Ucwaningo Olusha Lembula Ubungozi Kumathuluzi Adumile Okuvikela Ubuciko Ngokumelene Nokwebiwa kwe-AI

Ithebula Lezixhumanisi

Abstract kanye 1. Isingeniso

2. Isendlalelo kanye Nomsebenzi Ohlobene

3. Usongo Model

4. I-Robust Style Mimicry

5. Ukusethwa Kokulinga

6. Imiphumela

   6.1 Okutholakele Okumqoka: Konke Ukuvikela Kuzungeza Kalula

   6.2 Ukuhlaziya

7. Ingxoxo kanye Nomphumela Obanzi, Ukubonga, Nezikhombo

A. Izibonelo Zobuciko Eziningiliziwe

B. Robust Mimicry Generations

C. Imiphumela enemininingwane

D. Umehluko nge-Glaze Finetuning

E. Okutholakele ku-Glaze 2.0

F. Okutholakele ku-Mist v2

G. Izindlela Zokulingisa Isitayela

H. Isitayela Esikhona Ukuvikela Ukulingisa

I. Izindlela Zokulingisa Eziqinile

J. Ukusethwa Kokulinga

K. Isifundo somsebenzisi

L. Compute Resources

Abstract

Amaciko aya ngokuya ekhathazeka ngentuthuko yamamodeli okukhiqiza izithombe angaphinda aphindaphinde izitayela zawo zobuciko ezihlukile. Ukuphendula, amathuluzi amaningana okuvikela ngokumelene nokulingisa isitayela athuthukisiwe ahlanganisa ukuphazamiseka okuncane emisebenzini yobuciko eshicilelwe ku-inthanethi. Kulo msebenzi, sihlola ukusebenza kokuvikela okudumile—okunezigidi zokulanda—futhi sibonisa ukuthi kunikeza kuphela umuzwa wokuvikeleka okungamanga. Sithola ukuthi umzamo ophansi kanye namasu "ongekho eshalofini", njengokukhushulwa kwesithombe, anele ukudala izindlela zokulingisa eziqinile ezehlisa kakhulu ukuvikela okukhona. Ngocwaningo lomsebenzisi, sibonisa ukuthi konke ukuvikela okukhona kungadlulwa kalula, kushiye amaciko esengozini yokulingisa isitayela. Siyaxwayisa ngokuthi amathuluzi asekelwe ekuphazamisekeni kwezitha awakwazi ukuvikela amaciko ngendlela ethembekile ekusebenziseni kabi i-AI ekhiqizwayo, futhi sikhuthaze ukuthuthukiswa kwezinye izixazululo zokuvikela.

1 Isingeniso

Ukulingisa isitayela wuhlelo lokusebenza oludumile lwamamodeli okukhiqiza umbhalo uye esithombeni. Ngokunikezwa kwezithombe ezimbalwa ezivela kumdwebi, imodeli ingalungiswa ukuze kukhiqizwe izithombe ezintsha ngaleso sitayela (isb, umkhumbi-mkhathi ngesitayela sikaVan Gogh). Kodwa ukulingisa kwesitayela kunamandla okubangela umonakalo omkhulu uma kusetshenziswe kabi. Ikakhulukazi, abaculi abaningi besimanje bakhathazeka ngokuthi abanye manje sebengakhiqiza izithombe ezikopisha isitayela sabo sobuciko esihlukile, futhi okungenzeka bantshontshe amakhasimende (Heikkila¨, 2022). Njengempendulo, ukuvikela okuningana kuye kwathuthukiswa ukuvikela amaciko ekulingiseni isitayela (Shan et al., 2023a; Van Le et al., 2023; Liang et al., 2023). Lokhu kuvikela kwengeza ukuphazamiseka okuphikisayo ezithombeni abaculi abazishicilela ku-inthanethi, ukuze kuvinjwe inqubo yokuthuthukisa. Lezi zivikelo zithole ukunakwa okubalulekile kwabezindaba—ezinezici ku-New York Times (Hill, 2023), CNN (Thorbecke, 2023) kanye ne-Scientific American (Leffer, 2023)—futhi zilandwe izikhathi ezingaphezu kwe-1M (Shan et al. , 2023a).

Nokho, akucaci ukuthi la mathuluzi empeleni avikela kangakanani amaciko ekulingiseni isitayela, ikakhulukazi uma othile ezama ukuwagwema (Radiya-Dixit et al., 2021). Kulo msebenzi, sibonisa ukuthi amathuluzi asezingeni eliphezulu okuvikela isitayela—Glaze (Shan et al., 2023a), Mist (Liang et al., 2023) kanye ne-Anti-DreamBooth (Van Le et al., 2023) -ayisebenzi uma ibhekene nezindlela ezilula zokulingisa eziqinile. Izindlela zokulingisa eziqinile esizicabangelayo zisukela kumaqhinga omzamo omncane—njengokusebenzisa isikripthi sokulungisa esihlukile, noma ukungeza umsindo we-Gaussian ezithombeni ngaphambi kokuqeqeshwa—kuya kumasu wezinyathelo eziningi ahlanganisa amathuluzi angekho eshalofini. Siqinisekisa imiphumela yethu ngocwaningo lomsebenzisi, oluveza ukuthi izindlela zokulingisa eziqinile zingakhiqiza imiphumela engaqondakali ngekhwalithi kuleyo etholwe emisebenzini yobuciko engavikelekile (bona Umfanekiso 1 ukuze uthole isibonelo esinemifanekiso).

Sibonisa ukuthi amathuluzi akhona okuvikela ahlinzeka nje ngomuzwa ongamanga wokuphepha. Izindlela zethu zokulingisa eziqinile azidingi ukuthuthukiswa kwamathuluzi amasha noma izindlela zokushuna kahle, kodwa ngokucophelela

ukuhlanganisa amasu ajwayelekile okucubungula izithombe abesevele ekhona ngesikhathi la mathuluzi okuvikela ethulwa okokuqala!. Ngakho-ke, sikholelwa ukuthi ngisho nabaqambi abanamakhono aphansi bebengaweqa kalula la mathuluzi kusukela asungulwa.

Nakuba sihlola amathuluzi athile okuvikela akhona namuhla, imikhawulo yokuvikelwa kokulingisa kwesitayela ingokwemvelo. Amaciko asesimweni esibi ngempela njengoba kufanele athathe isinyathelo kuqala (okungukuthi, uma othile elanda ubuciko obuvikelwe, ukuvikela ngeke kusashintshwa). Ukuze asebenze kahle, amathuluzi okuvikela abhekana nomsebenzi oyinselele wokudala ukuphazamiseka okudlulisela kunoma iyiphi indlela yokuqinisa, ngisho nalezo ezikhethwe ngokuzivumelanisa nezimo esikhathini esizayo. Isiphetho esifanayo sithathwe nguRadiya-Dixit et al. (Radiya-Dixit et al., 2021), owaphikisa ngokuthi ukuphazamiseka okuphikisayo akukwazi ukuvikela abasebenzisi kumasistimu okubona ubuso. Ngakho siyaxwayisa ukuthi amasu okufunda emishini ephikisanayo ngeke akwazi ukuvikela amaciko ngendlela ethembekile ekulingiseni isitayela esikhiqizayo, futhi sikhuthaze ukuthuthukiswa kwezinye izindlela zokuvikela amaciko.

Sidalule imiphumela yethu kumathuluzi okuvikela athintekile ngaphambi kokushicilelwa, ukuze anqume inkambo engcono kakhulu yesenzo kubasebenzisi abakhona.