OpenAI Point E: Gadzira 3D point gore kubva kune yakaoma waveforms mumaminitsi pane imwe GPU

Muchinyorwa chitsva Point-E: Iyo sisitimu yekugadzira 3D point makore kubva kune yakaoma masaini, OpenAI timu yekutsvagisa inosuma Point E, ye3D point Cloud text conditional synthesis system inoshandisa diffusion modhi kugadzira akasiyana uye akaomarara maumbirwo e3D anofambiswa nemavara akaoma. cues.mumaminetsi pane imwe GPU.
Kuita kunoshamisa kwemamodheru emazuva ano ekugadzirwa kwemifananidzo kwakakurudzira tsvakiridzo muchizvarwa chezvinyorwa zve 3D.Nekudaro, kusiyana nemhando dze2D, dzinogona kuburitsa mumaminitsi kana kunyange masekonzi, mamodhi ekugadzira zvinhu anowanzoda maawa akati wandei ebasa reGPU kugadzira sampuli imwe chete.
Muchinyorwa chitsva Point-E: Iyo sisitimu yekugadzira 3D point makore kubva kune yakaoma masaini, OpenAI timu yekutsvagisa inopa Point · E, mameseji conditional synthesis system ye3D point makore.Iyi nzira nyowani inoshandisa modhi yekuparadzira kugadzira akasiyana uye akaomarara maumbirwo e3D kubva kune yakaoma masaini masaini muminiti imwe chete kana maviri paGPU imwechete.
Chikwata chakatarisana nedambudziko rekushandura mameseji kuenda ku3D, iyo yakakosha kugutsa demokrasi kugadzirwa kwezvinhu zve3D zvemashandisirwo epasirese echokwadi kubva pachokwadi chaicho uye kutamba kune dhizaini yemaindasitiri.Nzira dziripo dzekushandura zvinyorwa ku 3D dzinowira muzvikamu zviviri, chimwe nechimwe chine zvipingamupinyi zvayo: 1) maitiro ekugadzira anogona kushandiswa kugadzira sampuli zvakanaka, asi haagoni kukwira zvakanaka kune zvakasiyana-siyana uye zvakaoma zvinyorwa zvemashoko;2) chimiro chekare-chakadzidziswa chinyorwa-chifananidzo chekubata zvakaoma uye zvakasiyana-siyana zvinyorwa zvemavara, asi nzira iyi yakakura computation uye modhi yacho inogona kunamira zviri nyore mune minima yemuno isingaenderane nezvainoreva kana zvinoenderana 3D zvinhu.
Nokudaro, boka racho rakatsvaga imwe nzira iyo inotarisira kubatanidza simba remaitiro maviri ari pamusoro apa, uchishandisa chinyorwa-ku-chifananidzo chekuparadzanisa modhi yakadzidziswa pane yakakura seti yezvinyorwa-mifananidzo pairs (ichibvumira kubata zviratidzo zvakasiyana uye zvakaoma) uye 3D mufananidzo diffusion modhi yakadzidziswa pane diki seti yemavara-mifananidzo peya.mufananidzo-3D pair dataset.Iyo yemavara-kune-mufananidzo modhi inotanga sample mufananidzo wekuisa kuti ugadzire imwe chete yakagadzirwa inomiririra, uye mufananidzo-ku-3D modhi inogadzira 3D point Cloud zvichienderana nemufananidzo wakasarudzwa.
Murairo wekugadzira stack wakavakirwa pane ichangoburwa yakarongwa gadziriro yekugadzira mifananidzo kubva kune zvinyorwa (Sohl-Dickstein et al., 2015; Rwiyo & Ermon, 2020b; Ho et al., 2020).Ivo vanoshandisa GLIDE modhi ine 3 bhiriyoni GLIDE paramita (Nichol et al., 2021), yakanatswa-yakatarwa pane yakapihwa 3D modhi, seyavo-ku-ku-mufananidzo yekushandura modhi, uye seti yemamodhiyo ekuparadzanisa anoburitsa RGB point makore seyavo. shanduko modhi.mifananidzo kumufananidzo.3D mhando.
Kunyange basa rekare raishandisa 3D zvivakwa kugadzirisa makore, vaongorori vakashandisa yakapfava transducer-based modhi (Vaswani et al., 2017) kuvandudza kushanda zvakanaka.Mune yavo yekuparadzira modhi yedhizaini, mapikicha emakore ekutanga anotanga kupihwa mune yakadzidziswa ViT-L/14 CLIP modhi uye ipapo meshes anobuda anopihwa mushanduri semamaki.
Muchidzidzo chavo chehunyanzvi, timu yakafananidza nzira yakatsanangurwa yePoint · E nemamwe mamodhireni e3D pakuisa zvibodzwa kubva kuCOCO kuona chinhu, kupatsanura, uye siginecha dataset.Mhedzisiro yacho inosimbisa kuti Point · E inokwanisa kuburitsa akasiyana uye akaomarara maumbirwo e3D kubva kune yakaoma mavara masaini uye nekumhanyisa nguva yekufungidzira nemirairo imwe kusvika kuviri yehukuru.Chikwata chinotarisira kuti basa ravo richakurudzira kumwe kutsvagisa mu3D zvinyorwa synthesis.
Iyo pretrained point Cloud yekuparadzira modhi uye yekuongorora kodhi inowanikwa paGitHub yeprojekiti.Document Point-E: Iyo sisitimu yekugadzira 3D point makore kubva kune yakaoma clues iri paarXiv.
Tinoziva kuti haudi kupotsa chero nhau kana kuwanikwa kwesainzi.Nyorera kune yedu yakakurumbira Synced Global AI Weekly newsletter kuti ugamuchire vhiki nevhiki AI updates.