Iqembu le Abacwaningi base-Tel Aviv University basanda kwethula i-HyperStyle, okuyi inguqulo ehlehlisiwe yesistimu yokufunda yomshini I-NVIDIA StyleGAN2 eklanywe kabusha ukuze idale kabusha izingcezu ezingekho lapho uhlela izithombe zomhlaba wangempela.
I-StyleGAN ibonakala ngokuvumela ukuhlanganisa ubuso obusha babantu ngokubukeka okungokoqobo, ukusetha amapharamitha afana neminyaka, ubulili, ubude bezinwele, umlingiswa wokumamatheka, ukuma kwekhala, umbala wesikhumba, izibuko kanye ne-engeli yesithombe.
Ngakolunye uhlangothi, I-HyperStyle yenza kube nokwenzeka ukushintsha amapharamitha afanayo kulawo akhona, Ngamanye amazwi, ikuvumela ukuthi udale izithombe ngaphandle kokuguqula izici zazo zesici nokugcina ukubonakala kobuso bokuqala.
I-HyperStyle yethula ama-hypernetwork ukuze ifunde ukuthi zingacwengwa kanjani izisindo zejeneretha ye-StyleGAN eqeqeshwe ngaphambilini ngokuhlobene nesithombe sokufaka esinikeziwe. Ukwenza kanjalo kunika amandla ukwakhiwa kabusha kweleveli enezikhathi zokuqondisisa ezifana ne-encoder kanye nokuhleleka okuphezulu.
Isibonelo, uma usebenzisa I-HyperStyle, ingalingisa ushintsho eminyakeni yobudala bomuntu osesithombeni, shintsha isitayela sezinwele, engeza izibuko, intshebe noma amadevu, yenza isithombe sibukeke njengomlingiswa wekhathuni noma isithombe esidwetshwe ngesandla, veza ubuso obudabukile noma obujabulisayo.
Kulokhu, Uhlelo lungaqeqeshwa hhayi kuphela ukushintsha ubuso babantu, kodwa futhi nganoma iyiphi into, ngoba Isibonelo, ukuhlela izithombe zemoto.
Iningi lemisebenzi efunda ukuguqulwa ibheka ikhodi efihlekile eyakha kabusha ngokunembe kakhulu isithombe esinikeziwe. Omunye umsebenzi wakamuva wenze isiphakamiso sokulungiswa kahle kwesithombe kwezisindo ezikhiqizayo ukuze kuzuzwe ukwakhiwa kabusha kwekhwalithi ephezulu yesithombe esiqondiwe esinikeziwe. Nge-HyperStyle, sihlose ukuletha lezi zindlela zokushuna ijeneretha endaweni yezinhlelo zokusebenza ezisebenzisanayo ngokuzivumelanisa nendlela esuselwe kusifaki khodi.
Siqeqeshe i-hypernetwork eyodwa ukuze sifunde ukuthi singacwengisiswa kanjani izisindo ezikhiqizayo uma ziqhathaniswa nesithombe esiqondisiwe. Ngokufunda le mephu, i-HyperStyle ibikezela kahle izisindo eziqondiwe zejeneretha ngaphansi kwamasekhondi angu-2 ngesithombe ngasinye, iyenze isebenze ezinhlobonhlobo zezinhlelo zokusebenza.
Indlela ehlongozwayo ihlose ukuxazulula inkinga yokwakha kabusha izingxenye ezilahlekile zesithombe ngesikhathi sokuhlela. Amasu ahlongozwe ngenhla abhekane nebhalansi phakathi kokwakha kabusha nokuhlela ngokulungisa kahle umthwebuli wesithombe ukuze athathele izingxenye zesithombe esiqondiwe kuyilapho kwakhiwa kabusha izifunda ezihlelekayo ebezingekho ekuqaleni. Ububi bezindlela ezinjalo yisidingo sokuqeqeshwa okuhlosiwe kwesikhathi eside kwenethiwekhi ye-neural yesithombe ngasinye.
Indlela esekelwe ku-algorithm ye-StyleGAN ivumela ukusebenzisa imodeli evamile, iqeqeshelwe kusengaphambili emaqoqweni avamile ezithombe, ukukhiqiza izici zesici zesithombe sangempela ngezinga lokuzethemba eliqhathaniswa nama-algorithms adinga ukuqeqeshwa komuntu ngamunye kwemodeli yesithombe ngasinye.
Enye yezinzuzo zendlela entsha yikhono lokuguqula izithombe ngokusebenza eduze nesikhathi sangempela, ngaphezu kweqiniso lokuthi imodeli isilungele ukuqeqesha okulungiselelwe labo bantu, izimoto nezilwane ngokusekelwe kumaqoqo kusukela ku-Flickr-the Faces-HQ (FFHQ, izithombe ze-PNG zekhwalithi ephezulu ezingu-70,000 zobuso babantu), I-Stanford Cars (izithombe ezingu-16 zezimoto) kanye ne-AFHQ (izithombe zezilwane).
Futhi, isethi yamathuluzi inikezwa ukuqeqesha amamodeli akhokanye namamodeli aqeqeshiwe asekulungele ukusetshenziswa ezishumeki ezijwayelekile namajeneretha alungele ukusetshenziswa nazo. Isibonelo, kukhona amajeneretha atholakalayo okudala izithombe zesitayela se-Toonify, izinhlamvu ze-Pstrong, ukudala imidwebo, ngisho nokwenza isitayela njengamakhosazana e-Disney.
Okokugcina kulabo abathanda ukwazi okwengeziwe Mayelana naleli thuluzi, ungabheka imininingwane Kulesi sixhumanisi esilandelayo.
Kubalulekile futhi ukusho ukuthi ikhodi ibhalwe nge-Python kusetshenziswa uhlaka lwe-PyTorch futhi inelayisensi ye-MIT. Ungahlola ikhodi ku isixhumanisi esilandelayo.