
Namuhla i-MiniMax imemezele ngokusemthethweni ukuthi ivula umthombo wemodeli yayo entsha yesizukulwane samavidiyo ejwayelekile, i-MiniMax H3.
Ngokwethulwa kwayo, i-MiniMax H3 iwuhlelo olujwayelekile lokukhiqiza olusebenzisa zonke izinhlobo zemidiya. Iyakwazi ukuqonda ngasikhathi sinye umongo wemodi ehlukahlukene ohlanganisa umbhalo, izithombe, amavidiyo nomsindo, futhi ikhiqize amavidiyo anokulungiswa okufika ku-2K, ubude obufika kumasekhondi angu-15, nomsindo we-stereo ongowomdabu. Ngenxa yokwakheka kohlelo okugxile ekwenzeni imisebenzi ehlukahlukene, i-H3 isivele inamandla abanzi okuqonda nokukhiqiza umongo wemodi ehlukahlukene kusukela esigabeni sokuqeqeshwa kwangaphambili. Ngakho-ke iyakwazi ukulandela kahle imiyalelo eyinkimbinkimbi yemodi ehlukahlukene.
I-H3 isekela lezi zici zokufaka nokukhipha:
Ubude bokukhipha: amasekhondi angu-4–15
Isilinganiso sobubanzi nobude bokukhipha: isekela izilinganiso eziningi, ezihlanganisa kodwa ezingagcini ku-21:9, 16:9, 4:3, 1:1, 3:4 kanye no-9:16
Ukulungiswa kokukhipha: isekela osayizi abaningi bokulungiswa; ngokuzenzakalelayo uhlangothi olufushane lusethwa kumaphikseli angu-768. Ukusebenzisa i-H3-Regenerate-2K kwenza kube nokukhiqiza ngokulungiswa kwe-2K
Izinga lamafreyimu okukhipha: 24 FPS
Umsindo wokukhipha: 32 kHz stereo
Izilimi zengxoxo ezisekelwayo: kusekelwa ngokuqinile izilimi ezingu-11: isi-Arabhu, isiShayina, isiNgisi, isiFulentshi, isiJalimane, isiNtaliyane, isiJapane, isiKorea, isiPutukezi, isiRashiya nesiSpanishi. Ezinye izilimi nazo ziyasekelwa ngamazinga ahlukahlukene.
Izinhlobo zemodeli nezici zokufaka:
① H3-Base-FL2VA
Imodi yamafreyimu okuqala nawokugcina: isekela ukufaka izithombe ezingu-0, 1 noma 2:
Uma kungafakwa isithombe, kuba yimodi yokukhiqiza ividiyo ngombhalo;
Uma kufakwa isithombe sefreyimu yokuqala, kuba yimodi yokukhiqiza ividiyo kusukela kufreyimu yokuqala;
Uma kufakwa isithombe sefreyimu yokugcina, kuba yimodi yokukhiqiza ividiyo kusukela kufreyimu yokugcina;
Uma kufakwa izithombe zefreyimu yokuqala neyokugcina ngasikhathi sinye, kuba yimodi yokukhiqiza ividiyo kusukela kumafreyimu okuqala nawokugcina
② H3-Base-Ref2VA
Imodi yesithenjwa esebenzisa zonke izinhlobo zemidiya:
Izithombe: kufika kwezi-9;
Amavidiyo: kufika ku-3; ividiyo ngayinye kufanele ibe namasekhondi angu-2–15, kanti ubude obuphelele akumele budlule amasekhondi angu-15;
Umsindo: kufika ku-3. Umsindo kufanele ufakwe kanye nesithombe noma ividiyo, awukwazi ukuba yiwona kuphela ofakwayo; umsindo ngamunye kufanele ube namasekhondi angu-2–15, kanti ubude obuphelele akumele budlule amasekhondi angu-15;
Ukufaka okuxubile: ingqikithi yamafayela okufaka azo zonke izinhlobo ingafika ku-12

Uhlelo oluphelele lwe-H3 lwakhiwe ngamamojula amathathu alandelayo:
H3-Contex t-IR: Njengoba okufakwayo kuya ngokuya kuba yinkimbinkimbi, sakhe uhlelo olukhethekile lokuqonda nokucubungula ngokujulile imiyalelo yemodi ehlukahlukene efakiwe, bese luyiguqula ibe yifomethi i-H3 eyiqonda kalula nengasetshenziswa ngokuqondile ekukhiqizeni, ebizwa nge-Context Intermediate Representation (ukumelwa komongo okuphakathi). I-H3-Context-IR ibaluleke kakhulu kwikhwalithi yomphumela wokugcina. Ngakho-ke sincoma kakhulu ukuthi uhlanganise i-H3-Context-IR API nenqubo yokukhiqiza. Kungenjalo, ungabheka umhlahlandlela wamaphrompthi ukuze wakhe uhlelo lwakho lokucubungula umongo.
H3-Base: Ikhiqiza umsindo nevidiyo ngokusekelwe emiphumeleni ye-H3-Context-IR, bese ikhipha imiphumela enokulungiswa kwe-768p.
H3-Regenerate-2K: Ibuyisela umphumela wokukhiqiza we-768p ku-H3 kanye nomongo wokuqala, ukuze iphinde ikhiqize umphumela ngokulungiswa kwe-2K. Lokhu kusebenzisa amandla amakhulu okukhiqiza e-H3, ngesikhathi esifanayo kusizakala ngolwazi olunothile oluqukethwe kumongo wokuqala, ngaleyo ndlela kukhiqizwe imiphumela enokunemba okuphezulu kwemininingwane nokwethembeka okukhulu kokubonakalayo.
I-MiniMax H3 ikhishwa ngaphansi kwe-MiniMax H3 Community License.
Ikheli lomthombo ovulekile: huggingface.co/MiniMaxAI/MiniMax-H3
