acceptodds
Under review as a conference paper at ICLR 2027

Skeletons in Flow: Graph Structured Flow Matching for Human Motion Prediction

35 views

Abstract

Human motion prediction requires diverse future trajectories that remain consistent with observed motion and the articulated structure of the body. Skeletal constraints restrict individual poses, while coordinated motion also depends on spatial interactions (between connected joints) and temporal interactions (between time instants). We introduce Graph Structured Flow Matching (GSFM), which transports the complete future skeletal trajectory through one conditional velocity field. The trajectory forms a spatiotemporal skeleton graph, and spatial and temporal attention couple its evolution according to skeletal relations and physical time offsets. Bone directions lie on unit spheres relative to a root joint, and tangent evolution preserves prescribed bone lengths throughout generation. We train the field through conditional flow matching along geodesic paths connecting random trajectories centred on the last observed pose to recorded futures. Experiments on the archive of motion capture as surface shapes (AMASS) dataset evaluate prediction accuracy, diversity calibration, and motion statistics. We demonstrate the contributions of spatial and temporal message passing in the proposed architecture through an ablation study. GSFM models trained on AMASS also perform competitively on the Human3.6M skeleton without parameter updates or retraining, demonstrating applicability to an unseen skeletal structure.

Primary area:
generative models
Keywords:
human motion prediction, skeleton, 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_ux7XOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes Graph Structured Flow Matching (GSFM) for stochastic human motion prediction. The main idea is to formulate the entire future skeletal trajectory as a single object transported by a conditional flow matching process. The authors represent skeletons on a constrained geometric space using root positions and unit bone directions, and design a graph-based velocity network with spatial and temporal attention to model skeletal interactions. Experiments on AMASS and zero-shot transfer to Human3.6M are provided to evaluate prediction accuracy, diversity, and skeletal consistency. The paper aims to combine flow matching, geometric skeleton representation, and graph-structured modeling into a unified framework for physically plausible human motion prediction. However, the overall presentation, logical justification of the proposed approach, and coverage of relevant literature are not sufficiently developed, which makes the contribution and significance of the work less convincing.

Critical strengths
  1. Conceptually clean formulation of motion prediction using trajectory-level flow matching. The idea of modeling the whole future motion sequence in one shot, rather than generating poses sequentially, is principled and gives a unified probabilistic formulation.
  2. Geometry-aware skeleton representation. Representing bone directions on spherical manifolds offers an elegant way to preserve bone length constraints without extra post-processing.
Critical weaknesses

Weakness 1: Limited novelty and insufficient justification of the central methodological contribution.

The main contribution of this paper appears to be an integration of several existing techniques, including flow matching, geometric skeleton representations, and graph-based attention mechanisms. While the combination is sound, the paper does not convincingly demonstrate that Graph Structured Flow Matching introduces a substantial conceptual advance beyond existing human motion generation and prediction frameworks.

In particular, the central motivation of using a graph-structured velocity field is not sufficiently validated. The paper does not compare against a strong non-graph flow matching baseline, making it unclear whether the observed improvements are due to the proposed graph formulation, the geometric representation, or other architectural choices. This weakens the claim that the proposed framework provides a fundamentally better approach for motion prediction.

Weakness 2: Unconvincing empirical analysis to support the claimed advantages.

Although the paper reports extensive quantitative results, the experimental evaluation does not adequately analyze the proposed design choices. The reported improvements over previous methods are relatively limited, and the ablation studies mainly focus on removing spatial or temporal attention components rather than isolating the contribution of the key components of the framework.

For example, the paper lacks controlled comparisons between:

  • flow matching against diffusion-based alternatives under comparable settings;
  • graph-based velocity modeling against standard transformer architectures;
  • manifold-based skeleton representations against unconstrained representations.

Furthermore, given that the paper emphasizes generating diverse and physically plausible motions, the qualitative evaluation is limited. More detailed visualization of generated trajectories, covering diversity and failure cases, would better support the main claims.

Questions for authors
  1. The main contribution of the paper centers on the graph-structured velocity field trained with flow matching. Could the authors clarify how much of the gain is attributable to the flow matching formulation itself, for example by training the same network under a diffusion objective with comparable settings?

