paint-brush
I-ByteDance Ukubheja Okukhulu: Ukuphuma Okuyisu Kusuka Kusekisini Yezokuxhumana Ngevidiyo ye-AI?nge@bigmao
Umlando omusha

I-ByteDance Ukubheja Okukhulu: Ukuphuma Okuyisu Kusuka Kusekisini Yezokuxhumana Ngevidiyo ye-AI?

nge susie liu7m2024/09/27
Read on Terminal Reader

Kude kakhulu; Uzofunda

I-ByteDance yethule amamodeli amabili ashaya kakhulu, i-PixelDance ne-Seaweed, ngesithembiso esikhulu sokunyakazisa ukwakheka kwevidiyo ye-AI. Lokhu akuyona i-AI gimmick - kuyingxenye yoshintsho lwamasu. Ikusasa le-ByteDance alixhomekile kuma-hashtag adlulayo kanye nemali engenayo yezikhangiso kodwa ezandleni eziqinile, ezithembekile zokukhiqiza okubucayi—imodeli egxile ebhizinisini eklanyelwe ukukhula nokuqamba izinto ezintsha. Nakhu okudingeka ukwazi mayelana namamodeli abo akamuva.
featured image - I-ByteDance Ukubheja Okukhulu: Ukuphuma Okuyisu Kusuka Kusekisini Yezokuxhumana Ngevidiyo ye-AI?
susie liu HackerNoon profile picture
0-item
1-item
2-item


Lapho nje ucabanga ukuthi ungathatha umoya ku-AI, i-ByteDance sashays uye endaweni ebonakalayo - kodwa hhayi nge-saga ye-TikTok engapheli. Embukisweni we-AI ocwebezelayo e-Shenzhen, Injini ye-ByteDance ye-Volcano Engine iveze izinhlobo ezimbili ezishaya kakhulu, i-PixelDance ne-Seaweed , ngesithembiso esikhulu sokunyakazisa ukwakheka kwezwe levidiyo . I-PixelDance igxile ekukhiqizeni amavidiyo ashukumisayo, ekhwalithi ephezulu ngemibhalo nemiyalelo ebonakalayo—yicabange njengengcweti enikeza isifiso kubakhi bamavidiyo. I-Seaweed ingena ijule ezindaweni zokugqwayiza kwe-3D nokunikezwa kobuciko, ihlinzekela labo abafuna ukwenza okubonakalayo kwabo kugqame njenge-confetti ephathini.


Kunobuchwepheshe obuthile obungathí sina obufakwe kulawa mamodeli, futhi yingoba izethameli eziqondiswe kuzo akuwona umakhelwane wakho we-Gen Z—bangabadali abaqotho nezindlu zokukhiqiza.


Lesi isibuyekezo esimnandi esivela ku-Bytedance. Qhubeka ufunde ukuze uthole ukuthi kungani.


Lokho Abayikho Hhayi Ngiyakutshela


I-ByteDance ibilokhu ithandana ne-AI isikhathi eside. Emuva lapho beqala ukuthuthukisa izincomo ze-algorithmic, iningi lethu lahlobanisa isifinyezo esithi “AI” ne-“Adobe Illustrator”. Ungacabanga ukuthi bacija i-AI ngemisebenzi ebhekene nabathengi, kodwa yilokho abafuna ukuthi ukukholwe.


Babheja ukuthi i-AI izozisusa endaweni esetshenziswayo (ejwayele ukumangalelwa, ukucutshungulwa kwedatha, kanye nenqwaba yamatheyipu abomvu) iye ethuluzini elibalulekile .


Ikusasa le-ByteDance alixhomekile kuma-hashtag adlulayo kanye nemali engenayo yezikhangiso kodwa ezandleni eziqinile, ezithembekile zokukhiqiza okubucayi— imodeli egxile ebhizinisini eklanyelwe ukukhula nokuqamba izinto ezintsha. Ngokushintsha kwamasu ekugxilweni kwezinhlelo zokusebenza ezichwepheshile, i-ByteDance ngeke nje isinde esivunguvungwini; izophumelela ngokuchaza kabusha indima yayo ku-tech ecosystem .


