ParTY: Part-Guidance for Expressive Text-to-Motion Synthesis

KunHo Heo; SuYeon Kim; Yonghyun Gwon; Youngbin Kim; MyeongAh Cho

cs.CVMar 10, 2026

ParTY: Part-Guidance for Expressive Text-to-Motion Synthesis

Authors: KunHo Heo, SuYeon Kim, Yonghyun Gwon, Youngbin Kim, MyeongAh Cho

Organizations: Kyung Hee University

Abstract

Text-to-motion synthesis aims to generate natural and expressive human motions from textual descriptions. While existing approaches primarily focus on generating holistic motions from text descriptions, they struggle to accurately reflect actions involving specific body parts. Recent part-wise motion generation methods attempt to resolve this but face two critical limitations: (i) they lack explicit mechanisms for aligning textual semantics with individual body parts, and (ii) they often generate incoherent full-body motions due to integrating independently generated part motions. To overcome these issues and resolve the fundamental trade-off in existing methods, we propose ParTY, a novel framework that enhances part expressiveness while generating coherent full-body motions. ParTY comprises: (1) Part-Guided Network, which first generates part motions to obtain part guidance, then uses it to generate holistic motions; (2) Part-aware Text Grounding, which diversely transforms text embeddings and appropriately aligns them with each body part; and (3) Holistic-Part Fusion, which adaptively fuses holistic motions and part motions. Extensive experiments, including part-level and coherence-level evaluations, demonstrate that ParTY achieves substantial improvements over previous methods.

Figures & tables

Figure 1: (a) Holistic methods maintain coherence well but limited part-text alignment. In contrast, (b) Part-wise methods show enhanced part-text alignment (e.g., correctly performing the left leg lunge) but compromised coherence as a trade-off (e.g., neck distortion and misaligned arm and leg movements). (c) Our ParTY resolves this trade-off by achieving superior performance in both part-text alignment and coherence.

Figure 2: Architecture of the Temporal-aware VQ-VAE. Part VQ-VAE follows an identical architecture, where the sole distinction lies in processing part-level rather than full-body motion data.

Figure 3: Overview of ParTY. Text embeddings are processed through Part-aware Text Grounding, then part transformers generate Part Guidance for the holistic transformer to generate motion tokens, with Holistic-Part Fusion applied during generation. The notation { Part } indicates that the process is performed for both arms and legs.

Datasets	Method	R Precision $\uparrow$			FID $\downarrow$	MM-Dist $\downarrow$	Diversity $\rightarrow$	MultiModality $\uparrow$
Datasets	Method	Top 1	Top 2	Top 3	FID $\downarrow$	MM-Dist $\downarrow$	Diversity $\rightarrow$	MultiModality $\uparrow$
HumanML3D	Real motion	$0.511^{\pm.003}$	$0.703^{\pm.003}$	$0.797^{\pm.002}$	$0.002^{\pm.000}$	$2.974^{\pm.008}$	$9.503^{\pm.065}$	-
	MDM [ 36 ]	$0.320^{\pm.005}$	$0.498^{\pm.004}$	$0.611^{\pm.007}$	${0.544}^{\pm.044}$	$5.566^{\pm.027}$	${\underline{9.559}}^{\pm.086}$	$\mathbf{2.799}^{\pm.072}$
	T2M-GPT [ 44 ]	${0.491}^{\pm.003}$	${0.680}^{\pm.003}$	${0.775}^{\pm.002}$	${0.116}^{\pm.004}$	${3.118}^{\pm.011}$	${9.761}^{\pm.081}$	$1.856^{\pm.011}$
	ParCo [ 49 ]	${0.515}^{\pm.003}$	${0.706}^{\pm.003}$	${0.801}^{\pm.002}$	${0.109}^{\pm.005}$	${2.927}^{\pm.008}$	${9.576}^{\pm.088}$	$1.382^{\pm.060}$
	MMM [ 31 ]	${0.504}^{\pm.003}$	${0.696}^{\pm.003}$	${0.794}^{\pm.002}$	${0.080}^{\pm.003}$	${2.998}^{\pm.007}$	${9.411}^{\pm.058}$	$1.164^{\pm.041}$

Table 1: Quantitative comparison on HumanML3D and KIT-ML. Bold indicates the best result, while underlined refers the second-best. The right arrow

\rightarrow

indicates that closer values to ground truth are preferred.

