acceptodds
Under review as a conference paper at ICLR 2027

DyCoPart: Dynamic Body-Part Coordination for Text-to-Motion Generation

Abstract

Text-to-motion (T2M) generation aims to synthesize realistic and physically plausible human motion sequences that are semantically consistent with natural-language descriptions. Despite substantial progress, existing T2M methods still face two key limitations: holistic models tend to miss fine-grained local articulations, while fixed part-based models struggle to capture motion-dependent body coordination due to predefined body-part partitions. To address these limitations, we propose DyCoPart, a dynamic coordinated part-aware framework for text-to-motion generation. DyCoPart first encodes joint-level kinematic dependencies with a lightweight graph module and softly routes joints into six dynamic parts under anatomical-prior guidance, producing part tokens that capture both fine-grained local dynamics and motion-dependent coordination. These part tokens are then processed by a shared spatial Part State Space Model, through which whole-body spatial structure is further modeled in the dynamic-part space. Based on the coordinated part representations, a directed causal coordination rectifier is introduced, where temporally causal cross-part messages are propagated among dynamic-part slots so that whole-body coherence can be recovered from corrupted or misaligned local dynamics.Experiments on HumanML3D and KIT-ML demonstrate strong quantitative performance, and coherence-level evaluations further show that DyCoPart generates more fine-grained, physically plausible, and globally coordinated whole-body motions.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Text to Motion, Computer Vision, Multimodal models

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_C5pEOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

DyCoPart is a diffusion text-to-motion model built on a spatial-temporal Mamba backbone. Joint features from a two-layer skeleton GCN are softly routed into six part slots, using motion-conditioned logits combined with an anatomical distance prior (Dynamic Part Learning, DPL). The part tokens then go through a Part-SSM and a "directed causal coordination rectifier" (DCCR), where each part at frame t aggregates messages from the other parts at frame t−1. DCCR is trained with random part masking/freezing/noise and three regularizers on its attention matrix. Experiments report the standard HumanML3D/KIT-ML metrics plus author-defined coordination (TC/SC), physical-plausibility and part-level metrics.

Critical strengths

The same-backbone ablation (Table 4a, Table 6) is the most useful result in the paper. A hard fixed partition improves the part-level metrics but drops TC/SC from 0.603/0.742 to 0.433/0.591 and worsens all four physical metrics, while soft routing improves both sides. Provided the FP variant is a fair implementation (Q1), this turns the local-vs-global trade-off that part-based papers usually just assert into something measured. The modules are also cheap (+0.41M parameters, +5.9% latency, Table 10) and give similar gains on MDM (Table 7).

Critical weaknesses
  • Missing baselines, and gains credited to the wrong source. Table 1 leaves out the strongest recent methods, and these are exactly the joint-level / skeleton-aware models that the "holistic vs. fixed-part" framing ignores. MoGenTS (NeurIPS'24) reports FID 0.033 on HumanML3D and Top-1 0.445 / FID 0.143 on KIT-ML, all better than DyCoPart (0.041; 0.438 / 0.185). SALAD (CVPR'25) reports higher R-Precision (Top-1 0.581 vs. 0.565); it is used in the supplementary video but is neither cited nor in Table 1. LGTM, HiSTF Mamba, ParTY and BiPO, which Sec. 2.2 itself lists as the main part-based baselines, are also missing. Separately, the in-house baseline without any proposed module already reaches Top-3 0.832 (Table 4a), above every prior method in Table 1, yet Sec. 4.2 credits the +9.7% Top-1 gain over ParCo to dynamic routing and causal coordination. What the modules add is 0.832 → 0.848 Top-3 and 0.085 → 0.041 FID, reported as point estimates without seed variance.

  • Neither mechanism is shown to do what its name says. The ablation compares hard fixed parts with soft, motion-conditioned routing, so "soft" and "dynamic" are confounded. Static soft routing (prior π only) is the missing control, and time variation is shown for a single prompt (Table 9, own-slot weight 0.71 → 0.55). DCCR is claimed to recover coherent motion from corrupted or desynchronized parts (abstract, Fig. 1c), but nothing perturbs parts at inference, and neither the training perturbations nor the auxiliary losses are ablated. "Causal" amounts to a one-frame lag inside a denoiser that sees the whole sequence; "counterfactual" refers to random part masking/freezing/noise. The method description also contradicts itself in places (auxiliary losses, Part-SSM axis, part-to-joint mapping; see Q5).

Questions for authors
  • How is FP implemented: hard one-hot routing into the same slots and Part-SSM? Does it replace or run alongside the baseline's spatial Mamba? The FP collapse in TC/SC carries much of the argument, so this changes how I read Table 4.
  • Please add static soft routing as a DPL control, and an inference-time test that masks, freezes or noises one part slot, with TC/SC/FID reported with and without DCCR.
  • Table 1 should include MoGenTS, SALAD and the cited part-based methods, plus the baseline's full row (Top-1, MM-Dist, MModality). Since all coordination and part-level metrics in Tables 2-3 are defined or adapted by the authors, a small user study on coordination would also help.
  • p_t and ε_t are computed from root and foot-contact features of which signal: the noisy x_t, the predicted x_0, or something else? At high noise levels, contacts read from x_t carry no information. What is the "whole-motion branch" that produces c_t?
  • Clarifications: Sec. 3.3 says the auxiliary losses target intervention response, compensatory support and output stability, while App. A defines temporal-smoothness, entropy and reciprocity penalties on A. Which is correct, and what are the weights? Part-SSM is "spatial" in the abstract and Sec. 3.1-3.2 but runs along time within each part in Eq. 8. Which is it? How are part tokens mapped back to joints?
6 comments

Thank you for your careful review and valuable constructive feedback. We have carefully considered your comments and provide our detailed responses to each point below.


