acceptodds
Under review as a conference paper at ICLR 2027

ChainMo: Chain-Segment Motion Modeling across Arbitrary Skeleton Topologies

Abstract

Generating motion for articulated characters with diverse skeleton topologies remains challenging because existing motion representations are often tightly coupled to the skeletal structures on which they are defined. Position-based representations are relatively flexible across skeletons, but remain sensitive to rig discretization, as the same articulated chain may contain different numbers of joints across rigs. Moreover, converting generated joint trajectories into animation-ready local rotations requires inverse kinematics, introducing ambiguities in under-constrained degrees of freedom such as bone twist. Directly modeling local rotations avoids this post-processing ambiguity, but further entangles motion generation with rig-specific kinematics and local rotation conventions. We address these limitations with **ChainMo**, a unified framework that separates shared motion modeling from target-rig-specific rotation realization. ChainMo represents motion using a dedicated root token and a variable set of disjoint, non-branching kinematic chain-segment tokens, forming a shared motion abstraction that is robust to variations in within-chain joint resolution. A target-rig-conditioned decoder realizes the abstract motion as coherent local joint rotations without relying on external inverse kinematics. For motion generation, we further introduce a flow-matching model with global spatiotemporal attention over the variable set of chain-segment tokens, enabling shared generative modeling across heterogeneous articulated characters. Extensive experiments on characters with diverse and previously unseen skeletons demonstrate strong generalization across kinematic structures and rig resolutions, accurate reconstruction of local joint rotations, and effective generative modeling in the learned chain-segment latent space. See our anonymous project page for more demos: https://anonymous1778.github.io/chainmo/.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Motion Generation, Diffusion Model, Skeletal Animation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_HRbVOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ChainMo, a chain-based motion representation for modeling and generating motions across heterogeneous skeletons. It represents motion using non-branching kinematic chains in a shared latent space instead of individual joints, making it more robust to different joint discretizations. The method supports both text- and video-conditioned motion generation.

Critical strengths
  1. Intuitive design. The chain-based representation is simple and well-motivated, providing a natural way to handle different skeleton topologies and joint discretizations.

  2. Good results. The method achieves competitive results numerically and shows good performance on qualitative comparisons with other methods.

Critical weaknesses
  1. Insufficient quantitative evaluation of heterogeneous motion generation. The main quantitative results are reported on HumanML3D, which uses a fixed human skeleton and it is text-to-human motion dataset. For heterogeneous skeletons, the authors mainly rely on qualitative comparisons and user studies. Although there is no established cross-topology evaluator, more objective quantitative evaluations would strengthen the paper.

  2. Missing quantitative evaluation of video-conditioned generation. Since SkelMo provides paired video-motion data, it is unclear why the authors only show qualitative results.

  3. Insufficient ablation studies. The necessity of T5 joint-name embeddings is not validated.

  4. Inappropriate reconstruction metrics. Table 1 mainly reports FID and semantic alignment metrics, while direct reconstruction metrics such as MPJPE and rotation error would be more informative. semantic alignment seems not approprate for reconstruction evluation.

  5. Limited comparison with recent methods. Table 4 mainly compares against methods from 2023–2024, without including more recent text-to-motion approaches.

  6. Unclear text annotation protocol. Since SkelMo primarily contains video-motion pairs, the paper does not clearly explain how text annotations for heterogeneous motion generation are obtained.

  7. Questionable qualitative examples. Some examples in Figure 5 do not seem well suited , such as the shark-like LEGO character jumping with its arms spread, making the qualitative comparison less convincing.

Questions for authors
  1. Why is heterogeneous motion generation mainly evaluated through user studies? Could the authors provide more objective quantitative comparisons?

  2. Since SkelMo provides paired video-motion data, why is video-conditioned generation only evaluated qualitatively?

see Critical Weaknesses for other questions

3 comments
AuthorsOpenReview ↗

Response to Reviewer HRbV (1/2)

We thank the reviewer for the detailed feedback. We added quantitative evidence and clarified the evaluation protocol.

W1/Q1: Quantitative heterogeneous generation. Since no standard evaluator covers arbitrary-topology text-to-motion, we introduce joint-pooled kinematic FIDk\mathrm{FID}_k, diversity, minMPJPE, joint speed, foot skating, penetration, and jitter (Sec. 4.3; App. A.2).