Figure 4: Qualitative comparison on HumanML3D. Colored text in the descriptions corresponds to the colored body parts in the generated motions, with coherence-level (TC, SC) scores displayed for each sample.

Method	Part	R-Precision (Top-1) $\uparrow$	R-Precision (Top-3) $\uparrow$	FID $\downarrow$	MM-Dist $\downarrow$
MoMask [ 10 ]	Arms	$0.452^{\pm.003}$	$0.761^{\pm.002}$	$0.175^{\pm.003}$	$3.440^{\pm.006}$
MoMask [ 10 ]	Legs	$0.403^{\pm.003}$	$0.687^{\pm.003}$	$0.104^{\pm.003}$	$3.513^{\pm.009}$
ParCo [ 49 ]	Arms	$0.468^{\pm.003}$	$0.767^{\pm.003}$	$0.215^{\pm.003}$	$3.326^{\pm.008}$
ParCo [ 49 ]	Legs	$0.407^{\pm.003}$	$0.699^{\pm.002}$	$0.118^{\pm.003}$	$3.482^{\pm.011}$
Ours	Arms	$\mathbf{0.506}^{\pm.003}$	$\mathbf{0.802}^{\pm.002}$	$\mathbf{0.133}^{\pm.002}$	$\mathbf{3.079}^{\pm.005}$
Ours	Legs	$\mathbf{0.463}^{\pm.003}$	$\mathbf{0.755}^{\pm.003}$	$\mathbf{0.078}^{\pm.003}$	$\mathbf{3.122}^{\pm.008}$

Table 2: Quantitative comparison with part-level evaluation metrics on HumanML3D.

Method	Temporal Coherence (TC) $\uparrow$	Spatial Coherence (SC) $\uparrow$
ParCo [ 49 ]	$0.49^{\pm.062}$	$0.59^{\pm.057}$
MoMask [ 10 ]	$0.84^{\pm.047}$	$0.90^{\pm.044}$
Ours	$\mathbf{0.88}^{\pm.051}$	$\mathbf{0.92}^{\pm.041}$

Table 3: Quantitative comparison with coherence-level (TC, SC) scores on HumanML3D. We run each evaluation 20 times and report averages with 95% confidence intervals.

Method	Window size	Reconstruction			Generation
Method	Window size	# Params	FID $\downarrow$	MPJPE $\downarrow$	FID $\downarrow$	AIT
MoMask	4	19.44M	0.020	0.030	0.045	80ms
+ Ours	4	19.86M	0.003 (+85%)	0.011 (+63%)	0.033 (+26%)	80ms
MoMask	8	10.15M	0.042	0.055	0.094	43ms (+46%)
+ Ours	8	10.58M	0.005 (+88%)	0.014 (+74%)	0.039 (+58%)	43ms (+46%)
MoMask	12	7.67M	0.079	0.091	0.126	29ms (+64%)

Table 4: Porting Temporal-aware VQ-VAE to MoMask [ 10 ] . Reconstruction evaluates VQ-VAE performance, while Generation evaluates final performance including the transformer. Mean Per Joint Position Error (MPJPE) measures positional accuracy, and Average Inference Time (AIT) is averaged over 100 samples on an RTX A5000 GPU.

PG	PTG	HPF	R-Precision (Top-1) $\uparrow$	R-Precision (Top-3) $\uparrow$	FID $\downarrow$	MM-Dist $\downarrow$
			$0.494^{\pm.003}$	$0.780^{\pm.003}$	$0.158^{\pm.005}$	$3.087^{\pm.008}$
✓			$0.520^{\pm.002}$	$0.802^{\pm.003}$	$0.086^{\pm.003}$	$2.913^{\pm.010}$
✓	✓		$0.545^{\pm.003}$	$0.828^{\pm.003}$	$0.051^{\pm.003}$	$2.799^{\pm.008}$
✓	✓	✓	$\mathbf{0.550}^{\pm.003}$	$\mathbf{0.836}^{\pm.002}$	$\mathbf{0.035}^{\pm.002}$	$\mathbf{2.779}^{\pm.006}$

Table 5: Ablation studies of the proposed components. PG indicates Part Guidance.

