4-4 敏捷型风格化人像和虚拟形象创建.pdf

1、Agile Stylized Portrait and Avatar Creation敏捷型风格化人像和虚拟形象创建罗琳捷 数字人技术北美研究主管The Rising Popularity of Virtual Identities2Viola app/Snapchat cartoon filter Stylized PortraitsMemoji/ReadyPlayerMe/Bitmoji Stylized AvatarsVirtual Identities in ByteDance Products3抖音仔仔Manga StylePixar StyleTikTok Avatars4How

2、to automatically createlook-alike&appealing stylized virtual identities for usersExisting Approach Requires Laborious Annotations5large amount of paired stylized examplesFace shapeNose shapeHair lenghHair colorGlassesPearNubianLongYellowYesTriangleRomanShortBrownNoOvalSnubLongBlackYeslarge amount of

3、 annotationsKey ChallengesHow to reduce or eliminate the manual effort and increase agility?6Lower costFaster ShippingHigher QualityStylized Portraits7AgileGAN:Stylizing Portraits via Inversion-Consistent Transfer Learning SIGGRAPH 2021AgileAvatar:Stylized 3D Avatar Creation via Cascaded Domain Brid

4、ging SIGGRAPH Asia 2022Stylized AvatarsStylized Portraits8AgileGAN:Stylizing Portraits via Inversion-Consistent Transfer Learning SIGGRAPH 2021AgileAvatar:Stylized 3D Avatar Creation via Cascaded Domain Bridging SIGGRAPH Asia 2022Stylized AvatarsPrevious Work Manual Creation9Top quality,but very exp

5、ensivePrevious Work Style Transfer10InputGatys et al.2016Style exemplarLack semantic correspondencesPrevious Work Unsupervised Image Translation11InputCycleGAN Zhu et al.2017UGATIT Kim et al.2020UNIT Liu et al.2017Sub-optimal quality due to lack of style exemplarsPrevious Work ToonifyTransfer learni

6、ng from StyleGAN by very few(100)style exemplarsInputToonifyInputToonifyStyle exemplarsPrevious Work ToonifyTransfer learning from StyleGAN by very few(100)style exemplars13InputInputOur AgileGAN ResultsPortrait Stylization by Transfer Learning from StyleGAN14Z space(1 x 512)W space(1 x 512)MLPStyle