Amamodeli Okusabalalisa Okufihlekile
Amamodeli okusabalalisa acashile akhiqiza izithombe ngokusebenzisa inqubo yokusabalalisa endaweni efihlekile ecindezelwe esikhundleni samaphikseli aluhlaza, ukusika izindleko zokubala.
Uhlolojikelele
They are the engine behind Stable Diffusion and most modern open-source image generators.
I-Deep Dive
Imodeli yokusabalalisa evamile ifunda ukubuyisela emuva inqubo yomsindo: iqala emsindweni omsulwa futhi kancane kancane ihlengezele isithombe. Ukwenza lokhu ngokuqondile ngamaphikseli kuyabiza ngoba isithombe esingu-512x512 sinamanani angamakhulu ezinkulungwane. Ukusabalalisa okucashile, okwethulwa u-Rombach nozakwabo ngo-2022, kuqala kusetshenziswa i-autoencoder ehlukile eqeqeshwe kusengaphambili (VAE) ukuze iminyanise isithombe sibe yigridi encane ecashile (ngokuvamile engu-64x64x4, cishe ngama-48x amancane). I-Diffusion U-Net ibe ifunda ukwenza umsindo ngaphakathi kwaleso sikhala esifihlekile esihlangene, siqondiswa umbhalo ngokunaka okuphambene. Ekugcineni i-VAE decoder yakha kabusha amaphikseli anokulungiswa okugcwele. Lokhu kucindezelwa kombono kugcina ulwazi olunengqondo ngenkathi kulahla imininingwane engabonakali, okwenza ukukhiqizwa kwekhwalithi ephezulu kwenzeke kuma-GPU abathengi.
I-Technical Insight
Iqhinga elibalulekile ukwehlukanisa ukucindezelwa kombono kusukela ekumodeleni okukhiqizayo. I-VAE isingatha imininingwane yephikseli yefrikhwensi kanye, futhi i-U-Net imodeli kuphela ukusabalalisa okucashile kobukhulu obuphansi. Isimo sombhalo sijovwa ngezendlalelo zokunakwa okuphambanayo, lapho izici zendawo ze-U-Net zibhekelela ukushumeka kwamathokheni asuka kusifaki khodi sombhalo njenge-CLIP. Ngenxa yokuthi izinto ezifihlekile zincane cishe izikhathi ezingu-48 kunamaphikseli, isinyathelo ngasinye sokukhipha umsindo sishibhe ngokumangazayo kukho kokubili inkumbulo nama-FLOP.
I-Strategic Impact
Isivinini nesikali
I-Visual AI ingakwazi ukuhlola, ukutholwa, nokumaka imisebenzi esikalini.
Yakha ukukhetha
Amathimba aqanjiwe angakwazi ukulinganisa imiqondo ngokushesha ngezibuyekezo ezimbalwa ezenziwa mathupha.
Ithimba kanye nokusebenza komsebenzi
Imisebenzi ingasebenzisa amasiginali wesithombe nawevidiyo obekunzima ukuwenza ngaphambilini.
Ikusasa Lamamodeli Okusabalalisa Okufihlekile
Ukusabalalisa okucashile kunwebeka ngale kwezithombe kuya kuvidiyo (I-Stable Video Diffusion), amafa we-3D, nama-spectrogram alalelwayo, konke kusetshenziswa iresiphi efanayo yokucindezela-bese-denoise. Ucwaningo luphokophela ezinyathelweni ezimbalwa zesampula kusetshenziswa i-distillation kanye namamodeli afanayo, ama-VAE angcono agcina umbhalo nobuso obuhle, kanye nokwakheka kokugeleza okulungisiwe njengalawo aku-Stable Diffusion 3 eqondisa i-trajectory yesizukulwane ukuthola imiphumela esheshayo, ebukhali.
