Hi3D 1.0 Uyasebenza

I-Hi3D 1.0 iyimodeli yaseShayina yokukhiqiza i-AI 3D ekhishwe njenge-open source yi-Math Magic, ubuchwepheshe bayo buvela ku-Sparc3D ye-NeurIPS 2025. I-Hi3D 1.0 yaqeqeshwa cishe ngama-3D asset ayizigidi ezi-2 akhethiwe, isebenzisa uhlaka lokucwengwa kwejiyomethri oluncane, futhi isekela ukucwengwa kwezigaba 512³→1024³→1536³. Ingakhiqiza i-mesh engenamanzi enokulungiswa okuphezulu isebenzisa isithombe esisodwa, futhi ukunemba kokubuka okuningi kanye nobuningi bemininingwane kwayo kuhamba phambili phakathi kwezindlela ezifanayo. Izinga lokuphumelela lokusika lifinyelela ku-90%, ngakho ingena ngqo enqubweni yokuphrinta nge-3D.

Ukukhiqiza i-3D ngesithombe esisodwa: Faka isithombe esisodwa sereferensi ukuze ukhiqize imodeli ephelele enezinhlangothi ezintathu, uhlanganise ngokufanele ingemuva nezindawo ezivinjwe ezingabonisiwe.
Ukucwengwa kwejiyomethri enokulungiswa okuphezulu: Isekela ukucwengwa kwezigaba 512³→1024³→1536³, ukuze izakhi ezincane njengezimpaphe namaphethini amathayi zibonakale ngokucacile.
Ilungele ukuphrinta nge-3D: Ikhipha i-mesh engenamanzi, enezinga lokuphumelela lokusika elingu-90%, engangena ngqo enqubweni yokulungiselela ukuphrinta, futhi ingahlukanisa izingxenye, iklame izakhi zokuxhuma, iphinde ihlele imibala ngokuzenzakalelayo.
Ukuguqulwa okusizwa yi-Agent: Sebenzisa ingxoxo yolimi lwemvelo ukulungisa ukuma, izakhiwo ezithile, ukwakheka kwe-surface kanye ne-mesh, ulungise amaphutha ezindawo ezilengayo nezixhumanisi, futhi ulungiselele ukuphrinta, imidlalo, ukugqwayiza nezinye izinhloso.
Ukulethwa kwezicelo kuyo yonke inqubo: Imiphumela ekhiqiziwe ingalandwa iye ku-Blender noma kuma-game engine ukuze iqhubeke icutshungulwa, iboshwe kumathambo, yenziwe izenzo zobuso, noma iguqulwe ibe yimisebenzi ephathekayo efana nezithombe eziqoqwayo namathoyizi emfashini.
Amandla okuhlanganisa avulekile: Imodeli ye-open source ingafakwa ngaphakathi, i-interface ye-MCP isekela ukubizwa kwemisebenzi yokukhiqiza ama-Agent angaphandle, kanti i-Hi3D Agent isebenzela abadali ngokuqondile.

