acceptodds
Under review as a conference paper at ICLR 2027

QuanTopo: Topology-Directed Hierarchical Text-to-Motion Generation

Abstract

Text-to-motion generation has advanced in realism and text alignment, yet composite descriptions require global sequence organization alongside diverse, valid realizations. We introduce QuanTopo, a topology-directed hierarchical framework that learns the discrete branch of a hybrid generator as an explicit motion plan rather than only a reconstruction code. QuanTopo addresses a target–use gap: a useful plan needs both an objective that specifies what it should preserve and a structured role in synthesis. To define plan content, persistent homology summarizes multiscale relations along a trajectory of pose, root motion, and local change, emphasizing connectivity, revisitation, and cycles beyond framewise fidelity. To make this content actionable, the plan establishes anatomical context before continuous realization enters joint synthesis, while the continuous pathway models alternative timings, postures, and coordination patterns. A predictor conditioned on text and supplied duration connects the same trajectory target to inference. QuanTopo obtains FID scores of 0.027/0.122 on HumanML3D/KIT-ML and reaches 0.804 composite-action sequence satisfaction on a prespecified subset of composite HumanML3D captions, versus 0.782 for the strongest baseline. In this hybrid setting, ablations and interventions support aligning a plan's information target with its role in synthesis.

Primary area:
generative models
Keywords:
text-to-motion generation, 3D human motion synthesis, flow matching

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

This paper proposes QuanTopo, a hybrid text-to-motion generation framework that combines a discrete motion plan with continuous motion representations. The authors identify a "target–use gap" in existing hybrid models, where discrete representations are primarily learned for reconstruction without explicitly defining what information they should preserve or how they should be used during generation. To address this, QuanTopo introduces persistent homology-based trajectory supervision, plan-only reconstruction, and an anatomical MPS decoder, while using conditional flow matching to generate continuous realization latents. The method achieves FID scores of 0.027 and 0.122 on HumanML3D and KIT-ML, respectively, and reports a Seq. score of 0.804 on a prespecified subset of 180 composite-action captions from HumanML3D. The authors provide extensive ablations and controlled experiments to support their design.

Critical strengths

S1. Extensive experimental analysis.

The paper provides trajectory-summary comparisons (Table 3), capacity-matched controls (Table S9), topology-condition interventions (Table S6), plan permutation experiments (Table S15), and frozen representation probes (Table S17). These analyses help examine the roles of the proposed components from multiple perspectives.

S2. Competitive generation performance.

QuanTopo achieves strong results on HumanML3D and KIT-ML (Table 1). It also demonstrates a statistically significant improvement in composite-action sequence satisfaction over SALAD, including improved ordering accuracy on motions where both methods cover all required events (Table 5, Appendix C.2).

Critical weaknesses

W1. The contribution of topology-conditioned planning to temporal action organization is not sufficiently isolated.

  • Table 2 shows that topology modulation improves FID and R@1, but does not report Seq. or Order|Covered, which are more directly related to the paper's central motivation. The condition interventions in Tables S6 and S14 show that replacing the predicted topology descriptor with Zero, Mean, or Mismatched conditions reduces both metrics. This supports the conclusion that the trained planner uses the predicted descriptor. However, changing the condition of a model trained to rely on it does not establish that topology modulation outperforms a model trained without modulation.

  • Furthermore, the dataset-mean condition, which performs best among the alternative conditions, is not accompanied by a paired confidence interval against the predicted condition (Seq. 0.804 vs. 0.771; Order|Covered 0.904 vs. 0.884). A direct composite-action comparison with the "w/o Topology Modulation" model in Table 2 would better establish the contribution of this component.

W2. The improvement in composite-action generation is relatively modest, and qualitative verification is limited.

  • The improvement over SALAD in Seq. is +0.022 (95% CI [+0.006, +0.039]), which is statistically significant but relatively modest, based on one generated motion per caption across 180 captions (Table 5). In addition, the user study reports a Text Alignment win rate of 52.8% against SALAD, with a confidence interval that includes 50% (Appendix C.4).

  • Moreover, no supplementary videos are provided, making it difficult to directly assess whether temporal action ordering and transitions are accurately and naturally realized. Although Figures 1, 4, and S4 provide qualitative keyframes, these are insufficient to fully evaluate the temporal characteristics of the generated motions.

Questions for authors

Q1. Contribution of topology modulation.

  • Table 2 states that "w/o Topology Modulation" and QuanTopo share the same frozen Stage-A representation. Could the authors clarify whether the topology predictor's diagram loss and the soft-plan auxiliary loss are also retained during Stage B1?

  • Could the authors also report Seq., Coverage, and Order|Covered for these two models under the same composite-action evaluation protocol?

