Hoton da aka samar tare da Stable Diffusion 2.0
Kwanan nan Stability AI, an bayyana ta hanyar rubutun blog bugu na biyu na tsarin injin inji Tsayayyen Yaduwa, wanda ke da ikon haɗawa da gyaggyarawa hotuna dangane da samfuri da aka ba da shawara ko bayanin rubutun harshe na halitta.
Stable Yaduwa shine samfurin koyon injin Stability AI ya haɓaka don samar da ingantattun hotuna na dijital daga kwatancen harshe na halitta. Ana iya amfani da samfurin don ayyuka daban-daban, kamar samar da fassarar hoto-zuwa hoto mai jagorar rubutu da haɓaka hoto.
Ba kamar samfuran gasa kamar DALL-E ba, Stable Diffusion buɗaɗɗen tushe1 kuma baya iyakance hotunan da yake samarwa. Masu sukar sun tayar da damuwa game da ka'idodin AI, suna da'awar cewa za a iya amfani da samfurin don ƙirƙirar zurfafa tunani.
Ƙungiya mai ƙarfi ta Robin Rombach (Stability AI) da Patrick Esser (Runway ML) daga CompVis Group a LMU Munich wanda Farfesa Dr. Björn Ommer ya jagoranta, ya jagoranci fitowar asali na Stable Diffusion V1. Sun gina kan aikin lab ɗin su na baya tare da samfuran ɓoyayyiyar ɓoyayyiya kuma sun sami tallafi mai mahimmanci daga LAION da Eleuther AI. Kuna iya karanta ƙarin game da ainihin fitowar Stable Diffusion V1 a cikin gidan yanar gizon mu na baya. Robin yanzu yana jagorantar ƙoƙarin tare da Katherine Crowson a Stability AI don ƙirƙirar ƙarni na gaba na ƙirar watsa labarai tare da babbar ƙungiyarmu.
Stable Diffusion 2.0 yana ba da ɗimbin ingantattun abubuwa da fasali idan aka kwatanta da ainihin sigar V1.
Babban labarai na Stable Diffusion 2.0
A cikin wannan sabon sigar da aka gabatar An ƙirƙiri sabon ƙirar haɗin hoto dangane da bayanin rubutu "SD2.0-v", wanda ke goyan bayan samar da hotuna tare da ƙudurin 768×768. An horar da sabon samfurin ta amfani da tarin LAION-5B na hotuna biliyan 5850 tare da bayanin rubutu.
Samfurin yana amfani da saitin sigogi iri ɗaya kamar samfurin Stable Diffusion 1.5, amma ya bambanta ta hanyar canzawa zuwa amfani da maɓalli mai mahimmanci na OpenCLIP-ViT/H, wanda ya ba da damar haɓaka ingancin hotunan da aka samu.
An shirya A Sauƙaƙe sigar SD2.0-base, horar da hotuna 256 × 256 ta amfani da samfurin tsinkayar amo na gargajiya da kuma tallafawa tsararrun hotuna tare da ƙuduri na 512 × 512.
Baya ga wannan, an kuma yi nuni da cewa an ba da damar yin amfani da fasaha mafi girma (Super Resolution) don ƙara ƙuduri na ainihin hoton ba tare da rage inganci ba, ta amfani da sikelin sararin samaniya da cikakkun bayanai na sake ginawa.
Na sauran canje-canje wanda ya fice daga wannan sabon sigar:
- Samfurin sarrafa hoto da aka bayar (SD20-upscaler) yana goyan bayan haɓakar 4x, yana ba da damar ƙirƙirar hotuna tare da ƙudurin 2048 × 2048.
- Stable Diffusion 2.0 kuma ya haɗa da samfurin Diffusion na Upscaler wanda ke haɓaka ƙudurin hoto ta hanyar 4.
- An gabatar da samfurin SD2.0-depth2img, wanda ke la'akari da zurfin da tsarin sararin samaniya na abubuwa. Ana amfani da tsarin MiDaS don kimanta zurfin monocular.
- Sabon samfurin fenti na ciki wanda aka sarrafa rubutu, an daidaita shi akan sabon Stable Diffusion 2.0 rubutu-zuwa hoto tushe
- Samfurin yana ba ku damar haɗa sabbin hotuna ta amfani da wani hoto azaman samfuri, wanda zai iya bambanta sosai da asali, amma yana riƙe da gabaɗayan abun da ke ciki da zurfin. Misali, zaku iya amfani da matsayin mutum a cikin hoto don ƙirƙirar wani hali a cikin irin wannan matsayi.
- Samfuran da aka sabunta don gyara hotuna: SD 2.0-inpainting, wanda ke ba da damar amfani da alamun rubutu don maye gurbin da canza sassan hoton.
- An inganta samfuran don amfani akan tsarin yau da kullun tare da GPU.
A ƙarshe haka ne kuna sha'awar ƙarin sani game da shi, Ya kamata ku san cewa an rubuta lambar don horar da cibiyar sadarwar jijiyoyi da kayan aikin hoto a cikin Python ta amfani da tsarin PyTorch kuma an saki a ƙarƙashin lasisin MIT.
Samfuran da aka riga aka horar suna buɗe ƙarƙashin lasisin izini na Creative ML OpenRAIL-M, wanda ke ba da damar kasuwanci.
Source: https://stability.ai