Landela i-WeChat bese uphendula ngokuthi “开源” ukuze ujoyine iqembu lezingxoxo lamaphrojekthi e-AI open source
Ukumelwa kwejiyomethri okuncane (Sparcubes + Sparconv-VAE): I-Hi3D iqhubeka nokusebenzisa uhlaka oluyisisekelo lwe-Sparc3D, imele ukwakheka kwe-3D ngendlela encane: igcina kuphela ama-voxel asebenzayo aseduze nobuso bento. Izici ezifihlekile ziguquka kuma-coordinate obuso obumahhadlahhadla, bese kukhishwa i-mesh ngemva kokudecoda nge-VAE ezisika ngokwayo. Lokhu kugcina imininingwane yejiyomethri eyinkimbinkimbi futhi kwehlise kakhulu ukusetshenziswa kwememori ye-GPU nezindleko zokubala ekulungisweni okuphezulu.
Ukucwengwa okuqhubekayo kwamazinga amaningi: Imodeli ephezulu iqala ikhiqize i-mesh emangelengele isebenzisa isithombe, inqume umkhawulo jikelele nezingxenye ezinkulu; i-Hi3D ibhekisa esithombeni nakwi-mesh emangelengele ngesikhathi esisodwa ukuze yenze ukucwengwa kwejiyomethri okuncane. Ngemva kokuqeda izinga ngalinye, ikhipha i-mesh bese iyisebenzisa njengesisekelo somjikelezo olandelayo wokucwengwa, iqhubeke kancane kancane isuka ku-512³ iye ku-1024³ naku-1536³, ukuze “kuqale kunqunywe ukwakheka okuyisisekelo, bese kucwengwa imininingwane ezingeni ngalinye”.
Ukuqeqeshwa kwedatha enkulu nokuqinisekiswa kokukhiqizwa: Idatha yokuqeqesha yandiswe isuka kuma-asset angu-500,000 ephepheni le-Sparc3D yaba cishe ngama-3D asset ayizigidi ezi-2 akhethiwe. Lokhu kwenza imodeli ihlangane nezinhlobo eziningi zezakhiwo zezinto, kunikeze amasampula okufunda anele ukuze iqagele futhi igcwalise ingemuva elingabonakali esithombeni esisodwa. Ngaphambi kokuvulwa, iphinde yahlolwa enqubweni yangempela yokukhiqiza ukuze kuqinisekiswe ukuthi ama-asset akhiqiziwe angaqhubeka nokucutshungulwa futhi angene ezinhlelweni ezilandelayo.

Indlela yokuqala: Yizame mahhala ku-inthanethi Vula indawo yokuyizama: vakashela i-Hugging Face Space (https://huggingface.co/spaces/ilcve21/Sparc3D) noma ikhasi elisemthethweni le-open source https://www.hi3d.ai/hi3d-1-0-open-source.
Layisha isithombe sereferensi: Nikeza isithombe sento oyifunayo, imodeli izosisebenzisa njengesisekelo sokubuka ukuze ikhiqize imodeli ye-3D.
Lindela ukukhiqizwa bese ubuka: Uhlelo luqedela ngokuzenzakalelayo ukucwengwa kusuka ku-mesh emangelengele kuya ku-512³→1024³→1536³, bese ubuka imiphumela ezinhlangothini eziningi ku-inthanethi.
Landa futhi usebenzise: Landa i-mesh ekhiqiziwe njengefayela lemodeli, uyingenise ku-Blender, kuma-game engine nakwezinye izinhlelo ukuze uqhubeke nokuyicubungula noma uyisebenzise ngqo ekuphrinteni nge-3D.
Indlela yesibili: Faka imodeli ye-open source endaweni yangakini Thola ikhodi nama-weight: Klona ikhodi ku-GitHub (https://github.com/hi3d-official/Hi3Dv1.0), bese ulanda ama-model weight ku-Hugging Face.
Lungisa imvelo: Faka ama-dependency, ulungise imvelo yokusebenzisa i-GPU, bese ulayisha i-VAE ezisika ngokwayo nemodeli yokucwengwa okuncane.
Faka isithombe + i-mesh emangelengele: Shayela i-interface ye-inference, udlulisele isithombe sereferensi kanye ne-mesh emangelengele ekhiqizwe imodeli ephezulu, bese wenza ukucwengwa kwejiyomethri okuncane.
Khipha i-mesh: Khipha imiphumela ye-mesh engenamanzi ngokwamazinga. Ungama kunoma yiliphi izinga lokulungiswa, noma uyisebenzise njengesisekelo sokucwengwa okulandelayo.