  2. The paper does not compare against a non-graph flow matching baseline. Would it be possible to report such a comparison, so that the effect of the proposed graph-based design can be separated from the rest of the framework?

  3. The proposed framework combines flow matching with geometric skeleton representations. Can the authors quantify the contribution of the manifold-based representation by comparing it with a Cartesian-coordinate representation under the same architecture?

  4. The paper claims advantages in generating diverse and physically plausible motions, but the qualitative results are largely limited to Figure 1. Could the authors provide additional visualizations, including side-by-side comparisons with baselines and representative failure cases?

7 comments
AuthorsOpenReview ↗

Part 1. response to W1

We thank the reviewer for their feedback and suggestions.

We first clarify what the graph structure provides in the developed GSFM model. In GSFM, the kinematic tree (spatial graph) is used to define the state of the model during generation (which, due to the Riemannian framework always yields a valid skeletal configuration), structure the map from model state to joint positions (using forward kinematics), and as node features (tree depth, number of child nodes, end effector flag). The temporal graph structure (links between the same node across temporal frames) enables the model to enforce temporally consistent motion across generation. To clarify how the performance gains can be attributed to the graph structure, we recall Table 3 from the previously submitted manuscript.

Table 8: Change in ADE with respect to the "Full spatial and temporal attention" configuration.

ConfigurationΔ\DeltaADE
Spatial attention restricted to one hop+6.9%
Spatial attention removed+14.3%
Temporal attention restricted to one hop+27.4%
Temporal attention removed+52.4%

Table 8 precisely identifies the contribution of the graph structure to the predictive capabilities of the GSFM model. Restricting message passing to one hop in either the spatial or temporal dimension slightly increases ADE, and when message passing is eliminated altogether in the temporal domain, the ADE rises by 52.4%.

We further distinguish the contribution of the spatiotemporal graph to GSFM's performance by comparing to a non-graph baseline. We trained a spatio-temporal transformer which learns a joint identifier embedding in the place of the features which are provided by the kinematic tree. We refer to this model as the "Transformer FM" configuration. All other experimental parameters were held constant with respect to the GSFM-deep, ablation-scale model configuration. Transformer FM yields comparable ADE to GSFM-deep (9.358 ± 0.034 vs. 9.317 ± 0.063) on AMASS. However, Transformer FM is a fundamentally limiting approach in that its learned weights are tied to the joint identities in AMASS, and if it is to be used on any other skeleton, the model must be re-trained. In contrast, GSFM allows the same weights to run on skeletons with different kinematic trees with no retraining or retargeting. In the zero-shot transfer domain, GSFM provides more than a 50% improvement in ADE when compare to the best retargeting-based baseline method (0.491 ± 0.035 vs. 1.094).

Further, to better distinguish GSFM from related methods, we will revise the introduction around the two core contributions, move relevant references from Appendix F (Additional Related Work) into the Introduction, and report all additional comparisons and ablations.

AuthorsOpenReview ↗

Part 2. response to W2, Q1, Q2 and Q3

To further strengthen the positioning of GSFM with respect to other frameworks, we have trained the following variations:

  • A transformer which learns a joint identifier embedding in place of the kinematic tree features (Transformer FM),
  • A flow matching model which uses a Cartesian representation of node states without constraints (Cartesian-flow),
  • A diffusion model which uses a Cartesian representation of node states without constraints (Cartesian-Vdiff),
  • A Euclidean flow matching model which uses a single bone direction normalization after generation to yield joint positions (Euclidean FM).

Each variation is trained over three seeds at ablation scale, evaluated at 20 NFE on each seed, and each metric is averaged across the seeds.

Table 9: ADE and limb stretching for the GSFM model, Cartesian variations, and Euclidean flow matching configuration.

