Mwina mudamvapo zamphamvu zamtundu wa AI pazaka zingapo zapitazi. Koma kodi mumadziwa kuti ukadaulo womwewo ungathandize kudumpha kuchokera ku 2D kupita ku 3D?
Mitundu ya 3D yopangidwa ndi AI ili ndi vuto lalikulu pamawonekedwe amakono a digito. Masewera akanema ndipo filimu imadalira akatswiri aluso a 3D ndi mapulogalamu owonetserako monga Blender kuti apange katundu wa 3D kuti azidzaza zochitika zopangidwa ndi makompyuta.
Komabe, kodi n'zotheka kuti makampani angagwiritse ntchito makina ophunzirira kupanga zinthu za 3D popanda khama lochepa, mofanana ndi momwe akatswiri a 2D masiku ano akuyamba kutengera luso lamakono monga DALL-E ndi Ulendo wapakati?
Nkhaniyi iwunikanso ma algorithm atsopano omwe amayesa kupanga malembedwe amtundu wa 3D pogwiritsa ntchito omwe alipo zitsanzo zofalitsa.
Kodi Maloto?
Nkhani imodzi yayikulu pakupanga mtundu wophatikizika womwe umapanga zinthu za 3D mwachindunji ndikuti palibe zambiri za 3D zomwe zilipo. Mitundu ya 2D yofalitsa yakhala yamphamvu kwambiri chifukwa cha kuchuluka kwa zithunzi zomwe zimapezeka pa intaneti. Zomwezo sizinganenedwe ndi 3D assets.
Njira zina zopangira 3D zimagwira ntchito mozungulira kusowa kwa dataku pogwiritsa ntchito mwayi wochuluka wa 2D data.
Chithunzi cha DreamFusion ndi chitsanzo chopanga chomwe chingathe kupanga zitsanzo za 3D kutengera kufotokozera malemba omwe aperekedwa. Mtundu wa DreamFusion umagwiritsa ntchito fanizo lophunzitsidwa bwino losinthira mawu kupita pazithunzi kuti lipange zitsanzo zenizeni zamitundu itatu kuchokera pazambiri.
Ngakhale mulibe deta yophunzitsira ya 3D, njira iyi yapanga zida zogwirizana za 3D zowoneka bwino komanso zozama.
Kodi Ntchito?
The DreamFusion aligorivimu ili ndi mitundu iwiri ikuluikulu: 2D diffusion model ndi a neural network zomwe zimatha kusintha zithunzi za 2D kukhala mawonekedwe ogwirizana a 3D.
Google's Imagen Text-to-Image Model
Gawo loyamba la algorithm ndi mtundu wa diffusion. Chitsanzochi chili ndi udindo wotembenuza malemba kukhala zithunzi.
Imagen ndi chitsanzo chofalitsa chomwe chingathe kupanga chitsanzo chachikulu cha kusiyana kwa zithunzi za chinthu china. Pankhaniyi, kusiyanasiyana kwathu kwazithunzi kuyenera kuphimba ngodya zonse zomwe zaperekedwa. Mwachitsanzo, ngati tikufuna kupanga mtundu wa 3D wa kavalo, tingafune zithunzi za 2D za kavalo kuchokera kumakona onse otheka. Cholinga chake ndikugwiritsa ntchito Imagen kuti apereke zambiri momwe angathere (mitundu, mawonekedwe, kachulukidwe) pamtundu wotsatira mu algorithm yathu.
Kupanga Zithunzi za 3D ndi NeRF
Kenako, Dreamfusion amagwiritsa ntchito chitsanzo chotchedwa a Neural Radiance Field kapena NeRF kuti ipange mtundu wa 3D kuchokera pazithunzi zopangidwa. Ma NeRF amatha kupanga zojambula zovuta za 3D zopatsidwa deta ya zithunzi za 2D.
Tiyeni tiyese kumvetsetsa momwe NeRF imagwirira ntchito.
Chitsanzochi chikufuna kupanga mawonekedwe osalekeza a volumetric okongoletsedwa kuchokera pa data yoperekedwa ya zithunzi za 2D.
Ngati chitsanzo chimapanga ntchito, zolowetsa ndi zotuluka ndi zotani?
Ntchito yowonekera imatenga malo a 3D ndi njira yowonera 2D ngati cholowetsa. Ntchitoyi imatulutsa mtundu (mu mawonekedwe a RGB) ndi kuchuluka kwa voliyumu inayake.
Kuti mupange chithunzi cha 2D kuchokera kumawonedwe enaake, choyimiracho chipanga seti ya mfundo za 3D ndikuyendetsa nsongazo kudzera pachiwonetserochi kuti ibweze mitundu yamitundu ndi kuchuluka kwa voliyumu. Njira zoperekera voliyumu zidzasintha zikhalidwezo kukhala chithunzi cha 2D.