  • Additionally, a paired comparison between the Predicted and Mean conditions in Table S14 would be helpful. Seq. could be compared across all 180 captions, while Order|Covered should be evaluated on the subset where both conditions cover all required events.

Q2. Qualitative results.

  • Could the authors provide full motion videos for the examples in Figures 1, 4, and S4, and, if possible, the generated outputs used in the composite-action evaluation in Table 5? These would help assess temporal action ordering and transitions, which are central to the paper's contributions.

Q3. Clarification of the unified evaluation.

  • Appendix C.1 states that baselines with a single public checkpoint were reevaluated using three independent generation streams. However, several results for MoMask and MotionHiFlow in Table S3 match the previously reported values in Table 1 to three decimal places. For MotionHiFlow on KIT-ML, even the reported confidence intervals are identical. Could the authors clarify the evaluation procedure and explain these identical values? Providing the results for individual evaluation repeats would also help verify the reported paired confidence intervals.
7 comments
AuthorsOpenReview ↗

Part 1: topology-modulation contribution

Thank you for recognizing the experimental analysis and for asking us to isolate the contribution of topology modulation. The inference-time interventions alone left this question open. We have added a retrained control that retains the auxiliary losses and uses the same training budget. We have uploaded the revised manuscript, with the corresponding experimental additions and clarifications highlighted in blue.

Q1. What is retained in the retrained control? Within each training seed, the variants share the same frozen Stage-A encoders, codebook and decoder. Without Topology Modulation retains the diagram predictor loss (λ_z=1) and soft-plan auxiliary (λ_s=0.1/d_z, d_z=128) during Stage B1. Only the descriptor-to-logit pathway is removed: the planner uses language-only logits. The planner and realization model are trained under the ablated setting, with the same Stage-B2 budget.

The predictor loss trains the text-and-length topology head. The soft-plan auxiliary applies the plan readout to expected token embeddings, aligning the planner's distribution with the trajectory target. Keeping these losses in the control separates their supervision from the geometric logit-conditioning path. The constant-descriptor control also retains the modulation parameters and shared auxiliary losses, but supplies a caption-independent descriptor.

Training results. Your question calls for models that learn under each condition. We added retrained no-modulation and constant-descriptor controls in main-text Table 3. The same 180-caption task is evaluated for each of three checkpoints. The paired bootstrap resamples training checkpoints at the outer level and source-motion groups within checkpoints at the inner level, using identical draws for both variants. Values are mean±SD, with paired intervals for full minus variant:

Training conditionSeq.CoverageOrder given coverageFull minus variant: Seq. [95% CI]
Without modulation0.768±0.0150.848±0.0150.884±0.014+0.036 [0.020, 0.052]
Constant descriptor0.772±0.0150.854±0.0140.886±0.014+0.032 [0.016, 0.048]
Predicted descriptor0.804±0.0140.876±0.0130.905±0.012—

Compared with the no-modulation model, the predicted condition raises Seq. by .036, Coverage by .028 and conditional order by .021. Compared with the constant condition, the differences are .032, .022 and .019. Thus the training comparison addresses both completion and ordering, and the capacity control distinguishes caption-specific conditioning from simply retaining extra modulation parameters.

For clarity, Coverage is the fraction of outputs containing all required events. Seq. first measures the fraction of satisfied required DAG edges for each caption, then averages captions. An A→B edge is satisfied only when both events occur and A ends no later than B begins. Order given coverage averages these scores over fully covered outputs. Missing endpoints therefore reduce Seq. even when the surviving events are in order.

Paired Predicted–Mean and other interventions. To assess whether caption-specific geometry helps beyond the mean descriptor, Table S22 adds the requested paired comparison. Here one checkpoint is held fixed. The planner's geometric correction is a shared response difference, ρ(a,z/scale)−ρ(a,0), added to base language logits with scaling and clipping. Here a is the AR state from text, supplied length and token prefix; ρ is the shared modulator and scale normalizes descriptor coordinates. With z=0, the correction is exactly zero. Mean uses a common training-mean descriptor; Predicted uses text and supplied duration; Mismatched uses a donor reference descriptor; Reference is an oracle diagnostic.

AR random streams and Gaussian initial states are shared across conditions. Seq. uses all 180 captions, while paired conditional order uses only the jointly covered subset:

Paired contrastDelta Seq. [95% CI]Jointly covered NDelta conditional order [95% CI]
Predicted−Zero+.048 [.029,.067]148+.022 [.008,.036]
Predicted−Mean+.033 [.017,.049]151+.018 [.005,.031]
Predicted−Mismatched+.096 [.073,.119]139+.034 [.018,.050]
Reference−Predicted+.017 [.005,.029]155+.008 [−.002,.018]

The Predicted–Mean comparison favors the predicted descriptor in both Seq. and conditional order. Reference–Predicted measures the remaining gap under oracle geometry.