ConfigurationuADE [cm]Δ\DeltaADEStr. [%]
GSFM-deep9.317 ± 0.063−0.0
Transformer FM9.358 ± 0.034+0.4%0.0
Cartesian-flow9.542 ± 0.032+2.4%9.46
Cartesian-Vdiff9.529 ± 0.016+2.3%7.74
Euclidean FM12.380 ± 0.319+32.9%0.0

The results in Table 9 demonstrate that no tested alternative is able to yield GSFM-level performance in prediction accuracy, bone length preservation, and zero-shot transfer. Specifically, Cartesian-flow, Cartesian-Vdiff, and Euclidean FM all sacrifice performance in ADE relative to GSFM. Transformer FM achieves comparable ADE performance to GSFM, but its learned joint embeddings are tied to the representation of the skeleton in AMASS, so it cannot be applied in a zero-shot transfer setting without retraining. Next, we evaluate how much of GSFM's performance gains can be attributed to flow matching itself. To do so, we introduce another variations of a Cartesian diffusion model which use V predict and X predict, respectively. As above, the subsequent results are evaluated at 20 NFE over three seeds at ablation scale.

In addition to the uADE defined in the reply to Reviewer SVkr, Table 10 reports the unified final displacement error (uFDE), the uMAE, and the unified average pairwise distance (uAPD). Let tft_f denote the last frame in If\mathcal{I}_f. Following [2], the uFDE is the analogue of the uADE evaluated only at the final frame,

uFDE≜min⁡s∈{1,…,S}1J∑j=1J∥q^s,tf,j−qtf,j∗∥2,\mathrm{uFDE} \triangleq \min_{s \in \lbrace 1,\ldots,S\rbrace } \frac{1}{J} \sum_{j=1}^{J} \left\Vert \hat{q}_{s,t_f,j} - q^{\ast}_{t_f,j} \right\Vert _2 ,

and it is reported in centimeters. The uMAE is the MAE defined in the reply to Reviewer SVkr, reported in degrees. The uAPD is the unified version of the average pairwise distance (APD) introduced in [2], which measures the diversity of the predictions as the average distance between pairs of distinct samples among the SS samples generated for the same observation, where the distance between two samples is averaged over the joints as in the uADE, so that a larger uAPD indicates more diverse predictions.

Table 10: Comparison between GSFM across 3 trained seeds and Euclidean flow matching.

ConfigurationuADE [cm]uFDE [cm]uMAE [deg]uAPD [cm]
Cartesian-flow9.542 ± 0.03210.842 ± 0.0848.33 ± 0.0518.32
Cartesian-Vdiff9.529 ± 0.01610.905 ± 0.0277.82 ± 0.1015.26
Cartesian-Xdiff9.625 ± 0.00410.942 ± 0.0668.25 ± 0.2815.18
GSFM9.317 ± 0.06310.230 ± 0.0867.22 ± 0.1116.01

We see that GSFM outperforms the best performing Cartesian method in ADE (2.2% lower), FDE (6.2% lower), and MAE (7.7% lower), at the cost of a slight reduction in APD performance (5.2% higher). Further, GSFM yields perfect reproduction of bone lengths, while every other tested Cartesian variation has nonzero bone length error.

AuthorsOpenReview ↗

Part 2. continue response to W2, Q1, Q2 and Q3

Lastly, we compare the GSFM-deep model at full scale to a single seed of Cartesian-flow.

Table 11: Comparison between GSFM-deep and Cartesian-diff at full scale.

ConfigurationADEFDEMAEAPDECMDStr. %Jit. %
Cartesian-flow0.47680.55255.9402.725.411.090.23
GSFM-deep, three seed mean0.47680.54895.8362.815.360.00.0

Table 11 demonstrates that when compared to Cartesian-flow, GSFM achieves equivalent ADE, improved FDE (0.65% lower), and improved MAE (1.78% lower), improved CMD (0.92% lower), while retaining perfect bone length prediction. The cost of achieving bone length preservation through projection is shown in a reply to reviewer EKoq, where we demonstrate that Euclidean FM gives 0.0% Str. and Jit., but has 87% higher ADE than GSFM.

The above graph ablations and comparison to the Cartesian baseline demonstrate that the developed method achieves (i) improved prediction quality, (ii) perfect bone length preservation, and (iii) ability to transfer directly to unseen skeleton data without retraining or retargeting. We reiterate that, relative to SkeletonDiffusion (lowest ADE of all basline methods), GSFM-deep achieves lower ADE across all trained seeds, 4.7% lower MAE, and 53% lower CMD. Further, GSFM-tied achieves comparable performance with only 4.47M parameters, which is 43% fewer parameters than EquiFusion, the smallest comparable model in the baseline set.

