- 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
-
Intuitive design. The chain-based representation is simple and well-motivated, providing a natural way to handle different skeleton topologies and joint discretizations.
-
Good results. The method achieves competitive results numerically and shows good performance on qualitative comparisons with other methods.
Critical weaknesses
-
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.
-
Missing quantitative evaluation of video-conditioned generation. Since SkelMo provides paired video-motion data, it is unclear why the authors only show qualitative results.
-
Insufficient ablation studies. The necessity of T5 joint-name embeddings is not validated.
-
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.
-
Limited comparison with recent methods. Table 4 mainly compares against methods from 2023–2024, without including more recent text-to-motion approaches.
-
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.
-
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
-
Why is heterogeneous motion generation mainly evaluated through user studies? Could the authors provide more objective quantitative comparisons?
-
Since SkelMo provides paired video-motion data, why is video-conditioned generation only evaluated qualitatively?
see Critical Weaknesses for other questions
3 comments
Response to Reviewer HRbV (2/2)
W5: Recent methods. We expanded Tab. 5 with six recent methods and SAMoR:
| Method | FID ↓ | MM-Dist. ↓ | R@1 ↑ |
|---|---|---|---|
| BAMM (2024) | 0.055 | 2.919 | 0.525 |
| LaMP (2025) | 0.032 | 2.759 | 0.557 |
| Motion Anything (2025) | 0.028 | 2.859 | 0.546 |
| SegMo (2026) | 0.042 | 2.782 | 0.553 |
| FloodDiffusion (2026) | 0.057 | 2.887 | 0.523 |
| MotionHiFlow (2026) | 0.032 | 2.691 | 0.563 |
| SAMoR (2026) | 0.039 | 2.897 | 0.523 |
| ChainMo | 0.034 | 2.671 | 0.565 |
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.
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
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, 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):
ChainMo substantially reduces FIDk (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):
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.
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):
The two variants achieve comparable performance: the FIDk 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):
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.