FlashMotion: Ultra-Fast Long-Term 3D Motion Generation

AI summary

Overview

Research area: 3D human motion generation (computer vision / character animation), spanning text-to-motion, music-to-dance, and speech-to-gesture synthesis.

Technical level: Advanced. The paper combines flow-matching theory (MeanFlow identities, JVP-based training), selective state-space models (Mamba/BiMamba), and SMPL/SMPL-X motion representations. The high-level argument is accessible, but the method sections assume familiarity with diffusion/flow models and sequence backbones.

Scope (one sentence): The paper proposes FlashMotion, a roughly 60M-parameter framework that generates long, high-quality 3D human motion in 1–4 sampling steps at very high throughput across three conditional motion tasks.

What This Paper Is About

Practical 3D motion pipelines do not stop at generation: rigging, retargeting, and rendering all consume part of the same latency budget, so generation must be extremely fast to leave room for them. Existing methods are slow for two coupled reasons: high-quality generative strategies need many sampling steps, and the model architectures are expensive per step on long sequences. FlashMotion attacks both costs at once, pairing a few-step generative formulation with a lightweight non-autoregressive backbone so that long motion sequences can be produced in real time without sacrificing physical plausibility or semantic alignment.

Key Contributions

1. A co-designed framework for strict inference budgets. FlashMotion jointly optimizes trajectory-level efficiency (how many sampling steps) and per-step efficiency (cost of each step), and is applied to text-to-motion, speech-to-gesture, and music-to-dance with a separate model per task.

2. A 3D motion-space MeanFlow formulation with motion-specific constraints. Instead of predicting an interval-averaged velocity directly, the model predicts the clean motion and derives the velocity from it, which allows direct supervision in motion space: a reconstruction loss, a forward-kinematics-based kinematic loss on joint positions, and a dynamic loss on joint velocities and accelerations.

3. Adaptive Soft masking for training-free motion editing. A soft, time-varying mask replaces fixed hard masks during constraint-aware sampling, relaxing constraints gradually to reduce boundary discontinuities in the few-step regime. The paper states that 1-step editing is not supported; editing uses multiple sampling steps.

4. A lightweight non-autoregressive BiMamba architecture. The backbone uses bidirectional Mamba with linear-time complexity, condition-guided state selection (CoBiMamba) that lets condition features modulate the SSM's input-dependent parameters, and AdaIN-based channel-level condition fusion (CoFusion) with frame-wise projections rather than pairwise cross-attention.

The paper also states it extends the authors' prior FlowerDance (ECCV'26) in four respects: task coverage, generative strategy (clean-motion prediction), model architecture, and experiments.

Main Findings

• Text-to-motion on BABEL (long sequences, Mixed set of 311 validation sequences of 201–1030 frames): FlashMotion reports the best generated results in Table I. The 4-step variant reaches R@1 29.134, R@2 42.829, R@3 51.683, MM 4.340, FID 0.907, ΔPJ 0.00027, and ΔAUJ 0.01295; ground truth is R@1 28.504, MM 4.316, with ΔPJ and ΔAUJ at 0.00000. The 1-step variant reaches R@1 26.371, FID 1.630, ΔPJ 0.00663, ΔAUJ 0.21474. The strongest baseline FID in the table is FloodDiffusion at 0.542, and the best baseline R@1 is MMM at 26.936.

• Text-to-motion throughput: On 10,000-frame outputs at batch size 1, FlashMotion reports 173,172 FPS (1-step) and 46,303 FPS (4-step); at 1,000 frames it reports 29,767 and 10,711 FPS. For comparison, the table lists MoMask at 1,014 FPS@10K, MMM at 1,350, ST2M at 3,417, DART at 75, MotionStreamer at 10, and FloodDiffusion at 35. (ST2M is the fastest baseline in this table.)

