• mbendera

OpenAI Point E: Pangani mtambo wa 3D kuchokera pamafunde ovuta mumphindi pa GPU imodzi

M'nkhani yatsopano Point-E: Dongosolo lopangira mitambo ya 3D kuchokera kuzizindikiro zovuta, gulu lofufuza la OpenAI limayambitsa Point E, 3D point cloud text conditional synthesis system yomwe imagwiritsa ntchito mitundu yophatikizika kuti ipange mawonekedwe osiyanasiyana komanso ovuta a 3D oyendetsedwa ndi zolemba zovuta. zizindikiro.mumphindi pa GPU imodzi.
Kuchita modabwitsa kwa zitsanzo zamakono zamakono zopanga zithunzi zalimbikitsa kufufuza m'badwo wa zinthu za 3D.Komabe, mosiyana ndi mitundu ya 2D, yomwe imatha kutulutsa mphindi kapena masekondi, mitundu yopangira zinthu imafuna maola angapo a GPU kuti apange chitsanzo chimodzi.
M'nkhani yatsopano Point-E: Dongosolo lopangira mitambo ya 3D kuchokera kuzizindikiro zovuta, gulu lofufuza la OpenAI likupereka Point · E, njira yopangira zolemba zamtambo wa 3D.Njira yatsopanoyi imagwiritsa ntchito njira yofalitsira kuti ipange mawonekedwe osiyanasiyana komanso ovuta a 3D kuchokera kuzizindikiro zamawu ovuta pamphindi imodzi kapena ziwiri pa GPU imodzi.
Gululi likuyang'ana kwambiri pazovuta zosinthira zolemba kukhala 3D, zomwe ndizofunikira kuti pakhale demokalase yopanga zinthu za 3D pazogwiritsa ntchito zenizeni padziko lapansi kuyambira zenizeni zenizeni ndi masewera mpaka kupanga mafakitale.Njira zomwe zilipo zosinthira malemba ku 3D zimagwera m'magulu awiri, omwe ali ndi zovuta zake: 1) zitsanzo zopangira zingagwiritsidwe ntchito popanga zitsanzo bwino, koma sangathe kukweza bwino zizindikiro zosiyanasiyana ndi zovuta;2) chithunzi chazithunzi chophunzitsidwa kale kuti chigwirizane ndi zolemba zovuta komanso zosiyanasiyana, koma njira iyi ndi yowonjezereka kwambiri ndipo chitsanzocho chikhoza kukhazikika mu minima yapafupi yomwe sagwirizana ndi zinthu za 3D zatanthauzo kapena zogwirizana.
Choncho, gululo linafufuza njira ina yomwe cholinga chake ndi kuphatikiza mphamvu za njira ziwiri zomwe zili pamwambazi, pogwiritsa ntchito chitsanzo cha kufalitsa malemba ndi chithunzi chomwe chimaphunzitsidwa pamagulu akuluakulu azithunzi zazithunzi (zolola kuti zigwirizane ndi zizindikiro zosiyanasiyana komanso zovuta) ndi mtundu wa 3D wofalitsa zithunzi wophunzitsidwa pamagulu ang'onoang'ono azithunzi.chithunzi-3D awiriawiri deta.Mtundu wotengera mawu kupita ku chithunzi choyamba umapanga zitsanzo za chithunzi cholowetsamo kuti apange choyimira chimodzi chokha, ndipo chithunzi cha 3D chimapanga mtambo wa 3D potengera chithunzi chomwe mwasankha.
Kuchuluka kwa lamuloli kumachokera pamapangidwe opangira omwe angopangidwa kumene kuti apange zithunzi kuchokera pamawu (Sohl-Dickstein et al., 2015; Song & Ermon, 2020b; Ho et al., 2020).Amagwiritsa ntchito mtundu wa GLIDE wokhala ndi magawo 3 mabiliyoni a GLIDE (Nichol et al., 2021), okonzedwa bwino pamitundu yosinthidwa ya 3D, monga mawonekedwe awo osinthira malemba ndi zithunzi, komanso mitundu yofananira yomwe imapanga mitambo ya RGB ngati yawo. kusintha chitsanzo.zithunzi ku chithunzi.Zithunzi za 3D.
Ngakhale ntchito yam'mbuyomu idagwiritsa ntchito zomangamanga za 3D pokonza mitambo, ofufuzawo adagwiritsa ntchito njira yosavuta yopangira ma transducer (Vaswani et al., 2017) kuti athandizire bwino.M'mapangidwe awo amitundu yophatikizika, zithunzi zamtambo zimaperekedwa koyamba kukhala mtundu wophunzitsidwa kale wa ViT-L/14 CLIP ndiyeno ma meshes otuluka amalowetsedwa mu chosinthira ngati zolembera.
Pakafukufuku wawo woyeserera, gululo lidayerekeza njira ya Point·E ndi mitundu ina ya 3D yopangira zigoli kuchokera pakuzindikira kwa chinthu cha COCO, magawo, ndi ma siginecha.Zotsatira zimatsimikizira kuti Point · E imatha kupanga mawonekedwe osiyanasiyana ndi ovuta a 3D kuchokera ku zizindikiro zovuta zolemba ndikufulumizitsa nthawi yolondolera ndi maulamuliro amodzi kapena awiri a ukulu.Gululi likuyembekeza kuti ntchito yawo ilimbikitsa kufufuza kwina mu 3D text synthesis.
Njira yotsatsira mtambo ndi nambala yowunikira ikupezeka pa GitHub ya polojekitiyi.Document Point-E: Dongosolo lopangira mitambo ya 3D kuchokera pazowunikira zovuta lili pa arXiv.
Tikudziwa kuti simukufuna kuphonya nkhani zilizonse kapena zopezedwa zasayansi.Lembetsani kumakalata athu otchuka a Synced Global AI Weekly kuti mulandire zosintha za AI sabata iliyonse.


Nthawi yotumiza: Dec-28-2022