Diffusion qauv tau cheb lub ntiaj teb los ntawm cua daj cua dub nrog kev tso tawm ntawm Dall-E 2, Google's Imagen, ruaj khov diffusion, Thiab Ib nrab hnub, sparking innovation thiab stretching lub ciam teb ntawm kev kawm tshuab.
Cov qauv no tuaj yeem tsim cov duab yuav luag tsis txwv los ntawm cov lus qhia, suav nrog photorealistic, tej yam yees siv, futuristic, thiab, ntawm chav kawm, cov duab ntxim hlub.
Cov peev txheej no rov xav txog qhov nws txhais tau li cas rau tib neeg los cuam tshuam nrog silicon, muab peb lub peev xwm los ua cov duab uas peb tuaj yeem pom.
Raws li cov qauv no tsim los yog cov txheej txheem txuas ntxiv mus, tib neeg yuav tuaj yeem tsim cov duab, yeeb yaj kiab, thiab lwm yam kev paub dhau los nrog kev xav xwb.
Hauv tsab ntawv no, peb yuav tham txog qhov qauv diffusion, ruaj khov diffusion, nws ua haujlwm li cas, thiab tus qauv diffusion inpainting tutorial, thiab lwm yam.
Tus qauv Diffusion yog dab tsi?
Cov qauv kev kawm tshuab uas tuaj yeem tsim cov ntaub ntawv tshiab los ntawm cov ntaub ntawv kev cob qhia raug xa mus rau cov qauv tsim qauv. Lwm cov qauv tsim muaj xws li cov qauv ntws, hloov pauv hloov pauv hloov pauv hloov pauv, thiab cov kev sib txuas sib txuas sib txuas (GANs).
Txhua tus tuaj yeem tsim cov duab zoo heev. Diffusion qauv kawm kom rov qab tau cov ntaub ntawv los ntawm kev thim rov qab cov txheej txheem no nrov ntxiv tom qab ua kom puas cov ntaub ntawv kev cob qhia los ntawm kev ntxiv suab nrov. Txhawm rau muab tso rau lwm txoj hauv kev, cov qauv diffusion tuaj yeem tsim cov duab sib xyaw ua ke ntawm lub suab nrov.
Diffusion qauv kawm los ntawm kev qhia suab nrov rau cov duab, uas cov qauv tom qab masters qhov kev tshem tawm. Txhawm rau tsim cov duab tiag tiag, tus qauv tom qab ntawd siv cov txheej txheem denoising rau cov noob random.
Los ntawm kev kho cov txheej txheem tsim cov duab, cov qauv no tuaj yeem siv ua ke nrog kev taw qhia cov ntawv-rau-duab los tsim cov duab yuav luag tsis muaj kev txwv ntawm cov ntawv nyeem ib leeg. Cov noob tuaj yeem raug coj los ntawm cov khoom siv los ntawm embeddings zoo li CLIP los muab cov ntawv muaj peev xwm-rau-duab.
Diffusion qauv tuaj yeem ua ntau yam haujlwm, suav nrog kev tsim duab, duab denoising, inpainting, outpainting, thiab me ntsis diffusion.
Tam sim no, dab tsi yog qhov ruaj khov diffusion?
Stable Diffusion yog cov qauv kev kawm tshuab rau cov ntawv nyeem cov duab tsim los ntawm Stability.AI. Nws muaj peev xwm tsim cov duab los ntawm cov ntawv nyeem.
Cheebtsam ntawm ruaj khov diffusion
ruaj khov diffusion yog ib qho system uas muaj ntau yam thiab cov ntsiab lus. Nws tsis yog ib qho qauv. Thaum peb tshawb xyuas tom qab lub hood, thawj qhov peb pom yog tias muaj cov ntawv nyeem nkag siab uas hloov cov ntaub ntawv cov ntaub ntawv mus rau hauv tus lej sawv cev uas khaws cov ntawv nyeem cov ntsiab lus.
Peb tuaj yeem hu cov ntawv no encoder yog Transformer qauv lus (Technically: text encoder ntawm CLIP qauv). Nws siv cov ntawv sau thiab tsim cov npe ntawm cov lej (tus vector) rau txhua lo lus / token hauv cov ntawv nyeem. Cov ntaub ntawv ntawd yog tom qab ntawd muab rau Cov Duab Generator, uas yog tsim los ntawm ntau yam khoom.
