Matrix-Game 4.0 Uyasebenza

I-Matrix-Game 4.0 ngumdlalo wokuqala emhlabeni osebenzisa imodeli yevidiyo enkulu futhi osebenza ekhadini elilodwa lezithombe elitholakala kubathengi, owethulwe yi-Kunlun Tech. Usebenzisa ipharadayimu entsha ye-“Harness”, lapho injini yomdlalo isebenza njengomahluleli ogcina isimo nemithetho yomhlaba, i-Coding Agent iguqula indlela yokudlala isebenzisa ulimi lwemvelo, kanti imodeli yevidiyo isebenza njengomdwebi onikeza izithombe ezinemininingwane ngesikhathi sangempela. Ikhadi elilodwa le-RTX 5090 ku-Matrix-Game 4.0 linganikeza umuzwa womdlalo ongu-480p, 60 FPS, onokubambezeleka okungaphansi kuka-150 milliseconds, okwenza kube ngokoqobo ukuthi amaqembu amancane akhe imihlaba.

Ukusebenza ngesikhathi sangempela ngekhadi elilodwa: Ikhadi elilodwa le-RTX 5090 linganikeza umuzwa womdlalo odlalekayo ongu-480p, 60 FPS, onokubambezeleka okungaphansi kuka-150 milliseconds.
Ukushintsha indlela yokudlala ngolimi lwemvelo: Sebenzisa i-Coding Agent nomyalelo owodwa ukulungisa indawo, ukuziphatha kwama-NPC noma ukudizayina imithetho emisha ngokuphelele.
Ukuphathwa okusobala kwesimo somhlaba: Injini yomdlalo igcina ngesikhodi izimo ezifana nempilo, impahla ephethwe, ukushayisana, ukunqoba nokuhlulwa; izinto ziyaqhubeka nokushintsha noma seziphumile emkhakheni wokubona.
Ukusuka kumamodeli alula kuye ezithombeni ezinemininingwane: Imodeli yevidiyo iphinda inikeze ngesikhathi sangempela izithombe zejometri eziphuma enjinini, iziguqule zibe izithombe zezinga le-AAA ezinokukhanya, izithunzi nezinto ezibonakalayo.
Ukudlala okuvumelanayo isikhathi eside: Isekela ukudlala okungaphazamiseki imizuzwana engu-54-99, kuyilapho indawo, ubunikazi babalingiswa nemiphumela yokuxhumana kuhlala kuzinzile kungaphambuki.

Landela ku-WeChat bese uphendula ngokuthi “开源” ukuze ujoyine iqembu lokuxhumana ngamaphrojekthi e-AI avulekile
Ukwakhiwa kwe-Harness: Uhlelo lufaka i-Harness eyakhiwe yi-Coding Agent nenjini yomdlalo phezu kwemodeli yevidiyo. Injini iphethe ukuguquka kwesimo ngaphakathi kwemithetho futhi ikhiphe izithombe ezihlelekile; i-Coding Agent iguqula inhloso yomsebenzisi ibe imithetho emisha nemiyalelo enemininingwane; imodeli yevidiyo igxila kuphela ekunikezeni izithombe, ngaleyo ndlela imisebenzi emithathu ihlukaniswe.
Idatha eqondanisa izithombe ezihlelekile nezinemininingwane: Kusetshenziswa i-UE5 kwakhiwa iqoqo ledatha eliwela izizinda ezahlukene. Ipayipi lokuqoqa elenza lula ijometri liguqula izigcawu eziyinkimbinkimbi zibe izakhiwo zamamodeli alula ezigcina ukwakheka kwendawo, bese lizihlanganisa ngokunemba kwephikseli nefreyimu ngayinye yezithombe zamamodeli anemininingwane, liphinde liqoqe i-ID yesenzeko nemethadatha yemicimbi eyimbangela ngesikhathi esifanayo.
I-Caption eguqukayo: Uhlelo luhlobanisa i-ID yento nencazelo yayo enencazelo. Ngesikhathi sokusebenza, lufunda enjinini izinto ezibonakalayo njengamanje nezimo zazo zesikhathi sangempela, luzifake kuthempulethi yomyalelo. Lokhu kusebenza kanye nezithombe ezihlelekile ukuze kunikezwe imikhawulo yokuthi kudwetshweni nokuthi kudwetshwe kuphi.
Ukukhiqizwa okulawulekayo kwe-V2V: Kwenziwe ngokusekelwe ku-Wan2.2 5B. Izithombe ezihlelekile nevidiyo ezokhiqizwa zihlanganiswa zibe okokufaka kwamashaneli angu-148 ku-DiT eyimbangela enezingqimba ezingu-30. Ukunaka kwakho kusetshenziswa ukufunda inkumbulo yendawo ne-cache yomlando ye-KV, ukunaka okuphambanayo kufunda izimo zombhalo, bese kusetshenziswa indlela ye-autoregressive ukukhiqiza izingxenye ngokulandelana.

