I-Stability AI ishicilele uhlelo lwesibili lwesistimu yokufunda yomshini we-Stable Diffusion, ekwazi ukuhlanganisa nokulungisa izithombe ngokusekelwe kusifanekiso esihlongozwayo noma incazelo yombhalo wolimi lwemvelo. Ikhodi yokuqeqeshwa kwenethiwekhi ye-neural kanye namathuluzi okukhiqiza izithombe ibhalwe nge-Python kusetshenziswa uhlaka lwe-PyTorch futhi ishicilelwe ngaphansi kwelayisensi ye-MIT. Amamodeli asevele aqeqeshiwe avuliwe ngaphansi kwelayisensi ye-Creative ML OpenRAIL-M evumelayo, evumela ukusetshenziswa kwezentengiso. Ukwengeza, i-demo generator yesithombe esiku-inthanethi iyatholakala.
Ukuthuthukiswa okubalulekile kuhlelo olusha lwe-Stable Diffusion:
- Imodeli entsha yokuhlanganiswa kwesithombe esekelwe encazelweni yombhalo idaliwe - SD2.0-v, esekela ukukhiqizwa kwezithombe ngesinqumo esingu-768x768. Imodeli entsha iqeqeshwe kusetshenziswa iqoqo le-LAION-5B, elihlanganisa izithombe eziyizigidi eziyizinkulungwane ezingu-5.85 ezinezincazelo zombhalo. Imodeli isebenzisa isethi efanayo yamapharamitha njengemodeli ye-Stable Diffusion 1.5, kodwa iyahluka ekushintsheni ekusetshenzisweni kwesishumeki se-OpenCLIP-ViT/H esihluke kakhulu, esithuthukise kakhulu ikhwalithi yezithombe eziwumphumela.
- Inguqulo eyenziwe lula ye-SD2.0-base isilungisiwe, yaqeqeshwa ezithombeni ezingama-256Γ256 kusetshenziswa imodeli yokubikezela umsindo yakudala futhi isekela ukukhiqizwa kwezithombe ngesixazululo esingu-512Γ512.
- Kungenzeka ukusebenzisa ubuchwepheshe be-supersampling (I-Super Resolution) ukuze ukhuphule ukulungiswa kwesithombe sangempela ngaphandle kokunciphisa ikhwalithi, usebenzisa ukukala kwendawo kanye nama-algorithms wokwakha kabusha imininingwane. Imodeli yokucubungula isithombe enikeziwe (i-SD20-upscaler) isekela ukusondeza izikhathi ezine, okuvumela ukukhiqizwa kwezithombe ngesixazululo esingu-2048x2048.
- Kuhlongozwa imodeli ye-SD2.0-depth2img, kucatshangelwa ukujula nokuhlelwa kwendawo kwezinto. Ukulinganisa ukujula kwe-monocular, kusetshenziswa uhlelo lwe-MiDaS. Imodeli ikuvumela ukuthi uhlanganise izithombe ezintsha usebenzisa esinye isithombe njengesifanekiso, esingahluka kakhulu kwesasekuqaleni, kodwa ugcine ukwakheka nokujula sekukonke. Isibonelo, ungasebenzisa ukuma komuntu esithombeni ukuze wakhe omunye umlingisi endaweni efanayo.
- Imodeli yokulungisa izithombe ibuyekeziwe - i-SD 2.0-inpainting, ekuvumela ukuthi ushintshe futhi uguqule izingxenye zesithombe usebenzisa ukwaziswa kombhalo.
- Amamodeli alungiselelwe ukusetshenziswa kumasistimu avamile nge-GPU eyodwa.
Source: opennet.ru