Modèle Generatif yu sukkandiko ci poñ
Modèle generatif yu sukkandiko ci poñ yi dañuy sos done ci jàng gradient bi ci séddaleb done yi — yoon wi tax bépp misaal bu bari bruit di gëna niru ak done dëgg.
Résumé
This score-function view unifies diffusion models with stochastic differential equations and underpins many modern image generators.
Plongeur bu xóot
Duñu modele probabilite yi ci saasi, waaye model yi sukkandiko ci poñ yi dañuy jàng poñ yi: gradient bu densité probabilite log bi méngoo ak li ñuy dugal. Xam anam wi ñuy jaar ba nudge ab sample ngir yokk probabilite bi doyna ngir defar ay done yu bees. Liggéeyu Yang Song ak Stefano Ermon ci 2019 tàggat nañu benn reso ngir xayma poñ yii ci niveau bruit yu bari di jëfandikoo poñ yu denoising, ba noppi ñu defar ay misaal ak dinaamiku Langevin - di gëna jëm ci poñ yi ba noppi yokk tuuti bruit. Seenug këyitu poñ-SDE bu 2021 dafa wane ni diffusion ak xeetu poñ yi dañuy nuru ñaari kanam yu benn yoon buy wéy bu ñuy leeral ci equation différentielle stochastique. Lu gëna am solo mooy, SDE bu nekk amna 'flux de probabilité' ODE bu ko méngoo ak benn marginal, loolu mooy tax ñu mëna am probabilite yu gëna jubal ak jël misaal yu gaaw.
Gis-gis xarala
Xayma poñ yi done yu sell yi ci saasi lu jafe la fu done yi néew, moo tax dañu tàggat model bi ci done yu bruit Gaussian perturbé ci eskaal yu bari. Denoising score matching dafay joxe benn mébet buñu mëna jëfandikoo: poñ yi distribution noised bi tollu ci wàllu bruit bi nga xaaj ko ak variance bruit bi, kon wax luy waaja am ci bruit bi ak wax luy am ci poñ yi dañuy nuru luy benn. Sampling dafay saafara SDE biy dellu ginaaw (wala ODE biy méngoo) tàmbalee ci bruit Gaussian bu sell bi.
njeextalu pexe
Gaawaay ak yaatuwaay
Visual IA mën na otomatise saytu, gis ak etiketu liggéey ci eskaal.
Tabax tànneef
Ekipu kreatif yi mën nañu defar konsept yu gëna gaaw te duñu def lu bari ci loxo.
Ekip ak def liggéey
Liggéeyukaay yi mën nañu jëfandikoo siñaal nataal wala wideo yu jafewoon lool ci liggéey.
Ëlëgu model Generatif yu sukkandiko ci poñ
Kaadaru poñ-SDE mooy motër theorik bi ci ginaaw yokkute yu bari ci IA generatif. Solver numérique yu gëna gaaw, oraaru bruit yu gëna baax, ak ODE probabilite-flow ñooy tax ñu mëna defar ay jamono yu jege dëgg ak jàngat probabilite yu gëna jub. Benn xalaat biy méngale poñ yi dafay tasaaroo ginaaw nataal yi ci biir audio, jëmmal moleculaire ak protein, niir point, ak simulation scientifique, ci noonu la méngoo ak xeetu méngale flow di tabax ci kaw fondaasioŋ yu wéy yii ngir wàññi jikkoom ci ay jéego yu néew.
Doxal ci àdduna dëgg
Reseau poñ yu am sart (NCSN) di defar kanam yu fotorealist ci topp gradient poñ yiñ jàng jaaraleko ci dinaamiku Langevin.
Tabaxwaat nataalu pajum, lu melni MRI bu gaaw, fu poñ yi ñu jàng nekk lu njëkk ngir feesal done yu scanner yu néew.
Defar jumtukaayi molecule ak protein ci gis-gis drog, modele tabb atomik 3D ak diffusion bu sukkandiko ci poñ.
Formu onde synthese audio fu modelu partition yi di denoise ngir wax bu sell wala music, melni ci vocodeur yu lalu ci diffusion.
Risk yi ak balustrade yi
Yelleefi nataal ak nangu mën na nekk risku yoon sudee fi ñu bawoo leerul.
Performance model bi mën na wuute ci leeraay bi, demographie bi ak environmaa bi.
Njuumteg positive yi mën nañu dem te kenn duko seetlu fileek xool wuñu buntu wóolu sa bopp.
Roadmap ngir samp gi
Mandargal kritërium nangug njub, woowaat ak njëgu njuumte.
Saytu ak done yu méngoo ak anam yi ñuy liggéeyee dëgg.
Yokk jàngat nit ngir xam fu wóorul dara wala am njeexital yu rëy.
Toppal model drift bi nga baaxal ko ginaaw bi kamera bi wala done yi soppeekoo.
Weyal di banneexu
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 Score-Based Generative Models 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
Gis bi ci topp
Modèle yu sukkandiko ci energie
Laaj yi ñuy faral di laaj
What is Score-Based Generative Models?
Modèle generatif yu sukkandiko ci poñ yi dañuy sos done ci jàng gradient bi ci séddaleb done yi — yoon wi tax bépp misaal bu bari bruit di gëna niru ak done dëgg. Bii gis-gis fonction poñ dafay boole xeetu diffusion ak equation différentiel stochastique te dafay jàppale generatëri nataal yu bees yu bari.
Lan mooy 'point' bi model yooyu di jàng?
Njortu bi mooy gradient bi ci densité done log bi ci wàllu input bi — dafay wane wàll wi gëna yokk probabilite bi ci misaal bi.
Lan moo waral ñuy tàggat model yu lalu ci poñ ci done yu yàqu ak bruit ci eskaal yu bari?
Ci gox yi densité bi néew, poñ yi ci dëgg duñu leer bu baax; perturbing done ak niveau bruit yu bari def poñ méngoo tractable fépp.
Ban xeetu jël misaal la xeetu poñ yi njëkka jëfandikoo ngir defar ay done?
Dinamiku Langevin dafay toxal benn misaal ci wàllu poñ yi, ba noppi def bruit bu ndaw, di soppi bruit bu bari bi ci done yu dëggu.
Ban lëkkaloo bu mag la këyitu poñ-SDE 2021 taxawal?
Song ak ñeneen. diffusion buñ boole muy model yu lalu ci poñ ci suufu benn kaadaru equation différentiel stochastique, ak ODE buy méngoo ak probabilite.
Lan moo méngoo ak xam poñ yi ci denoising score matching?
Denoising poñ méngoo dafay wane poñ bi tollu ci bruit negatif bi xaaj ko ak variance bruit bi, def bruit bi ñuy wax ak poñ yi ñuy wax luy am benn mébet.