I-Point E yi-AI entsha ye-OpenAI yokwenza izithombe
I-OpenAI yandise amakhono esofthiwe yayo yombhalo ibe yisithombe kusukela ezilinganisweni ezimbili kuya kwezintathu ngokukhishwa kwe -Point-E, i -AI ekhiqiza izithombe ze-3D kusuka ezicelweni zombhalo. Ngokusho kwedokhumenti eshicilelwe ne-codebase, i-Point-E ayidingi ikhompyutha ephezulu ukuze isebenze futhi ingakhiqiza imodeli ngaphansi kwemizuzu emibili isebenzisa i-Nvidia V100 GPU eyodwa.
Ngalokhu singabona ukuthi ama-generator angamamodeli e-3D angaba impumelelo elandelayo ezothatha umhlaba we-AI ngesivunguvungu. I-OpenAI ibisebenza kakhulu kulo nyaka, njengoba ngokwesibonelo lapha kubhulogi sikhulume ngenkampani ngamaphrojekthi ayo amabili amakhulu: i-Dall-E 2 ne-ChatGPT.
Phakathi kwalezi zingxenyekazi ezimbili ezinkulu ze-AI, inkampani isize ekukhiqizeni izithombe kanye nemibhalo emide evela kumyalezo olula, futhi manje inkampani isibuyile nomqondo wesithathu, owasungulwa ngaphambi nje kukaKhisimusi ukuze udale umsindo. Lo mqondo wesithathu, obizwa ngokuthi i-Point-E, ulandela iphethini efanayo, udala okuqukethwe kwe-3D kusuka ezikhuthazweni ezilula.
Njengoba kumenyezelwe esihlokweni socwaningo esanyatheliswa yithimba le-OpenAI, i-Point-E isebenza ngezigaba ezimbili: okokuqala isebenzisa i-AI yombhalo ibe yisithombe ukuguqula isicelo sakho somlomo sibe isithombe, bese isebenzisa umsebenzi wesibili ukuguqula lesi sithombe sibe imodeli ye-3D.
I-Point-E ingumthombo ovulekile futhi ikhodi yomthombo iyatholakala ku-GitHub . Noma kunjalo, kungaba nzima kancane ukuqala, njengoba abasebenzisi kudingeka bajwayelane kahle namathuluzi omugqa wemiyalo futhi uhlelo kumele lube ne-Python efakiwe, ngokungafani ne-ChatGPT, lapho abasebenzisi bangabhalisa khona kuwebhusayithi futhi bahlole amakhono abo.
I-Point-E ayidali izinto ze-3D ngomqondo wendabuko; kunalokho, ikhiqiza amafu e-point, okungukuthi, amasethi ahlukene wamaphuzu edatha esikhaleni amelela isimo se-3D, yingakho isifinyezo esihlekisayo.
(I-"E" ku-Point-E imfushane ngokuthi "ukusebenza kahle" njengoba ibonakala ishesha kunezindlela zangaphambilini zokunikeza izinto ze-3D.) Ngokwabathuthukisi, amafu amaphuzu kulula ukuhlanganisa, kodwa awathathi ukuma noma ukuthungwa okuhle kwento, okuwumkhawulo omkhulu we-Point-E njengamanje.
Ukuze kunqotshwe lo mkhawulo , ithimba le-Point-E laqeqesha uhlelo olwengeziwe lwe-AI ukuguqula amafu e-Point-E abe ama-meshes. (Ama-Meshes—amaqoqo ama-vertice, imiphetho, kanye nobuso obuchaza into—avame ukusetshenziswa ekubumbeni nasekuklameni kwe-3D.) Kodwa abacwaningi baphawula ephepheni labo ukuthi imodeli ngezinye izikhathi ingaphuthelwa izingxenye zezinto , okuholela ezimeni ezisontekile noma ezivinjiwe. Ngaphezu kwemodeli ekhiqiza ama-mesh, equkethe ngokwayo, i-Point-E inamamodeli amabili: imodeli yombhalo-isithombe kanye nemodeli yesithombe se-3D.
Imodeli yombhalo-kuya-esithombeni, efana nezinhlelo zokukhiqiza ubuciko njenge-DALL-E 2 kanye ne-Stable Diffusion, yaqeqeshwa ngezithombe ezifakwe amalebula ukuze kuqondwe ubudlelwano phakathi kwamagama nemibono ebonakalayo . Imodeli yesithombe se-3D, ngakolunye uhlangothi, yondliwa isethi yezithombe ezihlotshaniswa nezinto ze-3D ukuze ifunde ukuthi ingazihumusha kanjani lezi zinto ezimbili ngempumelelo. Uma inikezwa umlayezo wombhalo, isibonelo, "igiya eliphrintwayo le-3D, igiya elilodwa elinobubanzi obungu-7 cm kanye nobukhulu obungu-1 cm," imodeli ye-Point-E yombhalo-kuya-esithombeni ikhiqiza into eyenziwe ngokwenziwa edluliselwa kumodeli yesithombe se-3D.
Lona wakamuva ube esekhiqiza ifu lamaphoyinti. Ngemva kokuqeqesha amamodeli kudathasethi yezinto ze-3D "eziyizigidi ezimbalwa" kanye nemethadatha ehlobene, i-Point-E ikwazile ukukhiqiza amafu amaphuzu anombala avame ukuhambisana nokwaziswa kombhalo, kusho abacwaningi be-OpenAI. Ayiphelele: Imodeli yesithombe se-3D ye-Point-E kwesinye isikhathi ayisiqondi isithombe esivela kumodeli yesithombe sombhalo, okuholela kumumo ongahambisani nomlayezo wombhalo. Noma kunjalo, ishesha kakhulu kunesimo sangaphambilini sobuciko, okungenani ngokusho kweqembu le-OpenAI.
Umphumela ukude nokufinyelela ikhwalithi yokuthengiswa kwe-3D kufilimu noma kugeyimu yevidiyo. Kodwa akufanele kube njalo. Uma sezifakwe kuhlelo lokusebenza lwe-3D njenge-Blender, zingaguqulwa zibe amameshi aqoshiwe abukeka njengezithombe ezijwayelekile ze-3D.
"Nakuba indlela yethu ingakafiki ekusetshenzisweni kwekhwalithi yesampula, ingumyalelo owodwa noma emibili yobukhulu ngokushesha, okuwukuyekethisa okusebenzayo kwamanye amacala okusetshenziswa," abacwaningi be-OpenAI bachaza ephepheni elichaza iphrojekthi.
Okokugcina, uma unesithakazelo sokufunda kabanzi nge-Point-E noma ikhodi yomthombo, ungathola imininingwane kusixhumanisi esilandelayo.