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

This paper proposes a method to achieve exact human bone structure (bone length) preservation through adapting Riemannian Flow Matching formulation with a skeletal manifold representation. Spatio-temperal graph attention are used to model the geometry. Experiments on AMASS show that the proposed method can achieve better accuracy than previous methods; zero-shot transfer experiments on Human3.6M shows improvement over some of the previous methods.

Critical strengths
  1. This paper shows that bone-length preservation in human motion prediction can be achieved through adapting Riemannian Flow Matching formulation with a skeletal manifold representation.
  2. The proposed method shows improved accuracy on AMASS dataset.
Critical weaknesses
  1. While the proposed method shows improvements over EquiFusion on AMASS, it has worse zero-shot transfer performance on Human3.6M. Since EquiFusion already supports arbitrary skeleton topologies and exact bone length preservation, it remains unclear what distinct advantages are provided by the proposed graph-structured Riemannian flow formulation beyond replacing latent diffusion with a manifold-based generative process.
  2. No experiments on (Euclidean Flow Matching + Bone Direction Normalization) vs. Riemannian flow formulation. Since the former approch can also preserve bone length, without this experiment, it is unclear whether the Riemannian flow formulation is necessary.
  3. The contribution of this paper is limited if the authors cannot show that the Riemannian flow formulation has distinct advantages over previous methods (like Equifusion) and Euclidean Flow Matching + Bone Direction Normalization baseline.
Questions for authors

See weaknesses.

3 comments
AuthorsOpenReview ↗

Part 1. response to W1 & W2

We thank the reviewer for a clearly framed review and the helpful insights.

First, we clarify the theoretical differences between EquiFusion and GSFM. EquiFusion trains a motion encoder and then a latent diffusion model, while GSFM transports the entire future skeleton trajectory as a single state. These different choice results in the trade-off between zero-shot transferring performance and other model performance, especially model parameters.

GSFM uses a single network conditioned on the observation and the kinematic tree, while latent methods train an autoencoder and a generator. The accuracy in the map from state to joint position for latent methods may be limited by the ability of the decoder to reconstruct the skeletal geometry.

Further, for GSFM, the Riemannian framework constrains the future state to valid skeletal configurations for every generation time. This enables us to reduce the number of generation steps, and still output a valid skeleton.

Lastly, we see benefits over EquiFusion in parameter efficiency. All tunable parameters are used to learn the velocity field, rather than also being used in the encoder/decoder structure, allowing us to train a competitive model using only 4.47M paramers (GSFM-tied).

Additionally, to clarify what the Riemannian flow matching formulation offers beyond EquiFusion and Euclidean flow matching with normalized bone directions, we have performed additional experiments and ablations. To begin, we trained GSFM-deep over two additional seeds to demonstrate reproducibility. We found that performance remained consistent across all trained seeds.

Table 5: Comparison between GSFM across 3 trained seeds and EquiFusion.

ConfigurationParameter CountADEFDEMAEAPDECMD
EquiFusion7.9M0.4960.5606.2142.51813.10
GSFM-tied, single seed4.47M0.4840.5515.9302.6834.78
GSFM-deep, three seed mean, 50 NFE30.49M0.47680.54895.8362.815.36
GSFM-deep, single seed, 10 NFE30.49M0.4830.5605.8543.106.25

In Table 5, "single seed" indicates that the model was trained over a single run, and "three seed mean" indicates that the same model was trained over three seeds, and the result was averaged across the three resulting models. We see that when trained and evaluated on the AMASS dataset, GSFM-deep beats EquiFusion in ADE, FDE, MAE, and CMD across all three training seeds. Further, GSFM-tied is 43% smaller than EquiFusion and better in ADE (+2.4%), FDE (+1.6%), MAE (+4.6%), and CMD (+64%). GSFM-deep builds upon these performance gains, yielding better ADE (+4.0%), FDE (+2.0%), MAE (+6.5%), and CMD (+144.4%) than EquiFusion, but at the cost of a larger model size and reduced APDE performance.

