Human motion prediction requires diverse future trajectories that remain consistent with observed motion and the articulated physical structure of the body. Skeletal constraints restrict individual poses, while coordinated motion depends on spatial interactions (between connected joints) and temporal interactions (between time instants). To facilitate human motion prediction in light of these constraints and interactions, we introduce Graph Structured Flow Matching (GSFM), which transports the complete future skeletal trajectory through a single conditional velocity field. The trajectory produces 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 input bone lengths throughout generation. We train a learned velocity field through conditional flow matching along geodesic paths connecting random trajectories centered on the last-observed pose to recorded future trajectories. 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 developed 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.
Figures & tables
Figure 1: Graph structured flow matching for human motion prediction. Left: the sample with the lowest average displacement error (ADE) and two diverse samples selected by farthest point sampling among 50 generated futures for an AMASS test segment. The observed column shows the last pose of the 0.5s input window, and gray poses denote the ground truth. Right: AMASS ADE against single-instance parameter counts. Baseline architectures and their parameter counts are sourced from Curreli et al. (2026) . The GSFM-tied configuration reaches near-best precision with 4.5M parameters, while the GSFM-deep configuration matches the best published ADE.
Figure 2: Spatiotemporal skeleton graph with observed motion C and recorded future motion Y .
Prediction
Multimodal prediction
Diversity
Motion statistics
Bone length consistency
Method
ADE ↓
FDE ↓
MAE ↓
MMADE ↓
MMFDE ↓
APDE ↓
APD
CMD ↓
Str. ↓
Jit. ↓
ZeroVelocity
0.755
0.992
7.779
0.814
1.015
—
0.000
39.262
0.00
0.00
TPK
0.656
0.675
10.191
0.658
0.674
2.265
9.283
17.127
7.34
0.34
DLow
0.590
0.612
8.510
0.618
0.617
4.243
13.170
15.185
8.41
0.40
GSPS
0.563
0.613
9.045
0.609
0.633
4.678
12.465
18.404
6.65
0.29
DivSamp
0.564
0.647
8.027
0.623
0.667
15.837
24.724
50.239
11.17
0.82
Table 1: Prediction accuracy, diversity calibration, motion statistics, and bone length consistency on the AMASS dataset. Baseline results are reproduced from EquiFusion ( Curreli et al., 2026 ) . Str. and Jit. are reported as percentages, and 0.00 denotes zero at the displayed precision. Bold metric values denote the best performance in this table and following all tables.
Prediction
Multimodal prediction
Diversity
Motion statistics
Method
ADE ↓
FDE ↓
MAE ↓
MMADE ↓
MMFDE ↓
APDE ↓
APD
CMD ↓
ZeroVelocity
0.597
0.884
6.753
0.683
0.909
8.085
0.000
22.812
TPK + retarget
1.154
0.983
22.686
1.155
0.987
1.968
7.221
10.051
DLow + retarget
1.094
0.948
22.272
1.096
0.951
2.060
9.683
9.204
GSPS + retarget
1.193
1.051
11.439
1.194
1.053
2.373
9.985
7.409
BeLFusion + retarget
1.226
1.029
10.957
1.225
1.033
2.284
6.483
8.031
Table 2: Zero-shot transfer from the AMASS dataset to the Human3.6M dataset. The label “retarget” identifies learning pipelines which convert between representations for the training and target skeletons.
Prediction
Diversity
Motion statistics
Variant
ADE ↓
FDE ↓
APD
CMD ↓
Δ ADE ↓
Full spatial and temporal attention
9.474
10.33
18.4
8.6
—
Spatial one hop
10.129
11.35
18.9
15.0
6.9%
Spatial attention removed
10.824
12.65
17.2
14.6
14.3%
Temporal one hop
12.073
12.08
12.7
32.4
27.4%
Temporal attention removed
14.441
11.27
18.4
482.2
52.4%
Table 3: Contribution of spatial and temporal attention to trajectory accuracy and motion statistics on AMASS. Let Δ ADE denote the percentage increase in ADE relative to the ‘full spatial and temporal attention’ configuration.
Configuration
Prediction
Multimodal prediction
Diversity
Motion statistics
Model
Parameters (M)
σd
ADE ↓
FDE ↓
MAE ↓
MMADE ↓
MMFDE ↓
APDE ↓
APD
CMD ↓
Tied
4.47
0.5
0.484
0.550
5.917
0.582
0.595
2.638
8.554
5.246
Tied
4.47
0.7
0.484
0.551
5.930
0.583
0.598
2.683
8.528
4.776
Tied
4.47
0.9
0.492
0.558
6.020
0.588
0.601
2.213
8.890
4.903
Deep
30.49
0.3
0.485
0.553
5.856
0.582
0.597
2.822
8.263
6.521
Deep
30.49
0.5
0.477
0.549
5.815
0.577
0.596
2.876
8.187
5.943
Table 4: In-depth comparison between GSFM-deep and GSFM-tied over varying source scale. For each model, the source scale value is used during both training and generation. Parameter counts are in millions, denoted by M .