Q1: FP implementation and fairness of the routing comparison

Overview. To address this question, we:

- 1. Clarify the original FP implementation and its relationship to the backbone (Appendix D.3); - 2. Define matched controls that separate routing changes from feature-processing differences (Appendix C.9, manuscript Table 12).

1. Original FP implementation.

The original FP variant groups joints by fixed anatomical indices, concatenates their features, and applies part-specific projections. Its groups can overlap, and it does not use DPL’s joint encoder and skeleton GCN. Although it uses the same six slot identities and shared Part-SSM processing, it is not simply a one-hot replacement of DPL’s assignment matrix.

The part pathway runs alongside the retained whole-body pathway, and its output is fused back into the whole-body representation. Consequently, the original FP comparison differs from DPL in feature construction and aggregation as well as grouping.

The original FP degradation therefore cannot be attributed entirely to fixed grouping. We will revise the corresponding discussion to distinguish this architectural comparison from the matched routing experiments below.

2. Matched comparison protocol.

The four matched variants in Table 1 share the joint encoder and GCN, aggregation framework, Part-SSM, joint reconstruction, whole-body pathway, fusion, decoder, and training settings. Each variant is trained separately, with DCCR disabled throughout.

Only the routing rule changes: fixed anatomical one-hot assignments, prior-only static soft assignments, sequence-conditioned static soft assignments, or frame-dependent dynamic soft assignments. This comparison addresses routing under matched feature processing.

Q2: Static soft routing and inference-time part corruption

Overview. To address these requests, we:

- 1. Add matched prior-only and sequence-conditioned static-soft controls (Appendix C.9, manuscript Table 12; Appendix C.9, manuscript Table 12); - 2. Evaluate inference-time masking, freezing, and noise injection with and without DCCR (Appendix C.10, manuscript Table 14).

1. Matched static-soft routing controls.

Following the matched protocol described in Q1, we compare:

  • Matched hard routing: fixed anatomical one-hot assignments.
  • Prior-only static soft routing: the learned anatomical-prior matrix, with motion-conditioned logits removed. The prior remains learnable during training; within each trained block, its assignments are shared across samples and frames.
  • Sequence-conditioned static soft routing: routing logits are averaged over valid frames, and the resulting assignment matrix is shared across the sequence.
  • Dynamic soft routing: assignments are predicted separately for each motion frame.

Table 1. Routing comparison. The last four rows use matched processing; the first two are architectural references. Best in bold; second-best underlined.

VariantFID ↓TC ↑SC ↑
Backbone baseline0.0850.603‾\underline{0.603}0.742‾\underline{0.742}
Original FP0.069‾\underline{0.069}0.4330.591
Matched hard routing0.0790.4780.636
Prior-only static soft routing0.0770.5310.681
Sequence-conditioned static soft routing0.0760.5200.669
Dynamic soft routing / DPL0.0580.6140.769

DPL improves FID, TC, and SC over both static-soft controls under matched processing. These results support contributions from motion-conditioned assignments and their frame dependence beyond soft allocation alone.

2. Inference-time part corruption.

We corrupt one part slot at the Part-SSM output immediately before the final DCCR block, or at the corresponding location without DCCR. Corruption is injected once, at the 16th step of the 20-step sampler, over a contiguous window covering 20% of valid motion frames.

Masking zeros the selected features; freezing repeats the state immediately preceding the window; noise injection adds Gaussian noise with per-channel standard deviation equal to half the corresponding validation feature scale. Prompts, initial sampling noise, affected slots, and windows are paired across models, with affected slots balanced across the test set. Both classifier-free guidance branches receive the same intervention.

Table 2. Inference-time corruption. Each cell reports without DCCR / with DCCR; the better value within each pair is bold.

ConditionTC ↑SC ↑FID ↓
Clean0.614 / 0.6230.769 / 0.7770.058 / 0.041
Mask0.568 / 0.6100.736 / 0.7630.094 / 0.051
Freeze0.547 / 0.5970.717 / 0.7590.109 / 0.063
Noise0.577 / 0.6040.740 / 0.7680.089 / 0.058

The DCCR-trained model achieves better corrupted-generation scores and smaller deterioration from its own clean performance across all three metrics and corruption types. This supports robustness to the tested part-feature corruptions beyond its clean-generation advantage.

These comparisons evaluate DCCR together with its intervention training and auxiliary objectives; they do not separately isolate the effects of training perturbations and individual auxiliary losses.

We will also revise the “causal” and “counterfactual” terminology. DCCR uses directed previous-frame messages within a denoiser that processes the complete sequence, and the training interventions are random feature perturbations. The evidence supports directed coordination and corruption robustness, rather than causal identification or temporal causality of the complete denoiser.

Q3: Expanded comparisons, gain attribution, and human evaluation

Overview. To address these concerns, we:

- 1. Expand the HumanML3D and KIT-ML comparisons to include the requested methods (Appendix C.1, manuscript Tables 4–5); - 2. Separate same-backbone improvements from cross-method differences (Sec. 4.3; Appendix C.4, manuscript Table 8); - 3. Report a coordination-focused human evaluation (Appendix C.12, manuscript Table 17).

1. Expanded benchmark comparisons.

The tables add published results for joint-level, skeleton-aware, and part-based methods, which we discuss in Sec. 2.2.

Table 3. Expanded HumanML3D comparison. Baseline methods use their published results; the internal backbone and DyCoPart use our experimental results. Best in bold; second-best underlined. Diversity is reported without ranking. NR: not reported in the cited paper.

