TACD: Distilling Efficient Text-to-Motion Models via Terminal Amplification Control

Wei-Jin Huang; Yuan-Ming Li; Kun-Yu Lin; Wang Luo; Yinlin Zhu; Yue Yu; Shenghao Ye; Junbin Yuan

cs.AIOct 2, 2026

TACD: Distilling Efficient Text-to-Motion Models via Terminal Amplification Control

Authors: Wei-Jin Huang, Yuan-Ming Li, Kun-Yu Lin, Wang Luo, Yinlin Zhu, Yue Yu, Shenghao Ye, Junbin Yuan,

Organizations: Sun Yat-sen University · Nanyang Technological University · Wuhan University · University of Science and Technology of China ·

Abstract

Recent text-to-motion models have improved motion quality and instruction following, yet many-step denoising and large model components make deployment slow and memory-intensive. We present Terminal-Amplification-Controlled Distillation (TACD), an on-policy approach for training efficient motion generators from text prompts and pretrained teachers, without real-motion training data. Building on segmented on-policy flow distillation, we supervise clean-motion predictions along student-generated trajectories. We identify a failure mode in which velocity matching on a fixed supervision grid repeatedly overweights errors near the denoising endpoint, degrading few-step generation. TACD ties the latest teacher query to the student's step size, bounding the effective loss weights in clean-motion space without changing inference. Experiments on HumanML3D and KIT-ML demonstrate improved few-step generation, including a 58% reduction in eight-step HY-Motion student FID relative to distillation without this bound. For diffusion teachers, the endpoint-matching form of TACD yields four-step students with lower FID and matched or improved text-motion retrieval relative to their 50-step teachers on HumanML3D. On HY-Motion and Kimodo, eight-step students with compact components achieve 7.7-11.9x end-to-end speedups and reduce peak GPU memory by 3.8-6.7x relative to their teachers. Project page: https://vkgo.github.io/TACD/

Figures & tables

Figure 1: Motion examples. Teacher, the same teacher truncated to fewer steps, and TACD. Rows share prompt, noise, frames, and camera ( Section K.1 ).

Figure 2: Controlling terminal amplification improves few-step motion distillation. (a) The known weight

(1-s)^{-2}

on our supervision grid. (b) Bounding it improves FID; hollow bars enforce the same bound on the original grid (dots: seeds).

Figure 3: Flow training and terminal amplification control. The rollout EMA generates detached supervision states, while gradients update the online student. TACD moves the supervision times

s_{k,j}

so that every coefficient is bounded; endpoint matching (

\ell_{x}

) enforces the same bound at the original queries (teal marks what each changes). Both preserve the inference schedule.

HY-Motion flow teacher; students at eight steps, three-seed means
Method	NFE	R@1 $\uparrow$	R@2 $\uparrow$	R@3 $\uparrow$	MM-Dist $\downarrow$	FID $\downarrow$	Div $\rightarrow$	CLIP $\uparrow$
Real motions	–	.505	.702	.801	3.241	–	10.33	.638
HY-Motion teacher	50	.415	.601	.708	4.056	3.119	9.47	.604
HY-Motion Lite (released)	50	.394	.578	.685	4.243	3.741	9.29	.589
Truncated teacher	8	.369	.544	.646	4.721	6.816	8.92	.510
Uncapped student (DX, $N{=}1$ )	8	.343	.528	.647	4.562	7.559	8.24	.569

Table 2: MARDM evaluator ( Meng et al., 2025b ) : TACD leads the truncated teacher and every eight-step student in R@

k

, MM-Dist, FID, and CLIP score; four-step MLD TACD matches real motions in R-precision.

