Jiilka Sawirka Gadaashiisa
Jiilka sawirka Autoregressive wuxuu dhisaa sawirro hal mar, isagoo saadaaliyay calaamad kasta wax kasta oo la soo saaray ka hor.
Dulmar
It matters because the same next-token machinery powering language models can produce coherent, controllable images.
quusid qoto dheer
Jiilka sawirka autoregressive wuxuu sawirka ula dhaqmaa sida isku xigxiga oo u saadaaliya curiye ahaan, halkaas oo shay kasta oo cusubi uu shuruud ku xidho dhammaan kuwii hore. Shaqadii hore sida PixelRNN iyo PixelCNN waxay saadaaliyeen sawirada hal pixel ceeriin markiiba, oo sawiraya saf saf, kaas oo ahaa mid gaabis ah laakiin aragti ahaan nadiif ah. Nidaamyada casriga ah waxay marka hore ku cadaadinayaan sawirka calaamado kala duwan iyagoo isticmaalaya codeer qaabka VQ-VAE, ka dib Transformer-ku wuxuu saadaaliyaa calaamadahaas bidix-ilaa-midig. OpenAI's DALL-E 1 iyo Google's Parti ayaa raacay habkan, soo saarista calaamado muuqaal ah oo shuruudaysan isla markiiba qoraalka ka hor inta aan dib loogu celin pixels. Faa'iidada ugu weyni waa qaabaynta saxda ah iyo qaab dhismeedka midaysan ee lala wadaago luqadda. Kharashku waa taxane, muunad gaabis ah.
Aragtida Farsamada
Qaabku waxa uu soo saara itimaalka wadajirka ah ee dhammaan calaamaduhu una noqdaan badeecad shuruudo leh: p(x) = sheyga p(x_i la siiyay x_1...x_{i-1}). Transformer-ka leh dareenka sababa (masjirada) ayaa xoojinaya in boos kasta uu arko kaliya calaamadihii hore. Inta lagu jiro tababarka waxa ay saadaalisaa calaamad kasta oo barbar socota iyada oo la adeegsanayo qasab macalinka, laakiin marka la eego waa in ay hal mar muunad ku muujiso, mid walbana dib ugu soo celiso. Buug-codebook oo la bartay ayaa dib ugu celinaya balastar sawireed, kaas oo decoder-ku soo koobo pixels final.
Saamaynta Istiraatijiyadeed
Xawaaraha iyo miisaanka
Visual AI wuxuu si otomaatig ah u samayn karaa baadhista, ogaanshaha, iyo sumadaynta hawlaha miisaanka.
Xulashada dhismayaasha
Kooxaha hal-abuurka leh waxay hindise karaan fikradaha si dhakhso leh iyagoo leh dib-u-eegis buugeed yar.
Kooxda iyo socodka shaqada
Hawlgalladu waxay isticmaali karaan calaamadaha muuqaalka iyo muuqaalka kuwaas oo markii hore adkeyd in la farsameeyo.
Mustaqbalka Jiilka Sawirka Autoregressive
Xawaaruhu waa goobta dhexe ee dagaalka. Farsamooyinka sida isbarbardhigga iyo calaamadaynta waji-xidhka (MaskGIT, Muse) waxay abuuraan calaamado badan hal mar, iyo dejinta malo-awaalka ah ee laga soo amaahday moodooyinka luqadda ayaa la waafajiyay sawirrada. Cilmi-baarayaashu waxay sidoo kale mideynayaan qoraalka iyo calaamadaha sawirka ee hal laf dhabarta ah si uu hal nooc u akhriyo oo u sawiro, sida lagu arkay hababka multimodal. Filo fikradaha is-daba-joogga ah iyo faafinta inay sii wadaan isku-dhafka, moodooyinka isku-dhafan ee qaadaya xakamaynta calaamadaha iyo tayada fidinta.
Dhaqangelinta Adduunka-dhabta ah
DALL-E 1 waxay soo saartay sawiro iyadoo si toos ah u saadaalinaysa shabagga calaamooyinka sawirka gaarka ah ee qoraalka qoraalka ah.
Google's Parti ayaa cabiray isbedbedel qoraal-u-sawir ah oo toos ah ilaa 20 bilyan oo cabbir si faahfaahsan, muuqaalo daacad ah oo degdeg ah.