Part	PG	PTG	HPF	R-Precision (Top-1) $\uparrow$	R-Precision (Top-3) $\uparrow$	FID $\downarrow$	MM-Dist $\downarrow$
Arms				$0.433^{\pm.003}$	$0.736^{\pm.002}$	$0.232^{\pm.004}$	$3.347^{\pm.014}$
	✓			$0.470^{\pm.003}$	$0.769^{\pm.003}$	$0.166^{\pm.002}$	$3.251^{\pm.006}$
	✓	✓		$0.501^{\pm.003}$	$0.798^{\pm.003}$	$0.152^{\pm.002}$	$3.102^{\pm.007}$
	✓	✓	✓	$\mathbf{0.506}^{\pm.003}$	$\mathbf{0.802}^{\pm.002}$	$\mathbf{0.133}^{\pm.002}$	$\mathbf{3.079}^{\pm.005}$

Table 6: Ablation studies with part-level evaluation metrics.

Figure 5: Visualization of cross attention map of HPF. Rows correspond to body parts and columns represent temporal frames. We visualize the normalized attention weights between the holistic motion token and each part motion token.

Figure 6: Embedding selection ratios in PTG. Mean and standard deviation of weights are computed over semantically similar text descriptions that share common motion patterns.

Figure S1 : Body part division for HumanML3D and KIT-ML.

Method	Part	R-Precision (Top-1) $\uparrow$	R-Precision (Top-3) $\uparrow$	FID $\downarrow$	MM-Dist $\downarrow$
MoMask [ 10 ]	Arms	$0.452^{\pm.003}$	$0.761^{\pm.002}$	$0.175^{\pm.003}$	$3.440^{\pm.006}$
MoMask [ 10 ]	Legs	$0.403^{\pm.003}$	$0.687^{\pm.003}$	$0.104^{\pm.003}$	$3.513^{\pm.009}$
AttT2M [ 48 ]	Arms	$0.431^{\pm.003}$	$0.738^{\pm.002}$	$0.264^{\pm.002}$	$3.489^{\pm.010}$
AttT2M [ 48 ]	Legs	$0.375^{\pm.003}$	$0.656^{\pm.003}$	$0.190^{\pm.003}$	$3.541^{\pm.011}$
ParCo [ 49 ]	Arms	$0.468^{\pm.003}$	$0.767^{\pm.003}$	$0.215^{\pm.003}$	$3.326^{\pm.008}$
ParCo [ 49 ]	Legs	$0.407^{\pm.003}$	$0.699^{\pm.002}$	$0.118^{\pm.003}$	$3.482^{\pm.011}$

Table S1 : Quantitative comparison with part-level evaluation metrics on HumanML3D.

Part	PG	PTG	HPF	R-Precision (Top-1) $\uparrow$	R-Precision (Top-3) $\uparrow$	FID $\downarrow$	MM-Dist $\downarrow$
Legs				$0.397^{\pm.003}$	$0.691^{\pm.003}$	$0.169^{\pm.003}$	$3.416^{\pm.012}$
	✓			$0.422^{\pm.003}$	$0.715^{\pm.003}$	$0.114^{\pm.003}$	$3.328^{\pm.011}$
	✓	✓		$0.456^{\pm.002}$	$0.744^{\pm.003}$	$0.095^{\pm.003}$	$3.175^{\pm.006}$
	✓	✓	✓	$\mathbf{0.463}^{\pm.003}$	$\mathbf{0.755}^{\pm.003}$	$\mathbf{0.078}^{\pm.003}$	$\mathbf{3.122}^{\pm.008}$

Table S2 : Ablation studies with part-level evaluation metrics on HumanML3D.

