AI论文资料包文件19 Fix the Noise_ Disentangling Source Feature for Transfer Learning of StyleGAN.pdfVIP

FixtheNoise:DisentanglingSourceFeatureforTransferLearningofStyleGAN

DongyeunLeeJaeYoungLeeDoyeonKimJaehyunChoiJunmoKim

KAIST

ledoye,mcneato,doyeonkim,chlwogus,junmo.kim@kaist.ac.kr

3

2

0

2

r

a

M

2

2

Transfer

]Learning

V

C

.

s

c

[

3

vGeneratedfromMoreSourceGeneratedfromMoreTarget

9Figure1.Givenasourcemodel,ourmethodcansmoothlycontrolthedegreeofsourcedomainfeaturesinafine-tunedmodel.

7Imagesineachrowaregeneratedfromthesamelatentcodebyand.

0

4

1Abstract1.Introduction

.

4TransferlearningofStyleGANhasrecentlyshowngreatGenerativeadversarialnetworks(GANs)haverecently

0

2potentialtosolvediversetasks,especiallyindomaintrans-madetremendousstridesinthefieldofhigh-qualityimage

2lation.Previousmethodsutilizedasourcemodelbyswap-synthesis,especiallyusingstyle-basedarchitecture[8–10].

:pingorfreezingweightsduringtransferlearning,how-Manystudieshaveutilizedpre-trainedStyleGAN[9,10]in

v

iever,theyhavelimitationsonvisualqualityandcontrol-variousfields,suchasimageencoding[20,25,28],discov-

Xlingsourcefeatures.Inotherwords,theyrequireadditionaleringlatentsemantics[4,21,23],imageediting[1,22]and

rmodelsthatarecomputationallydemandingandhavere-transferlearning[7,16,17].Inparticular,StyleAlign[24]