MLD teacher ( Chen et al., 2023 ) and MotionLCM-V1 ( Dai et al., 2024b )
Method	Training signal	NFE	R@1 $\uparrow$	R@3 $\uparrow$	MM-Dist $\downarrow$	FID $\downarrow$	Div $\rightarrow$	MMod $\uparrow$
MLD teacher (DDIM)	Reference	50	.502	.795	3.048	.387	9.53	–
Truncated MLD teacher	None	4	.459	.768	3.295	.585	9.49	–
MotionLCM-V1	Real motions	4	–	.798	–	.286	–	–
TACD	Teacher outputs	4	.525	.818	2.909	.222	9.80	–
MotionLCM-V2 teacher and student ( Dai et al., 2024a )

Table 6: Diffusion teachers on HumanML3D. Four-step TACD has lower FID than its 50-step teacher and the released MotionLCM student (our evaluation); bold: best at four steps in each group.

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Purpose	Student / encoder	$K$	Loss; sampling grid; margin $\delta$	Replication
Flow controls	1B LoRA-32 / 8B	4	Uncapped/grid cap/ last-point; uniform; $.001/.125$	3 each
Lite controls	Lite full / 0.6B	8	Original/grid cap/ endpoint/last-point; uniform; $.001/.0625$	3 each
Coefficient interventions	Lite full / 0.6B	8	Weight cap on original grid; $C{=}10^{6}$ on capped grid	3 each
Gradient probe	Lite full / 0.6B	8	Uncapped/grid cap; uniform; $.001/.0625$	32 batches $\times$ 2 probe seeds
DDIM transfer	Teacher-size / native	4	Endpoint matching; native DDIM grid	MLD/V2: 3

Appendix

Table 7: Experimental configurations. HY-Motion 1B is the teacher for the first four rows; replication counts denote training seeds except for gradient probes; text encoders remain frozen. Teal rows contain the TACD recipes: grid capping for the flow teacher, endpoint matching for diffusion teachers.

Model	NFE	CFG	FID $\downarrow$	R@3 $\uparrow$	MM-Dist $\downarrow$
Teacher (1B), 50 steps	50	2.0	0.9937	0.7052	3.7308
Teacher (1B), 50 steps	50	3.0	0.6989	0.7312	3.5619
Teacher (1B), 50 steps	50	5.0	0.5407	0.7431	3.4706
Teacher (1B), 50 steps	50	7.5	0.5487	0.7408	3.4893
Teacher (1B), 8 steps (truncated)	8	5.0	3.2503	0.6516	4.2250
Student (0.46B), uncapped (s42)	8	5.0	2.5631	0.7056	3.7131

Appendix

Table 8: Retrieval–quality trade-off: eight-step Lite vs. the 50-step HY-Motion teacher. Student rows: Qwen3-0.6B encoder, full 0.46B Lite backbone; teacher rows: its 8B encoder. Mean rows average three seeds; CFG

5.0

is the teacher’s best FID and R@3 of four tested. Rose rows: terminal coefficient

10^{6}

Method	NFE	FID $\downarrow$	R@3 $\uparrow$	Div $\rightarrow$
Teacher (MLD, 50)	50	0.387	0.795	9.53
Truncated teacher	4	0.585	0.768	9.49
MotionLCM-V1 (consistency)	4	0.286	0.798	–
TACD	4	0.222 $\pm$ .002	0.818	9.80

Appendix

Table 9: MotionLCM comparison : (a) the MLD teacher, (b) the MotionLCM-V2 teacher. MotionLCM rows: released checkpoints, our harness. TACD rows: three-seed mean

\pm

std.

(a) Published generators (real-motion training)
Method	NFE	R@1 $\uparrow$	R@3 $\uparrow$	MM-Dist $\downarrow$	FID $\downarrow$	Div $\rightarrow$
Real motions ( Guo et al., 2022 )	–	.424	.779	2.788	.031	11.08
T2M ( Guo et al., 2022 )	–	.361	.681	3.488	3.022	10.72
MotionDiffuse ( Zhang et al., 2024 )	–	.417	.739	2.958	1.954	11.10
MDM † ( Tevet et al., 2023 )	$\sim$ 1000	.404	.731	3.096	.513	10.73
MLD ( Chen et al., 2023 )	50	.390	.734	3.204	.404	10.80

Appendix

Table 10: KIT-ML, MDM teacher. TACD beats the truncated teacher at every NFE in R@1, R@3, MM-Dist, and FID; TACD rows: three-seed means after 10k updates; intervals in Table 11 . † Post-fix evaluation endorsed by the MDM authors (repository issue #211).