We retrain AnyTop-T with text and first-frame conditioning on the same split. On unseen skeletons (Tab. 4a):

MethodFIDk\mathrm{FID}_k ↓minMPJPE ↓Joint spd. →Jitter →
GT––0.411127.0
AnyTop-T7.590.2360.494312.3
ChainMo2.220.1290.404165.6

ChainMo substantially reduces FIDk\mathrm{FID}_k (non-overlapping 95% CIs: 1.86–2.88 vs. 6.94–8.55), lowers minMPJPE, and better preserves motion dynamics.

For mesh-animation methods (100 cases, skeleton fitting error ≤0.3%; Tab. 4b):

MethodFIDk\mathrm{FID}_k ↓Joint spd. →Jitter →VS2-Align. ↑
GT–0.38651.92.85
AnimateAnyMesh21.980.1275.12.79
BiMotion45.100.0202.22.35
AnyTop-T5.480.44868.22.79
ChainMo1.670.36748.12.86

The mesh baselines are nearly static (5–33% of GT joint speed); ChainMo better preserves dynamics and text alignment. These metrics complement our user study (91.86% semantic and 91.71% naturalness preference, App. A.4), rather than replacing a learned cross-topology semantic evaluator.

W2/Q2: Video-conditioned evaluation. We compare with SkelMo on 100 shared clips, frames 0–31, matching SkelMo's 32-frame window (Tab. 13). Both receive the rig and initial pose; SkelMo additionally uses a calibrated camera and projected first-frame joints.

MethodMPJPE ↓Root-rel. MPJPE ↓Root err. ↓Jitter →
GT–––109
SkelMo0.0620.0590.027377
ChainMo0.0660.0530.03496

ChainMo has comparable overall MPJPE, lower articulation error and jitter, but SkelMo recovers root trajectories and ground contact more accurately. ChainMo directly produces local rotations without external IK. Video conditioning is a demonstration of modality flexibility, not our primary task.

W3: T5 joint-name embeddings. We retrained the flow model without joint-name inputs, keeping the pretrained Chain-DAE, data, and training schedule fixed. We compare the two variants at four matched training checkpoints on unseen skeletons (Tab. 12):

VariantFIDk\mathrm{FID}_k ↓minMPJPE ↓Diversity ↑
With names (ours)2.74 ± 0.400.132 ± 0.0020.119 ± 0.008
Without names2.53 ± 0.620.132 ± 0.0010.112 ± 0.014

The two variants achieve comparable performance: the FIDk\mathrm{FID}_k difference is smaller than the variation across checkpoints, and minMPJPE is identical, while names yield slightly higher diversity. This indicates that joint-name inputs are not essential for the flow model in this setting and serve as a complementary cue. The pretrained Chain-DAE remains fixed and retains its own name conditioning.

W4: Reconstruction metrics. The original paper already reported heterogeneous MPJPE (0.0287/0.0347) and rotation error (6.99°/6.96°) on seen/unseen rigs (Tab. 2). We now additionally report HumanML3D MPJPE alongside established tokenizer metrics (Tab. 1):

MethodFID ↓MPJPE ↓R@1 →MM-Dist. →
GT––0.5132.969
MoMask0.0190.03000.5082.999
SAMoR0.0080.01730.5142.977
ChainMo0.0100.02740.5142.969

ChainMo achieves competitive reconstruction fidelity (MPJPE 0.0274 m, outperforming MoMask's 0.030 m), while closely preserving ground-truth semantic statistics (R@1 0.514 vs. GT 0.513, MM-Dist. 2.969 vs. GT 2.969).

While SAMoR achieves lower reconstruction error, ChainMo prioritizes topology-adaptive compression over maximizing reconstruction fidelity on a fixed human skeleton. Importantly, lower reconstruction error does not necessarily translate into better downstream generation: ChainMo achieves a better HumanML3D generation FID (0.034 vs. 0.039; Tab. 5).

Competing tokenizers do not report rotation errors, so a direct comparison is unavailable; ChainMo reports 5.77° in Tab. 6. We retain R-Precision and MM-Dist as complementary measures of semantic preservation, while MPJPE and rotation error directly evaluate reconstruction fidelity.

AuthorsOpenReview ↗

Response to Reviewer HRbV (2/2)

W5: Recent methods. We expanded Tab. 5 with six recent methods and SAMoR:

MethodFID ↓MM-Dist. ↓R@1 ↑
BAMM (2024)0.0552.9190.525
LaMP (2025)0.0322.7590.557
Motion Anything (2025)0.0282.8590.546
SegMo (2026)0.0422.7820.553
FloodDiffusion (2026)0.0572.8870.523
MotionHiFlow (2026)0.0322.6910.563
SAMoR (2026)0.0392.8970.523
ChainMo0.0342.6710.565

Thus ChainMo remains competitive on HumanML3D while supporting heterogeneous skeletons; surpassing human-specific SOTA is not its main aim.

