- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper presents TopX, a text-conditioned flow-matching Transformer for motion generation across 311 animal rigs with varying joint counts and topologies. Its TOPX-17 representation combines joint positions, continuous rotations, velocities, contact indicators, and root trajectory information, with normalization specific to each rig, joint, and channel. Skeleton conditioning uses a clean rest-pose frame, joint descriptions, structural features, and graph-dependent attention biases. Experiments evaluate generation on held-out clips of the training rigs, examine architectural and representation choices, and test transfer to unseen rigs with and without LoRA adaptation.
Critical strengths
- One model generates directly playable motion across a large collection of heterogeneous rigs. The 303M model handles 311 rigs spanning 34–102 joints and 108 distinct trees, and outputs joint rotations and world-space root trajectories without a separate inverse kinematics fit. On held-out clips of these rigs, R@1 reaches 0.867 compared with 0.904 for real motion under the authors’ evaluator (Table 1). At the 36M scale, the per-joint model also outperforms the flat padded-joint baseline by approximately 0.19 in R@1 under a matched training budget (Table 4).
- The motion representation is designed with animation in mind and provides useful geometric checks. TOPX-17 uses continuous 6D rotations relative to the rest orientation, rather than G-MDM’s Euler-angle parameterization, and separates body pose from root travel. Root trajectory and heading channels are used only at the root and masked elsewhere. Decoding does not require integrating predicted velocities. The independent position and rotation-based decodings expose internal inconsistencies through the FK–pose gap, which complements foot-sliding and articulation measurements where retrieval scores do not clearly distinguish the representation variants (Table 5 and Appendix K).
Critical weaknesses
- The abstract’s zero-shot and few-shot claims are broader than the unseen-rig results support. The zero-training setup requires motion statistics from the target rig. Table 9 shows similar results when these statistics are computed from adaptation clips alone, but target-rig motion data are still needed. In Table 3, the zero-training FK–pose gap ranges from 0.574 to 4.458 bone lengths, and jitter is approximately 2.5–13.2 times that of real motion. LoRA substantially reduces these errors, but articulation ratios remain at 0.24–0.77 relative to real clips. The four per-rig experiments also contain only 2–6 held-out clips each, and text fidelity on unseen rigs is not quantitatively evaluated. The adaptation results are useful, but the capability claim should reflect these limitations.
- The skeleton-conditioning contribution is not clearly distinguished from AnyTop. Both methods use a clean rest-pose frame, joint identity embeddings, tree-dependent spatial attention, and factorized attention over joints and frames. TopX differs in its use of LLM2Vec descriptions, an additional structural vector, and revised graph biases. Its text conditioning and directly playable output are also meaningful system-level differences. However, Section 3.2 does not clearly explain the relationship to AnyTop or establish what the revised conditioning mechanisms add. Removing joint descriptions slightly improves R@1 and FID in the reported 36M ablation (Table 4), and the graph-bias formulation is not directly compared with AnyTop’s.
Questions for authors
- What target-rig motion data are required to deploy the zero-training model? Would the Truebones results change if the rigs were processed with the current codec rather than the earlier build with different root smoothing, heading, and contact definitions described in Appendix A?
- Is there evidence that unseen-rig outputs respond reliably to different captions beyond the two prompt pairs in Figure 6? The current quantitative evaluation mainly measures geometric consistency, articulation, and jitter.
- Which conditioning choices are intended as methodological contributions beyond AnyTop? A clearer explanation in Section 3.2, supported by existing evidence or a small comparison on the 36M model, would help assess the structural vector and revised attention bias.
- How were deformation bones selected for the 311 training rigs, and what preparation is required for a new rig? Please clarify which steps are automated, whether twist or roll bones are retained, and how sensitive pruning and generation are to non-standard joint names or reworded descriptions.
- Was the evaluator trained exclusively on the training split? Appendix O states that checkpoints are selected using validation loss and that the reported results use the same validation set. Please clarify whether a separate test set was used and how stable the main ablation differences are across training runs.