Muaj ob kauj ruam hauv lub tshuab hluav taws xob duab:
1. Duab Cov ntaub ntawv tus tsim
Cov khoom tseem ceeb hauv Stable Diffusion yog cov khoom no. Nws yog qhov uas feem ntau ntawm kev txhim kho hauv kev ua tau zoo tshaj li yav dhau los versions tau ua.
Cov khoom no dhau los ntawm ntau theem los muab cov ntaub ntawv duab. Tus tsim cov ntaub ntawv duab ua haujlwm tsuas yog nyob rau hauv qhov chaw duab cov ntaub ntawv (lossis qhov chaw latent).
Nws yog nrawm dua li ua ntej diffusion qauv uas ua haujlwm hauv pixel qhov chaw vim yog tus yam ntxwv no. Technically hais lus, cov khoom no yog tsim los ntawm kev teem caij algorithm thiab UNet neural network.
Cov txheej txheem uas tshwm sim hauv cov khoom no yog hu ua "diffusion". Ib daim duab zoo kawg nkaus yog tsim los ntawm cov ntaub ntawv ua tiav hauv cov kauj ruam (los ntawm cov khoom txuas ntxiv, cov duab decoder).
2. Duab Decoder
Siv cov ntaub ntawv nws tau txais los ntawm cov ntaub ntawv tsim tawm, cov duab decoder tsim ib daim duab. Nws tsuas yog ua tiav ib zaug los tsim cov duab pixel tiav thaum xaus ntawm kev ua haujlwm.
Stable Diffusion Impainting tutorial
Stable Diffusion duab inpainting yog cov txheej txheem ntawm kev sau hauv qhov ploj lossis puas ntawm daim duab. Lub hom phiaj ntawm daim duab inpainting yog los zais qhov tseeb tias daim duab tau rov qab los.
Cov txheej txheem no nquag siv los tshem tawm yam tsis xav tau los ntawm cov duab lossis los kho thaj chaw puas ntawm cov duab keeb kwm. Stable Diffusion Inpainting yog ib txoj hauv kev tsis ntev los no ntawm inpainting uas ua rau muaj txiaj ntsig zoo.
Ua raws li cov lus qhia hauv qab no yuav ua rau koj pib tshawb nrhiav inpainting thiab hloov kho cov duab uas twb muaj lawm yog tias koj xav sim inpainting nrog kev ruaj khov diffusion:
- Mus rau Huggingface Stable Diffusion Impainting
- Upload koj tus kheej duab
- Tshem tawm ib feem ntawm koj daim duab uas yuav tsum tau hloov.
- Nkag mus rau koj qhov kev ceeb toom ntawm no (qhov koj xav ntxiv rau qhov chaw koj tshem tawm)
- Xaiv "khiav"
Hauv daim vis dis aus saum toj no, peb muab ib daim duab nrog peb txiv qaub thiab sib pauv rau cov txiv apples. Kuv tus kheej xav kom sim nws nrog koj tus kheej cov duab thiab cov lus ceeb toom.
xaus
Feem ntau, tsis tu ncua diffusion inpainting yog ib txoj hauv kev zoo rau kev tsim cov duab cuav lossis cov yeeb yaj kiab uas zoo li muaj tiag. Raws li peb txav mus rau kev nce qib tshiab thev naus laus zis, nws yuav nyuaj dua thiab nyuaj rau qhov txawv ntawm qhov tseeb thiab kev dag ntxias raws li thev naus laus zis nce qib.
Swahir
Thawj ib nrab yog kiag li tsis cuam tshuam rau qhov thib ob ib nrab. Nws yuav yog qhov txias tiag tiag yog tias tus sau yuav tau piav qhia yuav ua li cas inpaint ua haujlwm nyob rau hauv lub moj khaum ntawm tus qauv uas nws tau piav qhia ua ntej, tuaj yeem muab kev nkag siab. Tab sis tsis muaj! Qhov ntawd yuav xav tau kev nkag siab tiag tiag, tsis yog sau thiab ua cov ntawv sau tsis raug.