MethodR@1 ↑R@2 ↑R@3 ↑FID ↓MM-Dist ↓DiversityMModality ↑
MoGenTS0.5290.7190.8120.033‾\underline{0.033}2.8679.570NR
SALAD0.5810.7690.8570.0762.6499.6961.751
LGTM0.4900.6890.7880.2183.0139.638NR
HiSTF Mamba0.5370.7240.8130.0392.8849.6891.701
ParTY0.5500.7440.8360.0352.7799.5342.155
BiPO0.5230.7140.8090.0302.8809.5561.374
ParCo0.5150.7060.8010.1092.9279.5761.382
Internal backbone0.5470.7450.8320.0852.7749.6182.561
DyCoPart0.565‾\underline{0.565}0.761‾\underline{0.761}0.848‾\underline{0.848}0.0412.682‾\underline{2.682}9.6612.529‾\underline{2.529}

The internal backbone is the implementation used in our ablations, rather than the published Light-T2M result. HiSTF Mamba values follow arXiv v3.

Table 4. Expanded KIT-ML comparison. Reporting conventions follow Table 3; —: unavailable in this response.

MethodR@1 ↑R@2 ↑R@3 ↑FID ↓MM-Dist ↓DiversityMModality ↑
MoGenTS0.4450.6710.7970.1432.71110.918NR
SALAD0.4770.7110.8280.2962.58511.0971.004
HiSTF Mamba0.458‾\underline{0.458}0.6780.7980.2172.73711.1601.058
ParTY0.4490.680‾\underline{0.680}0.804‾\underline{0.804}0.155‾\underline{0.155}2.69411.2101.166‾\underline{1.166}
BiPO0.4440.6740.8030.1642.658‾\underline{2.658}10.8331.098
ParCo0.4300.6490.7720.4532.82010.9501.245
Internal backbone0.4210.6480.7720.2042.82410.951—
DyCoPart0.4380.6670.7970.1852.71911.0730.962

These literature results retain each paper’s evaluation setup and are not presented as unified reruns.

SALAD reports stronger text alignment on both datasets. MoGenTS, BiPO, and ParTY report lower FID on both datasets, while HiSTF Mamba reports lower FID on HumanML3D. DyCoPart therefore does not lead across all standard metrics.

2. Attribution to the proposed modules.

The 9.7% relative R@1 improvement over ParCo is a cross-method difference and cannot be attributed entirely to DPL and DCCR. We assess their contribution against our internal backbone:

  • HumanML3D: R@1 improves by 0.018, R@3 by 0.016, and FID decreases by 0.044.
  • KIT-ML: R@1 improves by 0.017, R@3 by 0.025, and FID decreases by 0.019.

The same-backbone results support improvements in text alignment and generation fidelity on both datasets. MModality does not improve on HumanML3D.

3. Coordination-focused human evaluation.

We conducted a blinded pairwise study on 60 randomly selected HumanML3D test prompts, fixed before inspecting outputs. DyCoPart was compared with the internal backbone, SALAD, and ParCo. One output per prompt and method was selected using a predefined sampling rule without output screening.

Rendering and playback settings were identical; method identities were hidden, and presentation order and left/right placement were randomized. From a pool of 10 raters, five distinct raters assessed each prompt–comparison pair, yielding 300 judgments per comparison and 900 judgments. Raters selected the motion with more natural coordination among the arms, legs, and torso, with ties allowed.

Table 5. Coordination preferences. Tie-adjusted preference assigns half a vote to each tie.

ComparisonDyCoPart winsTiesComparator winsTie-adjusted preference
DyCoPart vs. internal backbone181447567.7%
DyCoPart vs. SALAD168379562.2%
DyCoPart vs. ParCo217245976.3%

DyCoPart received higher coordination preference than every comparator on the evaluated prompts, providing perceptual evidence alongside the automatic coordination metrics.

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

This paper proposes a dynamic coordinated part-aware framework for text-to-motion generation, aiming to balance fine-grained local motion quality and overall body coordination. The method introduces dynamic body-part learning and a directed causal coordination rectifier. Experiments on two benchmarks demonstrate its effectiveness across multiple evaluation metrics.

Critical strengths

The paper provides extensive experiments on widely used benchmarks and evaluates the proposed method using multiple metrics. The ablation studies also provide evidence supporting the effectiveness.

Critical weaknesses

Although the core idea of dynamic body-part modeling is clear and intuitive, its implementation is complicated, involving Kinematic Relation Learning, Anatomical-Prior-Guided Dynamic Routing, and Dynamic-Part Modeling. The ablation studies mainly validate these components as a whole, without sufficiently examining the necessity of their individual designs or comparing them with simpler alternatives.

Questions for authors
  1. In the second paragraph of Sec.1, the authors state that existing methods "still lack an explicit mechanism to model time-varying information flow and coordination across body parts during generation." However, fixed body partitions do not necessarily imply fixed information flow, and some existing methods have already introduced mechanisms for cross-part coordination. Therefore, I think this statement is not entirely accurate and should be clarified.
  2. The DCCR part of Figure 2 difficult for me to understand, even when referring to the corresponding paragraph. In particular, the meanings of the connections, gates, and different perturbation symbols are not sufficiently clear. I suggest that the authors improve the figure and its caption to better explain the information flow.
  3. In Sec. 3.2, Eq. (8) describes how joint features are aggregated into six dynamic-part tokens. However, it is unclear how these six tokens are subsequently mapped back to the original joint space, as mentioned in Sec 3.1. Although mapping part tokens back to joint space may be a common operation, the proposed method uses dynamic soft routing to aggregate 22 joint features into six part tokens. Therefore, I think it is necessary to clarify whether the reverse mapping reuses the learned soft assignment weights or employs a separately learned mapping. This implementation detail should be explicitly described in the paper.
  4. How is FP implemented in Table 4? To better demonstrate the importance of "motion-adaptive routing", I think the comparison should not simply assign each joint to its corresponding fixed body part. Instead, a static soft-routing baseline could use fixed weights based on anatomical distances while also incorporating global joint information into each part token.
  5. Appendix C.6 provides quantitative routing weights. However, I think more visualizations of the learned dynamic parts would be helpful. For example, visualizing the joint-to-part assignment weights over time for different motions could more intuitively demonstrate how the proposed dynamic routing adapts to motion states.
5 comments

Thank you for your thoughtful review and valuable suggestions. We address each point below. New results and clarifications are included in the revised manuscript and highlighted in blue.


