M'ndandanda wazopezekamo[Bisani][Show]
Mwinamwake mukudziwa kuti kompyuta ikhoza kufotokoza chithunzi.
Mwachitsanzo, chithunzi cha galu akusewera ndi ana anu chingamasuliridwe kuti 'galu ndi ana m'munda.' Koma kodi mumadziwa kuti njira yosiyanayo ndi yothekanso? Mumalemba mawu, ndipo makinawo amapanga chithunzi chatsopano.
Mosiyana ndi kusaka kwa Google, komwe kumafufuza zithunzi zomwe zilipo kale, zonsezi ndizatsopano. M'zaka zaposachedwa, OpenAI yakhala imodzi mwamabungwe otsogola, ikuwonetsa zotsatira zabwino.
Amaphunzitsa ma algorithms awo pazolemba zazikulu ndi zithunzi. Adasindikiza pepala pazithunzi zawo za GLIDE, zomwe zidaphunzitsidwa pazithunzi mamiliyoni mazana. Pankhani ya photorealism, imaposa chitsanzo chawo choyambirira cha 'DALL-E'.
Mu positi iyi, tiwona OpenAI's GLIDE, imodzi mwazinthu zochititsa chidwi zomwe cholinga chake ndi kupanga ndikusintha zithunzi zazithunzi ndi mitundu yosinthira mawu. Tiyeni tiyambe.
Kodi Tsegulani AI Glide?
Ngakhale zithunzi zambiri zimatha kufotokozedwa m'mawu, kupanga zithunzi kuchokera pamawu amafunikira chidziwitso chapadera komanso nthawi yayitali.
Kulola wothandizira wa AI kupanga zithunzi zowoneka bwino kuchokera ku zilankhulo zachilengedwe sikumangolola anthu kupanga zowoneka bwino komanso zosiyanasiyana mosavuta komanso kumathandizira kuwongolera kosavuta komanso kuwongolera bwino kwa zithunzi zomwe zidapangidwa.
GLIDE itha kugwiritsidwa ntchito kusintha zithunzi zomwe zilipo kale pogwiritsa ntchito chilankhulo chachilengedwe kuti muyike zinthu zatsopano, kupanga mithunzi ndi zowunikira, kuchita. kujambula zithunzi, ndi zina zotero.
Itha kusinthiranso zojambula zoyambira kukhala zithunzi zowoneka bwino, ndipo ili ndi zida zapadera zopanga ndi kukonza ziro pakanthawi zovuta.
Kafukufuku waposachedwa wasonyeza kuti mitundu yotengera njira yosinthira imatha kupanganso zithunzi zapamwamba kwambiri, makamaka zikaphatikizidwa ndi njira yolondolera yomwe imayang'anira kusiyanasiyana ndi kukhulupirika.
OpenAI idasindikiza a motsogolere kufalitsa chitsanzo m'mwezi wa Meyi, zomwe zimalola kuti mitundu yosiyanasiyana ikhale yogwirizana ndi zilembo za gulu. GLIDE imachita bwino pakuchita bwino kumeneku pobweretsa kufalikira motsogozedwa ndi vuto lopanga zithunzi zovomerezeka.
Pambuyo pophunzitsa mtundu wa 3.5 biliyoni wophatikizika wa GLIDE pogwiritsa ntchito encoder yamawu kuti agwirizane ndi mafotokozedwe a chilankhulo chachilengedwe, ofufuzawo adayesa njira ziwiri zowongolera: chitsogozo cha CLIP ndi chitsogozo chopanda magulu.
CLIP ndi njira yowonjezereka yophunzirira zoyimira zolumikizana za mawu ndi zithunzi zomwe zimapereka chigoli kutengera momwe chithunzi chiliri pafupi ndi mawu ofotokozera.
Gululi linagwiritsa ntchito njirayi m'mawonekedwe awo ophatikizika polowetsa gululo ndi CLIP yachitsanzo yomwe "imatsogolera" zitsanzo. Pakadali pano, chitsogozo chopanda magulu ndi njira yolondolera mitundu yogawa yomwe simaphatikizapo maphunziro a gulu lapadera.