• Music-to-dance on FineDance (7.7-hour subset, 84 test clips from 19 recordings): The 4-step model reports FID_k 15.15, FID_g 21.13, DIV_k 9.55, DIV_g 6.58, BAS 0.240, and 40,545 FPS@10K; the 1-step model reports FID_k 18.32, FID_g 28.48, DIV_k 9.48, DIV_g 7.54, BAS 0.237, and 159,800 FPS@10K. The closest prior few-step method, FlowerDance, reports FID_k 29.73, FID_g 19.59, BAS 0.232, and 4,145 FPS@10K. Ground-truth references are DIV_k 9.94, DIV_g 7.54, BAS 0.201. Note that the 4-step variant has a higher FID_g than FlowerDance (21.13 vs 19.59) while improving FID_k and throughput.

• Speech-to-gesture on BEAT2 (265 test recordings, official English split): The 4-step model reports FGD 0.3242, BC 0.6006, DIV 9.1950, and 34,655 FPS@10K; the 1-step model reports FGD 0.4756, BC 0.4801, DIV 6.8688, and 135,179 FPS@10K. The best baseline FGD is SemTalk at 0.3560 (358 FPS), with EchoMask at 0.5660 (510 FPS) and EMAGE at 0.6920 (1,839 FPS). Ground truth BC is 0.4445 and DIV is 8.3714; FlashMotion's 4-step BC exceeds the ground-truth value reported in the table.

• Compact model size: The paper describes the model as approximately 60M parameters and frames the results as leading throughput at that size.

• Ablation signal on the training objective: The paper states that training with the MeanFlow objective alone "remains unstable and fails to converge" in their SMPL-based experiments, citing Sec. IV-E for corroboration; the detailed ablation tables are not included in the provided content, so specific ablation numbers are not reported here.

• Editing behavior: Motion editing is described as training-free and constraint-based (joint-wise or temporal constraints), supporting human-in-the-loop authoring and streaming via temporal inpainting, but it requires multiple sampling steps. Quantitative editing results are referenced (Fig. 6) rather than reported numerically in the provided text.

Methodology in Plain English

The problem framing. The authors argue end-to-end speed is the product of two things: the number of sampling steps and the cost of each step. Making only one of them cheap gives limited real speedup, so they optimize both.

Step 1 — Fewer steps by predicting clean motion. Flow-matching models learn an instantaneous velocity field and integrate it, which needs many small steps because real trajectories curve. MeanFlow instead learns the average velocity over a time interval, matching what a large-step solver actually does. The authors' twist is to have the network predict the clean motion for the noisy input and a chosen time interval, then compute the interval-averaged velocity algebraically from that prediction. This matters because the prediction can be supervised directly as motion: a reconstruction loss to the ground-truth motion, a loss on forward-kinematics joint positions, and a loss on joint velocities and accelerations. Their stated motivation is that SMPL pose parameters have explicit frame-to-frame continuity but only implicit intra-frame joint coupling, so large-step errors show up as jitter, root drift, foot sliding, and implausible articulation. At inference, each step predicts clean motion and interpolates between it and the current noisy state — no explicit velocity computation or noise injection.

Step 2 — Cheaper steps with a bidirectional Mamba backbone. Human motion needs strong local temporal continuity, and Mamba's sequential inductive bias suits that while costing linear rather than quadratic time. The model runs forward and backward Mamba branches, fuses them additively, and refines with a multiplicative skip connection. In the "CoBiMamba" variant, condition features are injected into the state-space parameters themselves — governing how much memory is retained, how much new information is written, and how the state is read out — so motion updates follow changing action instructions or rhythmic cues. Conditions are otherwise fused by AdaIN-style frame-wise scaling and shifting, which has no length-specific parameters and avoids pairwise cross-attention matrices, letting the model generate sequences longer than those seen in training in one non-autoregressive pass.

Step 3 — Editing with a soft mask. For constrained editing, known regions are blended into the sample using a mask multiplied by the current time, so constraints loosen as sampling proceeds, smoothing the seam between constrained and generated motion.