PixelCNN iyo PixelRNN waxay soo bandhigeen jiilka pixel-by-pixel cayriin waxaana wali loo isticmaalaa sidii wax lagu bari lahaa moodooyinka ku saleysan suurtagalnimada.
MaskGIT iyo Muse waxay adeegsadaan calaamado isbarbar socda oo waji-xidhan si ay u dedejiyaan isku-xidhka sawirka calaamooyinka ku salaysan iyadoo la ilaalinayo tababbarka qaab-is-daba-marineed.
Khatarta & Dariiqyada Ilaalada
Xuquuqda sawirka iyo ogolaanshaha waxay noqon kartaa khataro sharci ah haddii caddayntu aanay caddayn.
Waxqabadka moodeelku wuu ku kala duwanaan karaa iftiinka, tirakoobka, iyo deegaanka.
Wanaagga beenta ah waxa laga yaabaa inaan la dareemin ilaa xadka kalsoonida aan la kormeerin.
Qorshe Hawleedka Dhaqangelinta
Qeex shuruudaha aqbalida ee saxnaanta, dib u celinta, iyo kharashyada khaladka.
Ku tijaabi xogta ku habboon xaaladaha wax soo saarka dhabta ah.
Ku dar dib u eegis bini'aadamka si aad u hesho kalsoonida hoose ama saameeynta sare.
Lasoco moodeel dhaqaaqa oo dib u cusboonaysii kamarada ama xogta kaydinta ka dib.
Sii wad Sahaminta
Free newsletter
Get the daily AI briefing
Three verified AI stories every weekday morning, written in plain English. Free forever, no ads.
One email each weekday. Unsubscribe in one click. We never sell or share your address.
Test yourself
Take the Autoregressive Image Generation quiz
Instant feedback on every answer, and a shareable certificate with a verifiable ID once you pass a course.
Support free AI education. AI Understanding is a 501(c)(3) nonprofit — no ads, no paywall, ever. Make a donation
Hagaha xiga
Jiilka Sawirka AI
Su'aalaha soo noqnoqda
What is Autoregressive Image Generation?
Jiilka sawirka Autoregressive wuxuu dhisaa sawirro hal mar, isagoo saadaaliyay calaamad kasta wax kasta oo la soo saaray ka hor. Waa muhiim sababta oo ah mashiinnada ku xiga ee awoodda ku leh moodooyinka luqadda ayaa soo saari kara sawirro isku xiran, la xakameyn karo.
Waa maxay macnaha 'autoregressive' ee macnaha guud ee jiilka sawirka?
Moodooyinka 'Autoregressive' waxay u soo saaraan sawirka si isku xigta, iyagoo saadaalinaya calaamad kasta ama pixels iyadoo lagu saleynayo dhammaan walxaha la soo saaray ka hor.
Sidee moodooyinka casriga casriga ah ee autoregressive sida DALL-E 1 caadi ahaan u matalaan sawir ka hor inta aan la saadaalin?
Nidaamyada casriga ah waxay ku dhejiyaan sawirka calaamado kala duwan (badanaa iyada oo loo marayo codeer qaab VQ-VAE ah), ka dibna saadaaliya taxanaha calaamadda Transformer.
Waa kuwee moodooyinka hore ee keenay sawirada hal pixel ceeriin markiiba?
PixelRNN iyo PixelCNN waxay ahaayeen hormuudka moodooyinka is-difaacida kuwaas oo sawiray oo saadaaliyay sawirada pixel by pixel.
Farsamaynkee ayaa hubiya in Transformer-ku uu kaliya tago calaamadihii hore inta lagu jiro jiilka autoregressive?
Maaskarada sababa ayaa ka xannibaysa boos kasta ka qaybgalka calaamadaha mustaqbalka, iyada oo xoojinaysa habka bidix-ilaa-midig.
Waa maxay cillada ugu weyn ee muunad sawireedka autoregressive?
Sababtoo ah calaamad kastaa waxay ku xiran tahay kuwii hore, faragelinta waa inay ku socotaa token by token, taasoo ka dhigaysa jiilka mid la barbardhigo mid gaabis ah.