MAI-Image-2.6-Flash Uyasebenza

I-MAI-Image-2.6-Flash iyimodeli ka-Microsoft yokukhiqiza izithombe ngesivinini esikhulu. Isebenzisana ne-MAI-Image-2.6 esezingeni eliphezulu, igxile ezimweni zokukhiqiza ezidinga ukubambezeleka okuphansi kanye nokuphuma okuphezulu. Ijubane lokukhiqiza lemodeli liphindwe izikhathi ezingu-2.8 kune-GPT-Image-2-Medium, kanti ukusebenza kahle kwe-GPU kukhuphuke ngo-72%, kuyilapho igcina ikhwalithi yokubuka eseduze neyesigaba esiphezulu. I-MAI-Image-2.6-Flash isekela izithombe eziningi eziyizinkomba, i-Web Grounding kanye nezilinganiso zamafreyimu eziguqukayo, ngesinqumo esifinyelela ku-1.5K.

Ukukhiqiza okushesha kakhulu: Isivinini sokukhiqiza sikhuphuka cishe izikhathi ezingu-2.8 uma siqhathaniswa ne-GPT-Image-2-Medium, kanti ukusebenza kahle kwe-GPU kukhuphuka ngo-72%, okuvumela impendulo enokubambezeleka okuphansi.
Ikhwalithi yesithombe esezingeni eliphezulu: Nakuba ijubane lenyuswe kakhulu, igcina ikhwalithi yokubuka esondelene nenguqulo ephambili ye-MAI-Image-2.6 kanye namandla ayo ayinhloko okuhlela.
Ukuhlela kusetshenziswa izithombe eziningi eziyizinkomba: Isekela ukufakwa ngesikhathi esisodwa kwezithombe eziningi eziyizinkomba zabantu, imikhiqizo, izitayela nezigcawu, bese ihlanganisa izimo eziningi zokubuka ukuze ikhiqize noma ihlele.
Web Grounding: Ingahlanganisa ulwazi lwesikhathi sangempela oluvela ku-inthanethi ukuze inikeze ukubambelela kolwazi olunembile kokuqukethwe okubukwayo okuphathelene nemikhiqizo yangempela, izigameko zezindaba nezinye izinto.
Uhlaka lwesithombe oluguqukayo nokulungiswa okuphezulu: Ikhetha ngokuzenzakalelayo isilinganiso sobubanzi nobude esifanele kakhulu ngokwesigcawu, futhi isekela okukhiphayo okunokulungiswa okufika ku-1.5K.
Ukukhiqizwa ngezinga elikhulu: Idizayinelwe ngokukhethekile ukukhiqizwa kwezinto zokuthengisa ngobuningi, okune-throughput ephezulu futhi okubheka izindleko, kanti intengo ye-API yehliswe ngo-50%-69% uma iqhathaniswa nenguqulo ephambili.

Isakhiwo sokusabalalisa se-Flow Matching: Imodeli isekelwe esakhiweni se-Diffusion futhi iqeqeshwa kusetshenziswa i-Flow Matching Loss, ifunde indlela eqondile eqhubekayo esuka emsindweni iye esithombeni. Ingakhiqiza izithombe ezicacile ngezinyathelo ezimbalwa kakhulu zokucubungula, okuyisisekelo sezibalo sokwandisa isivinini ngo-2.8× kuyilapho kugcinwa ikhwalithi yesithombe.
Amapharamitha angu-20B nomongo ongu-32K: Imodeli inamapharamitha ayizigidi eziyizinkulungwane ezingu-20 angafaki ama-embedding kanye newindi lomongo lama-Token angu-32K. Isekela ukuqonda iziyalezo ezinde eziyinkimbinkimbi nokufakwa kwezithombe eziningi, kanti ukulungiswa okuphezulu kokuphumayo cishe kungu-1536×1536, okunikeza umthamo owanele wemodeli ekukhiqizeni kwekhwalithi ephezulu nasekuhleleni kusetshenziswa izithombe eziningi eziyizethenjwa.
Ukuthuthukiswa kobunjiniyela bokucubungula: Inguqulo ye-Flash ayiyona imodeli ehlungiwe noma ecindezelwe, kodwa ithuthukiswa ngokubambisana kwama-algorithm nobunjiniyela phezu kwesisekelo esifanayo samapharamitha angu-20B. Lokhu kunciphisa isikhathi sokucubungula isithombe esisodwa sibe sezingeni lama-millisecond kuya kumasekhondi angu-1-2, kuyilapho kukhululwa u-72% womthamo we-GPU ongasebenzi, kufinyelelwe ekuphatheni okuphezulu kokukhiqiza okunokuncintisana okuphezulu.
Ukwakhiwa kabusha kokuhlela kwe-multimodal: Amandla okuhlela akhiwe kabusha kusukela ekuqaleni. Imodeli inekhono lokucabanga ngomongo obonakalayo oluhlanganisa izinto ezihlukahlukene, ukwakheka kwesigcawu, ukukhanya, isikali nendawo yendawo. Ngisho nalapho ibhekene neziyalezo ezingacacile, igcina imiphumela yokuhlela engaguquki futhi isekele imisebenzi eyinkimbinkimbi efana nokushintsha izinto, ukulungisa izici nokubuyekeza umbhalo.