W6: Text annotations. Truebones Zoo captions come from Lee et al. (2025). For SkelMo motions, we render four views, sample frames over time, generate captions with a VLM (DeepSeek-V4.1-Flash), and manually review/correct them (App. A.2).

W7: Fig. 5 examples. We thank the reviewer for this observation. We would like to clarify that the original character was a humanoid shark-like LEGO figure, whose articulated structure supports human-like actions such as jumping with arms spread.

Nevertheless, to avoid visual ambiguity and make the qualitative comparison more convincing, we have replaced it with a reindeer performing a head-strike and added an unseen-topology insect flapping its wings. These examples more clearly demonstrate motions involving non-humanoid body parts. All methods are compared using the same mesh, camera, and text prompt for each case.

AuthorsOpenReview ↗

Follow-up on our response

Dear Reviewer HRbV,

Thank you again for your detailed review. As the discussion period is coming to a close, we would be grateful to know whether our responses address your concerns, or if there is anything else we can clarify.

Best regards,
The Authors

Reviewer_VfZYOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper targets the interesting problem of motion generation for articulated characters with different skeletons. In this paper, it presents a new method called, ChainMo, which includes a unified framework to spearate shared motion modeling from target-rig specific rotation realization. The paper is well motivated and the experiments show reasonable performance compared with the baselines.

Critical strengths
  • The paper targets the important problem of motion generation for articulated characters with different skeletons, which is of great importance to the industry community.

  • The paper is well motivated. Replacing "joint count" with "branch" as the unit count is a small conceptual shift, but it solves the problems in the target problem.

  • The experimental results on HumanML3D obtain better results compared with the baselins accroding to Table 1. Also, a set of ablations are provided to validate effectiveness of the proposed approach.

Critical weaknesses
  • The experiments should include more baselines which are published in the recent two years. Currently, the baselines are published in 2023 and 2024, which are not SOTA in the current benchmark like HumanML3D.

  • In Section 2, the paper includes the related work of TopoCap (SIGGRAPH'26), SAMoR (2026), and MotionS (2024). How about the comparison with these algorithm based on the experiments?

  • One of the important claims of a "shared motion space" is never quantitatively tested.

  • In the training details, the paper reports it operates on 196-frame sequences. How to handle longer sequence?

Questions for authors

Please mainly address the issues raised in the weakness section.

5 comments
AuthorsOpenReview ↗

Response to Reviewer VfZY

We thank the reviewer for the constructive feedback. We added the following comparisons and analyses.

W1: Recent baselines. We expanded Tab. 5 with six methods from 2024–2026 and the concurrent cross-topology method SAMoR:

MethodFID ↓MM-Dist. ↓R@1 ↑
BAMM (2024)0.0552.9190.525
LaMP (2025)0.0322.7590.557
Motion Anything (2025)0.0282.8590.546
SegMo (2026)0.0422.7820.553
FloodDiffusion (2026)0.0572.8870.523
MotionHiFlow (2026)0.0322.6910.563
SAMoR (2026)0.0392.8970.523
ChainMo0.0342.6710.565

ChainMo maintains competitive HumanML3D fidelity (FID 0.034 vs. best 0.028) and achieves the best alignment among these comparisons, while supporting heterogeneous skeletons rather than only a fixed human rig.

W2: TopoCap, SAMoR, and MotionS. We address the different task settings and provide available comparisons:

  • TopoCap primarily addresses monocular video-to-animation. Nevertheless, we evaluated its official code/checkpoint on 100 unseen clips, matching camera renders, correcting asset orientation, and providing the first frame as rest pose:

    MethodRoot-rel. MPJPE ↓
    TopoCap0.185
    SkelMo0.059
    ChainMo0.053

    TopoCap is designed around front-view inputs, whereas our evaluation uses random views; this viewpoint mismatch may contribute to the gap.

  • SAMoR is closely related, but a publicly usable implementation was unavailable for our protocol. We therefore include its published HumanML3D reconstruction and generation results (Tabs. 1 and 5).

  • MotionS addresses cross-skeleton stylization of an input motion, not text-conditioned motion generation.