Ukuqaliswa Komhlaba Wangempela
I-Stable Diffusion ekhiqiza umsebenzi wobuciko nemiklamo yomqondo evela emiyalweni yombhalo ku-GPU yomthengi oyedwa
I-Adobe ne-Canva inika amandla umbhalo uye esithombeni kanye nezici zokugcwalisa ngokukhiqizayo ezakhelwe phezu komgogodla ocashile wokusabalalisa
Izitudiyo zegeyimu ezikhiqiza amamephu wokuthungwa, ama-sprites, nobuciko bomqondo wemvelo ukuze kusheshiswe ukukhiqizwa kwangaphambili
Amaqembu ezithombe zesitoko nawokumaketha adala izithombe zemikhiqizo esezindaweni zomkhiqizo kanye nezithombe ezibukwayo ngaphandle kokuthwebula izithombe
Izingozi & Guardrails
Amalungelo ezithombe kanye nemvume kungaba ubungozi bezomthetho uma ukuvela kungacacile.
Ukusebenza kwemodeli kungahluka kukho konke ukukhanya, izibalo zabantu, kanye nezindawo.
Okuhle okungelona iqiniso kungase kungabonakali ngaphandle uma izinga lokuzethemba liqashelwa.
Ukuqalisa Umhlahlandlela
Chaza indlela yokwamukela yokunemba, ukukhumbula, nezindleko zamaphutha.
Hlola ngedatha efana nezimo zangempela zokukhiqiza.
Engeza isibuyekezo somuntu ukuze uthole ukuzethemba okuphansi noma izibikezelo zomthelela omkhulu.
Landelela ukukhukhuleka kwemodeli bese uqinisekisa kabusha ngemva kwezinguquko zekhamera noma zesethi yedatha.
Qhubeka Uhlole
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 Latent Diffusion Models 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
Umhlahlandlela olandelayo
Amamodeli Okusabalalisa Ividiyo
Imibuzo evame ukubuzwa
What is Latent Diffusion Models?
Amamodeli okusabalalisa acashile akhiqiza izithombe ngokusebenzisa inqubo yokusabalalisa endaweni efihlekile ecindezelwe esikhundleni samaphikseli aluhlaza, ukusika izindleko zokubala. Bayinjini engemuva kwe-Stable Diffusion kanye namajeneretha ezithombe zomthombo ovulekile wesimanje.
Iyiphi indlela entsha eqanjiwe yamamodeli okusabalalisa acashile uma kuqhathaniswa nokusabalaliswa kwesikhala kwe-pixel?
Ukusabalaliswa okucashile kucindezela izithombe endaweni encane ecashile kusetshenziswa i-VAE, ngakho-ke ukuhlukanisa okubizayo kwenzeka ngamavelu ambalwa kakhulu kunamaphikseli agcwele.
Iyiphi ingxenye ecindezela isithombe esikhaleni esicashile futhi isisakhe kabusha ngemva kwalokho?
I-VAE eqeqeshwe kusengaphambili ibhala ngekhodi isithombe sibe yigridi efihlekile ehlangene futhi isiqophi saso sakha kabusha amaphikseli anokulungiswa okugcwele ekugcineni.
Umbhalo uvame ukujovwa kanjani ku-U-Net ekhipha umsindo ekusakazekeni okucashile?
Ukushumeka amathokheni okuvela kusifaki khodi sombhalo kufakwa izendlalelo zokunaka okushayisanayo, okuvumela izici zesikhala ukuthi zinakekele ukwaziswa.
Kungani ukusebenza endaweni ecashile kunciphisa izindleko zokubala?
Linda — isizathu esifanele ukuthi igridi ecashile incane kakhulu (imvamisa engu-~48x) kunesithombe se-pixel, ngakho isinyathelo ngasinye sokwenza i-denoising sidinga ama-FLOP ambalwa kakhulu nenkumbulo.
Iyiphi imodeli yomthombo ovulekile esetshenziswa kabanzi edume ngokusabalaliswa okufihlekile?
I-Stable Diffusion, ekhishwe ngo-2022, yaletha ukusakazeka okucashile ku-hardware yabathengi kanye nokutholwa okukhulu.