Tawagar masu bincike daga Jami'ar Tel Aviv ta bayyana HyperStyle, wani juzu'i na tsarin koyon injina na NVIDIA's StyleGAN2 wanda aka sake tsara shi don sake ƙirƙirar sassan da suka ɓace yayin gyara hotuna na zahiri. An rubuta lambar a cikin Python ta amfani da tsarin PyTorch kuma an rarraba a ƙarƙashin lasisin MIT.
Idan StyleGAN ya ba ku damar haɓaka sabbin fuskokin mutane ta hanyar saita sigogi kamar shekaru, jinsi, tsayin gashi, ƙirar murmushi, siffar hanci, launin fata, tabarau da kusurwar hoto, to HyperStyle yana ba ku damar canza sigogi iri ɗaya a cikin data kasance. hotuna ba tare da canza halayen halayen su ba da kuma riƙe da ganewa na ainihin fuskar. Misali, ta amfani da HyperStyle, zaku iya kwaikwayi canjin shekarun mutum a hoto, canza salon gyara gashi, ƙara tabarau, gemu ko gashin baki, sanya hoto yayi kama da yanayin zane mai ban dariya ko hoton da aka zana da hannu, yin bakin ciki ko fara'a magana. A wannan yanayin, ana iya horar da tsarin ba kawai don canza fuskokin mutane ba, har ma ga kowane abu, alal misali, don gyara hotunan motoci.
Hanyar da aka tsara tana nufin magance matsalar tare da sake gina sassan hoton da suka ɓace lokacin gyarawa. A cikin hanyoyin da suka gabata, an warware sasantawa tsakanin sake ginawa da gyarawa ta hanyar daidaita janareta na hoto don maye gurbin sassan hoton da aka yi niyya lokacin sake ƙirƙirar wuraren da aka ɓace da farko. Rashin lahani na irin waɗannan hanyoyin shine buƙatar horo na dogon lokaci da aka yi niyya na cibiyar sadarwar jijiyoyi don kowane hoto.
Hanyar da ta danganci StyleGAN algorithm ta sa ya yiwu a yi amfani da samfurin na yau da kullum, wanda aka horar da shi a kan tarin hotuna na yau da kullum, don samar da abubuwan da suka dace na ainihin hoton tare da matakin amincewa da kwatankwacin algorithms waɗanda ke buƙatar horar da mutum na samfurin don kowane hoto. . Daga cikin fa'idodin sabuwar hanyar, ana kuma lura da yiwuwar gyaggyara hotuna tare da yin aiki kusa da ainihin lokacin.
An shirya samfuran da aka riga aka horar da su ga ɗan adam, mota da fuskokin dabba bisa tarin Flickr-Faces-HQ (FFHQ, 70k hotuna masu inganci na PNG na fuskokin mutum), Stanford Cars (hotunan motoci 16k) da AFHQ (hotuna). na dabbobi). Bugu da ƙari, an tanadar da kayan aikin don horar da ƙirar su, da kuma shirye-shiryen horar da ƙirar ƙira da janareta masu dacewa da amfani da su. Misali, ana samun injina don ƙirƙirar hotuna irin na Toonify, haruffan Pixar, zane-zane, har ma da salon su kamar gimbiya Disney.
source: budenet.ru