Iwe unogona kunge uchiziva kuti komputa inogona kutsanangura mufananidzo.
Somuenzaniso, mufananidzo wembwa iri kutamba nevana vako unogona kushandurwa kuti 'imbwa nevana vari mubindu.' Asi waiziva here kuti nzira yakapesana yave kukwanisika zvakare? Iwe unonyora mamwe mazwi, uye muchina unogadzira mufananidzo mutsva.
Kusiyana nekutsvaga kweGoogle, iyo inotsvaga mafoto aripo, izvi zvese zvitsva. Mumakore achangopfuura, OpenAI yanga iri imwe yemasangano anotungamira, ichishuma zvinokatyamadza mhedzisiro.
Ivo vanodzidzisa algorithms yavo pane yakakura zvinyorwa uye mifananidzo dhatabhesi. Vakaburitsa bepa pane yavo GLIDE mufananidzo modhi, iyo yakadzidziswa pamazana emamiriyoni emifananidzo. Panyaya yephotorealism, inodarika yavo yekutanga 'DALL-E' modhi.
Mune ino positi, isu tichatarisa kuOpenAI's GLIDE, imwe yeakati wandei anonakidza matanho ane chinangwa chekugadzira uye kushandura mafotorealistic mafoto ane mavara-anotungamirwa edhisheni modhi. Ngatitange.
Chii Vhura AI Glide?
Nepo mifananidzo mizhinji ichigona kutsanangurwa mumashoko, kugadzira mifananidzo kubva kune zvinyorwa kunoda ruzivo rwakanyanya uye nguva yakakosha.
Kubvumira mumiriri weAI kuburitsa mafotorealistic mapikicha kubva mumutauro wechisikigo zvinokurudzira hazvingobvumiri vanhu kugadzira akapfuma uye akasiyana ekuona zvinhu zviri nyore zvisati zvamboitika asi zvakare zvinobvumira kukwenenzverwa kudzokororwa uye kudzora kwakaringana kwemifananidzo yakagadzirwa.
GLIDE inogona kushandiswa kugadzirisa mafoto aripo nekushandisa mameseji emutauro wechisikigo kuisa zvinhu zvitsva, kugadzira mimvuri uye maratidziro, kuita. mufananidzo inpainting, zvichingoenda zvakadaro.
Inogona zvakare kushandura mitsara yedhirowa kuita mafotorealistic mafoto, uye ine yakasarudzika zero-sample yekugadzira uye kugadzirisa kugona kwemamiriro akaoma.
Tsvagiridzo ichangoburwa yakaratidza kuti mamodheru-akavakirwa pakupararira anogonawo kugadzira mifananidzo yemhando yepamusoro, kunyanya kana yakasanganiswa negwara rinoyera siyana uye kuvimbika.
OpenAI yakaburitsa a akatungamirirwa diffusion modhi muna Chivabvu, izvo zvinobvumira mamodheru ekuparadzira kuve ane zvirevo pamavara emugadziri. GLIDE inonatsiridza pabudiriro iyi nekuunza inotungamirwa kupararira kune dambudziko remavara-ane mamiriro ekugadzira mifananidzo.
Mushure mekudzidzisa 3.5 bhiriyoni parameter GLIDE diffusion modhi vachishandisa encoder yemavara kuti ienderane nerondedzero yemutauro wechisikigo, vaongorori vakaedza nzira mbiri dzekutungamira: kutungamira kweCLIP uye kutungamira-yemahara kutungamira.
CLIP inzira inokura yekudzidza mamiririro akabatanidzwa ezvinyorwa uye mapikicha anopa mamakisi zvichienderana nekuti mufananidzo uri padyo sei nemusoro.
Chikwata chakashandisa zano iri mumamodheru avo ekuparadzira nekutsiva mugadziri neCLIP modhi "inotungamira" mamodheru. Zvichakadaro, kutungamira-yemahara kutungamira izano rekutungamira modhi yekuparadzira isingasanganisire kudzidziswa kweakaparadzana classifier.