I-PixelDance: Ukusuka Engqondweni Kuya Kumlingo


I-PixelDance ithembisa ukuba yileso sikhungo esingakwazi ukuthatha imisho yakho esakheke uhhafu kanye nokubhibhidla okuhambisana kancane bese silungisa lokho obukade ucabanga ngakho . Izici zayo eziyinhloko zihlanganisa:


Umbhalo Ohlanganisiwe Nokufakwa Kwesithombe

  • Ngokungafani namamodeli okukhiqiza amavidiyo endabuko, i-PixelDance ivumela abasebenzisi ukuthi banikeze kokubili umbhalo nezithombe (ikakhulukazi uzimele wokuqala nowokugcina). Lokhu kusho ukuthi uthola ividiyo eqala futhi ephetha ncamashi njengoba ubucabanga.
  • Isifinyezo sika-Layman : Ungumqondisi olalela imibono yakho futhi ayenze iphile. Futhi akaxoxi.

I-Latent Diffusion Model Architecture

  • Le ngane isebenzisa imodeli yokusabalalisa ecashile, isebenzisa ama-Variational Autoencoder (VAEs) aqeqeshwe kusengaphambili kanye nesishumeki sombhalo . Okokufaka kwesithombe kufakwa nge-VAE, ehlanganiswe nezinto eziguquguqukayo zevidiyo ukuze kuqinisekiswe ukunyakaza nokuvumelana.
  • Isifinyezo sika-Layman: Kuyi-blender yobuchwepheshe obuphezulu ethuthukisa amaphupho akho evidiyo ngaphandle kwezigaxa.

Izigaba Zevidiyo Eziqhubekayo

  • I-PixelDance ingakhiqiza iziqeshana zevidiyo eziqhubekayo kuyilapho igcina ukuhambisana kwesikhashana . Isebenzisa uhlaka lokugcina lwesegimenti eyodwa njengohlaka lokuqala lwelandelayo.
  • Isifinyezo sika-Layman: Valelisa kulokho kusikeka okungajwayelekile.

Ukuhlelwa kwevidiyo eyi-Zero-Shot

  • Lesi sici sivumela abasebenzisi ukuthi bahlele amavidiyo ngaphandle kwesidingo sokuqeqeshwa okuthile. Ungaqondisa ukudalwa kwevidiyo ngokulungisa uzimele wokuqala nowokugcina.
  • Isifinyezo sika-Layman: Ungakwazi ukuhlela kabusha yonke ifilimu ye-blockbuster ngokulungisa izigcawu zokuvula nokuvala—umlingo wesinema omsulwa.

Ukusekelwa Kwesitayela Esibanzi

  • Kusukela kokumnyama nokumhlophe kuye kopopayi be-3D nezitayela zokudweba zesiShayina ezivamile, i-PixelDance isekela ububanzi obuningi bobuhle, okufaka phakathi izici ezifana no-1:1 kanye no-3:4.
  • Isifinyezo sika-Layman: Akuyena u-Anna Wintour, noma esinye isishicileli sesitayela. Le nto ayinazo izintandokazi zobuciko, iqoqo nje lamathuluzi.

Amasethi edatha okuqeqesha

  • I-PixelDance yaqeqeshwa ku-WebVid-10M , isethi yedatha yamavidiyo amafushane angaba yizigidi ezingu-10 (ubude obujwayelekile: imizuzwana engu-18) enokulungiswa okungu-336 x 596. Ukwengeza, basebenzise iziqeshana zamavidiyo ezingenawo ama-watermark angu-500,000 ukuze baqinisekise ukuphuma kwekhwalithi ephezulu.
  • Isifinyezo sika-Layman: Kubukwe amavidiyo amaningi ongawabuka. Kuyi-encyclopedia yezinkomba zesinema. Themba ubungcweti bayo.


I-Seaweed: I-Alchemist yokuhlela


I-Seaweed iyimpendulo ye-ByteDance kumaye wokuhlela akhungethe abadali. Idizayinelwe ukuhambisana ne-PixelDance , I-Seaweed isebenzisa i-AI ukuze iqondise inqubo yokuhlela ngezici ezihlaba umxhwele:


Ukunikezwa kwe-3D

  • I-Seaweed isebenzisa amasu okunikezela athuthukisiwe ukuze ikhiqize okubonakalayo okungekona nje okuyisicaba kodwa kunokujula nokuba ngokoqobo, okuyenza ifanelekele ukugqwayiza kwekhwalithi ephezulu nezinkulumo zobuciko.
  • Isifinyezo sika-Layman: Unesitudiyo sopopayi se-Pixar ezandleni zakho.


