I-Stability AI ishicilele amamodeli aselungile esistimu yokufunda yomshini we-Stable Diffusion, ekwazi ukuhlanganisa nokulungisa izithombe ngokusekelwe encazelweni yombhalo ngolimi lwemvelo. Amamodeli anikezwe ilayisense ngaphansi kwelayisensi ye-Creative ML OpenRAIL-M evumelekile ukuze asetshenziselwe ukuthengisa. Ukuze uqeqeshe isistimu, kusetshenziswe iqoqo lama-4000 NVIDIA A100 Ezra-1 GPUs kanye neqoqo le-LAION-5B, okuhlanganisa izithombe eziyizigidi eziyizinkulungwane ezingu-5.85 ezinezincazelo zombhalo, zisetshenzisiwe. Ngaphambilini, ikhodi yamathuluzi okuqeqesha inethiwekhi ye-neural kanye nokukhiqiza izithombe yayivuliwe ngaphansi kwelayisensi ye-MIT.
Ukutholakala kwemodeli eyenziwe ngomumo kanye nezidingo zesistimu ezinesizotha ezivumela umuntu ukuthi aqale izivivinyo ku-PC enama-GPU ajwayelekile kuholele ekuveleni kwenani lamaphrojekthi ahlobene:
- textual-inversion (ikhodi) - isengezo esikuvumela ukuthi uhlanganise izithombe ngohlamvu olunikeziwe, into noma isitayela. Ku-Stable Diffusion yasekuqaleni, izinto ezisezithombeni ezihlanganisiwe azihleliwe futhi azilawuleki. Isengezo esihlongozwayo sikuvumela ukuthi ungeze izinto zakho ezibonakalayo, uzibophe kumagama angukhiye futhi uwasebenzise ekuhlanganiseni.
Isibonelo, ku-Stable Diffusion evamile ungacela isistimu ukuthi ikhiqize isithombe "ngekati esikebheni". Ukwengeza, ungakwazi ukucacisa izici zekati kanye nesikebhe, kodwa akubikezeli ukuthi iyiphi ikati nesikebhe esizohlanganiswa. Ukuguqulwa kombhalo kukuvumela ukuthi uqeqeshe isistimu esithombeni sekati lakho noma isikebhe futhi uhlanganise isithombe ngekati noma isikebhe esithile. Ngendlela efanayo, ingakwazi futhi ukufaka esikhundleni sezinto zesithombe ngezinto ezithile, ibeke isibonelo sesitayela esibonakalayo sokuhlanganiswa, futhi icacise imiqondo (isibonelo, kuzo zonke izinhlobonhlobo zodokotela, ungasebenzisa ukukhetha okunembe kakhudlwana nokusezingeni eliphezulu. ngesitayela osifunayo).
- stable-diffusion-animation - ukudalwa kwezithombe ezigqwayizayo (ezinyakazayo) ezisuselwe ekuhumusheni phakathi kwezithombe ezikhiqizwe ku-Stable Diffusion.
- stable_diffusion.openvino (ikhodi) - imbobo ye-Stable Diffusion, esebenzisa i-CPU kuphela ukubala, evumela ukuhlolwa kumasistimu angenawo ama-GPU anamandla. Idinga iphrosesa esekelwe kulabhulali ye-OpenVINO. Ngokusemthethweni, i-OpenVINO inikeza ama-plugin ama-Intel processors anezandiso ze-AVX2, AVX-512, AVX512_BF16 ne-SSE, kanye namabhodi we-Raspberry Pi 4 Model B, Apple Mac mini kanye ne-NVIDIA Jetson Nano. Ngokungafanele, kungenzeka ukusebenzisa i-OpenVINO kuma-processor we-AMD Ryzen.
- I-sdamd iyichweba le-AMD GPUs.
- Ukuqaliswa kokuqala kokuhlanganiswa kwevidiyo.
- i-stable-diffusion-gui, i-stable-diffusion-ui, i-Artbreeder Collage, i-diffuse-the-rest - izixhumanisi zesithombe zokukhiqiza izithombe kusetshenziswa i-Stable Diffusion.
- i-beta.dreamstudio.ai, Izikhala Zobuso Obugonayo, i-hlky Stable Diffusion WebUI - izixhumanisi zewebhu zokwenziwa kwezithombe kusetshenziswa i-Stable Diffusion.
- Ama-plugin okuhlanganisa i-Stable Diffusion ne-GIMP, i-Figma, i-Blender ne-Photoshop.
Ukwengeza, singabona ukushicilelwa yi-Google kwekhodi yesistimu yokufunda yomshini we-RawNeRF (RAW Neural Radiance Fields), evumela, ngokusekelwe kudatha evela ezithombeni ezimbalwa ze-RAW, ukuthuthukisa ikhwalithi yezithombe ezinomsindo omkhulu ezithathwe ebumnyameni nasemotweni. ukukhanya okubi. Ngaphezu kokuqeda umsindo, amathuluzi athuthukiswe iphrojekthi enza kube nokwenzeka ukwandisa imininingwane, ukuqeda ukukhanya, ukuhlanganisa i-HDR nokuguqula ukukhanya okuphelele ezithombeni, kanye nokwenza kabusha isimo sezinto ezintathu-ntathu sezinto kusetshenziswa izithombe ezimbalwa ezivela engeli ehlukene, shintsha umbono, lawula ukugxila futhi ukhiqize izithombe ezihambayo.
Source: opennet.ru