The retrained control tests whether geometric modulation improves a model that can adapt throughout training. The fixed-weight interventions test how a trained planner responds to its descriptor.

Both comparisons favor text-specific geometric conditioning: the retrained variants test its benefit during learning, and the Predicted–Mean intervention tests its benefit over a caption-independent descriptor within one trained model.

AuthorsOpenReview ↗

Part 2: effect size and full motion videos

W2. Magnitude of the SALAD comparison. Thank you for drawing attention to the size of the improvement and the text-alignment interval. We agree that the advantage over SALAD is modest.

The absolute single-output results are:

MethodSeq. [95% CI]Coverage [95% CI]Order given coverage [95% CI]
ActionPlan.769 [.740,.795].844 [.803,.882].888 [.856,.918]
SALAD.782 [.759,.804].856 [.806,.894].886 [.855,.915]
QuanTopo.804 [.783,.825].872 [.824,.911].904 [.875,.930]

QuanTopo has higher completion and conditional-order point estimates, with paired intervals that include zero.

Seq. averages the fraction of satisfied precedence edges per caption. Coverage requires every requested event to appear, and conditional order evaluates the fully covered outputs. The failure decomposition in revised Table 6 and Table S18 helps explain the aggregate Seq. difference:

MethodMissingOverlapReversed
SALAD.120.025.073
QuanTopo.112.028.056

These are required-edge fractions averaged over captions. The difference combines fewer missing endpoints and reversals with slightly more overlap. This locates the observed benefit in completion and event progression; the matched retraining comparison in Q1 tests the contribution of modulation under shared supervision.

Our user study uses 60 captions and 30 valid participants. For each video pair, participants judge text alignment and motion realism, with a no-preference option. The decisive win rate is W/(W+L), excluding ties from its denominator; the intervals use participant/caption cluster bootstrap. Against SALAD, the results are:

CriterionQuanTopo win rate [95% CI]
Text alignment52.8% [47.3,58.2]
Motion realism60.7% [55.5,65.8]

The text-alignment interval includes 50%, while the motion-realism interval is above 50%. The clearer perceptual benefit is therefore motion naturalness, with text alignment close to SALAD. The event comparison also favors QuanTopo in completion and ordering point estimates.

The matched training control in Q1 provides the more direct evidence for the role of geometric modulation. Retaining the representation, auxiliary supervision and training budget while removing modulation lowers both event coverage and conditional order. This connects the component to the intended mechanism: the geometric condition helps the planner realize a complete multi-stage motion, while language and ordered tokens specify its semantic progression.

Q2. Full motion videos. Thank you for emphasizing the need to inspect action progression and transitions in the full sequences. The anonymous materials page provides the full videos; Figures S5–S8 add rendered keyframes to show action progression and coordination. We invite you to view the full clips to assess the action progression and transitions. 01_figure_videos/figure_examples contains the 32 clips for the eight captions in Figures 1, 4 and S4, with four methods per caption. 01_figure_videos/composite180_quantopo contains the 180 QuanTopo composite-motion examples. Caption lists are included, so examples can be matched to the figure descriptions and roster. 02_composite180 additionally provides the five-method comparison videos and associated event records, organized by caption.

AuthorsOpenReview ↗

Part 3: reevaluation and corrected Table S3

Q3. Why S3 entries matched Table 1. Thank you for checking the consistency between Tables 1 and S3 so carefully. We apologize for the error and the confusion it caused; we should have checked the assembled table more carefully. When assembling Table S3, we mistakenly entered published values in cells intended for our reevaluation results. The corrected baseline results, including SALAD, are shown below as mean±sample SD across three repeats:

Dataset / methodFIDR@1Matching distance
HumanML3D / MoMask0.059±0.0030.523±0.0032.945±0.012
HumanML3D / MotionHiFlow0.038±0.0020.563±0.0042.705±0.015
HumanML3D / SALAD0.071±0.0060.587±0.0072.649±0.020
KIT-ML / MotionHiFlow0.154±0.0150.482±0.0112.538±0.017

Reevaluation protocol. Within a repeat, all methods use the same record list, requested ground-truth lengths, motion evaluator and retrieval candidates. Each baseline uses its fixed public checkpoint with three independently seeded generation streams. The HumanML3D repeats each contain 4,640 records; the KIT-ML repeats each contain 704. The input lists retain the sampling and crop choices for each repeat. These shared conditions give the methods a common evaluation setting.

For a fixed-checkpoint baseline, differences between repeats reflect generation and evaluation-stream variability. The standard deviation therefore describes variation across these three repeated evaluations. FID compares generated and reference motion-embedding distributions; R@1 measures rank-1 text-to-motion retrieval; matching distance averages paired text–motion embedding distances.

