I-Stable Diffusion 2.0 inkqubo yokwenziwa kwemifanekiso yaziswa

Ukuzinza kwe-AI kuye kwapapasha uhlelo lwesibini lwenkqubo yokufunda yomatshini we-Stable Diffusion, ekwazi ukudibanisa kunye nokuguqula imifanekiso esekelwe kwi-template ecetywayo okanye inkcazo yolwimi lwendalo. Ikhowudi yoqeqesho lwenethiwekhi ye-neural kunye nezixhobo zokuvelisa umfanekiso ibhalwe kwiPython usebenzisa isakhelo sePyTorch kwaye ipapashwe phantsi kwelayisenisi ye-MIT. Imifuziselo esele iqeqeshiwe ivuliwe phantsi kwelayisensi evumayo ye-Creative ML OpenRAIL-M, evumela ukusetyenziswa kwezorhwebo. Ukongeza, i-demo ye-intanethi ye-generator yemifanekiso iyafumaneka.

Uphuculo oluphambili kuhlelo olutsha lweStable Diffusion:

• Imodeli entsha yokwenziwa komfanekiso esekelwe kwinkcazo yombhalo yenziwe - SD2.0-v, exhasa ukuveliswa kwemifanekiso kunye nesisombululo se-768x768. Imodeli entsha iqeqeshwe ngokusebenzisa iqoqo le-LAION-5B, elibandakanya i-5.85 yezigidigidi zemifanekiso kunye neenkcazo zetekisi. Imodeli isebenzisa isethi efanayo yeeparitha njengemodeli ye-Stable Diffusion 1.5, kodwa ihluke kwinguqu ekusebenziseni i-encoder ye-OpenCLIP-ViT / H ehluke kakhulu, ephucule kakhulu umgangatho wemifanekiso ebangelwayo.
• Uguqulelo olulula lwe-SD2.0-base lulungiselelwe, luqeqeshwe kwimifanekiso ye-256 × 256 usebenzisa imodeli yokubikezela ingxolo yeklasi kunye nokuxhasa isizukulwana semifanekiso ngesisombululo se-512 × 512.
• Kuyenzeka ukuba usebenzise itekhnoloji ye-supersampling (i-Super Resolution) ukwandisa isisombululo somfanekiso wokuqala ngaphandle kokunciphisa umgangatho, usebenzisa ukukala kwendawo kunye ne-algorithms yokwakhiwa kwakhona kweenkcukacha. Imodeli yokucwangcisa umfanekiso obonelelweyo (i-SD20-upscaler) isekela ukusondeza kane, okuvumela ukuveliswa kwemifanekiso ngesisombululo se-2048x2048.
• Kucetywa imodeli ye-SD2.0-depth2img, kuthathelwa ingqalelo ubunzulu kunye nokuhlelwa kwendawo yezinto. Kuqikelelo lobunzulu be-monocular, inkqubo ye-MiDaS isetyenziswa. Imodeli ikuvumela ukuba udibanise imifanekiso emitsha usebenzisa omnye umfanekiso njenge template, enokuthi yahluke kakhulu kwi-original, kodwa ugcine ukubunjwa ngokubanzi kunye nobunzulu. Umzekelo, ungasebenzisa ukuma komntu kwifoto ukwenza omnye umlinganiswa kwindawo efanayo.
• Imodeli yokuguqula imifanekiso ihlaziywe - i-SD 2.0-inpainting, evumela ukuba utshintshe kwaye utshintshe iinxalenye zomfanekiso usebenzisa i-text prompts.
• Iimodeli zilungiselelwe ukusetyenziswa kwiinkqubo eziqhelekileyo ezineGPU enye.

umthombo: opennet.ru