Sistimi e tsitsitseng ea ho ithuta mochini oa Diffusion e ikamahanyang le motsoako oa 'mino

Morero oa Riffusion o nts'etsapele mofuta oa mochini oa ho ithuta oa Stable Diffusion, o lokiselitsoeng ho hlahisa 'mino ho fapana le litšoantšo. 'Mino o ka kopanngoa ho tloha tlhaloso ea mongolo ka puo ea tlhaho kapa e thehiloeng ho thempleite e reriloeng. Likarolo tsa motsoako oa 'mino li ngotsoe ka Python ho sebelisoa moralo oa PyTorch mme li fumaneha tlasa laesense ea MIT. Khokahano e tlamang e kengoa tšebetsong ho TypeScript hape e ajoa tlasa laesense ea MIT. Mefuta e koetlisitsoeng e filoe laesense tlas'a laesense e lumelletsoeng ea Creative ML OpenRAIL-M bakeng sa tšebeliso ea khoebo.

Morero ona o thahasellisa hobane o ntse o tsoela pele ho sebelisa mefuta ea "text-to-image" le "setšoantšo-setšoantšo" ho hlahisa 'mino, empa e laola li-spectrogram e le litšoantšo. Ka mantsoe a mang, "Stable Diffusion" ea khale e koetliselitsoe eseng lifoto le litšoantšo, empa litšoantšong tsa li-spectrogram tse bonts'ang liphetoho tsa maqhubu le boholo ba leqhubu la molumo ha nako e ntse e ea. Ka hona, spectrogram e boetse e etsoa tlhahisong, e ntan'o fetoloa setšoantšo sa molumo.

Sistimi e tsitsitseng ea ho ithuta mochini oa Diffusion e ikamahanyang le motsoako oa 'mino

Mokhoa ona o ka boela oa sebelisoa ho fetola lipina tse teng tsa molumo le ho kopanya 'mino ho tloha sampuling, ho tšoana le ho fetola setšoantšo ho Stable Diffusion. Ka mohlala, moloko o ka etsa mohlala oa li-spectrogram ka mokhoa oa litšupiso, oa kopanya mekhoa e fapaneng, oa etsa liphetoho tse bonolo ho tloha ho mokhoa o mong ho ea ho o mong, kapa oa etsa liphetoho ho molumo o teng ho rarolla mathata a kang ho eketsa molumo oa liletsa ka bomong, ho fetola morethetho le ho fetola. lisebelisoa. Mehlala e boetse e sebelisoa ho hlahisa lipina tse bapalang nako e telele, tse entsoeng ka letoto la litemana tse arohaneng tse fapaneng hanyane ka nako. Litemana tse entsoeng ka karohano li kopanngoa hore e be molatsoana o tsoelang pele ho sebelisoa tlhaloso ea likarolo tse ka hare tsa mohlala.

Sistimi e tsitsitseng ea ho ithuta mochini oa Diffusion e ikamahanyang le motsoako oa 'mino

Fourier transformed e fenstere e sebediswa ho bopa spectrogram ho tswa ho modumo. Ha ho tsosolosoa molumo o tsoang ho spectrogram, bothata bo hlaha ka ho khetholla mohato (feela maqhubu le amplitude li teng ho spectrogram), bakeng sa ho tsosolosoa ha algorithm ea Griffin-Lim approximation.



Source: opennet.ru

Eketsa ka tlhaloso