W1 and Q1: Clarifying the motivation and relationship to prior work

Overview. We refine the motivation and acknowledge existing cross-part coordination mechanisms (Sec. 1, second paragraph; Sec. 2.2).

Thank you for pointing out this important issue. We agree that fixed body partitions do not necessarily imply fixed information flow, and that previous methods have already explored cross-part coordination. Our original statement was overly broad and did not sufficiently distinguish existing coordination mechanisms from the specific capability targeted by DyCoPart.

We have revised the second paragraph of Sec. 1 to explicitly acknowledge these advances. In particular, ParCo introduces communication among part-specific generators, while LGTM employs an attention-based full-body optimizer to coordinate independently generated local motions. AttT2M and HiSTF Mamba also explore body-part modeling and local-global information fusion.

We have therefore refined our motivation to focus on a more specific research gap rather than suggesting that existing methods lack cross-part interaction altogether. Specifically, DyCoPart investigates the joint modeling of (1) motion-conditioned joint-to-part reassignment under semantically anchored part slots, (2) temporally directed and asymmetric information propagation from historical source-part states to current target-part states, and (3) condition-aware coordination rectification under unreliable local representations.

We have also revised Sec. 2.2 to provide a more balanced discussion of existing part-aware methods. These changes better clarify the distinction between previous cross-part coordination and the particular combination of adaptive routing and temporally causal coordination investigated in DyCoPart.


W1 and Q2: Clarifying the DCCR illustration and information flow

Overview. We clarify directed connections, target-specific gates, and training-only perturbations (Sec. 3.3; Fig. 2 caption; Appendix D.2).

Thank you for this helpful suggestion. We agree that the original description of DCCR could be clearer, particularly regarding the directed connections, gates, and perturbation symbols.

In this revision, we have expanded the explanation in Sec. 3.3 and clarified the corresponding Figure 2 caption and Appendix D.2. Specifically:

  1. Directed connections: Each edge s→rs \rightarrow r represents information propagation from the historical state of source part ss at t−1t-1 to the current state of target part rr at tt. The edge weights determine the relative contributions of different source parts.

  2. Gates: We clarify that gt,rg_{t,r} is a target-specific gate rather than an edge-specific gate. It controls the overall strength of the aggregated cross-part message injected into the target representation, while ϵt\epsilon_t further modulates message injection around motion events.

  3. Perturbations: We explicitly explain the three training-only perturbations: masking removes selected part information, freezing holds selected part representations temporally unchanged, and noise injection perturbs selected part features. These operations are disabled during inference.

Due to the limited time available for this revision, we have prioritized clarifying the textual descriptions and mathematical roles of these components. We will further refine the visual presentation and legend of DCCR in the subsequent manuscript revision to make the information flow more intuitive.

Q3: Mapping part representations back to joint space

Overview. We explicitly describe joint-space reconstruction and reuse of the current block's routing weights (Sec. 3.3, Mapping parts back to joints; Eq. (15)).

Thank you for pointing out this missing implementation detail. We agree that the mapping from part representations back to joint space should be explicitly clarified.

In our implementation, joint reconstruction reuses the joint-to-part assignment weights produced by the same DyCoPart block. No separately learned reverse-routing matrix is introduced.

After DCCR, each refined part representation is projected into a 64-dimensional feature space. For each non-root joint at each frame, we reconstruct its feature as a weighted sum of the six projected part representations. The weights are that joint’s assignments to the six part slots, obtained from the same block and frame.

A shared affine decoder then applies a learned linear transformation and bias to produce a 12-dimensional joint descriptor containing position, rotation, and velocity information. Root and foot-contact channels are assembled separately. The reconstructed motion representation is subsequently projected to the backbone width and fused with the whole-body pathway.

We have clarified this reconstruction procedure in Sec. 3.3, explicitly stating that the motion-adaptive assignment weights are reused to distribute refined part information back to individual joints.


W1 and Q4: FP implementation and static soft-routing controls

Overview. We clarify FP feature construction and its parallel whole-body pathway (Appendix D.3) and report the additional anatomical-prior-only static-soft control below for inclusion in the matched routing comparison (Appendix C.9, manuscript Table 12).

Thank you for this suggestion. We clarify FP and provide matched static-soft controls to isolate motion-adaptive routing.

1. FP implementation.

The original FP groups joints using fixed anatomical indices, concatenates their features, and applies part-specific projections before Part-SSM. It does not use DPL’s joint encoder and skeleton GCN. The part pathway runs alongside the retained whole-body pathway, and its output is fused back into the whole-body representation.

FP differs from DPL in feature encoding and aggregation, so its coordination degradation cannot be attributed solely to fixed assignments.

2. Matched static-soft controls.

We evaluated routing variants with identical joint encoding, aggregation, Part-SSM, reconstruction, whole-body pathway, fusion, and training settings, with DCCR disabled throughout.

The anatomical-prior-only control retains the distance-based prior and its learnable corrections during training; its trained assignments are shared across samples and frames. Importantly, it aggregates the same root-conditioned, GCN-encoded features over the full joint set, preserving global root context and skeletal information.

We also evaluate sequence-conditioned, time-invariant routing, which averages routing logits over valid frames to produce one assignment matrix per sequence.

Table R1. Matched comparison of routing strategies on HumanML3D. All variants use identical feature processing with DCCR disabled. Best results are in bold.

Routing strategyFID ↓TC ↑SC ↑
Matched hard0.0790.4780.636
Anatomical-prior-only static soft0.0770.5310.681
Sequence-conditioned, time-invariant soft0.0760.5200.669
Dynamic soft / DPL0.0580.6140.769

DPL outperforms both static-soft controls across all three metrics, supporting the benefit of frame-dependent assignments beyond soft aggregation while retaining the same joint information. We will clarify FP’s implementation and include these matched comparisons in the revised manuscript.


