acceptodds
Withdrawn from ICLR 2027

TopX: Skeleton-Aware In-Context Text-to-Motion Diffusion Across Heterogeneous Topologies

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Text-to-motion generation has mostly been built for one body: the joint count, joint order and pose parameterisation of a single skeleton are baked into the architecture, so every new character means a new model or a retargeting step. Serving hundreds of production rigs with one network is hard for several reasons: their animations share no common representation, the network must be told what each joint of an unfamiliar tree is, a loss summed over bodies of 34 to 102 joints lets the largest dominate, and a rig outside the training library must be served with no clips or a handful. We present TopX, a text-conditioned motion diffusion transformer that serves 311 heterogeneous animal skeletons with one set of weights, built on three design choices that answer these difficulties. TopX-17, a per-joint representation we design, converts any rigged animation from its kinematic tree, rest pose and motion alone and standardises every (rig, joint, channel) cell by its own statistics, so that rigs of any proportion share one network input without a template, retargeting or inverse kinematics. The skeleton enters as a prompt—its rest pose leads the sequence as a clean frame and its tree distances bias the attention—so nothing in the model is specific to a skeleton. A grouped flow-matching objective with calibrated channel weights keeps every channel group's share of the gradient independent of the number of cells it spans. Scored by a skeleton-aware text–motion evaluator, TopX reaches R@1 0.867, where the real clips score 0.904, and an FID of 0.005, and returns per-joint rotations and a world-space root trajectory that the rig plays directly. For rigs outside the library, the same model has both zero-shot and few-shot capability.

Primary area:
generative models
Keywords:
motion generation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_GnqAOpenReview ↗
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.
Reviewer_eYhHOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TOPX, a text-to-motion diffusion framework for generating motions across heterogeneous skeleton topologies. The method introduces TOPX-17, a per-joint representation with per-cell normalization, and incorporates skeleton information through rest-pose conditioning, joint descriptions, and graph-aware attention. A calibrated grouped flow-matching objective is further employed to balance supervision across motion channels. Experiments on 311 animal rigs demonstrate promising generation performance, along with zero-shot and few-shot adaptation capabilities for unseen skeletons.

Critical strengths
  • The paper addresses an important and challenging problem of text-conditioned motion generation across heterogeneous skeletons. The unified representation and skeleton-aware architecture offer a practical way of generalization.

  • The experiments cover a relatively large collection of animal rigs and provide extensive ablations on normalization, model scaling, motion representations, and training objectives. The few-shot adaptation results are also promising.

Critical weaknesses

Although the paper discusses several relevant approaches, the quantitative evaluation primarily focuses on internal variants and a simplified baseline. More direct comparisons with existing heterogeneous-skeleton text-to-motion generation methods in the main paper, for example, OmniZoo [Chen et al., 2025], G-MDM [Lee et al., 2025], and SAMoR [Zhang et al., 2026], would strengthen the contribution. Although AnyTop is included as an in-distribution reference in Appendix, a more comparable evaluation of motion quality and generalization would better establish the advantages of TOPX. Furthermore, several closely related works are missing from the discussion, such as NECromancer [Xu et al., 2026], PUMPS [Mo et al., 2025], and AnyMoLe [Yun et al., 2025] could also be acknowledged to better contextualize the proposed framework within generalizable character motion generation.

Also, the unseen-skeleton evaluation focuses on only four rigs with relatively few validation clips, making it difficult to establish generalization across diverse topologies. Lastly, the qualitative results of the generated animations are not convincing. The generated motions exhibit noticeable artifacts and unnatural movements, raising concerns about the overall motion quality.

Questions for authors

Could the authors provide direct comparisons with existing skeleton augnostic text-to-motion generation methods under comparable settings?

How does TOPX perform on completely unseen skeletons without target-specific motion statistics? (Such as motion from Mixamo, or in-the-wild artistic models)

Reviewer_LuMdOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces a new framework for text-to-motion generation with arbitrary skeletons. The core components are TOPX-17 motion representation, rest-pose prompt and graph-biased attention.

Critical strengths
  1. The overall writing quality is good, ensuring that its content is easily understandable for readers.
  2. The topic is interesting, which focuses on the text-drive motion generation given arbitrary skeletons.
Critical weaknesses
  1. The lack of comparison. There are many works focus on the same or very closed topics, as mentioned in the related work section. However, in Table 1, there is none baseline method, which makes it hard to justify the real contribution of the proposed method. For example, AnyTop is a diffusion-based method that can model different motion. An intuitative modification is to introduce the text condition to the diffusion model by cross-attention (like MotionDiffuse) or token concatenation (like MDM). Besides, "How to Move Your Dragon: Text-to-Motion Synthesis for Large-Vocabulary Objects" is accepted by ICML 2025 and proposed the benchmark for the same topic. The authors should use the same benchmark or re-implement the method to their own one.
  2. The novelty is limited and the motivation behind it is not new. For example, the motivation, "the network needs no per-skeleton parameter", has already been achieved by many previous work, such as AnyTop, "How to Move Your Dragon", "NECromancer: Breathing Life into Skeletons via BVH Animation". The concept is not new and the solution are also similar to these previous works.
  3. The benchmark is not such convincing. The R@1 is abnormally too high for both real motion and the generated motion. For example, in HumanML-3D this accuracy should be around 50%. One possible reason is that the authors mix different skeleton into the same batch, so that the evlauator just need to distinguish between different species rather than understand the semantic structure of the motion sequences.
  4. The qualitative results are not good. The generated motion sequences in the demo video show severe foot skating, floating and jittering. The quality is not acceptable.
Questions for authors

Please refer to the weakness part.

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

Related papers

Discussion (0)

Sign in to comment.