Ukunemba kwezinga lokukhiqiza: Ukulungiswa kokukhipha okungu-1536³, izakhi ezincane njengezimpaphe namaphethini amathayi zibonakala ngokucacile, kanti ukunemba kokubuka okuningi nobuningi bemininingwane kuhamba phambili phakathi kwezindlela ezifanayo.
I-mesh engenamanzi yangempela: Okukhiphayo kungasikwa futhi kuphrintwe ngokuqondile, ngezinga lokuphumelela lokusika elingu-90%, ngaphandle kwesidingo sokulungisa imodeli isikhathi eside kusukela ekukhiqizweni kuya entweni ephathekayo.
I-open source futhi ingasetshenziswa kwezentengiselwano: Ikhodi nama-weight kuvuleke ngokuphelele, kuqinisekiswe ngenqubo yangempela yokukhiqiza, futhi amabhizinisi angayifaka ngaphakathi futhi ayithuthukise ngokuzethemba.
Ukuhlukaniswa komsebenzi phakathi kwamamodeli ochwepheshe: Ekuhlolweni okuqhathanisayo namamodeli amakhulu ajwayelekile afana ne-GPT-6 Astra, ihamba phambili ngokuphelele ekubuyiseleni umkhawulo, ubudlelwano bevolumu nemininingwane, okuqinisekisa ukuhlukaniswa okufanele kokuthi “i-LLM ilawula inhloso, imodeli yoqeqesho ilawula ukukhiqizwa”.
Ukumboza yonke inqubo: Kusukela ekukhiqizeni ngesithombe esisodwa, ekuguqulweni ngengxoxo, ekuboshweni kwamathambo, ekuhlanganisweni emidlalweni nasekuphrinteni izingxenye ezihlukanisiwe, ixhumanisa lonke uchungechunge lokudala “ukukhiqiza—ukuguqula—ukusebenzisa”.
Is生态 esivulekile sokuhlanganisa: Ihlinzeka nge-interface ye-MCP ukuze kubizwe ama-Agent angaphandle, futhi isebenzisa i-Hi3D Agent ukusebenzela abadali abajwayelekile, ihlanganise izidingo zonjiniyela nabasebenzisi bokugcina.

Iwebhusayithi esemthethweni yephrojekthi：https://www.hi3d.ai/zh/hi3d-1-0-open-source
Indawo yokugcina ye-GitHub：https://github.com/hi3d-official/Hi3Dv1.0
Indawo yamamodeli ye-HuggingFace：https://huggingface.co/spaces/ilcve21/Sparc3D

Ukuphrinta nge-3D namathoyizi emfashini nezithombe eziqoqwayo: Ngemva kokukhiqiza i-mesh engenamanzi, ihlukanisa izingxenye ngokuzenzakalelayo, iklame izakhi zokuxhuma futhi ihlele imibala, ukuze iphrinte ngokuqondile izinto eziphathwayo ezifana nezithombe eziqoqwayo namathoyizi emfashini.
Ukuthuthukiswa kwemidlalo: Khiqiza ngokushesha ama-asset abalingiswa, izinsiza nezindawo, uwafake endaweni yokusebenzisana ukuze kwakhiwe i-demo edlalekayo, unciphise kakhulu isikhathi sokuthuthukisa ama-prototype.
Ukugqwayiza nokuqukethwe kwesikhathi sangempela: Yenza lula abalingiswa abamile, ibabophe kumathambo, iphinde icubungule izenzo nezimpawu zobuso ukuze bangene enqubweni yokugqwayiza noma yokunikezwa kwesikhathi sangempela.
Ifilimu, ithelevishini nezithombe zedijithali: Yakha kabusha futhi ikhulise amamodeli ayinkimbinkimbi asezingeni lezithombe ngokulungiswa okuphezulu, igcine izakhi ezinemininingwane njengezimpaphe nokwakheka kwezikhali, futhi isize ekudaleni ubuciko bedijithali.
Izicelo ezime mpo ezizimele zamabhizinisi: Sebenzisa imodeli ye-open source efakwe ezinhlelweni zangaphakathi ukwakha izinqubo ezenzelwe izidingo ezithile, njengokuboniswa kwemikhiqizo ye-e-commerce ye-3D, izifundo zemfundo nokumodela izingxenye zezimboni.