Method	NFE	R@1 $\uparrow$	R@3 $\uparrow$	MM-Dist $\downarrow$	FID $\downarrow$	Div $\rightarrow$	FID by seed
Real motions (GT)	–	.426	.780 $\pm$ .005	2.786	.027 $\pm$ .004	10.99	–
MDM teacher	1000	.403	.725 $\pm$ .005	3.115	.506 $\pm$ .023	10.74	–
Truncated teacher	2	.395	.728 $\pm$ .005	3.193	1.530 $\pm$ .063	11.24	–
TACD	2	.410	.756	2.996	1.038	11.06	1.020 / 1.077 / 1.017
Truncated teacher	4	.410	.755 $\pm$ .006	3.014	.976 $\pm$ .064	11.38	–
TACD	4	.416	.761	2.932	.852	11.16	.840 / .857 / .859

Appendix

Table 11: KIT-ML measured rows with evaluator intervals (the rows of Table 10 ; Guo evaluator, 20 repetitions).

\pm

: 95% interval over the 20 repetitions. TACD rows average seeds

42/43/44

; the last column lists their per-seed FID.

Method	NFE	R@1 (%) $\uparrow$	R@3 (%) $\uparrow$	FID $\downarrow$	GT-R@1 (%) $\uparrow$	Lat.(s) $\downarrow$	VRAM
Teacher	100	73.15	88.63	0.045	75.66	2.96	15.3G
Truncated teacher	8	69.36	84.35	0.064	72.66	0.29	15.3G
Truncated teacher	4	67.05	82.37	0.079	70.55	0.18	15.3G
Truncated teacher	2	60.46	77.27	0.107	63.14	0.13	15.3G
TACD, 8B enc.	8	71.39 $\pm$ .50	86.55 $\pm$ .25	0.054	75.01 $\pm$ .37	0.29	15.3G
TACD, 8B enc.	4	68.42 $\pm$ .69	83.25 $\pm$ .38	0.067	72.13 $\pm$ .77	0.18	15.3G

Appendix

Table 12: Kimodo results (607-caption TMR evaluation; GT-R@1: R@1 restricted to the 567 prompts that have a ground-truth motion).

\pm

: SD over three seeds (

K{=}2

: two). TACD rows: teacher’s 8B encoder. Lat./VRAM: end to end, one RTX L40.

Method	Encoder	NFE	R@1 (%) $\uparrow$	R@3 (%) $\uparrow$	FID $\downarrow$	Time (s) $\downarrow$	Memory (GB) $\downarrow$
Teacher	8B	100	73.15	88.63	0.045	2.96	15.3
Truncated teacher	8B	8	69.36	84.35	0.064	0.29	15.3
TACD	8B	8	71.39	86.55	0.054	0.29	15.3
TACD (Kimodo)	1B	8	68.86	86.66	0.054	0.25	4.0
Truncated teacher	8B	4	67.05	82.37	0.079	0.18	15.3
TACD	8B	4	68.42	83.25	0.067	0.18	15.3

Appendix

Table 13: Kimodo teacher (607-caption TMR evaluation). TACD with the 8B encoder beats the truncated teacher at every NFE; at NFE 8, TACD (Kimodo) with a 1B text encoder keeps R@3 and FID with

3.8\times

less memory (paired runs in Table 14 ).

