• hae

OpenAI Point E: E hana i kahi ao kiko 3D mai nā nalu paʻakikī i nā minuke ma kahi GPU hoʻokahi

Ma kahi ʻatikala hou Point-E: He ʻōnaehana no ka hana ʻana i nā ao kiko 3D mai nā hōʻailona paʻakikī, hoʻolauna ka hui noiʻi OpenAI i ka Point E, kahi ʻōnaehana synthesis conditional 3D point cloud text conditional synthesis e hoʻohana ana i nā hiʻohiʻona diffusion e hana i nā ʻano like ʻole a paʻakikī hoʻi i alakaʻi ʻia e nā kikokikona paʻakikī. nā hōʻailona.i nā minuke ma kahi GPU hoʻokahi.
ʻO ka hana kupanaha o nā hiʻohiʻona kiʻi hou o kēia lā ua hoʻoulu i ka noiʻi ʻana i ka hana ʻana i nā mea kikokikona 3D.Eia naʻe, ʻaʻole e like me nā hiʻohiʻona 2D, hiki ke hoʻopuka i nā mea i loko o nā minuke a i ʻole kekona, koi maʻamau nā kumu hoʻohālike mea hoʻohālike i kekahi mau hola o ka hana GPU e hana i hoʻokahi laʻana.
Ma kahi ʻatikala hou Point-E: He ʻōnaehana no ka hoʻokumu ʻana i nā ao kiko 3D mai nā hōʻailona paʻakikī, hōʻike ka hui noiʻi OpenAI i Point·E, kahi ʻōnaehana hoʻonohonoho kūlana kikokikona no nā ao kiko 3D.Ke hoʻohana nei kēia ala hou i kahi hoʻohālike hoʻolaha e hana i nā ʻano 3D like ʻole a paʻakikī mai nā hōʻailona kikokikona paʻakikī i hoʻokahi minuke a ʻelua paha ma ka GPU hoʻokahi.
Hoʻokumu ka hui i ka paʻakikī o ka hoʻololi ʻana i nā kikokikona i 3D, he mea koʻikoʻi ia i ka democratizing 3D content haku no nā noi honua maoli mai ka ʻoiaʻiʻo maoli a me ka pāʻani ʻana i ka hoʻolālā ʻenehana.ʻO nā ʻano hana i kēia manawa no ka hoʻololi ʻana i ke kikokikona i 3D e hāʻule i loko o ʻelua mau ʻāpana, aia i kēlā me kēia me kāna mau hemahema: 1) hiki ke hoʻohana ʻia nā hiʻohiʻona generative no ka hana ʻana i nā laʻana me ka maikaʻi, akā ʻaʻole hiki ke hoʻohālikelike pono i nā hōʻailona kikokikona like ʻole a paʻakikī;2) he kŘkohu kiʻi kikokikona i hoʻomaʻamaʻa mua ʻia no ka mālama ʻana i nā kiʻi kikokikona paʻakikī a ʻano like ʻole, akā ʻoi aku ka ikaika o kēia ala a hiki ke hoʻopaʻa maʻalahi ke kumu hoʻohālike i ka minima kūloko i kūpono ʻole i nā mea 3D koʻikoʻi.
No laila, ua ʻimi ka hui i kahi ala ʻē aʻe e manaʻo nei e hoʻohui i nā ikaika o nā ala ʻelua ma luna, me ka hoʻohana ʻana i ke ʻano hoʻohālikelike kikokiko-i-kiʻi i hoʻomaʻamaʻa ʻia ma kahi pūʻulu nui o nā kiʻi kikokikona (e ʻae iā ia e mālama i nā hōʻailona like ʻole a paʻakikī) a he kumu hoʻohālike kiʻi 3D i hoʻomaʻamaʻa ʻia ma kahi pūʻulu liʻiliʻi o nā paʻa kiʻi kikokikona.kiʻi-3D hui pūʻulu waihona.Hoʻohālike mua ke kŘkohu kikokikona i ke kiʻi i ke kiʻi hoʻokomo no ka hana ʻana i hoʻokahi hōʻike synthetic, a hana ke kumu hoʻohālike kiʻi-i-3D i kahi ao kiko 3D ma muli o ke kiʻi i koho ʻia.
Hoʻokumu ʻia ka waihona generative o ke kauoha ma luna o nā ʻōnaehana generative i manaʻo ʻia no ka hoʻokumu ʻana i nā kiʻi mai ka kikokikona (Sohl-Dickstein et al., 2015; Song & Ermon, 2020b; Ho et al., 2020).Hoʻohana lākou i kahi hiʻohiʻona GLIDE me 3 billion GLIDE parameter (Nichol et al., 2021), i hoʻopaʻa maikaʻi ʻia i nā hiʻohiʻona 3D i hāʻawi ʻia, e like me kā lākou ʻano hoʻololi kikokikona-ki-kiʻi, a me kahi pūʻulu o nā hiʻohiʻona diffusion e hana i nā ao kiko RGB e like me kā lākou. hoʻohālike hoʻololi.nā kiʻi i ke kiʻi.Nā hiʻohiʻona 3D.
ʻOiai ua hoʻohana ka hana ma mua i ka hoʻolālā 3D e hoʻoponopono i nā ao kiko, ua hoʻohana nā mea noiʻi i kahi kumu hoʻohālike transducer maʻalahi (Vaswani et al., 2017) e hoʻomaikaʻi i ka pono.I loko o kā lākou hoʻolālā hoʻohālike hoʻopulapula, hānai mua ʻia nā kiʻi o ke ao i loko o kahi kumu hoʻohālike ViT-L/14 CLIP i hoʻomaʻamaʻa mua ʻia a laila hānai ʻia nā meshes i loko o ka mea hoʻololi ma ke ʻano he māka.
Ma kā lākou noiʻi empirical, ua hoʻohālikelike ka hui i ke ʻano Point·E i manaʻo ʻia me nā hiʻohiʻona 3D generative ʻē aʻe e pili ana i ka helu ʻana i nā hōʻailona mai ka ʻike ʻana o ka mea COCO, ka ʻāpana, a me nā ʻikepili pūlima.Hōʻoia nā hopena e hiki iā Point·E ke hana i nā ʻano 3D like ʻole a paʻakikī hoʻi mai nā hōʻailona kikokikona paʻakikī a wikiwiki i ka manawa inference e hoʻokahi a ʻelua mau kauoha o ka nui.Manaʻo ka hui i kā lākou hana e hoʻoikaika i ka noiʻi hou ʻana i ka synthesis kikokikona 3D.
Loaʻa kahi kumu hoʻohālike hoʻolaha ʻana o ke ao a me ka helu loiloi ma ka GitHub o ka papahana.Lae Palapala-E: Aia ma arXiv kahi ʻōnaehana no ka hana ʻana i nā ao kiko 3D mai nā hōʻailona paʻakikī.
ʻIke mākou ʻaʻole makemake ʻoe e poina i kekahi nūhou a i ʻole ʻike ʻepekema.E hoʻopaʻa inoa i kā mākou nūhou Synced Global AI Weekly e loaʻa ai nā mea hou AI i kēlā me kēia pule.


Ka manawa hoʻouna: Dec-28-2022