MAI-Image-2.7-Preview Uyasebenza

I-MAI-Image-2.7-Preview iyimodeli yokuhlela izithombe eyethulwe yi-Microsoft. Ibekwe endaweni yesibili emhlabeni wonke nge-Elo engu-1,193 ohlwini lokuhlela lwe-Artificial Analysis, ilandela kuphela i-GPT Image 2.5 Sunburst, futhi ihamba phambili ekuhleleni izigcawu nezitayela ngamaphuzu angu-24 e-Elo ngaphezu kwemodeli elandelayo ehamba phambili. I-MAI-Image-2.7-Preview inekhono ekushintsheni izitayela, ekulungiseni ukukhanya, ekuguquleni umbhalo nezimpawu, nasekulungiseni nasekuthuthukiseni izithombe. Okwamanje ingazanywa ezinkundleni ezifana ne-MAI Playground; izisindo zayo azikavulelwa umphakathi futhi iseyinguqulo yokubuka kuqala ye-Preview.

Ukuguqulwa kwezigcawu nezitayela: Dweba kabusha isitayela, ithoni nomoya wendawo yonke ngokuchofoza okukodwa. Leli yikhono elihamba phambili kakhulu, elidlula imodeli elandelayo ehamba phambili ngamaphuzu angu-24 e-Elo.
Ukulungisa ukukhanya ngobuhlakani: Lungisa ukukhanya nomuzwa wesithombe ukuze ukuqondiswa kwezithunzi nokubonakala kokukhanya ezintweni kuhambisane nemithetho ye-physics.
Ukuhlela umbhalo/izimpawu: Guqula umbhalo nezimpawu ezisesithombeni, uzihlanganise nokukhanya, umbono kanye nokwakheka kwesithombe sokuqala.
Ukulungisa nokuthuthukisa: Susa umsindo, lungisa izinto ezindala, uthuthukise ikhwalithi yesithombe, kuyilapho ugcina ukwakheka kwasekuqaleni ngokuzinzile.
Ukuhlela okusekelwe ekucabangeni: Qonda ubudlelwano phakathi kwezinto, ukukhanya nendawo ezikuyo, bese uphendula imiyalelo eyinkimbinkimbi edinga ukucabanga, njengokuthi “hambisa umkhiqizo uye endaweni entsha”.

Isakhiwo sokukhiqiza esisekelwe ku-diffusion: Imodeli isebenzisa indlela yokukhiqiza esekelwe ku-diffusion, ifunde ukuguquka okuqhubekayo phakathi kokusatshalaliswa komsindo nokusatshalaliswa kwedatha nge-flow-matching loss, ukuze kuqinisekiswe ukuqeqeshwa okuzinzile nokusebenza kahle.
Ukwakhiwa kabusha kokuhlela kwe-multimodal: Ilandela indlela “yokwakha kabusha kusukela ekuqaleni, egxile ekuhambeni komsebenzi wokuhlela kwe-multimodal”, kuyilapho icabanga ngokuhlanganyela ngezinto, ukwakheka kwendawo, ukukhanya, isikali nendawo ngokwezinhlangothi ezahlukene. Ngisho nangemiyalelo engacacile, ingakhiqiza imiphumela yokuhlela ehambisana nencazelo.
Ukuhlela okuhlanganyelwayo okunemikhawulo eminingi: Imodeli ingaphatha ngesikhathi esisodwa ukugcinwa kwesakhiwo, izinguquko zezinto, ukubuyekezwa kombhalo nokuvumelanisa nomongo, futhi isekele ukususwa, ukushintshwa nokuguqulwa kwezici zezinto kanye ne-inpainting. Lokhu akuphazamisi ukwakheka okuphelele futhi kugcina ukuvumelana okubonakalayo phakathi kokuhlela okuphindaphindiwe.
Indlela yokuhambisana nokuphepha: Idatha iyahlungwa, bese kunezelwa ezinye izindlela zokunciphisa ubungozi ngesikhathi sokusetshenziswa. Inhloso iwukuba kungakhiqizwa izithombe eziyingozi, ezinodlame, ezinegazi noma ezingcolile, ngisho noma umsebenzisi ezicela.

