Research project

FlashMotion

Ultra-Fast Long-Term
3D Motion Generation

Long-horizon human motion from text, speech, or music in just a few sampling steps.

Kaixing Yang, Xulong Tang‡, Ronghui Li‡, Xiangyue Zhang, Ziqiao Peng, Mingshuang Luo, Chenxi Zhao, Haoyu Wu, Ju Fan, Hongyan Liu†, Puwei Wang†, Zijiao Zeng†, and Jun He

† Corresponding authors · ‡ Project Leader

Watch the demo Paper Code BibTeX
FlashMotion quality and throughput comparisons with conditioned motion examples for text-to-motion, speech-to-gesture, and music-to-dance

01 / Overview

Abstract

3D human motion generation plays a crucial role in game development, extended reality (XR), digital humans, and embodied simulation. Beyond motion generation, practical pipelines involve other stages, such as rigging, retargeting, and rendering, making ultra-fast long-term 3D motion generation without compromising quality particularly valuable for real-time interactive applications. Despite promising progress, high-quality long-horizon motion generation under tight inference budgets remains challenging, owing to 2 coupled bottlenecks: (1) generative strategies that require many sampling steps for high-quality 3D motion generation and (2) model architectures with substantial per-step computational and memory costs for long sequences. To address these challenges, we present FlashMotion, a co-designed framework for high-quality long-horizon 3D motion generation under a strict inference budget, covering text-to-motion, music-to-dance, and speech-to-gesture tasks. To enhance trajectory-level efficiency, we introduce a 3D motion-space MeanFlow formulation with kinematic and temporal consistency constraints for few-step generation, together with a time-decayed soft-masking mechanism for temporally coherent motion editing. To improve per-step efficiency, we design a lightweight non-autoregressive BiMamba architecture with condition-guided state selection and AdaIN-based condition fusion mechanism, enhancing cross-modal alignment and generalization to longer motion sequences. Extensive experiments demonstrate that FlashMotion achieves overall state-of-the-art performance in both quality and efficiency across text-to-motion, speech-to-gesture, and music-to-dance generation.

02 / Demo

FlashMotion in motion

The full demo follows the original sequence, with chapters for each group of examples.

Chapters4:29 total
Introduction Applications Showcase Comparisons Ablation study Motion editing

03 / Method

Few steps, efficient steps

FlashMotion co-designs the sampling trajectory and the sequence model.

01

Motion-space MeanFlow

Direct prediction of clean 3D motion induces an interval-averaged velocity. Kinematic and temporal constraints regularize the generated poses and transitions.

02

Condition-guided BiMamba

Bidirectional state-space modeling processes the sequence non-autoregressively. Condition-guided state selection and AdaIN fusion inject time-aligned conditioning.

03

Time-decayed soft masking

The soft mask gradually relaxes fixed-region constraints during sampling to improve the transition between known and synthesized motion.

Generation throughput

10,000-frame sequences

Four-step FlashMotion versus the fastest external baseline for each task. Timing covers generation only.

13.6×Text-to-Motionvs. ST2M
4.4×Speech-to-Gesturevs. CaMN
9.8×Music-to-Dancevs. FlowerDance

04 / Results

Qualitative results

Play comparisons, sampling-step showcases, ablation studies, and motion editing drawn from the demonstration video.

01

Comparisons

Four text-to-motion methods responding to the same instructionsText-to-Motion · Example 1FloodDiffusion, MotionStreamer, and FlowMDM Four methods compared on a second sequence of text promptsText-to-Motion · Example 2BAMM, MMM, and MoMask Four gesture methods compared on a shared speech clipSpeech-to-Gesture · Example 1RAG-Gesture, SemTalk, and EchoMask Four gesture methods compared on a second speech clipSpeech-to-Gesture · Example 2EMAGE, CaMN, and DisCo Four dance methods generating motions for jazz musicMusic-to-Dance · JazzFlowerDance, MEGADance, and Lodge Four dance methods generating motions for breaking musicMusic-to-Dance · BreakingFlowerDance, MEGADance, and Lodge
02

Sampling-step showcase

Text-to-motion samples with one, two, and four stepsText-to-Motion1, 2, and 4 sampling steps Speech-to-gesture samples with one, two, and four stepsSpeech-to-Gesture1, 2, and 4 sampling steps Music-to-dance samples with one, two, and four stepsMusic-to-Dance1, 2, and 4 sampling steps
03

Ablation study

Generative strategy ablation with MeanFlow, diffusion, and flow matchingGenerative strategyMeanFlow, diffusion, and flow matching Comparison of conditional training variantsTraining objective IConditional training and MeanFlow loss Kinematic and dynamic constraint ablationTraining objective IIKinematic and dynamic constraints Motion backbone ablation with BiMamba, Mamba, and TransformerMotion backboneBiMamba, Mamba, and Transformer Comparison of full, scan-only, and AdaIN-only conditioningCondition fusionState selection and AdaIN fusion
04

Motion Editing

Motion editing comparison between a soft mask and a hard mask Soft-mask motion inpaintingA time-decayed soft mask blends known and generated motion during sampling. This excerpt compares its inpainting transitions with a fixed hard mask.

05 / Applications

From motion to interaction

Generated motion drives robot simulation, XR characters, and a mobile avatar.

A humanoid robot executing generated motion in simulationRobot motion generation Animated virtual characters in an XR sceneXR character animation A spoken motion request translated to an animated mobile avatarMobile motion interaction

06 / Reference

BibTeX

@misc{yang2026flashmotion,
  title={FlashMotion: Ultra-Fast Long-Term 3D Motion Generation},
  author={Yang, Kaixing and Tang, Xulong and Li, Ronghui and Zhang, Xiangyue and
          Peng, Ziqiao and Luo, Mingshuang and Zhao, Chenxi and Wu, Haoyu and
          Fan, Ju and Liu, Hongyan and Wang, Puwei and Zeng, Zijiao and He, Jun},
  year={2026},
  eprint={2611.13626},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2611.13626}
}

FlashMotion demo

Demo excerpt