MotionHiFlow: Text-to-motion via hierarchical flow matching

Abstract: Text-to-motion generation aims to generate 3D human motions that are tightly aligned with the input text while remaining physically plausible and rich in fine-grained detail. Although recent approaches can produce complex and natural movements, they usually operate at only one temporal scale, which limits both semantic alignment and temporal coherence. Inspired by the fact that complex motions are conceptualized hierarchically rather than at a single temporal scale in the human cognitive system, we propose \textit{MotionHiFlow}, a hierarchical flow matching framework to generate motion progressively by constructing flow path from low to high temporal scales. The flows at lower scales capture high-level semantics and coarse motion structures, while flows at higher scales refine temporal details. To link the flows across scales, we introduce a novel cross-scale transition process, ensuring continuity and preserving noise consistency. Furthermore, by integrating a Text-Motion Diffusion Transformer and a topology-aware Motion VAE, MotionHiFlow explicitly models structural dependencies among joints via joint-aware positional encoding and skeletal topology, enabling precise semantic alignment alongside fine-grained motion details. Extensive experiments on HumanML3D and KIT-ML benchmarks demonstrate state-of-the-art performance, with ablation studies confirming the effectiveness of the hierarchical design and key components. Code is available at https://github.com/ai-lh/MotionHiFlow.

• The paper introduces a hierarchical, multi-scale flow matching technique that decouples semantic alignment from fine motion details.
• It employs a deterministic cross-scale transition within a topology-aware Motion VAE, boosting retrieval performance with stable R-Precision.
• Empirical results on HumanML3D and KIT-ML show superior R-Precision and FID, establishing the method's effectiveness for realistic text-to-motion synthesis.

Hierarchical Flow Matching for Text-to-Motion Synthesis: MotionHiFlow

Motivation and Problem Setting

Text-to-motion generation, which targets the synthesis of 3D human motion that semantically and physically aligns with given natural language, remains a central problem for animation, VR, and robotics. Prior frameworks predominantly conduct single-scale generation, thereby entangling high-level semantic alignment with fine-grained temporal coherence within a sole modeling stage. This unified treatment limits both the semantic alignment to text and the long-range coherence of resulting motions. "MotionHiFlow: Text-to-motion via hierarchical flow matching" (2604.23264) departs from this paradigm by proposing a hierarchical, multi-scale flow matching methodology. It establishes a progressive coarse-to-fine generation schedule to decouple the modeling of semantics and motion details, integrating them via deterministic cross-scale flows.

Empirical Evidence for Hierarchy

MotionHiFlow's hierarchical design is justified by systematic empirical analysis demonstrating the preservation of semantic content even after aggressive temporal downsampling. As shown, models trained on heavily downsampled (coarse) motion maintain stable R-precision in text-to-motion retrieval, evidencing robust semantic alignment in the absence of high-frequency details. Notably, overemphasis on fine granularity can degrade text-motion consistency, supporting the hypothesis that hierarchical architectures—mirroring cognitive models of motion generation—offer theoretical and practical gains.

Figure 1: R-Precision remains stable under high temporal downsampling, indicating hierarchy preserves semantic alignment.

Methodology: MotionHiFlow Framework

MotionHiFlow operates via a hierarchical flow matching pipeline:

• Flow Matching at Multiple Scales: Motion generation is recast as a sequence of flow matching tasks operating at progressively increasing temporal resolutions. At each scale, a flow operator is learned via minimizing the squared deviation between network-predicted velocity and a target "oracle" velocity, as dictated by linear interpolation between noisy and clean latents.
• Cross-Scale Transition: Rather than naïve upsampling, a denoise–upsample–renoise process ensures noise consistency at scale transitions. This process enables the deterministic construction of the generative path as an ODE, solving for motion trajectories without intermediate noise injection.
• Latent Architecture: All flows operate in the latent space of a topology-aware Motion VAE, which encodes motion as a temporally and spatially downsampled set of latent body states.

  Figure 2: Overview of MotionHiFlow’s hierarchical multi-stage flow scheduling and cross-scale transition mechanism.

Text-Motion Diffusion Transformer and Joint RoPE

To operationalize hierarchical flow matching, the architecture incorporates two critical modules:

• Text-Motion Diffusion Transformer (TMDiT): A dual-stream transformer processes text and motion features through independent branches, employing context fusion and joint attention with partial parameter sharing. TMDiT is conditioned at both the word and sentence level, forestalling the bottleneck of collapsed global textual representations.
• Joint RoPE: Rotary positional encoding is extended to encode temporal, spatial, and kinematic tree information, imposing symmetry and bias reflecting human body structure. Each attention head splits its channels across axes for temporal index, joint 2D relative coordinates, and kinematic depth, fostering inductive alignment with skeleton-centric priors.

  Figure 3: Architectures of TMDiT and Joint RoPE, showing dual-stream design and multi-component positional encoding.

Experimental Results and Analysis

Quantitative and qualitative evaluations are conducted on HumanML3D and KIT-ML, two standard text-to-motion benchmarks. MotionHiFlow achieves:

• R-Precision: 0.563 (HumanML3D) and 0.482 (KIT-ML), exceeding prior SOTA (e.g., MoGenTS, BAMM).
• FID: 0.032 and 0.135 respectively, corresponding to the best distributional fidelity among all baselines.
• MultiModal Distance: 2.691 and 2.552, evidencing improved semantic and text-motion alignment.

The hierarchical flow matching significantly improves metrics relative to single-scale models, with ablation confirming the additive contributions of hierarchical scheduling, TMDiT, and the topology-aware Motion VAE.

Figure 4: Visualizations comparing MotionHiFlow to baselines, where correct motion directionality (green) consistently matches input text.

Figure 5: User study: MotionHiFlow preferred in realism and text alignment, often rivaling ground truth.

Implications and Future Directions

The work establishes a rigorously justified hierarchical framework, demonstrating that semantic and structural aspects of motion should be decoupled and modeled at appropriate scales. Practically, this architecture sets new performance standards for text-to-motion generation, with direct implications for interactive content creation, digital avatars, and robotics. Theoretically, it motivates further study of hierarchical, flow-matched modeling regimes for continuous, structured data.

Potential future work involves extending hierarchical flow matching to joint multimodal generation (e.g., text–audio–motion), integrating physics-based constraints for real-world deployment, and investigating richer modalities of cross-scale knowledge routing. Additionally, exploration of parameter-efficient tuning or online adaptation for downstream tasks (e.g., personalized avatars, dynamic control) is a promising trajectory.

Conclusion

MotionHiFlow (2604.23264) demonstrates that hierarchical flow matching, informed by both cognitive theory and empirical study, is highly effective for text-to-motion generation. By explicitly structuring the generation pipeline around temporally progressive flows and integrating topological priors, the system achieves superior semantic alignment, temporal coherence, and output realism compared to previous approaches. The results underline the importance of hierarchical modeling and inductive architectural bias for structured generative tasks in AI.