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 3: A block of the graph structured velocity network. For visual clarity, we omit residual connections and gates.
Setting
Deep
Tied
Feature width H
384
384
Block applications L
12
12
Attention heads
8
8
Time encoding dimension D
128
128
Attention and FFN parameters across depth
Distinct
Shared
Parameter count in millions
30.49
4.47
Appendix
Table 5: Structural parameters for the GSFM-deep and GSFM-tied configurations.
Sampling
Cost
Prediction
Diversity
Motion statistics
Midpoint steps
NFE
Total time (s)
Peak memory (MiB)
ADE ↓
FDE ↓
APDE ↓
CMD ↓
Inference performance at σd=0.9
25
50
15.659
3,533
0.4838
0.5522
2.513
4.964
10
20
6.261
3,533
0.4904
0.5556
2.117
4.928
Inference performance at σd=0.7
25
50
15.659
3,533
0.4775
0.5505
2.840
5.614
Appendix
Table 6: Sampling cost of GSFM-deep on AMASS for a single input observation and 50 generated predicted future trajectories having 120 frames each. Measurements use one NVIDIA B200 GPU and 32 bit floating point precision. The total time accounts for source sampling and numerical integration steps, but excludes data loading, conversion to Cartesian joint positions, and evaluation of performance metrics. Peak memory is reported in mebibytes (MiB).
Stochastic human motion prediction aims to forecast future motion distributions. Although recent studies have achieved strong performance in terms of accuracy and diversity, they often overlook plausibility (e.g., resulting in physically unrealistic predictions) and uncertainty quantification, both of which are essential for real-world applications and downstream tasks. To address these issues, we propose a latent flow-based model equipped with a data-driven Gaussian mixture prior that more effectively disentangles diverse human behaviors than conventional single-modal priors. This prior is derived from patterns in the training data without requiring additional annotations. Furthermore, the fully invertible nature of our model enables natural uncertainty quantification through tractable likelihood computation. Experiments on the Human3.6M and AMASS datasets demonstrate that our approach achieves state-of-the-art performance in both accuracy and plausibility.
Yue Ma, Frederick W. B. Li, Xiaohui Liang
Beihang University · Durham University · Zhongguancun Laboratory
3D multi-person motion prediction requires modeling both individual kinematics and inter-person interactions. While Flow Matching is effective for multi-hypothesis generation to improve prediction accuracy, directly predicting skeletal sequences from pure noise often compromises structural consistency and introduces unreliable cross-agent interactions during early noise-dominated integration steps. To address this, we propose a Prior-Guided Residual Flow Matching framework. First, a Deterministic Coarse Prior (DCP) establishes a kinematic anchor, formulating the generative process as a conditional flow over motion residuals to simplify the generative objective and preserve structural stability. Second, a Dynamic Cross-Interaction (DCI) mechanism temporally synchronizes inter-agent message-passing with the integration progress, ensuring the extraction of reliable social contexts and improving multi-person motion fidelity. Finally, a decoupled joint-motion architecture with bidirectional fusion effectively preserves fine-grained kinematic coherence. Extensive experiments demonstrate that our approach achieves state-of-the-art prediction accuracy across multiple datasets. Code is available at https://github.com/Wei-Wei-a/Residual-Flow-Matching-with-Dynamic-Cross-Interaction-for-3D-Multi-Person-Motion-Prediction.
Human motion prediction combines the tasks of trajectory forecasting and human pose prediction. For each of the two tasks, specialized models have been developed. Combining these models for holistic human motion prediction is non-trivial, and recent methods have struggled to compete on established benchmarks for individual tasks. To address this, we propose a simple yet effective transformer-based model for human motion prediction. The model employs a stack of self-attention modules to effectively capture both spatial dependencies within a pose and temporal relationships across a motion sequence. This simple, streamlined, end-to-end model is sufficiently versatile to handle pose-only, trajectory-only, and combined prediction tasks without task-specific modifications. We demonstrate that this approach achieves state-of-the-art results across all tasks through extensive experiments on a wide range of benchmark datasets, including Human3.6M, AMASS, ETH-UCY, and 3DPW.