AuthorsOpenReview ↗

Part 2. response to W3

Next, to further distinguish GSFM from alternate methods, we trained a Euclidean flow matching model which uses bone direction normalization to preserve the skeletal geometry. We refer to this model as "Euclidean FM" in the revised manuscript. The Euclidean FM model uses an identical training pipeline to GSFM-deep, varying only in the generation process. A single normalization of the bone direction is applied before the forward kinematics map.

Table 6: Comparison between GSFM across 3 trained seeds and Euclidean flow matching.

ConfigurationADEFDEMAEAPDECMD
Euclidean FM0.89220.899215.132.40157.2
GSFM-deep, three seeds0.4768 ± 0.00070.5489 ± 0.00145.836 ± 0.0062.81 ± 0.055.36 ± 0.31

The GSFM model demonstrates clear performance improvements over the Euclidean FM configuration in ADE (+46.6%), FDE (+39.0%), MAE (+159.3%), and CMD, while slightly sacrificing performance in ADPE (−14.6%). It should also be noted that in addition to the compromised predictive performance, the Euclidean FM configuration also yields up to 41% bone length error. Lastly, we examine how GSFM compares to Euclidean FM at varying NFE to quantify how GSFM affects generation time.

Table 7: Comparison between GSFM across 3 trained seeds and Euclidean flow matching while varying NFE.

ConfigurationADECMDTotal Time [s]
GSFM-deep, 50 NFE0.4775.6115.67
Euclidean FM, 50 NFE0.892157.215.64
GSFM-deep, 20 NFE0.4795.936.27
Euclidean FM, 20 NFE1.034426.36.27
GSFM-deep, 10 NFE0.4836.253.13
Euclidean FM, 10 NFE1.091580.23.13

It is clear from the results in Table 7 that the Riemannian framework adds no additional inference time, and preserves its performance gains over Euclidean FM across all tested NFE regimes.

We thank the reviewer again for the time and the clearly framed review!

The rebuttal has addressed my concerns. I would raise my rating.

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

The paper models future skeletal motion using conditional flow matching over bone directions and reconstructs positions through forward kinematics. It combines a geometric representation with structured attention and evaluates prediction accuracy, motion statistics, parameter sharing, and cross-skeleton transfer.

Critical strengths
  1. The spherical bone-direction construction and forward kinematics provide a clear bone-length preservation property.
  2. Competitive displacement results and improved motion statistics support the usefulness of the combined method.
Critical weaknesses
  1. The reported MAE has an unresolved definition mismatch. Appendix D.2 and mae_joint_angle measure adjacent-limb angle error, while the summary's mean_angle_error measures bone orientation with different sample selection. Which definition produced the tables remains unclear.
  2. The predictive benefit of the structural choices is not fully isolated. The attention ablations and geometric invariants do not separately establish prediction gains from explicit skeletal bias or the chosen spherical path.
Questions for authors
  1. Which function, units, and sample-selection rule produced the MAE tables, and do they match the baselines?
  2. Can an existing matched comparison isolate skeletal bias while retaining attention access, features, and budget?
  3. What training variation supports the small reported accuracy leads?
4 comments
AuthorsOpenReview ↗

Part 1. response to Q1 & W1

We thank the reviewer for the review and precise questions.

Q1 & W1

The mean angle error (MAE) provided in the manuscript uses the definition of MAE developed in SkeletonDiffusion [1]. Let X∈R(S,∣If∣,J−1,3)X \in \mathbb{R}^{(S,\vert \mathcal{I}_f\vert ,J-1,3)} denote the predicted trajectory and let Y∈R(∣If∣,J−1,3)Y \in \mathbb{R}^{(\vert \mathcal{I}_f\vert ,J-1,3)} denote the ground truth. Let bones $\ell = \lbrace t_\ell, \mathsf{p}(i), (t_\ell, i)\rbrace $ and $m = \lbrace t_m, \mathsf{p}(j), (t_m, j)\rbrace $ where ℓ,m∈Esp\ell, m \in \mathcal{E}_{\mathrm{sp}} be adjacent if p(i)=j\mathsf{p}(i) = j or p(j)=i\mathsf{p}(j) = i and tℓ=tmt_\ell = t_m. Then, define the set of adjacent bones

A≜{{ℓ,m}∈Esp×Esp: ℓ,m are adjacent}.\mathcal{A} \triangleq \left\lbrace \lbrace \ell, m\rbrace \in E_{\mathrm{sp}} \times E_{\mathrm{sp}} : \ \ell, m \text{ are adjacent} \right\rbrace .