Q5: Visualizations of learned dynamic parts

Overview. Additional temporal routing visualizations are in preparation and will be provided before the rebuttal period ends.

Thank you for this helpful suggestion. We are preparing additional visualizations of joint-to-part assignment weights over time across different motions to illustrate how dynamic routing adapts to motion states.

As the deadline approaches, we are submitting the other responses first while these figures are being finalized. We will update this response with the completed visualizations before the rebuttal period ends. We appreciate your patience and understanding.

Thank you for the clarifications and additional experiments. They have addressed my main concerns, and I will maintain my rating of 3 (weak acceptance).

Regarding the planned visualizations, I am particularly interested in whether the dynamic assignments exhibit interpretable, motion-dependent patterns, for example, whether actions requiring closer arm–shoulder coordination lead to corresponding changes in the routing weights. Such interpretable patterns would strengthen the evidence for the proposed dynamic routing and make the work more appealing to me.

Reviewer_MnHWOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DyCoPart, a text-conditioned diffusion framework that aims to improve local articulation while preserving whole-body coordination. Dynamic Part Learning uses a skeleton GNN and anatomical priors to softly route joints into six predefined part slots, with assignments varying across motion frames. The resulting tokens are refined by a shared Part-SSM and a Directed Causal Coordination Rectifier, which uses gated messages from previous-frame part tokens to address unreliable local dynamics. Training incorporates part-level perturbations and graph regularization. Experiments on HumanML3D and KIT-ML report improvements in motion quality and text–motion alignment over the evaluated baselines, complemented by part-level, coordination, and physical-plausibility evaluations on HumanML3D. Same-backbone ablations and MDM transfer experiments support the modules’ practical usefulness, while the overhead analysis reports modest additional computational cost.

Critical strengths

1. A useful part representation supported by controlled local–global evaluation.

Anatomically anchored soft routing offers a sensible way to retain recognizable body-part semantics while allowing joint contributions to change with motion. The same-backbone ablations in Tables 4 and 6 provide meaningful support for this design: fixed-part modeling improves local motion metrics but degrades coordination, whereas DPL improves both under the evaluated setup. Assessing local articulation and whole-body quality together is particularly valuable because it directly tests the trade-off motivating the paper, rather than relying solely on standard holistic metrics.

2. Evidence of practical usefulness across backbones with modest overhead.

The controlled MDM experiments in Table 7 suggest that the proposed modules can benefit a second backbone, extending the evidence beyond the authors’ Mamba-based architecture. Table 10 also reports modest increases in parameters, computation, and inference latency relative to the same-backbone baseline. Together, these results support the approach as a potentially practical architectural addition, even though further controls are needed to identify which specific mechanisms account for the gains.

Critical weaknesses

1. The module gains are supported, but the roles of dynamic routing and temporally directed rectification remain insufficiently isolated.

The same-backbone ablations (Tables 4 and 6) support improvements from DPL and DCCR, but do not fully isolate their defining mechanisms. The comparison with fixed-part modeling does not separate motion-dependent routing from soft assignment and possible differences in feature processing. FP’s feature processing is not described in sufficient detail, and a comparable time-invariant soft-routing control is missing. DCCR’s aggregation (Eqs. 9–12) has the form of cross-attention, augmented with conditional biases and the gating in Eqs. 13–14, yet no matched attention baseline isolates these additions. Asymmetric inter-part dependencies can be represented without a temporal lag. Although previous-frame sources are motivated as corrective historical context, their advantage over same-frame or bidirectional alternatives in offline full-sequence generation remains untested.

Direct evidence for the claimed recovery and compensatory behavior is also limited. Training uses part perturbations, but the manuscript reports no controlled perturbation-and-recovery comparison with and without DCCR. Table 6 supports better aggregate coordination rather than directly establishing recovery from corrupted part representations. Figures 3–4 compare complete methods rather than ablated variants, while Table 9 illustrates routing changes for one example at two frames without establishing their contribution to generation quality.

Finally, Part-SSM is described as spatial, whereas Eq. (8) and its explanation specify temporal processing. The response/compensation/output-stability description in Sec. 3.3 is not clearly reconciled with the adjacency regularizers in Appendix A, whose numerical weights are unspecified. These gaps limit confidence in the mechanism-level interpretation of the empirical gains.

2. The quantitative comparisons provide limited evidence of an advantage over recent, closely related methods.

The quantitative evaluation relies predominantly on methods from 2024 or earlier, but the more consequential issue is the omission of approaches directly relevant to the paper’s contributions. ParTY addresses the same challenge of combining part expressiveness with whole-body coherence and is cited as an inspiration for TC/SC, yet is absent from the quantitative comparisons. SALAD is mentioned as a supplementary-video comparator but has no quantitative results in the paper, while Light-T2M offers a relevant comparison as a lightweight Mamba-based model. Both report competitive results on the standard benchmarks. Among the six works described as “the main part-based baselines” in Sec. 2.2, only ParCo and LGTM are quantitatively evaluated. These omissions do not negate the same-backbone improvements, but leave DyCoPart’s advantage over contemporary approaches insufficiently established, particularly for its central objective of jointly improving local articulation and whole-body coordination.

Questions for authors
  1. Dynamic routing. How does FP’s feature encoding and aggregation differ from DPL? Can the authors provide a time-invariant soft-routing control with comparable processing to isolate the contribution of motion-dependent assignments?

  2. Temporal directionality. Beyond the stated motivation of corrective historical context, what advantage do previous-frame sources offer over same-frame or bidirectional context in offline full-sequence generation? A matched cross-part attention comparison varying the temporal connections would help establish the benefit of this choice.

  3. Qualitative evidence and rectification. Could the authors provide matched motion visualizations of the baseline, +FP, +DPL, and full model under the same prompts to illustrate changes in local articulation and whole-body coordination? More directly, a controlled perturbation-and-recovery comparison with and without DCCR, accompanied by quantitative results and visualizations, would help assess the claimed recovery and compensatory behavior.

