acceptodds
Under review as a conference paper at ICLR 2027

TRoPE: Topology-aware Relative Positional Encoding for 4D Generation

Abstract

Video-to-4D mesh generation has made substantial progress in recovering dynamic 3D geometry and motion from monocular videos. However, commonly used feed-forward approaches remain prone to sticking artifacts, where geodesically distant vertices move together incorrectly when they become spatially close. We attribute this failure to cross-part motion leakage: spatially nearby yet structurally distant vertices may retrieve similar motion information when motion decoding fails to distinguish spatial proximity from mesh connectivity. To address this issue, we introduce TRoPE, a topology-aware relative positional encoding that incorporates mesh structure into motion generation. TRoPE derives spectral features from the graph Laplacian and constructs pairwise topological relations from distances between multi-scale diffusion features, yielding a stable topology representation across meshes. We further introduce a vertex-wise tail-risk objective that emphasizes persistently high-error vertices to explicitly suppress sticking artifacts. To better characterize and quantify these failures, we introduce MeshSticking-4D, a dedicated benchmark for this failure mode. Experiments on Motion-80 and MeshSticking-4D show that TRoPE consistently suppresses sticking artifacts while maintaining comparable overall generation quality. Compared with the baseline, it reduces the worst-vertex error by 39.4% and lowers the separation failure rate from 49.35% to 34.90%. Additional evaluations across different experiments further demonstrate the general applicability of TRoPE. The code and benchmark will be made publicly available.

Primary area:
generative models
Keywords:
4D Generation, RoPE, Mesh 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_RuprOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces TRoPE, which constructs multi-scale topology-aware positional encodings to encourage appropriate motion sharing between topologically similar locations while distinguishing spatially close but structurally different locations, thereby reducing cross-part motion leakage and mitigating sticking artifacts.

Critical strengths
  1. The experimental evaluation is relatively thorough, with extensive ablation studies demonstrating TRoPE’s advantages in 4D generation over various positional encodings, including Euclidean Relative RoPE and topology-aware geodesic relative positional encoding.
  2. The motivation is clear. The paper identifies sticking artifacts as an existing problem, analyzes their causes, and mitigates them through topology-aware positional encoding. The motivation and experimental design are well aligned and reasonable.
Critical weaknesses
  1. TRoPE does not demonstrate a clear advantage over geodesic relative positional encoding. The ablation results show that geodesic relative positional encoding already substantially outperforms other encodings, as it is also topology-aware. However, the reported metrics do not indicate a substantial improvement of TRoPE over this alternative, and the paper lacks a discussion of why TRoPE would be preferable to geodesic relative positional encoding.
Questions for authors
  1. What is the computational cost of TRoPE compared with geodesic relative positional encoding and other positional encoding methods?
  2. Could the authors provide additional qualitative comparisons illustrating differences in 4D generation quality across different positional encoding methods?
  3. Could the authors provide insights into why TRoPE outperforms geodesic relative positional encoding, supported by experimental evidence?
Reviewer_7j4TOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces TRoPE, a topology-aware relative positional encoding for 4D generation conditioned by reference mesh and video. It constructs pairwise relations from multiscale spectral diffusion features and uses them to modulate anchor attention through rotary transformations. The method also incorporates local topology-aware aggregation and a vertex-wise tail-risk objective. Experiments on the proposed MeshSticking-4D benchmark demonstrate reduced sticking artifacts, with additional evaluation on Motion-80.