Finally, let θ^s,t,ℓ,m(X)\hat{\theta}_{s,t,\ell,m}(X) denote the predicted angle between two adjacent bones in degrees at sample ss and time tt, and let θt,ℓ,m∗(Y)\theta^{\ast}_{t,\ell,m}(Y) denote the true angle between two adjacent bones in degrees at time tt. The formula for the MAE is given as

MAE(X,Y)≜min⁡s∈{1,…,S}1∣If∣∣A∣∑t∈If∑{ℓ,m}∈A∣θ^s,t,ℓ,m(X)−θt,ℓ,m∗(Y)∣.\mathrm{MAE}(X, Y) \triangleq \min_{s \in \lbrace 1,\ldots,S\rbrace } \frac{1}{\vert \mathcal{I}_f\vert \vert \mathcal{A}\vert } \sum_{t \in \mathcal{I}_f} \sum_{\lbrace \ell,m\rbrace \in \mathcal{A}} \left\vert \hat{\theta}_{s,t,\ell,m}(X) - \theta^{\ast}_{t,\ell,m}(Y) \right\vert .

The unit of the MAE is degrees. In the codebase, we implement this metric in the function metrics.mae_joint_angle. The MAE values in Tables 1, 2, and 4 are reported using the formula above.

We acknowledge that the naming of the metrics metrics.mae_joint_angle and metrics.mean_angle_error was ambigious. In the previously submitted codebase, metrics.mean_angle_error differs from the MAE defined above in that it measures the difference between the predicted and actual bone directions for the sample having the minimum average displacement error (ADE). In the revised manuscript, we now refer to metrics.mean_angle_error as the bone direction angular error (BDAE). All references to this quantity will be updated in the revised manuscript.

AuthorsOpenReview ↗

Part 2. response to W2 & Q2

We agree with the reviewer that the ablation study in the previously submitted manuscript established the performance gains attributed to attention, but did not fully isolate the potential effect of the skeletal bias. To address this concern, we have added an additional experiment which retains the same configuration and training parameters as the "Full spatial and temporal attention" configuration, but having the learned bias terms (which are 136 scalars shared by all blocks) set to zero. We refer to this configuration as "Bias removed" in the revised manuscript. All configurations reported below correspond to those in Table 3 of the previously submitted manuscript.

Table 1: Change in ADE relative to full spatial and temporal attention.

ConfigurationΔ\DeltaADE
Spatial one hop+6.9%
Spatial attention removed+14.3%%
Temporal one hop+27.4%
Temporal attention removed+52.4%%
Bias removed+0.0%

We see from the results in Table 1 that the skeletal bias provides redundant information. Each node's feature vector already contains enough information to construct its position (e.g., tree depth, number of children, end effector flag), so the added bias is not required to obtain similar results.

Next, to isolate the effect of the spherical path on our results, we trained a Euclidean flow matching model which uses bone length normalization after generation to preserve the physical geometry of the skeleton. We refer to this model as the "Euclidean FM" in the revised manuscript. The Euclidean FM model targets Euclidean velocities and shares the same source scale and NFE as the GSFM model.

Table 2 reports, alongside the ADE, the unified average displacement error (uADE) introduced in EquiFusion [2]. Let JJ denote the number of joints, let q^s,t,j∈R3\hat{q}_{s,t,j} \in \mathbb{R}^{3} denote the predicted position of joint $j \in \lbrace 1,\ldots,J\rbrace $ at time t∈Ift \in \mathcal{I}_f in sample $s \in \lbrace 1,\ldots,S\rbrace $, let qt,j∗∈R3q^{\ast}_{t,j} \in \mathbb{R}^{3} denote the ground truth position of joint jj at time tt, and let ∥⋅∥2\Vert \cdot\Vert _2 denote the Euclidean norm on R3\mathbb{R}^{3}. The uADE is defined as