Vakashela iwebhusayithi esemthethweni: Vakashela iwebhusayithi esemthethweni ye-MAI-Image-2.7-Preview ku-https://playground.microsoft.ai/chat.
Ngena ku-akhawunti: Chofoza inkinobho yokungena ekhoneni eliphezulu kwesokudla, bese usebenzisa i-akhawunti ye-Microsoft ukuze ungene. Ungayizama mahhala.
Layisha isithombe sokuqala: Khetha i-MAI-Image-2.7-Preview ngaphakathi kwesigaba se-Images/ukuhlela izithombe, bese uchofoza ukulayisha noma uhudulele isithombe, idizayini noma isithombe-skrini osuvele unaso ukuze sibe yisisekelo sokuhlela.
Faka umyalo bese ukhiqiza: Bhala umyalo ocacile wokulungisa ngesiNgisi ebhokisini lokufaka, bese uchofoza inkinobho yokukhiqiza futhi ulinde imizuzwana embalwa ukuze kuvele isithombe.
Thuthukisa ngokuphindaphindiwe bese ulanda: Bheka umphumela okhiqiziwe. Uma ungagculisekile, lungisa umyalo bese uphinda ukuhlela. Uma umphumela usulungile, chofoza ukulanda ukuze uwugcine kudivayisi yakho.

Indawo yesibili ezingeni lomhlaba: Ibekwe endaweni yesibili emhlabeni wonke nge-Elo engu-1,193 ohlwini lokuhlela izithombe lwe-Artificial Analysis, ilandela kuphela i-GPT Image 2.5 Sunburst (max), futhi idlula i-Google, i-Meta, i-xAI kanye ne-GPT-Image-2.5-Flare.
Ukuhlela izigcawu nezitayela kuhamba phambili kakhulu: Esigabeni sokuhlela izigcawu nezitayela, idlula imodeli elandelayo ehamba phambili ngamaphuzu angu-24 e-Elo, okuyigebe elikhulu kunawo wonke kuwo wonke amakhono.
Idlula ngokuphelele isizukulwane sayo sangaphambilini: Uma iqhathaniswa ne-MAI-Image-2.6, ithuthukile kuzo zonke izimo zokusetshenziswa ezingu-10 nezinhlobo zezenzo zokuhlela ezingu-7 ku-Artificial Analysis, ngaphandle kobuthakathaka obubonakalayo.
Izinhlobo ezine zamakhono zisezingeni eliphambili: Ukuthuthukisa nokulungisa, ukuhlela okusekelwe ekucabangeni, izigcawu nezitayela, kanye nombhalo nezimpawu konke kusezingeni eliphambili emkhakheni.

Iwebhusayithi yephrojekthi: https://microsoft.ai/news/mai-image-2-7-debut/

Ukuguqula izithombe zemikhiqizo ye-e-commerce: Hambisa izithombe zemikhiqizo uye ezindaweni ezintsha ngokuchofoza okukodwa, ufanise ngokuzenzakalelayo ukukhanya nombono, futhi ukhiqize ngokushesha izithombe eziyinhloko zemikhiqizo nezinto zokumaketha ngezitayela ezahlukene.
Ukuphindaphinda amadizayini nezindawo ze-UI: Guqula isitayela samadizayini nezithombe-skrini, shintsha imibala yetimu futhi uhlole izakhiwo ezahlukene, ukuze abaklami bazame izindlela eziningi ngezindleko eziphansi ngaphambi kokuphothula idizayini.
Ukulungisa nokuthuthukisa izithombe nezinto ezindala: Lungisa umsindo nemihuzuko ezithombeni ezindala, uthuthukise ikhwalithi yezinto ezinokucaca okuphansi, futhi ugcine ukwakheka kwasekuqaleni ngokuzinzile. Lokhu kulungele ukufakwa kwedijithali kwamarekhodi nokusetshenziswa kabusha kwezinto.
Ukwenza imibhalo yezinto zokumaketha ibe ngeyendawo: Guqula ngokuqondile umbhalo nezimpawu kumaphosta, kumabhodi okukhangisa nasekupakisheni, uzihlanganise nombono nokukhanya kwasekuqaleni, ugweme isidingo sokuphinda uhlele umbhalo.