Hyperparameters	Temporal Coherence $\uparrow$
$\sigma=0.05$ , $\kappa=3$	0.92
$\sigma=0.05$ , $\kappa=5$	0.93
$\sigma=0.05$ , $\kappa=10$	0.93
$\sigma=0.05$ , $\kappa=15$	0.91
$\sigma=0.1$ , $\kappa=3$	0.95
$\bm{\sigma}=\mathbf{0.1}$ , $\bm{\kappa}=\mathbf{5}$	$\mathbf{0.96}$

Table S3: Ablation studies on

\sigma

and

\kappa

parameters.

Method	Temporal Coherence $\uparrow$
$L=20$ , $\tau_{\text{max}}=10$	0.95
$\bm{L}=\mathbf{20}$ , $\bm{\tau}_{\text{{max}}}=\mathbf{15}$	$\mathbf{0.96}$
$\bm{L}=\mathbf{20}$ , $\bm{\tau}_{\text{{max}}}=\mathbf{20}$	$\mathbf{0.96}$
$L=20$ , $\tau_{\text{max}}=30$	0.94
$L=30$ , $\tau_{\text{max}}=10$	0.94
$L=30$ , $\tau_{\text{max}}=15$	0.95

Table S4: Ablation studies on

L

and

\tau_{\text{max}}

parameters.

$\beta_{d}$	$\beta_{\theta}$	Spatial Coherence $\uparrow$
1.0	1.0	0.97
$\mathbf{1.5}$	$\mathbf{1.5}$	$\mathbf{0.99}$
2.0	2.0	0.98
3.0	3.0	0.94

Table S5: Ablation studies on

\beta_{d}

and

\beta_{\theta}

parameters.

Method	Temporal Coherence (TC) $\uparrow$	Spatial Coherence (SC) $\uparrow$
Real motion	$0.96^{\pm.032}$	$0.99^{\pm.026}$
ParCo [ 49 ]	$0.49^{\pm.062}$	$0.59^{\pm.057}$
LGTM [ 35 ]	$0.51^{\pm.048}$	$0.65^{\pm.039}$
T2M-GPT [ 44 ]	$0.85^{\pm.050}$	$0.87^{\pm.036}$
MoMask [ 10 ]	$0.84^{\pm.047}$	$0.90^{\pm.044}$
Ours	$\mathbf{0.88}^{\pm.051}$	$\mathbf{0.92}^{\pm.041}$

Table S6 : Quantitative comparison with coherence-level (TC, SC) scores on HumanML3D.

Method	Temporal Coherence (TC) $\uparrow$	Spatial Coherence (SC) $\uparrow$
Real motion	$0.87^{\pm.045}$	$0.91^{\pm.037}$
ParCo [ 49 ]	$0.32^{\pm.066}$	$0.44^{\pm.062}$
LGTM [ 35 ]	$0.40^{\pm.061}$	$0.49^{\pm.058}$
T2M-GPT [ 44 ]	$0.77^{\pm.057}$	$0.80^{\pm.049}$
MoMask [ 10 ]	$0.79^{\pm.042}$	$0.79^{\pm.065}$
Ours	$\mathbf{0.80}^{\pm.048}$	$\mathbf{0.81}^{\pm.054}$

Table S7 : Quantitative comparison with coherence-level (TC, SC) scores on KIT-ML.

PG	PTG	HPF	Temporal Coherence $\uparrow$	Spatial Coherence $\uparrow$
			$0.43^{\pm.056}$	$0.54^{\pm.044}$
✓			$0.81^{\pm.048}$	$0.86^{\pm.029}$
✓	✓		$0.83^{\pm.045}$	$0.89^{\pm.034}$
✓	✓	✓	$\mathbf{0.88}^{\pm.051}$	$\mathbf{0.92}^{\pm.041}$

Table S8 : Ablation studies with coherence-level evaluation metrics on HumanML3D.

Figure S2 : Qualitative comparison on part-text alignment on HumanML3D.

Figure S3 : Qualitative comparison on coherence on HumanML3D.

Figure S4 : Qualitative comparison on long and complex motions on HumanML3D.

