Autoregressive Image Generation
Isizukulwane sesithombe esizenzakalelayo sakha izithombe zibe ucezu olulodwa ngesikhathi, zibikezela ithokheni ngayinye kuyo yonke into ekhiqizwe ngaphambi kwayo.
Uhlolojikelele
It matters because the same next-token machinery powering language models can produce coherent, controllable images.
I-Deep Dive
Isizukulwane sesithombe esizenzakalelayo siphatha isithombe njengokulandelana futhi sisibikezele i-elementi ngayinye, lapho i-elementi entsha ngayinye ibekwe esimweni kuzo zonke ezidlule. Umsebenzi wangaphambi kwesikhathi onjenge-PixelRNN ne-PixelCNN ubikezele izithombe ngephikseli eyodwa eluhlaza ngesikhathi, iskena umugqa ngomugqa, obekunensa kodwa kuhlanzekile ngokombono. Amasistimu wesimanje esikhundleni salokho aqale acindezele isithombe kugridi yamathokheni ahlukene kusetshenziswa isishumeki sesitayela se-VQ-VAE, bese i-Transformer ibikezela lawo mathokheni ukusuka kwesokunxele kuye kwesokudla. I-DALL-E 1 ka-OpenAI kanye ne-Google's Parti balandele le recipe, bakhiqiza amathokheni esithombe abekwe emyalezweni wombhalo ngaphambi kokuwasusa amakhodi awabuyisele kumaphikseli. Inzuzo enkulu wukumodela okungenzeka kanye nezakhiwo ezihlanganisiwe ezabiwe nolimi. Izindleko ziyalandelana, zithatha isampula kancane.
I-Technical Insight
Imodeli ihlanganisa amathuba ahlangene awo wonke amathokheni abe umkhiqizo wemibandela: p(x) = umkhiqizo we-p(x_i onikezwe x_1...x_{i-1}). I-Transformer ene-causal (masked) ukunaka iphoqelela ukuthi indawo ngayinye ibona amathokheni angaphambili kuphela. Ngesikhathi sokuqeqeshwa ibikezela yonke ithokheni ngokuhambisana kusetshenziswa ukuphoqelela kukathisha, kodwa lapho kucatshangwa khona kufanele isampula ithokheni eyodwa ngesikhathi, inikeze ithokheni ngayinye. I-codebook efundiwe ibeka amamephu amathokheni abuyela kumapeshi esithombe, idikhoda eliwathathayo libe ngamaphikseli wokugcina.
I-Strategic Impact
Isivinini nesikali
I-Visual AI ingakwazi ukuhlola, ukutholwa, nokumaka imisebenzi esikalini.
Yakha ukukhetha
Amathimba aqanjiwe angakwazi ukulinganisa imiqondo ngokushesha ngezibuyekezo ezimbalwa ezenziwa mathupha.
Ithimba kanye nokusebenza komsebenzi
Imisebenzi ingasebenzisa amasiginali wesithombe nawevidiyo obekunzima ukuwenza ngaphambilini.
Ikusasa le-Autoregressive Image Generation
Ijubane liyinkundla yempi emaphakathi. Amasu anjengokukhipha amakhodi amathokheni ahambisanayo kanye namamaski (MaskGIT, Muse) akhiqiza amathokheni amaningi ngesikhathi esisodwa, kanye nokuqopha okucatshangelwayo okubolekwe kumamodeli olimi kulungiselelwa izithombe. Abacwaningi futhi bahlanganisa amathokheni ombhalo nesithombe kumgogodla owodwa ozenzakalelayo ukuze imodeli eyodwa ifunde futhi idwebe, njengoba kubonakala ezinhlelweni ze-multimodal. Lindela imibono ye-autoregressive neyokusabalalisa ukuze uhlale uxubene, namamodeli ayingxube abamba amandla okulawula amathokheni kanye nekhwalithi yokusabalalisa.
Ukuqaliswa Komhlaba Wangempela
I-DALL-E 1 ikhiqize izithombe ngokubikezela ngokuzenzakalelayo igridi yamathokheni esithombe ahlukene asuka kumagama-ncazo wombhalo.