Intelligent Cut Ukuthola

  • Isebenzisa ama-algorithms ayinkimbinkimbi, i-Seaweed ihlonza izikhathi ezibalulekile kuvidiyo yakho , ivumela ukusikeka okusheshayo ngaphandle kokulahlekelwa ukulandisa.
  • Isifinyezo sika-Layman: Umhleli womuntu siqu ongakhomba izikhathi ezinhle kuvidiyo yakho eluhlaza. Ngokushesha. Ngokushesha ngempela.


Ukugreda Kombala Okuthuthukisiwe kanye Nemiphumela

  • Leli thuluzi lenza ngokuzenzakalelayo ukulungiswa kombala futhi lisebenzisa imiphumela yesitayela ukuze kuqinisekiswe ukuthi ividiyo yakho ibukeka iphucuziwe ngaphandle kokulungiswa okuyisicefe okwenziwa ngesandla.
  • Isifinyezo sika-Layman: Ungathola imibala nezitayela ngisho no-Baz Luhrmann angavuma. Ngisho noma ungaboni umbala.


Isixhumi esibonakalayo esisebenziseka kalula esinosizo lwe-AI

  • I-Seaweed ihlanganisa amakhono e-AI athuthukisiwe nesixhumi esibonakalayo esilula, okwenza ifinyeleleke ngisho nakulabo abangekho ubuchwepheshe be-tech.
  • Isifinyezo sika-Layman: Kufana nokuthola i-Ferrari kodwa ngemvume yokufunda—ukugibela okubushelelezi kwawo wonke umuntu.


Imicabango Yokugcina: Ukwakhela I-Tech Elite


I-PixelDance ne-Seaweed akuwona nje amathuluzi; bayithikithi le-ByteDance lokuya endaweni yokudlala entsha, esezingeni eliphezulu lapho iziteki ziphezulu futhi umncintiswano ushubile. Nakuba abaningi benelisekile ukubhekelela abadali abavamile abafuna izifanekiso kanye nokuqukethwe “okwanele ngokwanele” kombukiso wesilayidi womshado womzala wakho, i-ByteDance ifuna abakwa-A.


I-ByteDance ikhombisa ngendelelo ukuthi ayinantshisekelo yokuba yiCanva yevidiyo ye-AI. Bahlose okuthile okufana kakhulu nengxube ye-Pixar-ILM eshayelwa yi-AI. Lokhu kusho ukuthi babeka phambili ikhwalithi yabasebenzisi babo ngaphezu kobuningi , abenzi bamafilimu ababhekile, abapopayi, ama-ejensi wokumaketha, namabhizinisi adinga amavidiyo aphucuzeke kakhulu, ayinkimbinkimbi agqama emakethe yokuqukethwe egcwele.


Kuhlakaniphile. Uhlakaniphe ngempela. Kodwa hhayi ngaphandle kwezingozi zayo.


Isu Lentengo Elinamandla: Inkemba Esika Kabili?


Isu labo lentengo elinolaka, ku-$0.002 ithokheni ngayinye—uma kuqhathaniswa ne-OpenAI engu-$0.03 ithokheni ngayinye —iklanyelwe ukuphazamisa imakethe. Kuyikholi ye-siren yamabhizinisi amancane naphakathi nendawo (ama-SME) nabadali be-indie, manje abangafinyelela amathuluzi aphambili abake bawathatha njengengatholakali.


Kodwa izindleko eziphansi ithokheni ngayinye ukubheja okuyingozi. I-ByteDance ingakwazi ukudlala le geyimu manje, kodwa ukugcina lezi zintengo isikhathi eside kungaba inselele , ikakhulukazi uma izindleko zokutholwa kwedatha nezikali zengqalasizinda zilinganiselwe. Nakuba intengo yokuqala ingase ihehe amakhasimende, ukulondoloza lokhu ngaphandle kokudela ikhwalithi noma ukusungula izinto ezintsha kungase kube ukuhamba ngentambo.