  4. Architecture and objective. Does Part-SSM operate across parts or across time, and in which scan direction? Please reconcile its spatial description with Eq. (8) and specify how part tokens are mapped back to joint space. While the conceptual sources of the whole-body condition, phase features, and event indicator are described, how are these signals computed at each denoising step during sampling? Please also clarify how the response/compensation/output-stability descriptions in Sec. 3.3 relate to the regularizers in Appendix A and provide their numerical weights.

  5. Relevant comparisons. Can the authors compare with ParTY on local articulation and whole-body coordination, and clarify the quantitative positioning relative to SALAD and Light-T2M? If comparable evaluation is not feasible, please explain the relevant protocol or implementation differences.

9 comments

Thank you for your detailed review and constructive feedback. We have carefully considered your comments and provide our detailed responses to each point below.


W1 and Q1: Isolating the contribution of dynamic routing

Overview. To address these concerns, we:

- 1. Clarify the feature-encoding and aggregation differences between the original FP and DPL (Appendix D.3); - 2. Add matched hard-routing and time-invariant soft-routing controls to separate the benefit of soft assignments from that of frame-dependent assignments (Appendix C.9, manuscript Table 12).

1. Feature encoding and aggregation in FP and DPL.

The original FP concatenates joint features within fixed anatomical groups and applies part-specific projections, without skeleton GNN encoding. DPL instead uses a skeleton GNN to encode joint relationships and aggregates joint features through soft assignment weights.

The original FP–DPL comparison therefore involves differences in both feature processing and routing. To isolate the contribution of routing, we provide the following matched controls.

2. Matched controls for routing.

We compare three variants with the same DPL joint encoder, aggregation framework, and downstream processing, including Part-SSM. The variants differ in their assignment rules:

  • Matched hard routing: fixed anatomical one-hot assignments.
  • Time-invariant soft routing: one soft-assignment matrix shared across all frames within each sequence.
  • Dynamic soft routing (DPL): soft assignments predicted for each frame.

Table 1. Routing ablations. The last three rows use matched feature processing; the baseline and original FP are included for reference. Best in bold; second-best underlined.

ModelR@3 ↑FID ↓KFID ↓TC ↑SC ↑
Baseline0.8320.0850.3340.603‾\underline{0.603}0.742‾\underline{0.742}
Original FP0.835‾\underline{0.835}0.069‾\underline{0.069}0.267‾\underline{0.267}0.4330.591
Matched hard routing0.8330.0790.2910.4780.636
Time-invariant soft routing0.8340.0760.2810.5200.669
Dynamic soft routing (DPL)0.8440.0580.2180.6140.769

Soft assignments improve coordination under matched processing. Compared with matched hard routing, time-invariant soft routing improves TC from 0.478 to 0.520 and SC from 0.636 to 0.669, while reducing KFID from 0.291 to 0.281.

Frame-dependent assignments provide additional gains beyond soft routing. Compared with time-invariant soft routing, DPL improves all five metrics, reducing KFID from 0.281 to 0.218 (22.4%) and increasing TC/SC from 0.520/0.669 to 0.614/0.769. The time-invariant control remains below the baseline in TC/SC, whereas DPL improves both local motion quality and coordination relative to the baseline.

Together, these matched controls support a distinct contribution from frame-dependent assignments: they improve local motion quality and whole-body coordination beyond the benefit of soft routing alone under the evaluated setting.

W1 and Q2: Temporal directionality and matched attention controls

Overview. We add matched comparisons of plain cross-part attention and DCCR under previous-frame, same-frame, and local bidirectional connections, using DPL without coordination as the reference (Appendix C.9, manuscript Table 13).

For each mechanism, queries correspond to the current frame, while source tokens come from the preceding frame, the current frame, or both the preceding and following frames, respectively. The DCCR variants retain conditional biases and gating. These controls compare temporal connections within each mechanism and compare message mechanisms under the same temporal connections.

Table 2. Message mechanisms and temporal connections. Best in bold; second-best underlined, including ties.

Message mechanismTemporal connectionsR@3 ↑FID ↓KFID ↓TC ↑SC ↑
Plain attentionPrevious-frame0.8430.0610.2240.6120.767
Plain attentionSame-frame0.8420.0640.2300.6090.764
Plain attentionLocal bidirectional0.8440.0590.2200.6130.768
DCCRSame-frame0.846‾\underline{0.846}0.0480.2030.6180.773
DCCRLocal bidirectional0.846‾\underline{0.846}0.046‾\underline{0.046}0.201‾\underline{0.201}0.619‾\underline{0.619}0.774‾\underline{0.774}
DCCR (ours)Previous-frame0.8480.0410.1910.6230.777

Previous-frame sources achieve the best reported results within DCCR across all five metrics. Compared with local bidirectional DCCR, the previous-frame variant reduces FID from 0.046 to 0.041 and KFID from 0.201 to 0.191, while increasing TC/SC from 0.619/0.774 to 0.623/0.777. It also outperforms same-frame DCCR on all five metrics. Thus, access to neighboring future-frame sources does not yield better results than previous-frame sources in the tested DCCR configuration.

Adding plain cross-part attention alone does not account for these gains. None of the tested plain-attention variants improves upon DPL without coordination. With previous-frame connections fixed, replacing plain attention with DCCR reduces FID from 0.061 to 0.041 and KFID from 0.224 to 0.191, while increasing TC/SC from 0.612/0.767 to 0.623/0.777. DCCR also outperforms plain attention under the other two connection patterns, supporting the combined contribution of its conditional biases and gating.

The preferred temporal connection depends on the message mechanism: plain attention performs best with local bidirectional context, whereas DCCR performs best with previous-frame context. These results support previous-frame sources as an effective temporal inductive bias specifically within DCCR under the evaluated offline-generation setting, with benefits over both same-frame and local bidirectional controls.