For heterogeneous text-to-motion, we instead retrain AnyTop-T with text and first-frame conditioning on the same training split. On unseen rigs, ChainMo reduces FIDk\mathrm{FID}_k 7.59→2.22 and minMPJPE 0.236→0.129 (Tab. 4).

W3: Shared motion space. We quantitatively evaluate whether the learned representation captures motion semantics rather than skeleton identity through cross-skeleton retrieval (Tab. 10). We compare frozen Chain-DAE latents against AnyTop's text-free denoiser features.

Both representations achieve comparable cross-rig retrieval performance (mAP: 0.297 vs. 0.299). However, they behave very differently when same-rig distractors are included:

RepresentationSame-rig confusion ↓Action > Identity ↑
AnyTop features37.7%57%
Chain-DAE (ours)5.2%74%

We evaluate these metrics on 77 queries whose rigs also perform other actions, using all test clips as the retrieval gallery. Same-rig confusion measures how often the top-1 match is a different action on the same rig. Action > Identity measures how often same-action motions on different rigs are more similar than different actions on the same rig.

ChainMo reduces same-rig confusion from 37.7% to 5.2%, while increasing Action > Identity from 57% to 74%, despite comparable cross-rig retrieval performance. These results provide direct quantitative evidence that our chain-segment representation is less biased toward skeleton identity than AnyTop's joint-level features, complementing the rig-densification consistency (Tab. 3) and cross-skeleton latent realization (Fig. 7).

W4: Longer sequences. ChainMo generates 196 frames in one pass (9.8 s at 20 fps on HumanML3D), covering many complete actions. This is substantially longer than the fixed windows of AnimateAnyMesh (16/32 frames) and SAMoR (64 frames). Longer sequences could potentially use autoregressive latent chunks, but additional boundary conditioning/training is needed for continuity; we clarify this in App. A.6.

AuthorsOpenReview ↗

Follow-up on our response

Dear Reviewer VfZY,

Thank you again for your constructive feedback. In our response we:

(1) added six recent methods and SAMoR to Tab. 5: ChainMo remains competitive in fidelity (FID 0.034 vs. best 0.028) and achieves the best text alignment (R@1 0.565, MM-Dist 2.671), while supporting heterogeneous skeletons;

(2) evaluated TopoCap with its official code on 100 unseen clips (root-relative MPJPE 0.185 vs. 0.053 for ChainMo), and compared against AnyTop retrained with text conditioning on unseen rigs (FID$_k$ 7.59→2.22, Tab. 4);

(3) quantitatively tested the shared motion space through cross-skeleton retrieval (Tab. 10): with cross-rig retrieval comparable to AnyTop (mAP 0.297 vs. 0.299), our latent is far less biased toward skeleton identity: same-rig confusion drops from 37.7% to 5.2%, and Action > Identity rises from 57% to 74%.

(4) clarified the remaining points: SAMoR has no publicly usable implementation for our protocol, so we include its published results (Tabs. 1 and 5); MotionS addresses stylization rather than text-to-motion; and ChainMo generates 196 frames (9.8 s) in a single pass, longer than the fixed windows of AnimateAnyMesh (16/32) and SAMoR (64), with extensions discussed in App. A.6.

As the discussion period is coming to a close, we would appreciate knowing whether our response resolves your concerns, or if there is anything further we can clarify.

Best regards,
The Authors

Dear Reviewer VfZY

As the discussion period is coming to an end, we would appreciate knowing whether our comprehensive response resolves your concerns. Looking forward to hearing from you.

Best regards,

The authors

Reviewer_JzPoOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ChainMo, a unified framework for motion generation across heterogeneous skeleton topologies. Its key contribution is a topology-adaptive chain-segment representation that abstracts joint-level motion into variable-length sets of kinematic chain tokens. A target-rig-conditioned decoder reconstructs local joint rotations, while a conditional flow-matching model generates motions in the learned latent space.

Critical strengths

I think this paper has two main strengths:

  1. Novel Topology-Adaptive Motion Representation. The paper introduces a chain-segment representation that effectively reduces the dependence of motion modeling on joint-level discretization. This provides a meaningful alternative to conventional joint-based representations and enables motion modeling across heterogeneous skeletons.
  2. Effective Separation of Motion Modeling and Rotation Realization. The proposed framework separates shared motion dynamics from target-rig-specific rotation reconstruction. The rig-conditioned decoder directly predicts local joint rotations, avoiding external inverse kinematics and providing animation-ready motion sequences.
Critical weaknesses

There are main weaknesses for this paper:

  1. The method seems hard to handle changes in skeleton branching structures. While chain-segment representations are robust to variations in joint discretization within a chain, adding or removing branches may alter the chain decomposition and disrupt the learned motion representation. Therefore, the claimed generalization to arbitrary skeleton topologies is not fully justified. I suggest the author can add experiments to check how the method performs when we add or remove branches.