Individual repeats. The scores for all three HumanML3D baselines and MotionHiFlow on KIT-ML are:

Dataset / methodRepeatFIDR@1Matching distance
HumanML3D / MoMask10.0624430.5256472.945664
HumanML3D / MoMask20.0567190.5198282.933410
HumanML3D / MoMask30.0591350.5239222.957130
HumanML3D / MotionHiFlow10.0383200.5584052.721844
HumanML3D / MotionHiFlow20.0409410.5627162.698353
HumanML3D / MotionHiFlow30.0359640.5670262.695044
HumanML3D / SALAD10.0778900.5788792.670606
HumanML3D / SALAD20.0669870.5931032.631299
HumanML3D / SALAD30.0675090.5879312.646237
KIT-ML / MotionHiFlow10.1575460.4801142.518497
KIT-ML / MotionHiFlow20.1675930.4943182.547204
KIT-ML / MotionHiFlow30.1374000.4715912.548428

Statistical records and materials. The bootstrap scripts resample generation streams at the outer level and source-motion groups within each selected stream, using shared draws across the available baseline methods. Captions from the same source motion remain grouped. FID is recomputed from the resampled embedding distributions; retrieval hits and matching distances are aggregated from their per-record scores. The reported percentile intervals use 10,000 bootstrap replicates.

The anonymous materials page links to S3_baselines.zip. Under 03_baseline_evaluation, runs/ contains the three repeats of MoMask, MotionHiFlow and SALAD on HumanML3D, plus MotionHiFlow on KIT-ML. Each repeat includes metric summaries, per-record evaluations, motion embeddings and a generation manifest. fixed_records/ and shared_embeddings/ contain the common evaluation inputs, while results/ and statistics/ contain the aggregates and bootstrap outputs and scripts.

We appreciate the time and care you have devoted to reviewing our work. Your comments helped us clarify the modulation controls, provide fuller qualitative evidence and correct the reevaluation report. We would be glad to answer any further questions.

Reviewer_H8skOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposed a hierarchical text-to-motion framework, which split the the motion generation into the descrete high-level motion and continuous low-level and detailed motion. The high-level motion is used to preserve the overall motion structure, such as epetition, loops, and returns. And then the continuous detailed motion is the different executions of the high-level motion. A decoder is used to combine the high-level motion and the details into the final motion. Experiment demonstrates the strong results on HumanML3D and KIT datasets.

Critical strengths
  1. The motivation is clear. The motion needs to be follow the text prompt with the correct order.
  2. Different from the explicit motion plan, the implicit motion plan should be more informative and do not need explicit plan annotation.
  3. The ablation is thorough and helps explain where the improvements come from
Critical weaknesses
  1. The paper does not discuss or compare with ActionPlan [1]. Clarifying the differences between frame-level text-based planning and topology-guided discrete planning would better establish the contribution of this paper.
  2. The topology descriptor does not directly encode semantic order and the order is primarily from the language and the sequential generator. I am a little bit confused about why it can improve semantic order.

[1] Eric Nazarenus, Chuqiao Li, Yannan He, Xianghui Xie, Jan Eric Lenssen, and Gerard Pons-Moll. ActionPlan: Future-Aware Streaming Motion Synthesis via Frame-Level Action Planning. In European Conference on Computer Vision (ECCV), 2026.

Questions for authors
  1. The proposed full system is slightly complex (multi-stage training, topology prediction, autoregressive planning, conditional flow, and hierarchical MPS decoding...). Any component can be simplified?
  2. How much additional runtime and memory do MPS decoding and point-set prediction require compared with the parameter-matched MLP and vector alternatives? Their accuracy gains are relatively small, so these costs would affect the assessment of their practical value.
5 comments
AuthorsOpenReview ↗

Part 1: ActionPlan comparison

Thank you for pointing us to ActionPlan and for raising the relationship between geometry and temporal order. We have added an ActionPlan comparison and expanded the explanation of how the geometric target supports the ordered plan. The revised PDF has been uploaded, with the added comparison and explanations highlighted in blue.

1. What is planned? ActionPlan uses frame-level action planning with frame-level semantic supervision. QuanTopo learns a low-rate ordered token sequence supervised by a trajectory-derived target. During training, each motion is represented by root-relative pose, root position and local displacement. Persistent homology summarizes how these states connect and revisit regions across scales; a shared readout turns the resulting birth–death points into a fixed-dimensional descriptor. This target is computed from the motion itself and requires no frame-level action labels.

At generation time, the inputs are text and supplied duration. They predict the geometric descriptor; the autoregressive planner samples q conditioned on that prediction and text; conditional flow samples a continuous realization r conditioned on q and text; D(q,r) produces motion. The geometric descriptor supplies global state organization, while the ordered q sequence supplies a temporally indexed plan. In particular, a persistence diagram and a token sequence are different objects: the former is an unordered geometric target and the latter is the ordered synthesis interface.