PG	PTG	HPF	R-Precision $\uparrow$			FID $\downarrow$	MM-Dist $\downarrow$	Diversity $\rightarrow$	MModality $\uparrow$
PG	PTG	HPF	Top-1	Top-2	Top-3	FID $\downarrow$	MM-Dist $\downarrow$	Diversity $\rightarrow$	MModality $\uparrow$
			$0.494^{\pm.003}$	$0.686^{\pm.003}$	$0.780^{\pm.003}$	$0.158^{\pm.005}$	$3.087^{\pm.008}$	$9.722^{\pm.067}$	$1.445^{\pm.051}$
✓			$0.520^{\pm.002}$	$0.711^{\pm.003}$	$0.802^{\pm.003}$	$0.086^{\pm.003}$	$2.913^{\pm.010}$	$9.688^{\pm.045}$	$1.841^{\pm.079}$
		✓	$0.506^{\pm.003}$	$0.695^{\pm.003}$	$0.791^{\pm.003}$	$0.132^{\pm.003}$	$3.019^{\pm.013}$	$9.710^{\pm.056}$	$1.860^{\pm.064}$
✓	✓		$\underline{0.545}^{\pm.003}$	$\underline{0.734}^{\pm.003}$	$\underline{0.828}^{\pm.003}$	$\underline{0.051}^{\pm.003}$	$\underline{2.799}^{\pm.010}$	$9.556^{\pm.042}$	$1.973^{\pm.068}$
✓		✓	$0.529^{\pm.002}$	$0.721^{\pm.003}$	$0.813^{\pm.003}$	$0.063^{\pm.002}$	$2.858^{\pm.008}$	$\underline{9.528}^{\pm.059}$	$\underline{2.101}^{\pm.071}$

Table S9 : Comprehensive ablation studies of the proposed component. Bold indicates the best result, while underlined refers the second-best. The right arrow

\rightarrow

indicates that closer values to ground truth are preferred.

Method	#Params	AIT	Part-level Evaluation (Avg.)
Method	#Params	AIT	R-Precision (Top-3) $\uparrow$	FID $\downarrow$	MM-Dist $\downarrow$
T2M-GPT [ 44 ]	247.5M	277ms	$0.706^{\pm.003}$	$0.198^{\pm.003}$	$3.534^{\pm.008}$
ParCo [ 49 ]	168.4M	643ms	$0.733^{\pm.003}$	$0.166^{\pm.003}$	$3.404^{\pm.010}$
MoMask [ 10 ]	44.8M	80ms	$0.724^{\pm.003}$	$0.139^{\pm.003}$	$3.476^{\pm.007}$
Ours	78.3M	158ms	$0.778^{\pm.003}$	$0.105^{\pm.003}$	$3.100^{\pm.007}$

Table S10 : Comparison of computational complexity. Part-level evaluation reports the average performance of arms and legs.

Method	Window size	Reconstruction			Generation
Method	Window size	# Params	FID $\downarrow$	MPJPE $\downarrow$	FID $\downarrow$	AIT
ParCo	4	6.35M	0.021	0.108	0.109	65ms
+ Ours	4	7.93M	0.005 (+76%)	0.034 (+68%)	0.077 (+29%)	65ms
ParCo	8	4.42M	0.047	0.166	0.172	33ms (+49%)
+ Ours	8	6.01M	0.009 (+81%)	0.041 (+75%)	0.085 (+50%)	33ms (+49%)
ParCo	12	3.59M	0.090	0.243	0.255	23ms (+64%)

Table S11 : Porting Temporal-aware VQ-VAE to ParCo [ 49 ] . AIT is averaged over 100 samples on an RTX A5000 GPU.

LTE	GTE	FID $\downarrow$	MPJPE $\downarrow$
		$0.038^{\pm.000}$	$0.121^{\pm.000}$
✓		$0.018^{\pm.000}$	$0.049^{\pm.000}$
	✓	$0.020^{\pm.000}$	$0.071^{\pm.000}$
✓	✓	$0.007^{\pm.000}$	$0.019^{\pm.000}$

Table S12 : Ablation studies on proposed components in Temporal-aware VQ-VAE.