uADE≜min⁡s∈{1,…,S}1∣If∣J∑t∈If∑j=1J∥q^s,t,j−qt,j∗∥2,\mathrm{uADE} \triangleq \min_{s \in \lbrace 1,\ldots,S\rbrace } \frac{1}{\vert \mathcal{I}_f\vert J} \sum_{t \in \mathcal{I}_f} \sum_{j=1}^{J} \left\Vert \hat{q}_{s,t,j} - q^{\ast}_{t,j} \right\Vert _2 ,

and it is reported in centimeters. For comparison, the conventional ADE is

ADE≜min⁡s∈{1,…,S}1∣If∣∑t∈If(∑j=1J∥q^s,t,j−qt,j∗∥22)1/2,\mathrm{ADE} \triangleq \min_{s \in \lbrace 1,\ldots,S\rbrace } \frac{1}{\vert \mathcal{I}_f\vert } \sum_{t \in \mathcal{I}_f} \left( \sum_{j=1}^{J} \left\Vert \hat{q}_{s,t,j} - q^{\ast}_{t,j} \right\Vert _2^{2} \right)^{1/2} ,

which takes a single Euclidean norm over all joints of a frame, so its value depends on the number of joints JJ. The uADE takes the Euclidean norm separately for each joint and averages over the joints, which makes it equal to the mean per joint position error of the best sample and makes it comparable across skeletons with different numbers of joints [2].

Table 2: Comparison between GSFM and the new baselines.

ConfigurationADEuADE [cm]
GSFM-deep, 3 seeds0.4768 ± 0.00079.317 ± 0.063
Bias removed0.47739.317 ± 0.048
Euclidean FM0.892212.380 ± 0.319

It is evident from above table that the spherical path in the flow matching model is a key feature which drives model performance. The Euclidean FM model has an 87% higher ADE than the GSFM model at full scale, and a 33% higher ADE than the GSFM model at ablation scale.

AuthorsOpenReview ↗

Part 3. response to Q3

In the revised manuscript, we have trained GSFM-deep under two additional seeds to demonstrate repeatability. The following results are reported using the AMASS dataset.

Table 3: Comparison between GSFM across 3 trained seeds and existing results.

ConfigurationADEFDEMAECMD
GSFM-deep, Seed 0 (submitted)0.47750.55055.8405.61
GSFM-deep, Seed 10.47700.54845.8305.45
GSFM-deep, Seed 20.47610.54785.8405.01
Mean ± SD0.4768 ± 0.00070.5489 ± 0.00145.836 ± 0.0065.36 ± 0.31
SkeletonDiffusion0.4800.5456.12411.42
EquiFusion [2]0.4960.5606.21413.10

The results in Table 3 indicate that across multi seeds, the variation in reported results is small. Specifically, we see that the standard deviation of the ADE is 0.15% of the reported mean. We also see that every trained seed outperforms SkeletonDiffusion and EquiFusion [2] in ADE, MAE, and CMD. Next, we evaluate zero-shot transfer to Human3.6M.

Table 4: Comparison between GSFM across 3 trained seeds and existing results.

ConfigurationADEFDEMAECMD
GSFM-deep, Seed 0 (submitted)0.4550.6456.1625.06
GSFM-deep, Seed 10.52600.7336.79211.96
GSFM-deep, Seed 20.4930.6896.6998.50
Mean ± SD0.491 ± 0.0350.689 ± 0.0446.551 ± 0.3408.51 ± 3.45
Best performing baseline, per column1.0940.94810.967.42

While less consistent across seeds in zero-shot transfer, the broad conclusions made with respect to Table 3 still hold. The GSFM model demonstrates a 55.1% improvement in ADE performance over the best performing baseline, and shows clear improvements in FDE and MAE across all three seeds.

We thank the reviewer for their helpful insights, and hope that the added clarification and additional experiments make the contribution of the GSFM model clear over the baselines.

References

[1] C. Curreli, D. Muhle, A. Saroha, Z. Ye, R. Marin, and D. Cremers, "Nonisotropic gaussian diffusion for realistic 3d human motion prediction," in 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2025, pp. 1871–1882.

[2] C. Curreli, F. Hofherr, D. Muhle, A. Saroha, R. Marin, and D. Cremers, "EquiFusion: Kinematics-agnostic human motion prediction via equivariant latent diffusion," in European Conference on Computer Vision (ECCV), 2026, arXiv:2607.10984.