2. Comparison setting. Following your suggestion, we evaluated the official ActionPlan checkpoint in offline mode on the same 180 captions and requested durations, using its official TAE decoder. Native temporal quantization produces small frame-count differences. The event comparison uses the same caption-specific required events, precedence DAG and scoring rule for each method; it evaluates event completion and precedence from the generated motions. Table 6 now includes this comparison, and Table S16 reports the paired differences.

MethodSeq. [95% CI]Coverage [95% CI]Order given coverage [95% CI]
ActionPlan.769 [.740,.795].844 [.803,.882].888 [.856,.918]
SALAD.782 [.759,.804].856 [.806,.894].886 [.855,.915]
QuanTopo.804 [.783,.825].872 [.824,.911].904 [.875,.930]

Coverage counts outputs containing all required events. Seq. averages the fraction of satisfied required edges over every caption; an edge requires both endpoints and predecessor-end ≤ successor-start. Order given coverage evaluates those edges on fully covered outputs. This separates completing the requested actions from ordering them.

For the paired QuanTopo–ActionPlan comparison, ΔSeq. is +.036 [.007,.066]. On the 139 captions fully covered by both methods, ΔOrder is +.021 [−.014,.056]. The conditional-order interval includes zero. QuanTopo’s higher Seq. score accompanies higher event coverage.

The anonymous materials page provides videos and event-level scoring records under 02_composite180, including results/paired_comparisons.json.

AuthorsOpenReview ↗

Part 2: why geometry affects event satisfaction

3. Division of responsibilities. The persistence descriptor is direction-agnostic. Connectivity and cycle persistence describe how motion states group, revisit regions and form loops; language and ordered tokens provide temporal direction and semantic progression. A geometric summary of a walk-and-return trajectory can express recurrence, while the text and token sequence specify which phase comes first. Topology serves as the global geometric condition; language and sequential generation specify semantic event order.

The training objectives make this division operational. A readout of pooled q embeddings predicts the trajectory descriptor, giving the plan a sequence-level information target. A plan-only anchor reconstructs the ordered motion through D(q,0), so q must support synthesis before continuous detail is supplied. The encoder, decoder and autoregressive planner retain positional order even though the pooled descriptor readout is permutation-invariant. The continuous r channel then supplies within-plan differences in timing, posture and coordination.

4. Why this affects Seq. Seq. counts satisfied required precedence edges. An edge A→B fails when either event is missing, when their intervals overlap, or when B precedes A. Thus Seq. combines event completion with precedence. For A→B→C, missing A while realizing B→C gives Seq.=1/2, Coverage=0 and exclusion from Order given coverage. Geometric organization and recurrence are relevant to realizing a complete multi-stage motion, while language and the ordered plan govern its semantic progression.

To identify which event failures account for the Seq. difference, we decomposed failed edges into missing endpoints, overlap and reversal. These rates are included in Table 6 and detailed in Table S18:

MethodSatisfiedMissingOverlapReversed
MoMask0.7240.1600.0200.096
MotionHiFlow0.7590.1410.0250.076
ActionPlan0.7690.1340.0260.071
SALAD0.7820.1200.0250.073
QuanTopo0.8040.1120.0280.056

Fractions are computed over required edges within each output and then averaged over captions. Relative to SALAD, the rounded difference comprises .008 fewer missing-edge failures and .017 fewer reversals, offset by .003 more overlap, summing to the .022 Seq. difference. These are observed differences across full methods; the matched training control below addresses the modulation pathway itself.

5. Training-time modulation control. The retrained no-modulation model retains the frozen Stage-A representation, topology-predictor loss and soft-plan supervision, but learns without descriptor-to-logit modulation; its flow stage is retrained with the same budget. The constant condition retains modulation capacity while removing caption-specific descriptor variation. Mean±SD over three checkpoints is:

Training conditionSeq.CoverageOrder given coverageFull minus variant: Seq. [95% CI]
Without modulation0.768±0.0150.848±0.0150.884±0.014+0.036 [0.020, 0.052]
Constant descriptor0.772±0.0150.854±0.0140.886±0.014+0.032 [0.016, 0.048]
Predicted descriptor0.804±0.0140.876±0.0130.905±0.012—

Predicted conditioning improves both coverage and conditional order in this matched setting. This is a different comparison from replacing the condition of one model already trained with geometric input: the retrained alternatives can adapt their planner and flow during training.