Setup details. Text uses UMT5-XXL features (4096-dim) with 263-dim HumanML3D motion at 20 FPS; speech uses 1024-dim HuBERT plus 300-dim text features with 333-dim motion (55 SMPL-X joints, 6D rotations, plus 3 root components) at 30 FPS; music uses 35-dim Librosa features with 147-dim motion (3D root plus 24-joint 6D rotations) at 30 FPS. Evaluation metrics are R@1/2/3, MM, FID, ΔDiv, ΔPJ, ΔAUJ (text); FID_k, FID_g, DIV_k, DIV_g, BAS (dance); FGD, BC, DIV (gesture). Timing uses an NVIDIA RTX 3090 with an Intel Xeon Gold 5218 (2.30 GHz), batch size 1, lengths of 1k or 10k frames, averaged over 5 GPU-synchronized runs after one warm-up, excluding external modules.

Why This Matters

For research: The paper reframes efficiency as a two-factor problem rather than a single knob, and shows that reformulating a flow-matching objective in the native representation space of the data (3D motion rather than generic velocity) enables domain-specific physical constraints that generic formulations cannot easily impose. It also extends the small but growing line of work on few-step flow methods for motion, which the related-work section notes has been stuck at roughly 50 steps (FlowMotion, Match-TTSG, UMF's prior stage) or 10 steps (Motion Flow Matching, FloodDiffusion, FlowerDance).

Real-world applications (as identified by the paper):

• Game development and interactive non-player characters that respond to player input or incoming conditions in real time.
• Extended reality (XR) character animation and mobile motion interaction.
• Digital humans and conversational avatars, including streaming gesture generation where each new window is conditioned on the previous window's tail.
• Embodied simulation and robot motion generation.
• Human-in-the-loop content authoring, where users fill temporal gaps between dance clips or retarget selected hand trajectories.

Industry relevance: The work was conducted at Tencent IEG Motus Group and is explicitly framed around production pipelines — rigging, retargeting, and rendering — rather than generation in isolation. The argument is that generation must be cheap enough to leave compute for the rest of the pipeline; the reported throughput on 10,000-frame outputs is far above every baseline in the three comparison tables, which is the relevant claim for deployment rather than benchmark scores alone.

Future Directions

• One-step editing. The paper explicitly states that 1-step editing is not supported, so extending the Adaptive Soft mask to the single-step regime — or characterizing why it fails — is an open problem.

• Closing the residual quality gap at four steps. On FineDance, the 4-step model's FID_g (21.13) is worse than FlowerDance's (19.59) even as FID_k improves, and on BABEL the gap to ground-truth semantic alignment (R@1 29.134 vs 28.504 ground truth, but MM 4.340 vs 4.316) is small but not closed; how quality scales between 1 and 4 steps is only partly characterized in the provided content.

• Robustness of the training objective. Since MeanFlow alone failed to converge in their SMPL-based experiments, the relative weightings of the reconstruction, kinematic, and dynamic losses appear to carry real burden. The provided content does not report the ablation values, leaving the sensitivity to these weights an open question.

• Generalization beyond the evaluated tasks and datasets. The paper claims length-agnostic generation from short training windows, but the provided content reports evaluations on BABEL, FineDance, and BEAT2 only; behavior on other motion representations, unseen skeleton topologies, or interaction/group settings is not established here.

• Streaming and long-horizon drift. Temporal inpainting is proposed for streaming, but sustained error accumulation over very long sequences is precisely the failure mode the kinematic and dynamic losses target; a systematic long-duration study is not reported in the provided text.

Target Audience

Researchers and engineers working on generative human motion, character animation, and real-time avatars will get the most from this paper, particularly those already familiar with diffusion or flow matching and with Mamba/state-space backbones. It is also relevant to graphics and game-industry practitioners who care about latency budgets shared across generation, rigging, retargeting, and rendering, and to readers tracking few-step generative modeling who want a case study in moving a flow-matching formulation into a structured, physically constrained representation space. Readers new to the area should expect to consult the cited background (Flow Matching, MeanFlow, Mamba, SMPL/SMPL-X) to follow the methodology sections.

Authors’ 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 include rigging, retargeting, and rendering, making ultra-fast long-term 3D motion generation without compromising quality valuable for real-time interactive applications. Despite promising progress, high-quality long-horizon motion generation under tight inference budgets remains challenging, owing…