notesum.ai

Published at December 6

SoPo: Text-to-Motion Generation Using Semi-Online Preference Optimization

cs.CV

Released Date: December 6, 2024

Authors: Xiaofeng Tan¹, Hongsong Wang¹, Xin Geng¹, Pan Zhou²

Aff.: ¹Southeast University, Nanjing, China; ²Singapore Management University

Arxiv: http://arxiv.org/pdf/2412.05095v1

[Uncaptioned image]

Methods	Time^∗	R-Precision $\uparrow$			MM Dist $\downarrow$	Diversity $\rightarrow$	FID $\downarrow$
Methods	Time^∗	Top 1	Top 2	Top 3	MM Dist $\downarrow$	Diversity $\rightarrow$	FID $\downarrow$
Real	-	.511^±.003	.703^±.003	.797^±.002	2.974^±.008	9.503^±.065	.002 ^±.000
MLD [3]	+0 X	-	-	.755^±.003	3.292^±.010	9.793^±.072	.459^±.011
$+$ MoDiPO-T [23]	+121 $K$ X	-	-	.758 ${{}^{\pm.002}}_{+0.40\%}$	3.267^±.010_+0.76%	9.747^±.073_+0.046	.303^±.031_+33.9%
$+$ MoDiPO-G [23]	+121 $K$ X	-	-	.753^±.003_-0.26%	3.294^±.010_-0.01%	9.702^±.075_+0.091	.281^±.031_+38.8%
$+$ MoDiPO-O [23]	-	-	-	.677^±.003_-10.3%	3.701^±.013_-12.4%	9.241^±.079_-0.018	.276^±.007_+39.9%^†
$+$ SoPo (Ours)	+20 X	-	-	.763^±.003_+1.06%	3.185^±.012_+3.25%^†	9.525^±.065_+0.222^†	.374^±.007_+18.5%
MDM [33]	+0 X	.418 ^±.005	.604^±.005	.703^±.005	3.658^±.025	9.546^±.066	.501^±.037
$+$ MoDiPO-T [23]	+121 $K$ X	-	-	.706^±.004_+0.42%	3.634^±.026_+0.66%	9.531^±.073_+0.015	.451^±.031_+9.98%
$+$ MoDiPO-G [23]	+121 $K$ X	-	-	.704^±.001_+0.14%	3.641^±.025_+0.46%	9.495^±.071_+0.035	.486^±.031_+2.99%
MDM (fast) [33]	+0 X	.455^±.006	.645^±.007	.749^±.004	3.304^±.023	9.948^±.084	.534^±.052
$+$ SoPo (Ours)	+60 X	.479^±.006_+5.27%^†	.674 ^±.005_+4.50%^†	.770^±.006_+2.80%^†	3.208^±.025_+2.91%	9.906^±.083_+0.042	.480^±.046_+10.1%