6. Where temporal order resides. The revised mechanism discussion draws on Table S21 to test how token order affects synthesis while pooled geometric content is held fixed. Permuting q before decoding leaves the pooled descriptor readout unchanged but changes the ordered input seen by the decoder. The reported Seq. falls from .804±.014 to .642±.020, and plan-only MPJPE rises from 409.6±10.8 to 623.7±16.9 mm. This intervention distinguishes aggregate plan content from token position: topology specifies a global organization target, while ordered synthesis determines the temporal realization.

The mechanism is therefore complementary: trajectory supervision gives the plan a global organization target, the anchor makes that plan usable for synthesis, and the sequential planner combines it with language to determine the action progression. The retraining and permutation controls examine these roles separately.

AuthorsOpenReview ↗

Part 3: simplification and practical costs

7. Which components can be simplified? The training stages give stable roles to the two latent channels. Stage A learns the encoders, codebook and decoder; Stage B1 learns text-conditioned descriptor prediction and ordered-plan generation; Stage B2 learns continuous realizations in that fixed latent coordinate system. Persistence targets are precomputed during training. At inference, the text-and-length head predicts the geometric condition.

The anatomical decoder first converts q into six shared part contexts: trunk, head, two arms and two legs. Fixed routing sends each joint its corresponding part context, and joint construction combines that context with r and joint identity. The parameter-matched hierarchical MLP preserves this routing and replaces the MPS (matrix product state/tensor-train) operator, separating the operator choice from the anatomical hierarchy.

Likewise, point-set prediction outputs birth–death coordinates and existence weights before applying the shared descriptor readout; the vector alternative predicts the descriptor directly. The former retains the unordered point structure of the target. Both prediction heads can be used with the same planner and flow model.

Gated roll-in occasionally trains the flow on planner-proposed plans with the continuous realization re-encoded for each proposal; a reconstruction-error gate accepts suitable pairs. “No gated roll-in” uses reference-encoded plans throughout. Table S12 addresses simplification by replacing the prediction head and decoder and removing roll-in, individually and jointly:

ConfigurationFIDSeq.Training GPU-hLatency s
Full.027.80441.2.48
Vector-standard head.031.78340.2.48
Hierarchical Conditional-MLP.031.79138.1.47
No gated roll-in.030.79639.6.48
Lite.034.77635.5.47
Lite, 10 steps.039.77135.5.30

Lite combines vector prediction, the Conditional-MLP decoder and no gated roll-in. Its training cost decreases from 41.2 to 35.5 GPU-hours, a reduction of about 14%, with FID changing from .027 to .034 and Seq. from .804 to .776. Reducing Lite to ten flow steps gives .30 s latency, with FID .039 and Seq. .771.

8. Component runtime and memory. For the practical overhead you asked about, Table S11 compares MPS and point-set prediction with parameter-matched alternatives. Reported settings are one A800, BF16, 196 frames and 1,000 synchronized trials. b1/b32 denote batch sizes; the memory column is the reported peak at batch 32. Training throughput refers to Stage A for decoders and Stage B1 for prediction heads.

ComponentParametersb1 msb32 msb32 peak GBTraining samples/s
Hierarchical MPS15.4M9.448.61.84860
Conditional-MLP15.3M6.231.41.52985
Point-set head1.623M2.18.4.42510
Wide vector head1.648M1.55.8.34575

Relative to Conditional-MLP, MPS adds 3.2 ms at batch 1 and 17.2 ms at batch 32 (about 52%/55%), and the reported peak rises by 0.32 GB. Throughput changes from 985 to 860 training samples/s. The parameters remain closely matched, while FID improves from .031 to .027 and Seq. from .791 to .804.

Relative to the wide vector head, point-set prediction adds 0.6/2.6 ms at batch 1/32 (about 40%/45%) and 0.08 GB reported peak memory. Training throughput changes from 575 to 510 samples/s. FID improves from .029 to .027 despite the point-set head having slightly fewer parameters (1.623M versus 1.648M).

9. Why the end-to-end difference is smaller. Descriptor prediction and final motion decoding run once per generated sequence. The flow vector field runs at each integration step; the default uses 20 Euler steps. Replacing MPS with the matched MLP changes the reported full-generation latency from .48 to .47 s, although the local decoder cost differs more substantially. Component latency and memory therefore describe the local overhead, while the end-to-end column describes its impact on complete generation.

The full configuration favors generation quality, while Lite offers lower training cost and its ten-step version reduces sampling latency further. These alternatives allow the same planning-and-realization framework to be used with a simpler head and decoder.

The implementation is available through the anonymous materials page, under 04_code.

We appreciate your thoughtful review and the suggestions on related work, mechanism and practical costs. Thank you for helping us explain these aspects more clearly. Please let us know if any point would benefit from further explanation.

Reviewer_1GErOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

QuanTopo combines topology-guided discrete planning with conditional flow to generate diverse motions from text and a given duration. It achieves FID scores of 0.027/0.122 on HumanML3D/KIT-ML and improves composite-action ordering on the evaluated HumanML3D subset.

