Stability AI ya buga samfuran shirye-shiryen don tsarin koyo na inji Stable Diffusion, wanda zai iya haɗawa da canza hotuna dangane da kwatancen rubutun harshe na halitta. An buɗe samfuran a ƙarƙashin lasisin izini na Creative ML OpenRAIL-M, yana ba da damar kasuwanci. An horar da tsarin ta amfani da gungu na 4000 NVIDIA A100 Ezra-1 GPUs da kuma bayanan LAION-5B, wanda ya ƙunshi hotuna biliyan 5.85 tare da bayanin rubutu. Kayan aikin horar da hanyar sadarwar jijiya da samar da hotuna an riga an buɗe su ƙarƙashin lasisin MIT.
Samun samfurin da aka shirya da kuma ƙayyadaddun buƙatun tsarin, ba da izinin gwaji akan PC tare da daidaitattun GPUs, ya haifar da bullar ayyukan da yawa masu alaƙa:
- inversion textual (ladi) ƙari ne wanda ke ba ka damar haɗa hotuna tare da abin da aka ba da hali, abu, ko salo. A cikin asalin Stable Diffusion, abubuwan da ke cikin hotunan da aka haɗa ba su da tsari kuma ba a sarrafa su ba. Wannan add-on yana ba ku damar ƙara abubuwan gani naku, haɗa su zuwa kalmomi, da amfani da su a cikin haɗin gwiwa.
Misali, a daidaitaccen Stable Diffusion, zaku iya tambayar tsarin don samar da hoton "cat a cikin jirgin ruwa." Kuna iya ƙara ƙayyadaddun halaye na cat da jirgin ruwa, amma ainihin cat da jirgin ruwan da za a haɗa ba shi da tabbas. Juyawar rubutu tana ba ku damar horar da tsarin akan hoton cat ko jirgin ruwan ku sannan ku haɗa hoton takamaiman cat ko jirgin ruwa. Hakazalika, yana iya maye gurbin abubuwa na hoto tare da takamaiman abubuwa, ƙayyade misali na salon gani don haɗawa, da kuma tsaftace ra'ayoyi (alal misali, daga likitoci iri-iri, za ku iya zaɓar samfurin da ya fi dacewa da inganci a cikin salon da ake so).
- stable-disfusion-animation - ƙirƙirar hotuna masu rai (motsi) dangane da haɗin kai tsakanin hotunan da aka haifar a cikin Stable Diffusion.
- stable_diffusion.openvino (ladi) tashar jiragen ruwa ce ta Stable Diffusion wacce ke amfani da CPU kawai don ƙididdigewa, ba da izinin gwaji akan tsarin ba tare da GPUs masu ƙarfi ba. Ana buƙatar na'ura mai sarrafawa wanda ɗakin karatu na OpenVINO ke tallafawa. A bisa hukuma, OpenVINO yana ba da plugins don masu sarrafawa na Intel tare da AVX2, AVX-512, AVX512_BF16, da kari na SSE, da kuma na Rasberi Pi 4 Model B, Apple Mac mini, da allunan NVIDIA Jetson Nano. Ba bisa hukuma ba, Hakanan ana iya amfani da OpenVINO akan na'urori na AMD Ryzen.
- sdamd tashar jiragen ruwa ce don AMD GPUs.
- Farkon aiwatar da haɗin bidiyo.
- stable-diffusion-gui, stable-diffusion-ui, Artbreeder Collage, difffuse-the-raura - musaya mai hoto don samar da hotuna ta amfani da Stable Diffusion.
- beta.dreamstudio.ai, Hugging Face Spaces, hlky Stable Diffusion WebUI - musaya na yanar gizo don haɗin hoto ta amfani da Stable Diffusion.
- Plugins don haɗa Stable Diffusion tare da GIMP, Figma, Blender, da Photoshop.
Bugu da ƙari, Google ya buga lambar don tsarin ilmantarwa na injin RawNeRF (RAW Neural Radiance Fields), wanda ke amfani da bayanai daga hotuna RAW da yawa don inganta ingancin hotuna masu yawan hayaniya da aka ɗauka a cikin duhu da ƙananan haske. Baya ga raguwar amo, kayan aikin aikin suna ba masu amfani damar haɓaka daki-daki, kawar da kyalkyali, haɗa HDR, da canza hasken gabaɗaya a cikin hotuna. Hakanan za su iya amfani da hotuna da yawa daga kusurwoyi daban-daban don sake gina matsayi mai girma uku na abubuwa, canza ra'ayi, sarrafa mayar da hankali, da samar da hotuna masu motsi.
source: budenet.ru