Conditioning encoder	NFE	R@1 (%) $\uparrow$	R@3 (%) $\uparrow$	FID $\downarrow$	GT-R@1 (%) $\uparrow$
teacher’s own 8B encoder	8	71.72 $\pm$ .50	86.66 $\pm$ .33	0.0542 $\pm$ .0002	75.25 $\pm$ .10
1B encoder, trained in	8	68.86 $\pm$ .44	86.66 $\pm$ .28	0.0544 $\pm$ .0003	72.60 $\pm$ .67
paired $\Delta$ (1B $-$ 8B)	8	$-2.86$ [ $-3.30$ , $-2.47$ ]	$-0.00$ [ $-0.66$ , $+0.49$ ]	$+0.0002$ [ $-0.0001$ , $+0.0004$ ]	$-2.65$ [ $-3.18$ , $-1.94$ ]

Appendix

Table 14: Training on the 1B text encoder vs. the teacher’s own 8B encoder (Kimodo, 607-prompt bench,

K{=}8

; the two conditions share seeds

42/43/44

\pm

: SD over three seeds; brackets on the

\Delta

rows: range of the three paired differences.

Table 16

Table 17

Figure 7: Eight-step text-to-motion with TACD (I). One prompt per row; each row is a single eight-step sample of TACD (HY-Motion), the Qwen3-0.6B/Lite student, shown at evenly spaced time points ( Section K.2 ).

Figure 8: Eight-step text-to-motion with TACD (II). Continued from Fig. 7 ; same student, same rendering.

Model / steps	end to end	sampling	text enc.	peak GB
1B teacher $+$ Qwen3-8B, 50 steps	$2077$ ms	$2026$	$54$	$17.63$
0.46B Lite $+$ Qwen3-8B, 50 steps	$1476$ ms	$1422$	$52$	$16.54$
TACD (HY-Motion): 0.46B $+$ Qwen3-0.6B, 8 steps	$268$ ms	$227$	$41$	$2.62$

Appendix

Table 19: Batch-1 cost of the HY-Motion systems on one RTX L40 ( bf16 ; medians over five round-robin passes of 30 motions of

6.5

s). Text encoding is per caption.

Diffusion Model Distillation

Diffusion model distillation transfers a large model’s generative behavior to faster or smaller models that need fewer sampling steps.

See all 191 papers

MixiMotion: One-Step Text-to-Motion Generation via Asymmetric Set Distillation
Sep 19, 2026Hung Dinh, Binh Mai, Tran Quoc Bao Le +2Text-to-Motion Generation Human Motion Generation
Posts and Telecommunications Institute of Technology, Hanoi, Vietnam
MotionRFT: Unified Reinforcement Fine-Tuning for Text-to-Motion Generation
Mar 28, 2026Xiaofeng Tan, Wanjiang Weng, Hongsong Wang +3Text-to-Motion Generation Human Motion Generation
School of Computer Science and Engineering, Southeast University, Nanjing 211189, China, and also with Key Laboratory of New Generation Artificial Intelligence Technology and Its Interdisciplinary Applications (Southeast University), Ministry of Education, China · State Key Laboratory for Novel Software Technology, School of Intelligence Science and Technology, Nanjing University, Nanjing 210023, China · New Laboratory of Pattern Recognition (NLPR), State Key Laboratory of Multimodal Artificial Intelligence Systems (MAIS), Institute of Automation, Chinese Academy of Sciences (CASIA), and also with School of Artificial Intelligence, University of Chinese Academy of Sciences
Reconstruction-Anchored Diffusion Model for Text-to-Motion Generation
Jan 21, 2026Yifei Liu, Changxing Ding, Ling Guo +2Diffusion Model Guidance Text-to-Motion Generation
South China University of Technology · Joy Future Academy

Sep 19, 2026J/K move · Enter open · S save

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

Hung Dinh, Binh Mai, Tran Quoc Bao Le, Lam Nguyen+1

Posts and Telecommunications Institute of Technology, Hanoi, Vietnam

Text-to-Motion Generation Human Motion Generation One-Step Generative Modeling Knowledge Distillation

Open paper AskPDF arXiv

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. These results demonstrate an effective quality--efficiency trade-off for strict one-step text-to-motion generation.

Abstract

Diffusion Model Distillation

Explore similar work

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

MotionRFT: Unified Reinforcement Fine-Tuning for Text-to-Motion Generation

Reconstruction-Anchored Diffusion Model for Text-to-Motion Generation