Critical strengths
  1. A well-motivated design connecting topology-guided planning with continuous motion generation.
  2. Thorough ablations and a carefully designed human evaluation.
  3. Strong benchmark performance and improved composite-action ordering.
Critical weaknesses
  1. Ablation definitions could be clearer. The components retained in Base-Hybrid could be stated more explicitly. Table 2 also mixes text-conditioned generation metrics with encoded-reference reconstruction and probe diagnostics without grouping the columns.
  2. Some reproduction details need clarification. The composite-evaluation records and validation-selection procedure are not fully covered by the supplied reproduction path. The PSD computation is also described differently in the paper and code. These are minor, fixable concerns.
Questions for authors
  1. What does Base-Hybrid retain? Please distinguish generation metrics from reference-code reconstruction/probe metrics in Table 2.
  2. Please provide anonymized composite-evaluation records and bootstrap scripts, separating single-checkpoint interventions from three-checkpoint summaries.
  3. Was PSD computed by FFT/interpolation or direct Fourier sums? Please align C.2 with the implementation.
3 comments
AuthorsOpenReview ↗

Part 1: ablation definitions and interpretation

Thank you for your careful reading and for identifying the ambiguities in our ablation definitions and evaluation protocol. We have clarified the retained components and grouped generation results separately from reference-code reconstruction and probe diagnostics. These clarifications are highlighted in blue in the revised manuscript, which we have uploaded.

1. What Base-Hybrid retains. Given text c and supplied duration T, the model first generates a low-rate discrete plan q, then samples a frame-rate continuous realization r conditioned on that plan and text. A shared decoder produces D(q,r). Base-Hybrid retains the discrete/continuous encoders, codebook, text-conditioned autoregressive planner, conditional flow, and shared decoder. It jointly removes directed-plan supervision, geometric conditioning, and anatomical routing. It is therefore a joint simplified baseline; the individually named ablations remove their corresponding components from the full model. At loss level, Base-Hybrid sets the Stage-A topology/anchor weights and Stage-B1 diagram/soft-plan weights to zero, retaining reconstruction, velocity, VQ commitment, AR and flow objectives. In contrast, no-modulation retains the full Stage-A objective and the Stage-B1 diagram/soft-plan auxiliaries; only the descriptor-to-logit path is disabled.

The plan and realization have different computational roles. The ordered q sequence establishes motion organization before r supplies within-plan variation. Topology supervision specifies global trajectory information that q should retain, while the plan-only anchor trains D(q,0) to produce a motion before continuous detail is supplied. These objectives distinguish an operational plan from a discrete code learned only through full reconstruction.

2. Generation metrics and reference-code diagnostics. Table 2 now labels generation and reference diagnostics separately to make their input conditions and interpretation explicit. FID measures motion-distribution agreement, and R@1 is rank-1 text-to-motion retrieval accuracy. Together with Seq., they evaluate text-conditioned generation: q and r are sampled from the text and supplied duration. The reconstruction and probe diagnostics encode a reference motion into q* and r* and measure the information retained in these codes.

Generation results:

VariantFIDR@1Seq.
Base-Hybrid.061.541.664
Without plan topology.043.568.721
Without anchor.039.573.744
Without modulation.035.578.768
Full.027.586.804

Reference-encoding diagnostics:

VariantProbe errorPlan-only MPJPEFull MPJPE
Base-Hybrid.284712.4231.6
Without plan topology.219428.9228.7
Without anchor.149812.5215.9
Without modulation.141409.6224.8
Full.141409.6224.8

Values are three-checkpoint means; MPJPE is in mm. Full MPJPE measures D(q*,r*), plan-only MPJPE measures D(q*,0), and the probe predicts the trajectory descriptor from the frozen encoded plan. Thus full reconstruction and plan usefulness are distinct measurements. For example, removing the anchor improves full MPJPE from 224.8 to 215.9 but worsens plan-only MPJPE from 409.6 to 812.5: the two channels reconstruct well together while q alone carries less usable motion information. Removing plan-topology supervision increases probe error from .141 to .219.

Within each training seed, the full and no-modulation variants share the same frozen Stage-A representation. Their reference diagnostics are identical because they use the same frozen encoders and decoder. Descriptor-to-logit conditioning acts in the downstream generation stage.

3. Complementary training control. Table 3 tests caption-specific geometry with retrained no-modulation and constant-descriptor controls. The following table reports mean±SD over three trained checkpoints, using 180 captions per checkpoint:

Training conditionSeq.CoverageOrder given coverageFull minus variant: Seq. [95% CI]
Without modulation0.768±0.0150.848±0.0150.884±0.014+0.036 [0.020, 0.052]
Constant descriptor0.772±0.0150.854±0.0140.886±0.014+0.032 [0.016, 0.048]
Predicted descriptor0.804±0.0140.876±0.0130.905±0.012—

The constant-descriptor control retains the modulation capacity and shared auxiliary losses while supplying the same descriptor across captions. The predicted condition performs better under the matched training setting. These event metrics complement FID/R@1 by directly reporting event completion and required precedence.