Critical strengths
  1. Distinguishing spatial proximity from mesh connectivity addresses an important failure mode in articulated motion reconstruction.
  2. MeshSticking-4D provides explicit contact-to-separation evaluation and separates both character and source animation identities between training and testing.
  3. The encoding comparisons disable CVaR and match the training protocol. TRoPE improves worst vertex error and separation failure rate over content-only and Euclidean-relative attention, supporting the usefulness of connectivity information.
Critical weaknesses
  1. Appendix A.3.3 describes directly adding the distance relation to attention logits. Without a calibrated sign, scale, or learned mapping, this is a limited baseline and may favor structurally distant anchors.
  2. The contribution of TRoPE under tail-risk training remains insufficiently isolated. The headline 39.4% WVE reduction combines several architectural and training changes. Although Table 3 isolates the encoding without CVaR, a matched content-only model with CVaR is missing, leaving the encoding’s additional benefit under the final training objective unclear.
  3. On Motion-80, TRoPE performs worse than Motion 3-to-4 on all four metrics: CD increases by approximately 8.7% and HD95 by 12.7%. These results warrant a more explicit discussion of the trade-off between sticking robustness and overall reconstruction quality.
Questions for authors
  1. Does rotary modulation still outperform a learned or appropriately scaled topology-relative additive bias?
  2. How does TRoPE compare with content-only anchor attention when both use the same CVaR objective?
  3. Which components cause the Motion-80 regression, and can the robustness and accuracy trade-off be quantified?
  4. AnimateAnyMesh++ already proposes topology-aware aggregation to mitigate trajectory sticking. Could the authors clarify whether sticking artifacts persist in AnimateAnyMesh/AnimateAnyMesh++, and how much additional benefit TRoPE provides over an AnimateAnyMesh++-style decoder under matched training conditions? What's more, representative cases where TRoPE resolves failures of this baseline would help clarify its contribution. If the authors can answer the questions above, I'm willing to revise my recommendation to positive.
Reviewer_9yG3OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper addresses sticking artifacts in video-to-4D generation, where spatially close but geodesically distant vertices exhibit incorrectly coupled motion. The authors propose sampling anchor points to aggregate information from motion tokens, and then leveraging multi-scale graph diffusion features to modulate vertex-to-anchor attention through rotary transformations. The method also introduces a vertex-wise tail-risk training objective.

Critical strengths
  1. Sticking artifacts are a practical issue in current video-to-4D mesh generation methods.

  2. The proposed design is generally reasonable, and the experiments provide some evidence supporting its effectiveness.

Critical weaknesses

I have several concerns regarding the experimental setup, particularly its fairness and the generalizability of the results. 1) Limited dataset quality and diversity. The proposed MeshSticking-4D dataset appears limited in both visual quality and diversity. The rendered videos seem overexposed, with some image regions appearing completely white. Moreover, the dataset focuses exclusively on human motion, leaving the method's generalizability to other object categories insufficiently demonstrated. 2) Concerns about the fairness of the comparisons. Although Table 1 shows that the proposed method achieves the best performance across multiple metrics, the evaluation is limited to MeshSticking-4D, on which the proposed method is explicitly fine-tuned. Therefore, the improvements may partly reflect better adaptation to this specific data distribution rather than the effectiveness of the proposed design. Notably, Motion 3-to-4, which also appears to be fine-tuned on MeshSticking-4D according to the appendix, achieves considerably closer performance than several other baselines. For a fairer comparison, it would be important to evaluate the other methods after comparable fine-tuning. 3) Degraded general reconstruction performance. Table 2 shows that the proposed method performs worse than the original Motion 3-to-4 baseline across all reported reconstruction metrics. This raises concerns about whether the proposed design improves sticking robustness at the expense of general reconstruction quality, leaving its overall effectiveness uncertain.

Questions for authors
  1. Anchor-to-motion attention. The authors use sampled anchor points to aggregate information from the motion tokens. Could the authors clarify the implementation details of this attention mechanism? Specifically, is RoPE or any other positional encoding used in this stage?

  2. Necessity of anchor points. The core idea behind addressing sticking artifacts is to incorporate mesh topology into motion decoding. A more straightforward alternative would be to directly incorporate topological features into each vertex query and perform cross-attention with the motion tokens, without introducing intermediate anchor points. How would such a simpler approach perform, and what motivates the additional anchor-based design?

  3. Distinction from prior topology-aware methods. The novelty relative to existing topology-aware approaches should be better clarified. Incorporating mesh topology to mitigate sticking or incorrect motion coupling has already been explored in works such as AnimateAnyMesh (2025) and the recent mesh registration method MeshLoom (2026). What are the key technical differences between the proposed method and these prior approaches?

  4. Qualitative visualization quality. The qualitative comparisons would benefit from more clearly presented ground-truth results alongside the predictions. Additionally, the current visualizations have relatively low resolution, making it difficult to examine local artifacts when zooming in.

  5. Clarity of the pipeline figure. Figure 2 contains a substantial amount of information, making the overall pipeline difficult to follow at first glance. A clearer separation of the main processing stages and their interactions would improve readability.

Comments on the paper

1 comment
Reviewer_RuprOpenReview ↗

Thanks for the clarification

Thank you for the detailed response and additional results. I have no further concerns and will maintain my positive rating.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.