人工智能论文英文版-MOGO：Residual Quantized Hierarchical Causal Transformer for High-Quality and Real-Time 3D Human Motion Generation.pdfVIP

MOGO:ResidualQuantizedHierarchicalCausal

TransformerforHigh-QualityandReal-Time3D

HumanMotionGeneration

DongjieFu1*TengjiaoSun2*PengchengFang2*

dongjie.fu@mogo.ait.s.sun@soton.ac.ukp.fang@soton.ac.uk

5

2XiaohaoCai2HansungKim2†

0x.cai@soton.ac.ukh.kim@soton.ac.uk

2

n1MogoAI,China

u2

JUniversityofSouthampton,UnitedKingdom

*Equalcontribution†Correspondingauthor

6

]

V

C

.

s

c

[

1

vapersonwalkslikeazombie

2

5

9

5

0apersonwalksinashapeofSApersonisdancingballetgracefullyApersonbouncingaround

.(13s,260frames)whilethrowingjabsanduppercuts

6

0Figure1:IllustrationofMOGO,aGPT-stylemodelforhigh-qualitytext-to-motiongeneration.The

5modelsupportsopen-vocabularypromptsandcangeneratelong,lifelike3Dmotionsequencesfrom

2naturallanguageinputs.

:

v

i

XAbstract

r

a

Recentadvancesintransformer-basedtext-to-motiongenerationhaveledtoim-

pressiveprogressinsynthesizinghigh-qualityhumanmotion.Nevertheless,jointly

achievinghighfidelity,streamingcapability,real-timeresponsiveness,andscalabil-