SOICT 2026 · Under Review
Iterative text-to-motion generation delivers high-quality and semantically aligned motions but requires multiple network evaluations, resulting in substantial inference latency. We present MixiMotion, a strict one-step text-to-motion generation framework based on offline set distillation. Instead of distilling a single teacher trajectory for each text prompt, MixiMotion constructs an offline bank of multiple teacher motions and aligns teacher and student sample sets through asymmetric bidirectional matching. The teacher-to-student direction promotes coverage of diverse teacher-supported motions, while the student-to-teacher direction suppresses unsupported generations. We further introduce differentiable decoded-space kinematic supervision to complement normalized representation matching with constraints in the decoded motion space. At inference, MixiMotion generates a complete motion sequence with a single network evaluation, without teacher queries, iterative sampling, or candidate ranking. On ViMoGen, MixiMotion achieves a semantic alignment score of 0.835, outperforming the evaluated one-step baselines and approaching the 0.858 score of its 50-step HY-Motion-1.0-Lite teacher. In blinded human evaluation, MixiMotion obtains an overall rating of 4.33, compared with 4.50 for the teacher, while outperforming the evaluated one-/few-step baselines. Meanwhile, generation latency is reduced from 829.58 ms to 9.30 ms, corresponding to an \(89.2\times\) speedup.
A frozen multi-step teacher is used offline to construct a multi-candidate motion bank for each text prompt, while the one-step student generates a set of motion candidates from independently sampled noise. Training aligns the two sets using asymmetric bidirectional matching in both representation and decoded motion spaces, with gradients propagated only through the student.
Randomly sampled evaluation clips under a shared renderer, camera, and lighting. Not selected by output quality.
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-1.0-Lite |
|---|---|---|---|---|
| "A person walks forward." | ||||
| "A person shoots a basketball." |
Question-level correctness on ViMoGen. Overall pools all judgments across 2{,}000 prompts rather than averaging category scores. MixiMotion versus the 50-NFE HY-Motion-1.0-Lite teacher is renderer-controlled.
| Method | NFE ↓ | Daily ↑ | Fit. ↑ | Game ↑ | Loco. ↑ | Social ↑ | Sports ↑ | Overall ↑ |
|---|---|---|---|---|---|---|---|---|
| One-step methods | ||||||||
| MixiMotion (Ours) | 1 | 0.842 | 0.721 | 0.886 | 0.944 | 0.903 | 0.806 | 0.835 |
| MotionLCM-1 | 1 | 0.801 | 0.709 | 0.881 | 0.951 | 0.911 | 0.772 | 0.825 |
| Naive 1-step teacher | 1 | 0.681 | 0.604 | 0.743 | 0.823 | 0.768 | 0.642 | 0.711 |
| Multi-step methods | ||||||||
| HY-Motion-1.0-Lite (teacher) | 50 | 0.874 | 0.738 | 0.901 | 0.958 | 0.918 | 0.823 | 0.858 |
| MLD | 50 | 0.811 | 0.708 | 0.896 | 0.958 | 0.931 | 0.780 | 0.840 |
| MDM | 50 | 0.820 | 0.724 | 0.865 | 0.950 | 0.927 | 0.752 | 0.833 |
| MotionDiffuse | 1000 | 0.782 | 0.689 | 0.855 | 0.926 | 0.899 | 0.742 | 0.809 |
MixiMotion is best among one-step methods (0.835 vs 0.825 MotionLCM-1 and 0.711 naive 1-step). It remains 2.3 points below the 50-step teacher (0.858).
| Method | Params (M) ↓ | NFE ↓ | Gen. ms ↓ | E2E ms ↓ |
|---|---|---|---|---|
| HY-Motion-1.0-Lite | 460.0 | 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 versus the 50-NFE HY-Motion-1.0-Lite teacher (\(829.58\) ms \(\rightarrow\) \(9.30\) ms). MotionLCM-1 remains smaller and faster.
MixiMotion reuses the teacher backbone, so this is not an architecture result. All systems generate fixed 120-frame clips. Semantic score is 0.835 versus 0.858 for the 50-step teacher; human overall is 4.33 versus 4.50. Nearest-neighbor set matching does not recover the teacher distribution. Variable-length generation is untested.
Independently sourced prompts, a common cross-family renderer, and variable-length generation would test whether the one-step quality–latency trade-off generalizes beyond this teacher family.
@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}
}