Iqela labaphandi abavela kwiDyunivesithi yaseTel Aviv thaca iHyperStyle, inguqulelo eguqulweyo ye-NVIDIA's StyleGAN2 inkqubo yokufunda ngomatshini ohlaziywe ngokutsha ukuze enze iindawo ezingekhoyo xa uhlela imifanekiso yokwenyani. Ikhowudi ibhalwe kwiPython usebenzisa isakhelo sePyTorch kwaye ihanjiswa phantsi kwelayisenisi yeMIT.
Ukuba i-StyleGAN ikuvumela ukuba wenze ubuso obutsha obunembonakalo yokwenyani yabantu ngokuchaza iiparameters ezinje ngobudala, isini, ubude beenwele, umlinganiswa woncumo, imilo yempumlo, umbala wolusu, iiglasi kunye ne-angle yefoto, emva koko i-HyperStyle yenza kube lula ukutshintsha iiparameter ezifanayo kwizinto ezikhoyo. iifoto ngaphandle kokutshintsha iimpawu zazo ngelixa zigcina ukuqondwa kobuso bokuqala. Ngokomzekelo, usebenzisa i-HyperStyle, unokwenza utshintsho kwiminyaka yobudala bomntu kwifoto, utshintshe i-hairstyle, ungeze iiglasi, iindevu okanye amadevu, unike umfanekiso wokubonakala komlingiswa wekhathuni okanye umzobo owenziwe ngesandla, wenze. imbonakalo yobuso elusizi okanye echwayitileyo. Ngaphezu koko, inkqubo inokuqeqeshwa kungekhona nje ukutshintsha ubuso babantu, kodwa nakweziphi na izinto, umzekelo, ukuhlela imifanekiso yeemoto.
Indlela ecetywayo ijolise ekusombululeni ingxaki yokwakhiwa kwakhona kweendawo ezilahlekileyo zomfanekiso ngexesha lokuhlela. Kwiindlela ezicetywayo ngaphambili, ukurhweba phakathi kokwakhiwa ngokutsha kunye nokuhlelwa kwasombululwa ngokulungelelanisa i-generator yemifanekiso ukuze ithathe indawo yomfanekiso ojoliswe kuyo xa uphinda uphinda ulahle imimandla ehlelekayo ekuqaleni. Ukungalungi kweendlela ezinjalo yimfuneko yoqeqesho olujoliswe kwixesha elide lwe-neural network kumfanekiso ngamnye.
Indlela esekelwe kwi-algorithm ye-StyleGAN yenza kube nokwenzeka ukuba kusetyenziswe imodeli eqhelekileyo, eqeqeshwe kwangaphambili kwiingqokelela eziqhelekileyo zemifanekiso, ukuvelisa iimpawu zeempawu zomfanekiso wokuqala kunye nenqanaba lokuthembeka elithelekiseka kwii-algorithms ezifuna uqeqesho lomntu ngamnye kwimodeli nganye. umfanekiso. Enye inzuzo yendlela entsha kukukwazi ukuguqula imifanekiso kunye nokusebenza ngokusondeleyo kwixesha langempela.
Iimodeli eziqeqeshiweyo esele zenziwe zilungiselelwe ubuso babantu, iimoto kunye nezilwanyana ezisekelwe kwiingqokelela zeFlickr-Faces-HQ (FFHQ, i-70 lamawaka eemifanekiso ezikumgangatho ophezulu wePNG wobuso babantu), iimoto zaseStanford (i-16 amawaka emifanekiso yeemoto) kunye AFHQ (iifoto zezilwanyana). Ukongezelela, izixhobo zokuqeqesha iimodeli zakho zibonelelwe, kunye neemodeli eziqeqeshiweyo esele zenziwe ze-encoders eziqhelekileyo kunye neejeneretha ezifanelekileyo zokusetyenziswa nazo. Umzekelo, iijenereyitha ziyafumaneka ekudaleni imifanekiso ye-Toonify-style, abalinganiswa bePixar, ukuvelisa imizobo, kunye nokuhombisa amakhosazana kwiikhathuni zeDisney.
umthombo: opennet.ru