GLIDE Architecture
Zomangamanga za GLIDE zimakhala ndi zigawo zitatu: Ablated Diffusion Model (ADM) yophunzitsidwa kuti ipange chithunzi cha 64 × 64, mtundu wamawu (transformer) womwe umakhudza kutulutsa zithunzi kudzera pamawu achangu, ndi chitsanzo chowonjezera chomwe chimasintha 64 × 64 yathu yaying'ono. Zithunzi zomveka bwino za 256 x 256 pixels.
Zigawo ziwiri zoyambirira zimagwirira ntchito limodzi kuti ziwongolere njira yopangira zithunzi kuti ziwonetsere moyenerera mawuwo, pomwe chomalizacho chimafunikira kuti zithunzi zomwe timapanga zikhale zosavuta kuzimvetsetsa. Ntchito ya GLIDE idalimbikitsidwa ndi a lipoti lofalitsidwa mu 2021 zomwe zinawonetsa kuti njira za ADM zapambana kwambiri pakali pano zodziwika bwino, zamakono zopanga zopanga malinga ndi khalidwe lachitsanzo chazithunzi.
Kwa ADM, olemba GLIDE adagwiritsa ntchito chitsanzo chomwecho cha ImageNet 64 x 64 monga Dhariwal ndi Nichol, koma ndi mayendedwe a 512 m'malo mwa 64. Chitsanzo cha ImageNet chili ndi magawo pafupifupi 2.3 biliyoni chifukwa cha izi.
Gulu la GLIDE, mosiyana ndi Dhariwal ndi Nichol, linkafuna kukhala ndi ulamuliro wolunjika pa ndondomeko yopangira zithunzi, motero anaphatikiza chitsanzo chowoneka ndi chosinthira chothandizira chidwi. GLIDE imakupatsani chiwongolero panjira yopangira chithunzi pokonza zomwe mawu akulowetsa.
Izi zimatheka pophunzitsa mtundu wa thiransifoma pagulu lalikulu loyenera la zithunzi ndi mawu ofotokozera (zofanana ndi zomwe zidagwiritsidwa ntchito mu DALL-E).
Zolembazo poyamba zimayikidwa mu mndandanda wa zizindikiro za K kuti zikhazikike. Pambuyo pake, zizindikirozo zimayikidwa mumtundu wa transformer. Kutulutsa kwa transformer kumatha kugwiritsidwa ntchito m'njira ziwiri. Pachitsanzo cha ADM, kuyika komaliza kumagwiritsidwa ntchito m'malo moyika kalasi.
Chachiwiri, gawo lomaliza la zoyika zizindikiro - mndandanda wa ma vectors - amawonetsedwa mopanda miyeso ya gawo lililonse lachitsanzo cha ADM ndikulumikizana ndi chidwi chilichonse.
Zoona zake, izi zimathandiza kuti chitsanzo cha ADM chipange chithunzi kuchokera kumagulu atsopano a zizindikiro zofananira zofananira m'njira yapadera komanso yojambula zithunzi, pogwiritsa ntchito kumvetsetsa kwake kwa mawu olowetsamo ndi zithunzi zawo zogwirizana. Transformer-encoding iyi ili ndi magawo 1.2 biliyoni ndipo imagwiritsa ntchito midadada 24 yotsalira ndi 2048 m'lifupi.
Pomaliza, mtundu wa upsampler wophatikizika umaphatikizapo magawo ozungulira 1.5 biliyoni ndipo umasiyana ndi mtundu woyambira chifukwa mawu ake osindikiza ndi ocheperako, okhala ndi makulidwe a 1024 ndi 384 base, poyerekeza ndi mtundu woyambira. Chitsanzochi, monga momwe dzinalo likusonyezera, chimathandizira kukweza kwachitsanzo kuti chimveke bwino pamakina ndi anthu.