Thank you for the rebuttal and revised manuscript. The response addresses my main questions.

At the same time, I share the concern that the specific contribution of the graph structure is not fully demonstrated in the revised manuscript, so this remains borderline for me.

Overall, I still lean slightly toward acceptance and will keep my rating at 3.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of rebuttal section

We thank Reviewers SVkr, EKoq, and ux7X for their careful reading and constructive questions. Their concerns primarily involve (i) the contribution of the graph structure and Riemannian representation, (ii) the differences between GSFM and existing methods, particularly EquiFusion, and (iii) the robustness of the reported performance. We address these shared concerns as follows.

(i) Graph structure and Riemannian representation (SVkr, EKoq, and ux7X).

We clarify how the graph structure defines the skeletal state, supplies structural conditioning, and organizes spatial and temporal interactions. The original ablation study shows that removing spatial or temporal attention increases ADE by 14.3% or 52.4%, respectively for GSFM. The additional comparisons use the "Full spatial and temporal attention" configuration and study how the skeletal relation biases and structural node features affect performance. Removing the relation biases has little effect on accuracy, and a transformer with learned joint identifiers achieves comparable accuracy on AMASS.

To address the concern about the importance of the geometric formulation, we trained a Euclidean flow matching model which normalizes bone lengths once before using the forward kinematics relation to generate output joint positions. Both GSFM and the Euclidean flow matching model preserve output bone lengths, but GSFM achieves an ADE of 0.4768 compared with 0.8922 for Euclidean flow matching. Its advantage persists at 10, 20, and 50 function evaluations (NFE). We also provide the Cartesian representation and diffusion comparisons requested by reviewer ux7X.

(ii) Differences from EquiFusion (EKoq and ux7X).

We clarify that GSFM uses a single velocity network conditioned on the observed, input motion and the supplied kinematic tree. Its state explicitly represents all future bone directions throughout sampling. Every intermediate state therefore specifies a complete candidate trajectory with prescribed bone lengths. This candidate trajectory can be reconstructed in joint space by using the forward kinematics relation. This formulation removes the requirement to separately learn a model for motion reconstruction stage, which is accomplished by an autoencoder in EquiFusion. The GSFM tied model uses only 4.47M parameters, compared with 7.9M for EquiFusion, and achieves lower AMASS ADE, FDE, MAE, and CMD. We view this comparison as a design trade-off, with GSFM retaining competitive performance in the zero-shot transfer setting on the Human3.6M dataset.

(iii) Robustness and evaluation consistency (SVkr).

We trained GSFM with two additional seeds. Across three runs, the ADE on the AMASS dataset is 0.4768 \pm 0.0007, which denotes the mean and standard deviation. Across each seed, GSFM achieves lower ADE, MAE, and CMD than the published SkeletonDiffusion and EquiFusion results. We also report the larger variation observed on zero-shot transfer to Human3.6M. Finally, we clarify the benchmark MAE definition, its units and sample selection rule, and its distinction from bone direction angular error.

Summary of revisions.

  • Discussion of EquiFusion (Section 6.2). We state that EquiFusion is accurate in zero-shot transfer and describe the trade-off with model size (GSFM-tied, 4.47M parameters). This is inherited from the design choice.
  • New Figure 3 (Section 6.2). A qualitative comparison of bone length preservation between GSFM and SkeletonDiffusion on AMASS test windows, with bones whose length deviates by more than 10% highlighted.
  • Geometric formulation (Section 6.3, Table 5; Appendix D.6, Table 8). GSFM is compared with Euclidean flow matching with normalization and with Cartesian flow matching.
  • Training seeds (Appendix D.5, Table 7). GSFM-deep is reported as mean + standard deviation over three training seeds on AMASS and in zero-shot transfer to Human3.6M.
  • Flow matching versus diffusion (Appendix D.7, Table 9). Cartesian flow matching is compared with Cartesian diffusion under velocity and sample prediction at ablation scale.
  • Sampling budget (Appendix E, Table 11). GSFM-deep and Euclidean flow matching are compared at 50, 20, and 10 NFE.

We would like to thank all the reviewers and the Area Chair (AC) once again for their valuable time and insightful feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.