Isikhathi sangempela ekhadini elilodwa labathengi: Ikhadi elilodwa le-RTX 5090 likhiqiza i-480p, 60 FPS nokubambezeleka okungu-112ms, okwenza kube ngokokuqala ukuthi umdlalo wemodeli yevidiyo uphume emaqenjini amakhulu anamakhadi amaningi asebenzisa izindawo zokuhlola.
Isimo somhlaba singaphathwa ngokusobala: Injini yomdlalo igcina imithetho yomhlaba ngesikhodi; isimo singafundwa, siguqulwe futhi siqinisekiswe, kuqedwe inkinga yokuphambuka yesizukulwane sangaphambilini lapho umhlaba wawukhona kuphela enkumbulweni yemodeli.
Shintsha indlela yokudlala ngomusho owodwa: I-Coding Agent iguqula ngokuqondile imiyalelo yolimi lwemvelo ibe yimithetho nesikhodi yomdlalo, ngakho ukuthuthukisa imibono akusabambeki kulayini wokukhiqiza izimpahla ze-3D.
Ukuzinza nokuvumelana isikhathi eside: Inkumbulo yendawo + amafreyimu ayisithenjwa + indlela ye-Caption eguqukayo kuqinisekisa ukuthi ukwakheka kwendawo nobunikazi bezinto akuphazamiseki ngesikhathi sokudlala isikhathi eside.
Ukuhola okukhulu ekusebenzeni: Ukubambezeleka cishe kungu-1/90 kwe-Code World Model no-1/5 kwe-StreamRender-H3, kanti lesi yisona kuphela isixazululo esitholakala kubathengi.

Isikhathi sangempela ekhadini elilodwa labathengi: Ikhadi elilodwa le-RTX 5090 likhiqiza i-480p, 60 FPS nokubambezeleka okungu-112ms, okwenza kube ngokokuqala ukuthi umdlalo wemodeli yevidiyo uphume emaqenjini amakhulu anamakhadi amaningi asebenzisa izindawo zokuhlola.
Isimo somhlaba singaphathwa ngokusobala: Injini yomdlalo igcina imithetho yomhlaba ngesikhodi; isimo singafundwa, siguqulwe futhi siqinisekiswe.
Shintsha indlela yokudlala ngomusho owodwa: I-Coding Agent iguqula ngokuqondile imiyalelo yolimi lwemvelo ibe yimithetho nesikhodi yomdlalo, ngakho ukuthuthukisa imibono akusabambeki kulayini wokukhiqiza izimpahla ze-3D.
Ukuzinza nokuvumelana isikhathi eside: Inkumbulo yendawo + amafreyimu ayisithenjwa + indlela ye-Caption eguqukayo kuqinisekisa ukuthi ukwakheka kwendawo nobunikazi bezinto akuphazamiseki ngesikhathi sokudlala isikhathi eside.
Ukuhola okukhulu ekusebenzeni: Ukubambezeleka cishe kungu-1/90 kwe-Code World Model no-1/5 kwe-StreamRender-H3, kanti lesi yisona kuphela isixazululo esitholakala kubathengi.

Iwebhusayithi esemthethweni yephrojekthi: https://matrix-game-v4-0.github.io/page/
Indawo yokugcina ye-GitHub: https://github.com/SkyworkAI/Matrix-Game.git
Indawo yamamodeli ye-HuggingFace: https://huggingface.co/Skywork/Matrix-Game-4.0

Ukuthuthukisa ngokushesha ama-prototype emidlalo: Amaqembu amancane achaza indlela yokudlala ngolimi lwemvelo, bese i-Coding Agent ikhiqiza ngokuqondile i-demo yomdlalo esebenzayo, kuncishiswe kakhulu isikhathi sokuqinisekisa iphrojekthi.
Ukwenza imidlalo ibe ngeyakho (UGC): Abadlali bangashintsha imithetho nesitayela sokubukeka ngokwezinto abazithandayo, bakhe umuzwa womdlalo ohlukile kumuntu ngamunye.
Imfundo nokuzilolonga kohlelo lwe-AI: Njengephrojekthi yomthombo ovulekile, abathuthukisi bangafunda ukwakhiwa kwe-Harness “yenjini + imodeli yevidiyo”, bahlole indlela yokuhlanganisa amamodeli omhlaba nezinjini zemidlalo.
Ukukhiqizwa kokuqukethwe kwemidlalo ngezindleko eziphansi: Kunciphisa ukuthembela ezimpahleni ze-3D ezinemininingwane nasezinjinini ezinkulu. Abadali abazimele bangasebenzisa izakhiwo zamamodeli alula ukukhiqiza izithombe zezinga le-cinema, behlise izindleko zokukhiqiza.