Chitsanzo cha kufalikira
GLIDE imapanga zithunzi pogwiritsa ntchito mtundu wake wa ADM (ADM-G wa "wotsogolera"). Mtundu wa ADM-G ndikusintha kwa mtundu wa U-net wofalikira. Mtundu wophatikizika wa U-net umasiyana kwambiri ndi njira zodziwika bwino zophatikizira zithunzi monga VAE, GAN, ndi zosinthira.
Amamanga unyolo wa Markov wa masitepe ophatikizika kuti pang'onopang'ono amabaya phokoso lachisawawa mu data, kenako amaphunzira kubweza njira yolumikizirana ndikumanganso zitsanzo za data zomwe zimafunikira kuchokera kuphokoso lokha. Zimagwira ntchito m'magawo awiri: kutsogolo ndi kubwerera kumbuyo.
Njira yoyatsira patsogolo, yopatsidwa malo a data kuchokera pakugawa kwenikweni kwachitsanzo, imawonjezera phokoso laling'ono pachitsanzo pamasitepe okonzedweratu. Pamene masitepe akuwonjezeka kukula ndi kuyandikira mopanda malire, chitsanzocho chimataya makhalidwe onse odziwika ndipo ndondomekoyi imayamba kufanana ndi piritsi la isotropic Gaussian.
Pa m'mbuyo kufalikira gawo, chitsanzo cha kufalikira amaphunzira kubweza chikoka cha phokoso anawonjezera pa zithunzi ndi kutsogolera fano opangidwa kubwerera mawonekedwe ake oyambirira ndi kuyesa kufanana choyambirira athandizira chitsanzo kugawa.
Mtundu womalizidwa ukhoza kutero ndi phokoso lenileni la Gaussian komanso mwachangu. Njira ya ADM-G imasiyanasiyana ndi yoyambayo chifukwa chakuti chitsanzo, CLIP kapena chosinthira makonda, chimakhudza gawo lakumbuyo lakumbuyo pogwiritsa ntchito ma tokeni ofulumira omwe amalowetsedwa.
Glide luso
1. Mbadwo wa Chithunzi
Kugwiritsa ntchito GLIDE kodziwika kwambiri komanso kogwiritsidwa ntchito kwambiri mwina kungakhale kaphatikizidwe kazithunzi. Ngakhale zithunzizo ndi zochepetsetsa ndipo GLIDE imavutika ndi mitundu ya nyama/anthu, kuthekera kopanga chithunzi chimodzi kumakhala kosatha.
Itha kupanga zithunzi za nyama, anthu otchuka, malo, nyumba, ndi zina zambiri, ndipo imatha kuzichita m'njira zosiyanasiyana zamaluso komanso zithunzi zenizeni. Olemba ofufuzawo akuti GLIDE imatha kumasulira ndikusintha zolemba zambiri kuti zikhale zowoneka bwino, monga zikuwonekera m'zitsanzo zomwe zili pansipa.
2. Glide inpeinting
Kujambula kwazithunzi kwa GLIDE mosakayikira ndiko kugwiritsa ntchito kosangalatsa kwambiri. GLIDE imatha kutenga chithunzi chomwe chilipo ngati cholowetsa, ndikuchikonza ndi mawu achangu m'malingaliro a malo omwe akufunika kusinthidwa, kenako kusintha magawowo mosavuta.
Iyenera kugwiritsidwa ntchito limodzi ndi njira yosinthira, monga SDEdit, kuti ipange zotsatira zabwinoko. M'tsogolomu, mapulogalamu omwe amapezerapo mwayi pazifukwa ngati izi zitha kukhala zofunika kwambiri popanga njira zosinthira zithunzi popanda ma code.
Kutsiliza
Tsopano popeza tadutsamo, muyenera kumvetsetsa zoyambira momwe GLIDE imagwirira ntchito, komanso kukula kwa kuthekera kwake pakupanga zithunzi ndikusintha pazithunzi.
Siyani Mumakonda