GLIDE Architecture
Iyo GLIDE dhizaini ine zvikamu zvitatu: Ablated Diffusion Model (ADM) yakadzidziswa kugadzira 64 × 64 mufananidzo, chinyorwa chinyorwa (transformer) chinopesvedzera kugadzirwa kwemifananidzo kuburikidza nekuchimbidza kwemavara, uye upsampling modhi inoshandura diki yedu 64 × 64. mifananidzo kune zvimwe kududzira 256 x 256 pixels.
Zvikamu zviviri zvekutanga zvinoshanda pamwe chete kudzora magadzirirwo emifananidzo kuitira kuti iratidze nenzira kwayo kukurumidza kwechinyorwa, nepo chekupedzisira ichidikanwa kuita kuti mifananidzo yatinogadzira zvive nyore kunzwisisa. Iyo GLIDE chirongwa chakafemerwa na chirevo chakaburitswa muna 2021 iyo yakaratidza kuti ma ADM maitiro akadarika iye zvino akakurumbira, mamiriro-e-the-art generative modhi maererano nemufananidzo wemhando yemhando.
KuADM, vanyori veGLIDE vakashandisa mufananidzo wakafanana weImageNet 64 x 64 seDhariwal naNichol, asi nematanho 512 panzvimbo ye64. Mufananidzo weImageNet une mabhiriyoni anosvika 2.3 semugumisiro weizvi.
Chikwata cheGLIDE, kusiyana naDhariwal naNichol, chaida kuve nekutonga kwakananga pamusoro pekugadzira mapikicha, nekudaro vakabatanidza iyo yekuona modhi neyekutarisisa-inogonesa transformer. GLIDE inokupa iwe kutonga pamusoro peiyo pikicha inogadzira process inobuda nekugadzirisa zvinyorwa zvekupinza.
Izvi zvinopedzwa nekudzidzisa modhi yeshanduko pane yakakura zvakakodzera dataset yemifananidzo uye zvinyorwa (zvakafanana neakashandiswa muchirongwa cheDALL-E).
Iwo mameseji anotanga encoded kuita nhevedzano yeK tokens kuitira kuti agadzirise. Mushure meizvozvo, zviratidzo zvinotakurwa mushanduro yemuenzaniso. Kubuda kwetransformer kunogona kushandiswa nenzira mbiri. Kune iyo ADM modhi, yekupedzisira tokeni yekumisikidza inoshandiswa pachinzvimbo chekumisikidza kirasi.
Chechipiri, chiratidzo chekumisikidza 'yekupedzisira layer - nhevedzano yemaficha mavheji - inofungidzirwa yakazvimiririra kune zviyero zvega rega rekutarisa mumuenzaniso weADM uye yakabatana kune yega yega yekutarisisa.
Muchokwadi, izvi zvinogonesa iyo ADM modhi kuti ibudise mufananidzo kubva mumisanganiswa mitsva yezvinyorwa zvakafanana zvemavara mune yakasarudzika uye photorealistic fashoni, zvichibva pakunzwisisa kwayo kwakadzidzwa kwemazwi ekuisa uye mifananidzo yavo yakabatana. Ichi chinyorwa-encoding transformer ine 1.2 bhiriyoni paramita uye inoshandisa makumi maviri nemana akasara mabhuroko ane hupamhi hwe24.
Chekupedzisira, iyo upsampler diffusion modhi inosanganisira yakatenderedza bhiriyoni 1.5 paramita uye inosiyana kubva kune yekutanga modhi pakuti zvinyorwa zvayo encoder idiki, nehupamhi hwe1024 uye 384 base chiteshi, zvichienzaniswa neiyo base modhi. Iyi modhi, sezita rinoratidza, inobatsira mukukwidziridzwa kwemuenzaniso kuitira kuvandudza kududzira kumichina uye nevanhu.
