Ithimba labacwaningi abavela eNyuvesi yase-Tel Aviv lethule i-HyperStyle, inguqulo ehlanekezelwe yesistimu yokufunda yomshini ye-NVIDIA's StyleGAN2 eklanywe kabusha ukuze idale kabusha izingxenye ezingekho lapho uhlela izithombe zangempela. Ikhodi ibhalwe nge-Python kusetshenziswa uhlaka lwe-PyTorch futhi isatshalaliswa ngaphansi kwelayisense ye-MIT.
Uma i-StyleGAN ikuvumela ukuthi uhlanganise ubuso obusha babantu obubukeka bungokoqobo ngokucacisa imingcele efana neminyaka, ubulili, ubude bezinwele, umlingiswa wokumamatheka, ukuma kwekhala, umbala wesikhumba, izibuko kanye ne-angle yesithombe, i-HyperStyle yenza kube nokwenzeka ukushintsha imingcele efanayo ekhona. izithombe ngaphandle kokushintsha izici zazo zesici ngenkathi zigcina ukuqashelwa kobuso bokuqala. Isibonelo, usebenzisa i-HyperStyle, ungakwazi ukulingisa ushintsho eminyakeni yobudala bomuntu osesithombeni, ushintshe isitayela sezinwele, ungeze izibuko, intshebe noma amadevu, unikeze isithombe ukubukeka komlingiswa wekhathuni noma umdwebo odwetshwe ngesandla, wenze isimo sobuso esidabukisayo noma esijabule. Ngaphezu kwalokho, uhlelo lungaqeqeshwa hhayi kuphela ukushintsha ubuso babantu, kodwa nanoma yiziphi izinto, isibonelo, ukuhlela izithombe zezimoto.
Indlela ehlongozwayo ihloselwe ukuxazulula inkinga yokwakha kabusha izingxenye ezingekho zesithombe phakathi nokuhlela. Ezindleleni ezihlongozwe ngaphambilini, ukuhwebelana phakathi kokwakhiwa kabusha nokuhleleka kwaxazululwa ngokulungisa kahle ijeneretha yesithombe ukuze ithathele izingxenye zesithombe esiqondiwe lapho idala kabusha izifunda ezingahleleka ekuqaleni. Ububi bezindlela ezinjalo yisidingo sokuqeqeshwa okuhlosiwe kwesikhathi eside kwenethiwekhi ye-neural yesithombe ngasinye.
Indlela esuselwe ku-algorithm ye-StyleGAN yenza kube nokwenzeka ukusebenzisa imodeli evamile, eqeqeshwe kusengaphambili emaqoqweni avamile ezithombe, ukuze kukhiqizwe izici zesici sesithombe sokuqala ngezinga lokwethembeka eliqathaniswa nama-algorithms adinga ukuqeqeshwa komuntu ngamunye kwemodeli ngayinye. isithombe. Enye inzuzo yendlela entsha yikhono lokuguqula izithombe ngokusebenza eduze nesikhathi sangempela.
Amamodeli aqeqeshiwe esenziwe ngomumo alungiselelwa ubuso babantu, izimoto nezilwane ngokusekelwe kumaqoqo e-Flickr-Faces-HQ (FFHQ, izithombe eziyizinkulungwane ezingama-70 zekhwalithi ephezulu ye-PNG yobuso babantu), Izimoto zaseStanford (izithombe eziyizinkulungwane eziyi-16 zezimoto) kanye AFHQ (izithombe zezilwane). Ukwengeza, amathuluzi okuqeqesha amamodeli akho anikeziwe, kanye namamodeli aqeqeshiwe enziwe kakade amakhodi ajwayelekile namajeneretha afanele ukusetshenziswa nawo. Isibonelo, amajeneretha ayatholakala ekudaleni izithombe zesitayela se-Toonify, izinhlamvu ze-Pixar, ukukhiqiza imidwebo, ngisho nokwenza isitayela amakhosazana avela emakhathuni e-Disney.
Source: opennet.ru