Google's Parti sikale i-autoregressive text-to-image Transformer yaba amapharamitha ayizigidi eziyizinkulungwane ezingama-20 ukuze uthole izigcawu ezinemininingwane, ezithembekile.
I-PixelCNN ne-PixelRNN zibonise isizukulwane se-pixel-by-pixel eluhlaza futhi zisasetshenziswa njengezisekelo zokufundisa zamamodeli asekelwe okungenzeka.
I-MaskGIT kanye ne-Muse zisebenzisa i-parallel masked-token decoding ukusheshisa ukuhlanganiswa kwesithombe okususelwe kumathokheni kuyilapho kugcinwa ukuqeqeshwa kwesitayela esizenzakalelayo.
Izingozi & Guardrails
Amalungelo ezithombe kanye nemvume kungaba ubungozi bezomthetho uma ukuvela kungacacile.
Ukusebenza kwemodeli kungahluka kukho konke ukukhanya, izibalo zabantu, kanye nezindawo.
Okuhle okungelona iqiniso kungase kungabonakali ngaphandle uma izinga lokuzethemba liqashelwa.
Ukuqalisa Umhlahlandlela
Chaza indlela yokwamukela yokunemba, ukukhumbula, nezindleko zamaphutha.
Hlola ngedatha efana nezimo zangempela zokukhiqiza.
Engeza isibuyekezo somuntu ukuze uthole ukuzethemba okuphansi noma izibikezelo zomthelela omkhulu.
Landelela ukukhukhuleka kwemodeli bese uqinisekisa kabusha ngemva kwezinguquko zekhamera noma zesethi yedatha.
Qhubeka Uhlole
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
Umhlahlandlela olandelayo
I-AI Image Generation
Imibuzo evame ukubuzwa
What is Autoregressive Image Generation?
Isizukulwane sesithombe esizenzakalelayo sakha izithombe zibe ucezu olulodwa ngesikhathi, zibikezela ithokheni ngayinye kuyo yonke into ekhiqizwe ngaphambi kwayo. Kubalulekile ngoba imishini efanayo elandelayo yamamodeli olimi anamandla angakhiqiza izithombe ezihambisanayo, ezilawulekayo.
Isho ukuthini i-'autoregressive' kumongo wokwenziwa kwesithombe?
Amamodeli we-autoregressive enza isithombe sibe ngokulandelana, sibikezela ithokheni ngayinye noma iphikseli ngokusekelwe kuzo zonke izici ezikhiqizwe ngaphambi kwaso.
Ingabe amamodeli ezithombe zesimanje ezizenzakalelayo afana ne-DALL-E 1 ngokuvamile amele isithombe ngaphambi kokusibikezela?
Amasistimu esimanje acindezela isithombe sibe amathokheni ahlukene (ngokuvamile ngesishumeki sesitayela se-VQ-VAE), bese sibikezela lokho kulandelana kwethokheni ne-Transformer.
Imaphi amamodeli okuqala akhiqize izithombe ngephikseli eyodwa eluhlaza ngesikhathi?
I-PixelRNN ne-PixelCNN bekungamamodeli e-autoregressive aqala ukuskena futhi abikezele izithombe ngephikseli ngephikseli.
Iyiphi indlela eqinisekisa ukuthi i-Transformer ibheka amathokheni angaphambili kuphela ngesikhathi sokukhiqiza okuzenzakalelayo?
Ukufihla okuyimbangela kuvimbela indawo ngayinye ekunakekeleni amathokheni esikhathi esizayo, kuphoqelela ukufakwa kwesici ukusuka kwesokunxele kuye kwesokudla.
Ithini i-drawback eyinhloko esebenzayo yesampula yesithombe esizenzakalelayo?
Ngenxa yokuthi ithokheni ngalinye lincike kwedlule, ukuqagela kufanele kusebenze ngethokheni, okwenza isizukulwane sihambe kancane uma kuqhathaniswa.