  2. The method needs more experiments on unseen skeletons to demonstrate its effectiveness. Table 2 only compares the proposed method between seen and unseen settings, without quantitative comparisons against existing methods such as AnyTop or SkelMo. The cross-topology motion generation evaluation also relies heavily on qualitative comparisons and user studies, making it difficult to assess the actual improvements over existing approaches.

Questions for authors

see Critical Weaknesses

4 comments
AuthorsOpenReview ↗

Response to Reviewer JzPo

We thank the reviewer for the positive assessment and constructive suggestions. We added targeted quantitative experiments addressing both concerns.

W1: Adding or removing branches. Chain decomposition follows the branching structure; edits may add/remove segment tokens or split/merge segments at branching joints. We test three structural edits (Tab. 11): removing a terminal chain, removing a limb subtree, and adding a copied leaf chain at an existing branching joint (e.g., a second tail). Addition is tested on unseen clips. All metrics are computed on joints shared by the original and edited skeletons. For minMPJPE, the two values denote generation errors before and after the edit, evaluated against the same ground-truth motion on shared joints.

EditΔJointsRecon. drift ↓Gen. drift ↓Seed spreadminMPJPE (before→after) ↓
Remove terminal chain−16%0.0090.0520.1090.141→0.143
Remove subtree (limb)−26%0.0110.0590.1060.135→0.139
Add copied chain+16%0.0060.0320.1010.125→0.126

The results demonstrate robust motion preservation: Reconstruction drift on shared joints is only 0.006–0.011, while generation drift is 3.2–5.9% of the character's bounding-box diagonal, substantially smaller than random-seed variation (10.1–10.9%). For 87–99.6% of clips, edit-induced drift is below the seed spread, and minMPJPE remains nearly unchanged.

These findings complement Tab. 3, where chain segments reduce decoded-motion changes by 61.6% compared with joint tokens under 3× densification. We do not claim invariance to arbitrary structural edits, especially when functionally important parts such as wings are affected (App. A.6).

W2: Quantitative comparison on unseen skeletons. We introduce a topology-agnostic protocol using kinematic-descriptor FIDk\mathrm{FID}_k, diversity, minMPJPE, and physical plausibility metrics (Sec. 4.3, Tab. 4; App. A.2). We extend AnyTop with text and first-frame conditioning (AnyTop-T) and retrain it on the same split, with identical prompts and initial poses:

MethodFIDk\mathrm{FID}_k ↓minMPJPE ↓Joint spd. →Jitter →
GT––0.411127.0
AnyTop-T7.590.2360.494312.3
ChainMo2.220.1290.404165.6

On unseen rigs, ChainMo improves FIDk\mathrm{FID}_k with non-overlapping 95% bootstrap CIs (1.86–2.88 vs. 6.94–8.55) and better matches GT motion dynamics.

We also compare with mesh-animation approaches on a 100-case subset, fitting the asset's rig to each generated mesh (re-skinning error ≤0.3%; Tab. 4b):

MethodFIDk\mathrm{FID}_k ↓Joint spd. →Jitter →VS2-Align. ↑
GT–0.38651.92.85
AnimateAnyMesh21.980.1275.12.79
BiMotion45.100.0202.22.35
AnyTop-T5.480.44868.22.79
ChainMo1.670.36748.12.86

AnimateAnyMesh and BiMotion are nearly static (5–33% of GT joint speed). ChainMo better preserves dynamics and text alignment. Finally, in the separate video-conditioned comparison with SkelMo (100 clips, frames 0–31; Tab. 13), ChainMo achieves comparable MPJPE (0.066 vs. 0.062), lower root-relative MPJPE (0.053 vs. 0.059), and much lower jitter (96 vs. 377), while predicting local rotations directly without external IK.

AuthorsOpenReview ↗

Follow-up on our response

Dear Reviewer JzPo,

Thank you again for your thoughtful review. In our response we:

(1) tested adding and removing branches directly (Tab. 11): on the joints shared by the original and edited skeletons, generation drift is only 3.2–5.9% of the character’s bounding-box diagonal, well below the variation from merely resampling the random seed (10.1–10.9%), and below it for 87–99.6% of clips, while minMPJPE remains nearly unchanged;

(2) added a quantitative comparison on unseen skeletons (Tab. 4): against AnyTop retrained with text and first-frame conditioning, ChainMo reduces FID$_k$ from 7.59 to 2.22 (non-overlapping 95% CIs) and minMPJPE from 0.236 to 0.129, while mesh-animation methods (AnimateAnyMesh, BiMotion) remain nearly static.