# of Emb.	R-Precision $\uparrow$			FID $\downarrow$	MM-Dist $\downarrow$
# of Emb.	Top-1	Top-2	Top-3	FID $\downarrow$	MM-Dist $\downarrow$
1	$0.531^{\pm.003}$	$0.722^{\pm.003}$	$0.815^{\pm.003}$	$0.060^{\pm.003}$	$2.842^{\pm.011}$
2	$0.538^{\pm.003}$	$0.731^{\pm.002}$	$0.827^{\pm.003}$	$0.052^{\pm.003}$	$2.825^{\pm.015}$
3	$0.541^{\pm.002}$	$0.735^{\pm.003}$	$0.829^{\pm.003}$	$0.047^{\pm.003}$	$2.811^{\pm.007}$
4	$\mathbf{0.550}^{\pm.003}$	$\mathbf{0.744}^{\pm.003}$	$\mathbf{0.836}^{\pm.003}$	$\underline{0.035}^{\pm.002}$	$\mathbf{2.779}^{\pm.006}$
5	$\underline{0.546}^{\pm.003}$	$\underline{0.739}^{\pm.002}$	$\underline{0.833}^{\pm.003}$	$\mathbf{0.033}^{\pm.002}$	$\underline{2.804}^{\pm.011}$

Table S13 : Quantitative results according to the number of embeddings in PTG.

Method	R-Precision $\uparrow$			FID $\downarrow$	MM-Dist $\downarrow$
Method	Top-1	Top-2	Top-3	FID $\downarrow$	MM-Dist $\downarrow$
Ours (Full model)	$0.550^{\pm.003}$	$0.744^{\pm.003}$	$0.836^{\pm.003}$	$0.035^{\pm.002}$	$2.779^{\pm.006}$
w/o aux. loss	$0.543^{\pm.003}$	$0.739^{\pm.003}$	$0.830^{\pm.003}$	$0.042^{\pm.003}$	$2.805^{\pm.011}$

Table S14 : Ablation studies on auxilary loss.

Method	Part	R-Precision (Top-1) $\uparrow$	R-Precision (Top-3) $\uparrow$	FID $\downarrow$	MM-Dist $\downarrow$
Ours (Full model)	Arms	$0.506^{\pm.003}$	$0.802^{\pm.002}$	$0.133^{\pm.002}$	$3.079^{\pm.005}$
Ours (Full model)	Legs	$0.463^{\pm.003}$	$0.755^{\pm.003}$	$0.078^{\pm.003}$	$3.122^{\pm.008}$
w/o aux. loss	Arms	$0.491^{\pm.003}$	$0.793^{\pm.002}$	$0.148^{\pm.003}$	$3.098^{\pm.008}$
w/o aux. loss	Legs	$0.444^{\pm.002}$	$0.732^{\pm.003}$	$0.092^{\pm.003}$	$3.167^{\pm.008}$

Table S15 : Ablation studies on auxilary loss with part-level evaluation metrics.

Types	R-Precision $\uparrow$			FID $\downarrow$	MM-Dist $\downarrow$
Types	Top-1	Top-2	Top-3	FID $\downarrow$	MM-Dist $\downarrow$
$\text{LLM}_{1}$	$0.549^{\pm.003}$	$0.744^{\pm.002}$	$\mathbf{0.836}^{\pm.003}$	$0.035^{\pm.002}$	$\mathbf{2.777}^{\pm.009}$
$\text{LLM}_{2}$	$\mathbf{0.551}^{\pm.003}$	$0.743^{\pm.003}$	$0.835^{\pm.003}$	$0.038^{\pm.003}$	$2.792^{\pm.009}$
$\text{LLM}_{3}$	$0.550^{\pm.003}$	$0.744^{\pm.003}$	$\mathbf{0.836}^{\pm.002}$	$0.035^{\pm.002}$	$2.779^{\pm.006}$
$\text{LLM}_{4}$	$0.550^{\pm.003}$	$\mathbf{0.745}^{\pm.003}$	$\mathbf{0.836}^{\pm.003}$	$0.036^{\pm.003}$	$2.788^{\pm.007}$
$\text{LLM}_{5}$	$0.548^{\pm.003}$	$0.743^{\pm.003}$	$\mathbf{0.836}^{\pm.002}$	$\mathbf{0.032}^{\pm.003}$	$2.790^{\pm.011}$

Table S16 : Quantitative results with multiple LLM-generated part texts.