Kugwiritsa Ntchito NeRF ndi 2D Diffusion Models Pamodzi
Tsopano popeza tadziwa momwe NeRF imagwirira ntchito, tiyeni tiwone momwe mtundu uwu ungapangire zolondola za 3D kuchokera pazithunzi zomwe tapanga.
Pachidziwitso chilichonse choperekedwa, DreamFusion imaphunzitsa NeRF yongoyambika mwachisawawa kuyambira poyambira. Kubwereza kulikonse kumasankha malo a kamera mwachisawawa mumagulu ozungulira. Ganizirani za chitsanzo chomwe chili mu galasi lozungulira. Nthawi iliyonse tikapanga chithunzi chatsopano cha mtundu wathu wa 3D, timasankha malo mwachisawawa mugawo lathu ngati malo owonera zomwe timatulutsa. DreamFusion idzasankhanso malo owala mwachisawawa l kugwiritsa ntchito popereka.
Tikakhala ndi kamera ndi malo opepuka, mtundu wa NeRF udzaperekedwa. DreamFusion idzasankhanso mwachisawawa pakati pa kumasulira kwamitundu, kumasulira kosasinthika, ndi kumasulira kwa albedo popanda shading.
Tidanenapo kale kuti tikufuna kuti chithunzi chathu chojambula zithunzi (Imagen) chipange zithunzi zokwanira kuti tipange chitsanzo choyimira.
Kodi Dreamfusion amakwaniritsa bwanji izi?
Dreamfusion amangosintha kulowetsako pang'ono kuti akwaniritse zomwe akufuna. Mwachitsanzo, titha kukhala ndi ma angles okwera kwambiri powonjezera "mawonekedwe apamwamba" kumayendedwe athu. Titha kupanga ma angles ena powonjezera mawu monga "kutsogolo", "mawonedwe am'mbali", ndi "mawonedwe akumbuyo".
Zithunzi zimaperekedwa mobwerezabwereza kuchokera pamakamera osasintha. Kumasulira uku kumadutsa mu ntchito ya distillation yotayika. Njira yosavuta yotsika pang'onopang'ono idzawongolera pang'onopang'ono Mtundu wa 3D mpaka igwirizane ndi zochitika zofotokozedwa ndi lembalo.
Titapanga mtundu wa 3D pogwiritsa ntchito NeRF, titha kugwiritsa ntchito Marching Cubes algorithm kutulutsa mauna a 3D amtundu wathu. Maunawa amatha kutumizidwa kuzinthu zodziwika bwino za 3D kapena pulogalamu yamachitsanzo.
sitingathe
Ngakhale kutulutsa kwa DreamFusion kumakhala kochititsa chidwi kwambiri chifukwa imagwiritsa ntchito mitundu yomwe ilipo yosinthira zithunzi ndi zithunzi m'njira yatsopano, ofufuzawo awona zofooka zingapo.
Ntchito yotayika ya SDS yawonedwa kuti ipange zotsatira zochulukirapo komanso zosalala. Mutha kuwona izi mumitundu yosakhala yachilengedwe komanso kusowa kwatsatanetsatane komwe kumapezeka pazotulutsa.
The DreamFusion algorithm ilinso ndi malire chifukwa cha kutulutsa kwa mtundu wa Imagen, womwe ndi mapikiselo 64 x 64. Izi zimapangitsa kuti mitundu yopangidwayo ikhale yopanda tsatanetsatane.
Pomaliza, ofufuzawo adawona kuti pali vuto lomwe limakhalapo popanga mitundu ya 3D kuchokera ku data ya 2D. Pali zitsanzo zambiri za 3D zomwe titha kupanga kuchokera pazithunzi za 2D, zomwe zimapangitsa kukhathamiritsa kukhala kovuta komanso kosamveka bwino.
Kutsiliza
Zomasulira za DreamFusion's 3D zimagwira ntchito bwino kwambiri chifukwa cha kuthekera kwamitundu yosiyanasiyana yazithunzi kuti ipange chinthu chilichonse kapena mawonekedwe. Ndizodabwitsa momwe neural network ingamvetsetsere zochitika mu 3D space popanda chidziwitso cha 3D. Ndikupangira kuwerenga pepala lonse kuti mudziwe zambiri zaukadaulo wa DreamFusion algorithm.
Tikukhulupirira, ukadaulo uwu ukhala bwino kuti pamapeto pake upange zithunzi zowoneka bwino za 3D. Ingoganizirani masewera onse apakanema kapena zoyerekeza zomwe zimagwiritsa ntchito malo opangidwa ndi AI. Itha kutsitsa chotchinga cholowera kwa opanga masewera apakanema kuti apange maiko ozama a 3D!
Kodi mukuganiza kuti mameseji kupita ku 3D adzakhala ndi gawo lanji mtsogolomu?
Siyani Mumakonda