MAI Playground: inkundla ye-inthanethi esemthethweni yakwa-Microsoft yokuhlola https://playground.microsoft.ai/chat, lapho ungahlola khona ngokuqondile ukuguqula umbhalo ube isithombe, isithombe sibe esinye isithombe, kanye nokuhlela izithombe kusetshenziswa izithenjwa eziningi. Ilungele ukuhlola ngokushesha imiphumela yemodeli.
Microsoft Foundry: isiteshi sokufinyelela esitholakala emphakathini esisesigabeni sokubuka kuqala, esenzelwe onjiniyela. Ngemva kokubhalisa nokuthola ukhiye we-API, ungahlanganisa imodeli nezinhlelo noma ukugeleza komsebenzi wakho, futhi isekela ukubizwa ngezinga elikhulu.

Ukukhiqiza ngesivinini esikhulu: Isivinini sokukhiqiza siphindwe cishe ngo-2.8 kune-GPT-Image-2-Medium, kanti ukusebenza kahle kwe-GPU kukhuphuke ngo-72%, okuvumela impendulo enokubambezeleka okuncane.
Ikhwalithi yesithombe esezingeni eliphezulu: Naphezu kokusheshisa kakhulu, igcina ikhwalithi yokubuka esondele ku-MAI-Image-2.6 kanye namakhono ayo abalulekile okuhlela.
Inani elihle kakhulu: Intengo ye-API yehliswe ngo-50%-69% uma iqhathaniswa nenguqulo esezingeni eliphezulu; ngokusho kwenkampani, inenani elihamba phambili emhlabeni le-price-per-Elo.
Ukuhlanganiswa kwezithombe eziningi eziyizethenjwa: Isekela ukuthunyelwa kwezithombe eziningi eziyizethenjwa ngesikhathi esisodwa, ezihlanganisa abantu, imikhiqizo, izitayela nezimo, ukuze kukhiqizwe ngokusekelwe ezimweni eziningi zokubuka.
Web Grounding: Ihlanganisa ulwazi lwakamuva oluvela ku-inthanethi ukuze inikeze ukuqondaniswa kolwazi okunembile kokuqukethwe okubukwayo okuphathelene nemikhiqizo yangempela, izenzakalo zezindaba nokunye.
Usayizi wesithombe oguqukayo nokulungiswa okuphezulu: Ikhetha ngokuzenzakalelayo isilinganiso sobubanzi nobude esifanele kakhulu ngokwesimo, futhi isekela okukhiphayo okufika ku-1.5K resolution.

Iwebhusayithi esemthethweni yephrojekthi: https://microsoft.ai/news/pushing-the-quality-cost-frontier-with-mai-image-2-6/

Ukukhiqizwa kwezinto eziningi ze-e-commerce: Khiqiza ngokushesha izithombe eziyinhloko eziningi zemikhiqizo, izithombe zamakhasi emininingwane namaphosta okukhangisa, unciphise izindleko ze-API zesithombe ngasinye.
Ukusebenzisana kwesikhathi sangempela nezikhangiso eziguqukayo: Thola impendulo yezithombe esheshayo ezingeni lama-millisecond kuma-chatbot, abasizi be-AI noma ezinhlelweni zezikhangiso eziguqukayo, uthuthukise ulwazi lwabasebenzisi.
Imboni yokuqukethwe ezinkundleni zokuxhumana: Nikeza abadali bokuqukethwe nezinhlangano ze-MCN amandla okukhiqiza ngobuningi izithombe nombhalo ngesivinini esikhulu, usekele ukugeleza komsebenzi okukhiqiza izithombe ezingafika kweziyikhulu ngosuku.
Ulayini wokubona we-AI Agent: Sebenza njengenodi yokukhiqiza izithombe ekugelezeni komsebenzi okuzenzakalelayo; isici sokubambezeleka okuphansi siqinisekisa ukuthi ukusebenza okuphelele kwe-Agent pipeline akuvinjwa.
Ukuphindaphinda okusheshayo kwama-prototype nemidwebo: Abaklami bakhiqiza ngokushesha izithombe zemibono ezinguqulweni eziningi ekuqaleni komkhiqizo, bese becwenga ngemva kokusebenzisa umsebenzi wezithombe eziningi eziyireferensi ukuze bagcine isitayela esifanele.