Architecture StyleGAN
StyleGAN ab reso adversarial la bu bawoo ci NVIDIA buy defar kanam yu dëggu ak mbir yu yéeme ci dugal ay leeral ci stil ci bepp layer.
Résumé
It matters because its design gives unprecedented, disentangled control over coarse and fine image attributes.
Plongeur bu xóot
StyleGAN, Karras ak ñeneen ñoo ko dugal. ci 2018, defaraat generatëru GAN ci xalaatu 'stil.' Du dugal benn vecteur aleatoire ci biir reso bi, dafay njëkka karte kode latente z jaaraleko ci benn MLP bu am 8 couche ci biir espace intermédiaire W, biy dindi fakteer yiy soppiku. Tensor bu fippu buñ jàng ñu ngi koy yokk ndànk-ndànk, ci dayo bu nekk vecteur stil bi dafay soppi kàrtu màndarga yi jaaraleko ci Normalisation Instance Adaptive (AdaIN), di saytu màndarga yi tàmbalee ci pose (couche yu dëgër) ba ci texture der (couche yu woyof yi). Dugalug bruit bu nekk dafay yokk ay detay stochastic yu melni freckles ak kawar yu réer. StyleGAN2 (2020) dafa wecci AdaIN ak demodulaasioŋ diisaay ngir dindi 'blob' yiñ defar, ba noppi StyleGAN3 (2021) dafa defar pseudo buy tafu ngir fexe ba man-man yi di toxal seen bopp ci diiru animaasioŋ.
Gis-gis xarala
Mekanism bi gëna am solo mooy modulasioŋ bu sukkandiko ci stil. Reseau kartograafi bi dafay soppi z ci w, te soppi affine yi ñu jàng soppi w ci eskaalu chaine bu nekk ak biais buñ def ci kàrtu màndarga yiñ normalisee ci dayo bu nekk. Ndax stil yi dañuy doxal couche par couche, mën nga jaxase w bu benn nataal ci couche yu dëgër ak beneen ci couche yu woyof yi ('jaxase stil') ngir weccoo pose te baña bàyyi texture bi. Demodulation bu StyleGAN2 dafay boole lim yii ci diisaayu konvolusioŋ, di dindi mbiri normalisasioŋ.
njeextalu pexe
Gaawaay ak yaatuwaay
Visual IA mën na otomatise saytu, gis ak etiketu liggéey ci eskaal.
Tabax tànneef
Ekipu kreatif yi mën nañu defar konsept yu gëna gaaw te duñu def lu bari ci loxo.
Ekip ak def liggéey
Liggéeyukaay yi mën nañu jëfandikoo siñaal nataal wala wideo yu jafewoon lool ci liggéey.
Ëlëgu architecture StyleGAN
Dema yaakaroon ni xeetu diffusion yi ñoo jiite generation text-to-image, StyleGAN's espace latent buñ mëna soppali (W ak W+) moo koy tëye ci diggu soppali kanam, jaay màndarga, ak synthese ci jamono dëgg, fu GANs di gëna gaaw. Xaarandi liggéey bu wéy ci GAN inversion (projection nataal dëgg ci W), 3D-xam anam yu melni EG3D yuy joxe gis-gis yu méngoo, ak ibrid yuy boole latents yuñ mëna saytu StyleGAN ak diffusion wala transformateur priors ngir gëna baax ci ñaari àdduna yi.
Doxal ci àdduna dëgg
Defar kanam nit ñu bari te amul àpp, ni ko thispersondoesnotexist.com wane.
Soppi kanam gu semantik: soppi at yi, melokaan bi, wala pose bi ci toxal ci yoon wi ci espace W bi.
Sosal ay done tàggat yaram ak ay avatar sudee nataal yu dëggu te wóor yu nëbbu bariwul.
Jumtukaayi art yuy boole wala 'jaxase stil' ci diggante nataal ngir jaxase jëmmal ak detay yu ndaw.
Risk yi ak balustrade yi
Yelleefi nataal ak nangu mën na nekk risku yoon sudee fi ñu bawoo leerul.
Performance model bi mën na wuute ci leeraay bi, demographie bi ak environmaa bi.
Njuumteg positive yi mën nañu dem te kenn duko seetlu fileek xool wuñu buntu wóolu sa bopp.
Roadmap ngir samp gi
Mandargal kritërium nangug njub, woowaat ak njëgu njuumte.
Saytu ak done yu méngoo ak anam yi ñuy liggéeyee dëgg.
Yokk jàngat nit ngir xam fu wóorul dara wala am njeexital yu rëy.
Toppal model drift bi nga baaxal ko ginaaw bi kamera bi wala done yi soppeekoo.
Weyal di banneexu
Free newsletter
Get the daily AI briefing
Three verified AI stories every weekday morning, written in plain English. Free forever, no ads.
One email each weekday. Unsubscribe in one click. We never sell or share your address.
Test yourself
Take the StyleGAN Architecture quiz
Instant feedback on every answer, and a shareable certificate with a verifiable ID once you pass a course.
Support free AI education. AI Understanding is a 501(c)(3) nonprofit — no ads, no paywall, ever. Make a donation
Gis bi ci topp
Architecture U-Net
Laaj yi ñuy faral di laaj
What is StyleGAN Architecture?
StyleGAN ab reso adversarial la bu bawoo ci NVIDIA buy defar kanam yu dëggu ak mbir yu yéeme ci dugal ay leeral ci stil ci bepp layer. Dafa am solo ndax jëmmal dafay may kàttan bu musul am, bu jaar yoon ci kaw màndarga nataal yu ñaaw ak yu rafet.
Luy njariñu reso kartograafi bu StyleGAN?
Benn MLP bu am 8 couche dafay màndargaal z ba W, di barab bu nëbbu ci digg bi, foofu lañuy gëna mëna tàqale fakteer yiy soppiku, suko defee ñu mëna doxal bu baax.
Ci StyleGAN bi njëkk, naka lañuy dugal stil ci kàrtu màndarga yi?
AdaIN dafay normalise kàrtu màndarga yi ba noppi di leen eskale ak di leen soppi ci jëfandikoo lim yu bawoo ci stil ci dayo bu nekk.
Lan la reso synthese bi di tàmbalee ci barabu vecteur latente bi?
StyleGAN mingi tàmbali ci dugal bu sax buñ jàng, ba noppi dugal stil ak bruit bi muy upsamples, moo gën di dundal z ci saasi.
Lan mooy njëkka saytu defar stil yi ci couche yu dëgër yi (resolution bu woyof)?
Couche yu dëgër yi ñoo yor jëmmal bu yaatu lu ci melni pose ak forme générale, waaye couche yu woyof yi ñooy yor texture bi ak ay detay yu ndaw.
Ban jafe-jafe la StyleGAN2 saafara buñu ko méngale ak li njëkk?
StyleGAN2 dafa wecci AdaIN ak demodulation diisaay, dindi tiflit/blob artefact ak gëna baaxal kalite nataal bi.