Window size	R-Precision $\uparrow$			FID $\downarrow$	MM-Dist $\downarrow$
Window size	Top-1	Top-2	Top-3	FID $\downarrow$	MM-Dist $\downarrow$
1	$0.545^{\pm.003}$	$0.735^{\pm.003}$	$0.829^{\pm.003}$	$0.051^{\pm.003}$	$2.822^{\pm.012}$
2	$0.547^{\pm.003}$	$0.740^{\pm.003}$	$0.833^{\pm.002}$	$0.040^{\pm.003}$	$2.801^{\pm.008}$
3	$0.550^{\pm.003}$	$\mathbf{0.744}^{\pm.003}$	$\mathbf{0.836}^{\pm.002}$	$\mathbf{0.035}^{\pm.002}$	$2.779^{\pm.006}$
4	$\mathbf{0.551}^{\pm.003}$	$0.741^{\pm.003}$	$0.835^{\pm.003}$	$0.036^{\pm.002}$	$\mathbf{2.754}^{\pm.009}$
5	$0.548^{\pm.003}$	$0.740^{\pm.002}$	$0.833^{\pm.003}$	$0.039^{\pm.003}$	$2.793^{\pm.007}$

Table S17 : Quantitative results for different window sizes in PG.

Figure S6 : Visualization of cross attention map of HPF.

Figure S8 : User evaluation interface for the user study.

Panel of Figure 38 — Figure S9 : User study results on HumanML3D dataset. Each bar represents the average score on a scale from 1 to 5.

Datasets	Method	R Precision $\uparrow$			FID $\downarrow$	MM-Dist $\downarrow$	Diversity $\rightarrow$	MultiModality $\uparrow$
Datasets	Method	Top 1	Top 2	Top 3	FID $\downarrow$	MM-Dist $\downarrow$	Diversity $\rightarrow$	MultiModality $\uparrow$
HumanML3D	Real motion	$0.511^{\pm.003}$	$0.703^{\pm.003}$	$0.797^{\pm.002}$	$0.002^{\pm.000}$	$2.974^{\pm.008}$	$9.503^{\pm.065}$	-
	TM2T [ 12 ]	$0.424^{\pm.003}$	$0.618^{\pm.003}$	$0.729^{\pm.002}$	$1.501^{\pm.017}$	$3.467^{\pm.011}$	$8.589^{\pm.076}$	$2.424^{\pm.093}$
	T2M [ 11 ]	$0.457^{\pm.002}$	$0.639^{\pm.003}$	$0.740^{\pm.003}$	$1.067^{\pm.002}$	$3.340^{\pm.008}$	$9.188^{\pm.002}$	$2.090^{\pm.083}$
	MDM [ 36 ]	$0.320^{\pm.005}$	$0.498^{\pm.004}$	$0.611^{\pm.007}$	${0.544}^{\pm.044}$	$5.566^{\pm.027}$	${9.559}^{\pm.086}$	$\mathbf{2.799}^{\pm.072}$
	MLD [ 6 ]	${0.481}^{\pm.003}$	${0.673}^{\pm.003}$	${0.772}^{\pm.002}$	${0.473}^{\pm.013}$	${3.196}^{\pm.010}$	$9.724^{\pm.082}$	${2.413}^{\pm.079}$

Table S18 : Quantitative comparison on HumanML3D and KIT-ML. Bold indicates the best result, while underlined refers the second-best. The right arrow

\rightarrow

indicates that closer values to ground truth are preferred.