Ukuncintisana okuvela ku-Unity kanye ne-Unreal Engine


I-ByteDance ingena enkundleni yempi kakade ebuswa imidondoshiya efana ne-Unity kanye ne-Unreal Engine abangamavulandlela emikhakheni yokuhlinzeka nge-3D, ukugqwayiza, ngisho nokwenza amafilimu esikhathi sangempela, nemiphakathi emikhulu, ethembekile nezinsiza ezibanzi. I-Unity ibilokhu yenza intuthuko ekukhiqizeni okubonakalayo nasekuxoxeni izindaba ngesikhathi sangempela, kuyilapho izibuyekezo zakamuva ze-Unreal Engine ziphendule amakhanda ngamakhono azo okunikeza angokoqobo. Bobabili manje sebehlanganisa izici ze-AI ukuze bathuthukise ukunikezwa kwabo, okubenza baqhudelane ngqo namakhono e-Seaweed okunikezela nge-3D kanye namavidiyo asebenzayo.


I-ecosystem emisiwe ye-Unity kanye ne-Unreal Engine iphinde ibanike umphetho obalulekile: isisekelo sabasebenzisi nokusekelwa komphakathi . Inselele ye-ByteDance izoba ukukholisa abadali ukuthi i-PixelDance ne-Seaweed zinikeza okuthile lezi zinkundla ezingakwazi—kungaba ukuhlanganiswa okungcono namathuluzi akhona, ukugeleza komsebenzi okusheshayo, noma ikhwalithi yokuphuma ephezulu. Kodwa ingabe ukuphakama kwezobuchwepheshe kunganele ukukholisa abasebenzisi ukuthi i-Unity ne-Unreal bekutshalwa iminyaka?


Izingibe ezingaba khona: Isifiso Sokuvelela Esiphezulu, Izibalo Eziphakeme


Izifiso eziphakeme ze-ByteDance ze-PixelDance kanye ne-Seaweed ziletha uxhaxha olukhazimulayo lwezingozi. Ukusetha ibha phezulu ngezithembiso zamathuluzi ebanga eliqeqeshiwe ukuze kuqhutshwe ama-titans embonini kuyisinyathelo esinesibindi, kodwa uma lezi zingxenyekazi zehluleka ukuletha , zizifaka engozini yokubhekana nolaka lwalabo chwepheshe abahlose ukubaheha. Ngeza ekuhlolweni kokulawula -ngoba ngubani ongakuthandi ukucwaningwa kwamabhuku kukahulumeni okuhle uma udlala ne-AI?—futhi unendlela yokubhekana nokungezwani, ikakhulukazi ngokukhathazeka kobumfihlo bedatha okuza kukhulu kanye ne-TikTok njengenzalo yabo. Futhi ngenkathi ukuqondisa kwabaphambili kuwumdlalo ohlakaniphile, u-Peter Thiel mhlawumbe angabuza: ingabe sikhona isikhala sokwakha ukuzimela lapha?


Isithombe Esikhulu: Sihlose Izinkanyezi, Kodwa Ingabe Zizonamathela Ekufikeni?


Njengoba i-PixelDance ne-Seaweed zilungiselela ukwethulwa okubanzi, i-ByteDance yenza isitatimende esinesibindi: abanelisekile ngokuba umdondoshiya wenkundla yezokuxhumana . Bahlose ukubamba iqhaza eliphezulu ekudalweni kokuqukethwe kwedijithali , futhi ukugxila kwabo kubadali bekhwalithi ephezulu ngaphezu kwevolumu enkulu kuwumnyakazo we-gutsy, ikakhulukazi emhlabeni lapho yonke inkundla igila izimanga ukuze ifinyeleleke ngangokunokwenzeka.


Bangakwazi yini ukubumba kabusha umhlaba wevidiyo ekhiqizwe yi-AI FUTHI bazijike basondelane Nesihlalo Sobukhosi Sensimbi sobuchwepheshe? Noma ingabe lokhu kuzoba esinye isilingo sokuvelela emibhalweni yomlando wobuchwepheshe?


Kungumfanekiso wenyanga, futhi njengoba noma yimuphi uchwepheshe wezobuchwepheshe azi, labo abahlali kahle njalo. Kodwa into eyodwa eqinisekile: benza ukwakheka kwevidiyo kube indawo eheha kakhulu.


Ngakho-ke, lungisa ama-popcorn akho. Lo mbukiso uyaqala.


Qaphela: Zombili izinkundla okwamanje zisesigabeni sokuhlola sokumema kuphela (ungazama ukufaka isicelo nge-Volcano Engine), ngokufinyelela okubanzi okulindeleke maduzane—nakuba usuku oluqondile lungaqondakali njengesignali yakho ye-Wi-Fi phakathi nekholi ye-Zoom).