W1 and Q3: Qualitative evidence and rectification.

Overview. We provide matched ablation visualizations (Appendix C.9, Fig. 5) and a controlled inference-time corruption comparison (Appendix C.10, manuscript Table 14 and Fig. 6).

Thank you for this suggestion. We have added two qualitative comparisons to the revised manuscript: (1) matched visualizations of the baseline, +FP, +DPL, and full model under the same prompt; and (2) a comparison with and without DCCR after Gaussian noise injection at the 16th sampling step. These visualizations illustrate differences in local articulation and whole-body coordination and complement the quantitative TC/SC/FID results for assessing robustness to part-feature corruption.

W1 and Q4: Architecture and objective

Overview. We clarify the Part-SSM scan axis (Sec. 3.2, Eq. (8)), joint reconstruction (Sec. 3.3, Eq. (15)), sampling-time conditions (Sec. 3.3, Conditions at each denoising evaluation), and active objectives (Sec. 3.3, Eq. (16); Appendix A).

Thank you for identifying these inconsistencies. We will correct the manuscript accordingly.

1. Part-SSM: scan axis and direction.

The six part slots are folded into the batch dimension and share SSM parameters. After projection, diffusion-step-token insertion, and dynamic convolution, the input has shape [BS,Dp,T+1][BS,D_p,T+1], where S=6S=6 and Dp=128D_p=128 (i.e., [6B,128,197][6B,128,197] for 196 frames). The SSM scans unidirectionally along the 128 latent-feature positions, not along the part or frame axis.

For the pre-SSM representation Fp∈R(T+1)×DpF_p\in\mathbb{R}^{(T+1)\times D_p}, the operation is

F~p=[SSM(Fp⊤)]⊤.\widetilde{F}_p=[\mathrm{SSM}(F_p^\top)]^\top.

The scan runs from feature position 11 to DpD_p, after which the diffusion-step token is removed. We will revise Eq. (8) to identify this as a shared part-wise, feature-axis SSM. DCCR computes cross-part messages, whereas Temporal-SSM scans motion time; DCCR's previous-frame source indexing is separate from the Part-SSM scan.

2. Mapping parts back to joints.

Each refined part token is projected to a 64-dimensional representation. For each non-root joint at each frame, a joint feature is reconstructed as a weighted sum of the six part representations. The weights are the corresponding joint-to-part assignments produced by the current block. Thus, joint reconstruction reuses the current block’s assignment matrix.

A shared affine decoder maps each reconstructed joint feature to a 12-dimensional descriptor containing position, rotation, and velocity information. Root and foot-contact channels are assembled separately. The motion representation is projected to the backbone width and fused with the whole-body branch.

3. Conditions at sampling time.

At each denoiser evaluation kk, conditions are recomputed from the current noisy sample x(k)x^{(k)} through the current block's hidden features. The whole-body condition ct(k)c_t^{(k)} is the frame-aligned whole-body-branch output, excluding the diffusion-step token.

A learned projection supplies seven root and four contact channels. Their concatenation is encoded into a 32-dimensional phase-conditioning feature using two temporal convolutions (kernel size 3, padding 1) with GELU between them.

The event indicator thresholds the four contact channels at 0.5, detects adjacent-frame changes, and extends each transition to its immediate neighbors. A learned gain controls its contribution to the gates.

These root/contact channels are internal proxies; sampling requires no ground-truth motion, phase, or contact annotations. Conditions are recomputed for both classifier-free guidance branches.

4. Objectives and weights.

The response, compensation, and output-stability losses compare reference and perturbed forward passes.

Response loss. We measure the mean absolute feature change in the perturbed slot over the intervention window. The loss equals the amount by which this change falls below a margin of 0.15, and is zero once the margin is reached.

Compensation loss. We similarly measure the mean absolute feature change in the other slots over the intervention window. The loss equals the amount by which this change falls below a margin of 0.05, and is zero once the margin is reached. This term encourages a response in the other slots, but does not by itself prescribe a corrective direction.

Output-stability loss. This loss is the mean absolute difference between the predictions produced by the perturbed and reference passes. The reference prediction is detached, so gradients from this loss propagate only through the perturbed prediction. While compensation encourages responses in other slots, output stability constrains the resulting prediction to remain close to the reference prediction.

Active objective and intervention probability. The active T2M/KIT objective assigns a weight of 3 to the denoising loss. When an intervention is applied, the response, compensation, and output-stability losses are additionally included, each with a weight of 0.3. Otherwise, these intervention losses do not contribute to the objective.

An intervention is applied with probability 0.1. Conditional on an intervention, masking, freezing, or noise injection is selected uniformly. Thus, 0.1 is the overall intervention probability, rather than the probability of each perturbation type.

Appendix A’s temporal-variation, incoming-entropy, and reciprocal-edge regularizers were not enabled and did not contribute to the reported objective. Their nominal coefficients of 0.05, 0.02, and 0.02 were not applied during training. We will revise Appendix A and the main text to report the actual training objective and remove inconsistent descriptions.

Comments on the paper

3 comments
AuthorsOpenReview ↗

Summary of changes in the revised PDF

Summary of changes in the revised PDF

We thank all reviewers for their detailed and constructive feedback. We have revised the manuscript with the following additional experiments and clarifications. New and revised material is highlighted in blue.

  • Matched controls for dynamic routing [MnHW, muXu, C5pE]. We add anatomical-prior-only and sequence-conditioned, time-invariant soft-routing controls alongside matched hard routing. These variants share DPL’s feature processing, with DCCR disabled, to distinguish the contributions of soft aggregation, motion conditioning, and frame-dependent assignments. We also clarify the original FP implementation and its parallel whole-body pathway (Appendix C.9, Table 12; Appendix D.3).