Explore similar work

Nov 28, 2024

BiPO: Bidirectional Partial Occlusion Network for Text-to-Motion Synthesis

Generating natural and expressive human motions from textual descriptions is challenging due to the complexity of coordinating full-body dynamics and capturing nuanced motion patterns over extended sequences that accurately reflect the given text. To address this, we introduce BiPO, Bidirectional Partial Occlusion Network for Text-to-Motion Synthesis, a novel model that enhances text-to-motion synthesis by integrating part-based generation with a bidirectional autoregressive architecture. This integration allows BiPO to consider both past and future contexts during generation while enhancing detailed control over individual body parts without requiring ground-truth motion length. To relax the interdependency among body parts caused by the integration, we devise the Partial Occlusion technique, which probabilistically occludes the certain motion part information during training. In our comprehensive experiments, BiPO achieves state-of-the-art performance on the HumanML3D dataset, outperforming recent methods such as ParCo, MoMask, and BAMM in terms of FID scores and overall motion quality. Notably, BiPO excels not only in the text-to-motion generation task but also in motion editing tasks that synthesize motion based on partially generated motion sequences and textual descriptions. These results reveal the BiPO's effectiveness in advancing text-to-motion synthesis and its potential for practical applications.

Seong-Eun Hong, Soobin Lim, Juyeong Hwang +2

Kyung Hee University Korea, Republic · NC Research, NCSOFT Corp. Korea, Republic · Korea University Korea, Republic

Text-to-Motion Generation Human Motion Generation Autoregressive Generation

Ask PDF

Oct 9, 2026

PartiGen: Part-Aware Representations for Language-Conditioned Humanoid Motion Generation

Translating language into humanoid motion requires coordinating body parts, determining motion duration, and producing references suitable for robot execution. We introduce PartiGen, a part-aware latent diffusion framework that incorporates anatomical structure into a shared whole-body motion representation. Part-specific attention processes anatomically masked frame features, and its outputs are fused to combine local limb information with whole-body context. Conditioned on text and motion history, the model generates motion primitives autoregressively. Forward-kinematics supervision encourages kinematic consistency, while a learned end-of-sequence predictor selects a frame-level endpoint after generation. On HumanML3D retargeted to the Unitree G1, PartiGen achieves a segment-level FID of 3.165, a 16.8% reduction over TextOp, and improves R@1 by 1.4 percentage points. In the duration study with 16-frame primitives, endpoint prediction reduces mean absolute duration error from 12.471 to 3.080 frames compared with a fixed-length cutoff. We further evaluate robot execution using a pretrained, frozen GRIT tracker: 409 of 500 generated references (81.8%) pass the posture-based stability screen in MuJoCo. Following simulation screening and manual selection, 26 of 27 attempted motions are judged successful on the real G1.

Chuxuan Huang, Zhanxiang Cao, Tianyue Zhan +14

Fudan University · Shanghai Innovation Institute · Shanghai Jiaotong University +5

Text-to-Motion Generation Human Motion Generation Robot Motion Generation Diffusion Models for Motion Generation Motion Representation Learning

Ask PDF

May 29, 2026

MultiAct: Text-to-Motion Generation from Composite Text via Tailored Attention Guidance

Text-to-motion generation has progressed rapidly in recent years, offering an expressive interface for animation and human-computer interaction. However, current models remain brittle when handling prompts that describe multiple actions occurring at the same time. Rather than realizing all components of a composite description, models frequently prioritize a single dominant action and neglect the rest, leading to incomplete or ambiguous motion. We present MultiAct, an unpaired, inference-time framework for compositional text-to-motion synthesis that operates directly on pretrained motion generators without retraining or architectural modification. Our method counteracts semantic collapse by adaptively amplifying cross-attention scores associated with underrepresented prompt components. We note that effective modulation depends on prompt-specific choices, such as which tokens and layers to target, and introduce a lightweight auxiliary decision scheme that determines the most effective attention-strengthening parametrization. Extensive quantitative and qualitative evaluations demonstrate that MultiAct consistently outperforms existing baselines on composite prompts, achieving improved semantic coverage while preserving motion realism. Project page: https://natsala13.github.io/multiact.github.io.

Nathan Sala, Ofir Abramovich, Ariel Shamir +3

Tel Aviv University, Israel · Reichman University, Israel · CYENS Centre of Excellence, Cyprus +1

Text-to-Motion Generation Cross-Modal Attention Attention Mechanisms

Ask PDF