As the discussion period is coming to a close, we would be grateful to know whether these address your concerns, or if there is anything else we can clarify.

Best regards,
The Authors

Thanks for your detailed explanations, I keep my weak acceptance score.

Thank you for your time and consideration. We appreciate your feedback.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response

We thank all reviewers for their constructive feedback. The reviewers found the chain-segment representation novel and well motivated (JzPo, VfZY, HRbV), and the separation of motion modeling from rotation realization effective (JzPo). The main shared concern was the lack of objective quantitative evaluation on heterogeneous skeletons. We addressed it with new experiments. All changes are highlighted in magenta in the revised PDF.

The key revisions are summarized below:

  1. Quantitative heterogeneous motion generation (Sec. 4.3, Tab. 4): We introduce a topology-agnostic evaluation protocol and provide quantitative comparisons with AnyTop-T, AnimateAnyMesh, and BiMotion.

  2. Robustness to topology changes (App. A.5, Tab. 11): We evaluate motion stability under both branch addition and removal.

  3. Shared motion space (App. A.5, Tab. 10): We introduce cross-skeleton retrieval experiments comparing ChainMo with AnyTop, demonstrating reduced sensitivity to skeleton identity.

  4. Additional quantitative comparisons (Tabs. 1, 5, 13): We add recent HumanML3D baselines, direct reconstruction metrics, and video-conditioned comparisons with SkelMo. We also provide a supplementary comparison with TopoCap using its official implementation in the reviewer response.

  5. Ablations and clarifications (App. A.2, A.5–A.6; Tab. 12; Fig. 5): We investigate joint-name conditioning through retraining, clarify the text annotation protocol and sequence-length limitations, and improve qualitative examples.

Detailed responses to each concern are provided in the corresponding reviewer discussions.

AuthorsOpenReview ↗

Rebuttal Summary for the Area Chair

Dear Area Chair,

Thank you for handling our submission. We sincerely thank the reviewers for their time and constructive feedback, and we kindly provide a brief summary of the rebuttal in the hope that it may assist your evaluation. We are grateful that all three reviewers found the chain-segment representation novel and well motivated. The main concerns and our responses (revised PDF) are as follows:

  • Quantitative evaluation on heterogeneous skeletons (JzPo, HRbV). The reviewers noted that results on unseen skeletons were mainly qualitative. We introduced a topology-agnostic protocol (kinematic FID$k$, minMPJPE, diversity, and physical plausibility) and retrained AnyTop with text conditioning on the same split. On unseen skeletons, ChainMo reduces FID$k$ from 7.59 to 2.22 (non-overlapping 95% CIs) and minMPJPE from 0.236 to 0.129, while mesh-animation methods remain nearly static (Tab. 4).
  • Shared motion space (VfZY). The reviewer asked whether the latent space is truly shared across skeletons. Using cross-skeleton retrieval with frozen latents, our representation retrieves the same action on a different skeleton rather than a different action on the same skeleton: same-rig confusion 5.2% vs. 37.7% for AnyTop features, at comparable cross-rig retrieval accuracy (Tab. 10).
  • Adding/removing branches (JzPo). The reviewer asked how structural edits affect the motion. When removing a chain or limb, or adding a new chain, the drift on the remaining joints is well below the variation from simply resampling the random seed, with nearly unchanged accuracy (Tab. 11).
  • Further analyses (HRbV, VfZY). Retraining the generator without joint names performs on par (Tab. 12); in video-conditioned generation, ChainMo achieves lower root-relative error and jitter than SkelMo and outperforms TopoCap (Tab. 13); we also added HumanML3D reconstruction metrics (Tab. 1) and six recent methods plus SAMoR, where ChainMo achieves the best text alignment (Tab. 5).

We are encouraged by the discussion:

  • Reviewer HRbV found that our response addresses their main concerns and raised their rating.
  • Reviewer JzPo maintained their positive rating.
  • Reviewer VfZY has not responded during the discussion period; we have addressed each of their points (recent baselines, TopoCap/SAMoR/MotionS, shared motion space, and sequence length) with new experiments or clarifications.

We sincerely appreciate your time and effort in evaluating our work.

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.