  • Temporal connections and matched attention comparisons [MnHW]. We compare plain cross-part attention and DCCR using previous-frame, same-frame, and local bidirectional sources. These controls evaluate temporal connections within each mechanism and compare the mechanisms under the same connections, supporting the previous-frame design specifically within the tested DCCR configuration (Appendix C.9; Table 13).

  • Inference-time corruption tests and qualitative ablations [MnHW, C5pE]. We add controlled part-slot masking, freezing, and Gaussian-noise tests, comparing the separately trained DPL reference with the Full model. We also provide same-prompt visualizations of the backbone, FP, DPL, and Full variants, together with a qualitative comparison after noise injection at the 16th sampling step (Appendices C.9–C.10; Table 14; Figures 5–6).

  • Expanded comparisons and corrected gain attribution [MnHW, C5pE]. We expand the HumanML3D and KIT-ML tables with published results from recent methods, including MoGenTS, SALAD, Light-T2M, ParTY, HiSTF Mamba, and BiPO, and include the internal backbone. We add common-protocol local-quality and coordination comparisons with SALAD and Light-T2M, explain the resource limitation on reevaluating ParTY, and distinguish same-backbone module gains from cross-method differences (Section 4.2; Appendices C.1, C.3–C.4; Tables 4–5 and 7–8).

  • Coordination-focused human evaluation [C5pE]. We add a blinded pairwise study on 60 randomly selected HumanML3D test prompts, comparing DyCoPart with the internal backbone, SALAD, and ParCo. The study includes 900 judgments with ties allowed and provides perceptual evidence alongside the automatic coordination metrics (Appendix C.12; Table 17).

  • Clarified motivation and implementation [MnHW, muXu, C5pE]. We acknowledge existing cross-part coordination and clarify the Part-SSM scan axis, joint reconstruction, DCCR information flow, conditioning sources, and active loss weights. We explicitly identify unused regularizers and add conditioning reliability and sensitivity experiments (Sections 1, 2.2, and 3.2–3.3; Figure 2 caption; Appendices A, C.11, and D.2; Tables 15–16).

Thanks for your clarification, but it seems to be intended for Reviewer C5pE :)

AuthorsOpenReview ↗

Summary of Discussion

We thank the Area Chair and all three reviewers for their time and constructive feedback. The discussion focused on whether frame-dependent joint-to-part assignments improve motion generation beyond static soft routing and whether DCCR preserves coordination better than plain attention when a part is perturbed at inference.

Score updates

  • MnHW — 2 (weak rejection) → 3 (weak acceptance)
  • muXu — 3 (weak acceptance) → 3 (weak acceptance)
  • C5pE — 2 (weak rejection) → pending reviewer update

Strengths recognized by the reviewers

  • Interpretable body parts. Anatomical anchors preserve recognizable part identities while each joint’s contribution can change with the motion. [MnHW]
  • Evaluation of local and whole-body motion. Reviewers highlighted the evaluation on HumanML3D and KIT-ML and the same-backbone ablations that measure local articulation and whole-body coordination together. [muXu, MnHW, C5pE]
  • Transfer and efficiency. The modules also improved generation on MDM. On the primary backbone, they only added 0.41M parameters and increased inference latency by 5.9%. [MnHW, C5pE]

Concerns acknowledged as resolved during discussion

MnHW recognized the matched routing and attention controls and raised their rating after the follow-up corruption test. muXu confirmed that the clarifications and experiments addressed their main concerns.

  • Contribution of frame-dependent routing. With feature processing matched and DCCR disabled, HumanML3D FID was 0.058 for frame-dependent routing, 0.077 for prior-only static soft routing, and 0.076 for sequence-conditioned soft routing held fixed across frames. (Appendix C.9, Table 12)
  • Robustness under matched intervention training. Both models used previous-frame part tokens and the same intervention protocol and auxiliary losses during training. Under masking, freezing and Gaussian noise, DCCR showed smaller drops in both coordination scores and smaller FID increases. Freezing raised FID by 0.022 with DCCR and 0.035 with plain attention relative to each model’s clean result. (Appendix C.9, Table 13 for the matched models and 18 Nov follow-up to MnHW for the corruption results)
  • Architecture and training objective. We clarified that Part-SSM scans the latent feature axis, joint reconstruction reuses the routing weights, and sampling conditions come from the current denoiser features. In DCCR, “causal” refers to messages from the previous frame within an offline denoiser. The denoising loss has weight 3, each active intervention loss has weight 0.3, and the adjacency regularizers were disabled. MnHW confirmed that these clarifications resolved the ambiguities they raised. (Sections 3.2–3.3, Appendix A)

Additional revisions in response to reviewer concerns

  • Arm and shoulder routing across poses. Four heatmaps show routing during alternating elbow-to-knee lifts and a deep squat. From the reference pose to the squat, the left shoulder’s weight on the right-arm slot rose from 0.0487 to 0.2598 as both arms extended. The left elbow and wrist shifted in the same direction, and the left-arm slot retained the largest weight for all three joints. (Appendix C.8, Figure 5) [muXu]
  • Benchmark coverage and attribution. We added published comparisons with MoGenTS, SALAD, LGTM, HiSTF Mamba, ParTY and BiPO, and evaluated SALAD and Light-T2M under the same coordination protocol. On HumanML3D, adding DPL and DCCR to our own backbone raised Top-3 R-Precision from 0.832 to 0.848 and lowered FID from 0.085 to 0.041. MoGenTS reports a lower FID of 0.033. (Appendices C.1 and C.3–C.4, Tables 4–5 and 7–8) [C5pE, MnHW]
  • Perceptual coordination. Ten raters provided 900 blinded judgments across 60 HumanML3D prompts. Counting ties as half a vote, preference for DyCoPart’s coordination was 67.7% against the internal backbone, 62.2% against SALAD and 76.3% against ParCo. (Appendix C.12, Table 17) [C5pE, MnHW]

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.