Ib pab pawg ntawm cov kws tshawb fawb los ntawm Tel Aviv University tau nthuav tawm HyperStyle, qhov hloov pauv ntawm NVIDIA's StyleGAN2 tshuab kev kawm uas tau raug kho dua tshiab los tsim qhov ploj lawm thaum kho cov duab hauv lub neej. Cov cai sau rau hauv Python siv PyTorch lub moj khaum thiab muab faib raws li MIT daim ntawv tso cai.
Yog tias StyleGAN tso cai rau koj los tsim kom muaj lub ntsej muag tshiab ntawm tib neeg los ntawm kev teeb tsa tsis xws li hnub nyoog, poj niam los txiv neej, plaub hau ntev, luag ntxhi, lub qhov ntswg zoo, xim tawv nqaij, tsom iav thiab yees duab lub kaum ntse ntse, HyperStyle ua rau nws muaj peev xwm hloov qhov tsis sib xws hauv cov uas twb muaj lawm. cov duab tsis hloov lawv cov yam ntxwv nta thiab tuav lub recognizability ntawm thawj lub ntsej muag. Piv txwv li, siv HyperStyle, koj tuaj yeem simulate qhov kev hloov pauv ntawm lub hnub nyoog ntawm tus neeg hauv daim duab, hloov cov plaub hau, ntxiv tsom iav, hwj txwv lossis hwj txwv, ua cov duab zoo li cov duab tas luav lossis kos duab kos duab, ua ib daim duab. kev tu siab lossis kev zoo siab. Hauv qhov no, lub kaw lus tuaj yeem raug cob qhia tsis tsuas yog hloov lub ntsej muag ntawm tib neeg, tab sis kuj rau txhua yam khoom, piv txwv li, rau kho cov duab ntawm lub tsheb.
Txoj kev npaj yog tsom rau kev daws qhov teeb meem nrog kev tsim kho ntawm qhov ploj ntawm daim duab thaum kho. Nyob rau hauv cov txheej txheem dhau los, qhov kev sib haum xeeb ntawm kev tsim kho thiab kho kom haum xeeb tau raug daws los ntawm kev kho kom zoo ntawm lub tshuab hluav taws xob duab los hloov qhov chaw ntawm lub hom phiaj duab thaum rov tsim dua qhov chaw uas tsis tuaj yeem kho tau. Qhov tsis zoo ntawm cov txheej txheem zoo li no yog qhov xav tau rau kev cob qhia mus sij hawm ntev ntawm neural network rau txhua daim duab.
Cov txheej txheem raws li StyleGAN algorithm ua rau nws muaj peev xwm siv tau tus qauv zoo, yav tas los tau kawm ntawm kev sau ntau ntawm cov duab, los tsim cov yam ntxwv ntawm cov duab qub nrog qib kev ntseeg siab piv rau cov algorithms uas xav tau kev cob qhia tus kheej ntawm tus qauv rau txhua daim duab. . Ntawm qhov zoo ntawm txoj kev tshiab, muaj peev xwm hloov kho cov duab nrog kev ua haujlwm ze rau lub sijhawm tiag tiag kuj tau sau tseg.
Cov qauv kev cob qhia ua ntej tau npaj rau tib neeg, lub tsheb thiab lub ntsej muag tsiaj raws li kev sau ntawm Flickr-Faces-HQ (FFHQ, 70k zoo PNG duab ntawm tib neeg lub ntsej muag), Stanford Cars (16k duab ntawm lub tsheb) thiab AFHQ (duab duab ntawm tsiaj). Tsis tas li ntawd, cov cuab yeej tau muab rau kev cob qhia lawv cov qauv, nrog rau kev npaj ua qauv ntawm cov encoders thiab generators tsim nyog siv nrog lawv. Piv txwv li, cov tshuab hluav taws xob muaj los tsim Toonify-style dluab, Pixar cim, sketching, thiab txawm styling lawv zoo li Disney Princesses.
Tau qhov twg los: opennet.ru