SOICT 2026 · Under Review

MixiMotion: One-Step Text-to-Motion Generation via Asymmetric Set Distillation

Hung Dinh1, Binh Mai1, Tran Quoc Bao Le1, Lam Nguyen1, Cong Tran1†
1 Posts and Telecommunications Institute of Technology, Vietnam
† Corresponding author
Paper Code Live Demo Human Eval

Abstract

Diffusion and flow-matching models synthesize high-quality 3D human motion from text, yet iterative samplers require tens of network evaluations. Collapsing the sampler to one step is not a scheduling problem: one prompt admits many valid motions, so regression onto a single teacher sample averages modes, and latent residuals still decode into skating or root drift. We present MixiMotion, a strict one-step generator distilled from an offline teacher cache. A frozen multi-step model is run once per prompt to store several motion candidates; the student maps independent noises to one-step motions and is trained by asymmetric bidirectional set matching plus decoded kinematic constraints. The teacher is never queried during optimization. On ViMoGen, MixiMotion scores 0.880 versus 0.886 for the 50-step teacher under a shared renderer, at one network evaluation and 9.3 ms versus 830 ms on an RTX 5090.

Method Overview

MixiMotion distills a frozen multi-step text-to-motion teacher into a strict one-step student using only an offline cache. Any multi-step generator that emits a motion sequence can serve as \(G_T\); any one-step map in the same output space can serve as \(G_\theta\). For each prompt we store \(K\) teacher motions, then train the student with asymmetric bidirectional set matching \(\mathcal{B}_{\boldsymbol{\alpha}}\) between teacher and student sets, plus decoded kinematic maps \(\Phi_r\) (contact, stride, root) after a frozen body decoder. Matching from teacher to student covers every cached target; the reverse direction penalizes unsupported samples. Optional staged reweighting and a heuristic ranking loss can be layered on the same bank. At inference the student uses one network evaluation.

Overview of the MixiMotion distillation framework
Figure 1: MixiMotion overview. A frozen multi-step teacher produces \(K\) cached targets per prompt once. The one-step student generates \(M\) candidates from independent noise and is optimized by latent and decoded-motion set matching; only the student receives gradients.
Staged training schedule for MixiMotion
Figure 2: One staged schedule used in our run. Arrows encode checkpoint order and do not imply monotonic metric improvement.

Motion Samples

Selected successful clips under a shared renderer, camera, and lighting. These examples illustrate temporal progression and are not a random sample.

HY-Motion-1.0 full (50 NFE)
→
MixiMotion (1 NFE) · “A person walks forward.”
HY-Motion-1.0 full (50 NFE)
→
MixiMotion (1 NFE) · “A person shoots a basketball.”
HY-Motion-1.0 full (50 NFE)
→
MixiMotion (1 NFE) · “A person waves.”
Qualitative frame comparison
Figure 3: Six temporally ordered frames from each 120-frame clip. Upper row: HY-Motion-1.0 full. Lower row: MixiMotion.

Comparison with Baselines

Compare MixiMotion with compact one- and few-step baselines on the same prompts. Cross-family clips keep native renderers.

Prompt MotionLCM-1 MotionHiFlow MixiMotion (Ours) HY-Motion full
"A person walks forward."
"A person shoots a basketball."

Experimental Results

ViMoGen Semantic Alignment

MixiMotion versus 50-NFE HY-Motion-1.0 full is renderer-controlled. Other rows keep native renderers and are contextual. The paired MixiMotion−HY-full gap is \(-0.0055\) with 95% CI \([-0.0121, 0.0011]\).

Method NFE ↓ Daily ↑ Fit. ↑ Game ↑ Loco. ↑ Social ↑ Sports ↑ Overall ↑
One-step and few-step
MixiMotion (Ours) 1 0.9009 0.7487 0.9205 0.9757 0.9332 0.8430 0.8802
MotionLCM-1 1 0.8065 0.7213 0.8977 0.9621 0.9211 0.7807 0.8415
MotionHiFlow 14 0.8617 0.7717 0.9187 0.9689 0.9452 0.8274 0.8764
Multi-step reference
HY-Motion-1.0 full 50 0.9075 0.7613 0.9215 0.9757 0.9419 0.8451 0.8856
MLD 50 0.8112 0.7079 0.8959 0.9582 0.9310 0.7796 0.8398
MDM 50 0.8196 0.7235 0.8648 0.9504 0.9266 0.7516 0.8331

Efficiency (RTX 5090, BF16)

Method Params (M) ↓ NFE ↓ Gen. ms ↓ E2E ms ↓
HY-Motion-1.0 full 1042.9 50 829.58 968.25
MixiMotion (Ours) 460.0 1 9.30 143.55
MotionHiFlow 33.1 14 77.32 83.80
MotionLCM-1 26.5 1 3.04 13.94

The \(89.2\times\) generation speedup is only versus 50-NFE HY-Motion full under matched-family timing. MotionLCM-1 remains smaller and faster.

Limitations and Future Work

Limitations

MixiMotion generates fixed 120-frame clips. SSAE is a same-family protocol with one VLM judge; only the HY-Motion comparison controls the renderer. The student reuses the teacher backbone, so this is not an architecture result. MBench generalizability (0.24 vs. 0.40 for HY-Motion full) shows that one-step set matching does not recover the teacher distribution. Staged training is non-monotonic on SSAE and skating.

Future Work

Independently sourced prompts, a common cross-family renderer, component ablations on the filtered checkpoint, and variable-length generation would test whether the one-step quality–latency trade-off generalizes beyond this teacher family.

Citation

@inproceedings{dinh2026miximotion,
  title={MixiMotion: One-Step Text-to-Motion Generation via Asymmetric Set Distillation},
  author={Hung Dinh and Binh Mai and Tran Quoc Bao Le and Lam Nguyen and Cong Tran},
  booktitle={SOICT},
  year={2026}
}