Together, these controls support two distinct roles: topology supervision and the anchor determine what the plan retains, while geometric modulation determines how that information guides text-conditioned plan generation.

AuthorsOpenReview ↗

Part 2: records, statistics and PSD

4. Composite records and scoring. The anonymous materials page links to the records in 02_composite180 and implementation in 04_code. The roster contains 180 caption–clip records, each with at least two required events and a DAG of required precedence edges. The caption, supplied length, event definitions, source-motion group, output mapping, onset/offset labels and scoring files identify what is evaluated for each output. All methods are evaluated against the same required events and edges for a caption.

For a required edge A→B, a positive vote requires both events to be present and A to end no later than B begins; equal boundary frames are allowed. A missing endpoint, reversed order or overlap fails that edge. Majority voting over the supplied label sets gives the edge outcome. A caption's Seq. is its fraction of satisfied required edges; the reported Seq. is the equally weighted mean of caption scores.

Figure S9 makes the completion–precedence distinction concrete with missing-event, overlap and reversal examples. Coverage is the fraction of outputs containing every required event. Order given coverage averages Seq. only over fully covered outputs. For an A→B→C chain, omitting A while completing B→C gives Seq.=1/2 and Coverage=0; it is excluded from the conditional-order average. Completing all three events in the wrong order has Coverage=1 and fails the corresponding precedence edge.

5. Repeated models versus fixed-weight interventions. The three-checkpoint training summary reports variation across trained models, each evaluated on the same 180-caption task. A condition intervention holds one checkpoint fixed and changes only the supplied descriptor. Zero, Mean, Mismatched, Predicted and Reference all use the same trained weights and vary the descriptor supplied at inference. AR random streams and Gaussian initial states are paired across conditions.

For the released single-output comparisons, score.py computes absolute event metrics and paired_score.py computes differences between methods. The 10,000 paired bootstrap resamples draw source-motion groups; all related records remain together. The roster contains 86 groups. Each replicate applies the same group draw to both methods before taking their difference. Conditional-order differences use only captions fully covered by both methods. For the three-checkpoint comparison, the outer level resamples training-seed/checkpoint indices with replacement; within each selected checkpoint, the inner level resamples source-motion groups, keeping related records together. Both variants use the same checkpoint and group draws. Each replicate averages the paired differences over selected checkpoints; the 95% interval is the 2.5th–97.5th percentile of 10,000 replicates.

6. Validation selection. Validation and saving occur every 5k updates. Stages A and B1 select the checkpoint with the lowest validation stage objective. Stage B2 selects by mean validation FID using fixed validation records and five generation seeds. Stage A learns the encoders, codebook and decoder; freezing it keeps the plan and realization coordinates fixed for subsequent planner and flow training. These are the selection rules specified in Appendix A, Table S2.

7. PSD implementation. The PSD experiments reported in the paper use the fixed-grid direct Fourier sums described in the manuscript. The originally supplied code did not match this implementation. We have updated the released code to implement the same computation and apologize for this inconsistency. For sampled normalized states s[n,c], N=min(T,32), and f_m=m/32 for m=1,...,16:

Pc(fm)=N−1∣∑n=0N−1sn,cexp⁡(−2πifmn)∣2P_c(f_m)=N^{-1}|\sum_{n=0}^{N-1}s_{n,c}\exp(-2\pi i f_m n)|^2.

It then applies log(1+P), flattens the channels and frequencies, and uses training-fitted 128-D PCA. DC is excluded; there is no windowing, detrending or spectral interpolation. At N=32, these frequencies and the 1/N normalization equal nonzero orthonormal real-FFT power. For shorter sequences, an ordinary N-point FFT would use a different grid m/N; direct sums retain the common m/32 grid without interpolation. The fixed grid makes each frequency coordinate represent the same frequency across sequence lengths.

PowerSpectrumSummary.features implements the sine/cosine sums directly. ACF describes temporal correlation, PSD oscillatory content and EDM pairwise geometry. The matched vector controls share sampled states, output dimension, prediction-head capacity, supervision locations and downstream model budget. The experimental variable is the trajectory summary.

Thank you again for the time and care you have put into reviewing our work. Your comments have helped us clarify the ablations and improve the reproducibility details. We would be happy to clarify any remaining questions.

Thank you for the response and for providing the additional retrained controls. These results help clarify the contribution of geometric conditioning. I will maintain my original rating of 3.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.