Diffusion model
GLIDE inogadzira mifananidzo ichishandisa vhezheni yayo yeADM (ADM-G ye"kutungamirwa"). Iyo ADM-G modhi igadziriso yekupararira U-net modhi. Iyo yekuparadzanisa U-net modhi inosiyana zvakanyanya kubva kune yakajairika nzira yekugadzira mifananidzo yakadai seVAE, GAN, uye transformers.
Ivo vanovaka cheni yeMarkov yematanho ekuparadzira zvishoma nezvishoma kupinza ruzha rusina kurongeka mu data, uye vobva vadzidza kudzoreredza maitiro ekuparadzira uye kuvakazve inodiwa data samples kubva kune ruzha chete. Inoshanda mumatanho maviri: mberi uye reverse diffusion.
Iyo yekumberi yekuparadzira nzira, yakapihwa nzvimbo yedata kubva mukugovera kwechokwadi kwemuenzaniso, inowedzera hushoma hweruzha kumuenzaniso pamusoro peakatemerwa nhanho. Sezvo nhanho dzichiwedzera saizi uye kuswedera kusingagumi, sampuli inorasikirwa nehunhu hwese hunozivikanwa uye kutevedzana kunotanga kuita kunge isotropic Gaussian curve.
Munguva yekudzokera shure chikamu, muenzaniso wekupararira inodzidza kudzoreredza pesvedzero yeruzha rwakawedzerwa pamifananidzo uye kutungamira chifananidzo chakagadzirwa kudzoka kune chimiro chayo chepakutanga nekuyedza kufanana neyekutanga yekugovera sampuli yekugovera.
Muenzaniso wakapedzwa unogona kuzviita neicho chaiyo yeGaussian ruzha rwekuisa uye nekukasira. Iyo nzira yeADM-G inosiyana kubva kune yakapfuura pakuti modhi, ingave CLIP kana yakagadziridzwa shanduri, inokanganisa chikamu chekudzokera kumashure nekushandisa mavara ekukurumidza tokens anoiswa.
Glide kugona
1. Chizvarwa cheMufananidzo
Iyo inonyanya kufarirwa uye inoshandiswa zvakanyanya kushandiswa kweGLIDE ingangove iri mufananidzo synthesis. Kunyangwe iyo mifananidzo ine mwero uye GLIDE ine kuomerwa nemhuka/vanhu mafomu, mukana weimwe-pfuti yekugadzirwa kwemufananidzo hauperi.
Inogona kugadzira mapikicha emhuka, vane mukurumbira, nzvimbo, zvivakwa, uye zvimwe zvakawanda, uye inogona kuzviita mumhando dzakasiyana siyana dzehunyanzvi pamwe nemifananidzo-chaizvo. Vanyori vevaongorori vanoti GLIDE inokwanisa kuturikira nekugadzirisa zvakasiyana-siyana zvemanyorero kuita chimiro chekuona, sezvinoonekwa mumasampuli ari pazasi.
2. Glide inpainting
GLIDE's otomatiki pikicha inpeinting ndiyo ine nharo yekushandisa inonyanya kunakidza. GLIDE inogona kutora pikicha iripo sekuisa, igadzirise nerunyoro rwekumisikidza kune nzvimbo dzinoda kuchinjwa, uye wobva waita shanduko inoshanda kune izvo zvikamu zviri nyore.
Inofanirwa kushandiswa pamwe chete neyekugadzirisa modhi, senge SDEdit, kuburitsa zvirinani mhedzisiro. Mune ramangwana, mapurogiramu anotora mukana wekuita seaya anogona kuve akakosha mukugadzira macode-emahara mapikicha ekuchinja nzira.
mhedziso
Zvino zvatapfuura nekuita, iwe unofanirwa kubata izvo zvakakosha zvemashandiro anoita GLIDE, pamwe nehupamhi hwekugona kwayo mukugadzira mifananidzo uye mu-image kugadziridzwa.
Leave a Reply