retargeted examples show why recovery quality matters when the character moves. We thank the reviewers for helping us test this contribution more directly.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.