I-OpenAI Point E: Dala ifu lephoyinti le-3D kusuka kumagagasi ayinkimbinkimbi ngamaminithi ku-GPU eyodwa

Esihlokweni esisha, i-Point-E: Uhlelo lokukhiqiza amafu wephoyinti le-3D kusuka kumasiginali ayinkimbinkimbi, ithimba labacwaningi be-OpenAI lethula i-Point E, isistimu ye-3D point cloud text cloud synthesis system esebenzisa amamodeli okusabalalisa ukudala izimo ze-3D ezihlukahlukene neziyinkimbinkimbi eziqhutshwa umbhalo oyinkimbinkimbi. izimpawu.ngemizuzu ku-GPU eyodwa.
Ukusebenza okumangalisayo kwamamodeli wanamuhla wokukhiqiza izithombe kukhuthaze ucwaningo ekukhiqizeni izinto zombhalo we-3D.Nokho, ngokungafani namamodeli e-2D, angakwazi ukukhiqiza okukhiphayo ngemizuzu noma imizuzwana, amamodeli akhiqiza izinto ngokuvamile adinga amahora ambalwa omsebenzi we-GPU ukuze enze isampula eyodwa.
Esihlokweni esisha, i-Point-E: Uhlelo lokukhiqiza amafu ephoyinti le-3D kusuka kumasiginali ayinkimbinkimbi, ithimba locwaningo le-OpenAI lethula i-Point·E, uhlelo lokuhlanganisa olunemibandela lombhalo lwamafu wephoyinti le-3D.Le ndlela entsha isebenzisa imodeli yokusakaza ukuze idale izimo ze-3D ezihlukahlukene neziyinkimbinkimbi kusukela kumasignali ombhalo ayinkimbinkimbi ngomzuzu nje noma amabili ku-GPU eyodwa.
Ithimba ligxile enseleleni yokuguqula umbhalo ube yi-3D, okubalulekile ekudalweni kwentando yeningi ekudalweni kokuqukethwe kwe-3D kwezinhlelo zokusebenza zomhlaba wangempela kusukela kokungokoqobo okubonakalayo kanye negeyimu kuya kumklamo wezimboni.Izindlela ezikhona zokuguqula umbhalo ube yi-3D ziwela ezigabeni ezimbili, ngayinye enezinkinga zayo: 1) amamodeli akhiqizayo angasetshenziswa ukukhiqiza amasampula ngendlela efanele, kodwa awakwazi ukukala ngokuyimpumelelo kumasignali ombhalo ahlukahlukene futhi ayinkimbinkimbi;2) imodeli yesithombe sombhalo esiqeqeshwe kusengaphambili ukuze isingathe izinkomba zombhalo eziyinkimbinkimbi nezihlukahlukene, kodwa le ndlela inamandla ngokwezibalo futhi imodeli ingabhajwa kalula ku-minima yasendaweni engahambisani nezinto eziphusile noma ezihambisanayo ze-3D.
Ngakho-ke, ithimba lihlole enye indlela ehlose ukuhlanganisa amandla alezi zindlela ezimbili ezingenhla, kusetshenziswa imodeli yokusabalalisa umbhalo kuya kwesithombe oqeqeshwe kusethi enkulu yamapheya ezithombe zombhalo (okuyivumela ukuthi iphathe amasignali ahlukahlukene futhi ayinkimbinkimbi) futhi imodeli yokusabalalisa isithombe se-3D eqeqeshwe kusethi encane yamapheya esithombe sombhalo.image-3D ipheya idathasethi.Imodeli yombhalo uye esithombeni iqala isampula yesithombe sokufakwayo ukuze idale isethulo esisodwa sokwenziwa, futhi imodeli yesithombe ukuya ku-3D idala ifu lephoyinti le-3D ngokusekelwe esithombeni esikhethiwe.
Isitaki sokukhiqiza somyalo sisekelwe kuzinhlaka zokukhiqiza ezihlongoziwe kamuva nje zokukhiqiza izithombe ezingokomthetho zisuka embhalweni (Sohl-Dickstein et al., 2015; Ingoma & Ermon, 2020b; Ho et al., 2020).Basebenzisa imodeli ye-GLIDE enamapharamitha we-GLIDE angamabhiliyoni angu-3 (uNichol et al., 2021), acushwe kahle kumamodeli anikeziwe e-3D, njengemodeli yawo yokuguqula umbhalo uye esithombeni, kanye nesethi yamamodeli okusabalalisa akhiqiza amafu e-RGB njengawo. imodeli yoguquko.izithombe esithombeni.Amamodeli we-3D.
Ngenkathi umsebenzi wangaphambilini usebenzisa izakhiwo ze-3D ukucubungula amafu ephuzu, abacwaningi basebenzisa imodeli elula esekelwe ku-transducer (Vaswani et al., 2017) ukuze bathuthukise ukusebenza kahle.Ekwakheni kwabo imodeli yokusatshalaliswa, izithombe zamafu amaphoyinti ziqale ziphakelwe imodeli ye-ViT-L/14 CLIP eqeqeshwe kusengaphambili bese amameshi okukhiphayo afakwa kusiguquli njengomaka.
Ocwaningweni lwalo lobuchwepheshe, ithimba liqhathanise indlela ehlongozwayo ye-Point·E namanye amamodeli e-3D akhiqizayo kumasiginali wokuthola amaphuzu asuka ekutholweni kwento ye-COCO, ukuhlukaniswa, namasethi wedatha wesiginesha.Imiphumela iqinisekisa ukuthi i-Point·E iyakwazi ukukhiqiza izimo ze-3D ezihlukene neziyinkimbinkimbi kusukela kumasignali ombhalo ayinkimbinkimbi futhi isheshise isikhathi sokunquma nge-oda elilodwa kuya kwamabili obukhulu.Ithimba lithemba ukuthi umsebenzi walo uzogqugquzela ucwaningo olwengeziwe ekuhlanganiseni umbhalo we-3D.
Imodeli yokusakaza yamafu eqeqeshelwe kusengaphambili kanye nekhodi yokuhlola kuyatholakala ku-GitHub yephrojekthi.I-Document Point-E: Uhlelo lokudala amafu ephoyinti le-3D kusuka kuzinkomba eziyinkimbinkimbi liku-arXiv.
Siyazi ukuthi awufuni ukuphuthelwa yinoma yiziphi izindaba noma ukutholwa kwesayensi.Bhalisela iphephandaba lethu elidumile le-Synced Global AI lamasonto onke ukuze uthole izibuyekezo zeviki zonke ze-AI.