Inkqubo yokufunda yomatshini woDiffusion eZinzileyo ilungiselelwe ukuhlanganiswa komculo

Iprojekthi yeRiffusion iphuhlisa uguqulelo lwenkqubo yokufunda koomatshini iStable Diffusion, elungiselelwe ukuvelisa umculo endaweni yemifanekiso. Umculo unokudityaniswa kwinkcazo yombhalo ngolwimi lwendalo okanye ngokusekwe kwitemplate ecetywayo. Amacandelo omculo abhalwe kwiPython usebenzisa isakhelo sePyTorch kwaye afumaneka phantsi kwelayisenisi yeMIT. I-interface ebophelelayo iphunyezwa kwi-TypeScript kwaye isasazwa phantsi kwelayisenisi ye-MIT. Imifuziselo eqeqeshiweyo inikwe ilayisenisi phantsi kwelayisensi evumelekileyo ye-Creative ML OpenRAIL-M yokusetyenziselwa urhwebo.

Iprojekthi inomdla kuba iyaqhubeka nokusebenzisa "isicatshulwa-kuya-kumfanekiso" kunye "nomfanekiso-kumfanekiso" weemodeli ukuvelisa umculo, kodwa ilawula i-spectrograms njengemifanekiso. Ngamanye amazwi, i-classic Diffusion eZinzileyo ayiqeqeshelwanga kwiifoto kunye nemifanekiso, kodwa kwimifanekiso ye-spectrogram ebonisa utshintsho kwi-frequency kunye ne-amplitude ye-wave yesandi ekuhambeni kwexesha. Ngokufanelekileyo, i-spectrogram nayo yenziwe kwimveliso, ethi iguqulwe ibe ngumboniso wesandi.

Inkqubo yokufunda yomatshini woDiffusion eZinzileyo ilungiselelwe ukuhlanganiswa komculo

Indlela inokusetyenziselwa ukuguqula iingoma ezikhoyo zesandi kunye nokudibanisa umculo kwisampuli, efana nokuguqulwa komfanekiso kwi-Stable Diffusion. Umzekelo, isizukulwana sinokusampula iispectrograms ngesimbo sokubhekisela, sidibanise izitayile ezahlukeneyo, senze utshintsho olugudileyo ukusuka kwesinye isimbo ukuya kwesinye, okanye senze utshintsho kwisandi esikhoyo ukusombulula iingxaki ezinjengokunyusa umthamo wesixhobo ngasinye, ukutshintsha isingqi kunye nokutshintsha isingqi. izixhobo. Iisampulu zikwasetyenziselwa ukuvelisa iingoma ezidlala ixesha elide, ezenziwe ngothotho lweepaseji ezisondeleleneyo ezihluka kancinci ngokuhamba kwexesha. Iindinyana eziveliswe ngokwahlukileyo zidibene zibe ngumlambo oqhubekayo usebenzisa i-interpolation yeeparamitha zangaphakathi zemodeli.

Inkqubo yokufunda yomatshini woDiffusion eZinzileyo ilungiselelwe ukuhlanganiswa komculo

Uguqulo lweFourier olunefestile lusetyenziswa ukwenza ispectrogram kwisandi. Xa uphinda uphinda uphendule isandi kwi-spectrogram, ingxaki ivela ngokumisela isigaba (kuphela i-frequency kunye ne-amplitude ikhona kwi-spectrogram), ukulungiswa kwakhona kwe-algorithm ye-Griffin-Lim approximation.



umthombo: opennet.ru

Yongeza izimvo