Abashakashatsi b'Abarusiya Batezimbere Umwandiko-Kuri-Ishusho Ubwubatsi Ukoresheje Ishusho Mbere, Gutandukana Bitinze

Imbonerahamwe

• Ibisobanuro n'Intangiriro
• Akazi kajyanye
• Sisitemu ya Demo
• Kandinsky Ubwubatsi
• Ubushakashatsi
• Ibisubizo
• Umwanzuro & Imipaka
• Ibitekerezo byimyitwarire, gushimira hamwe

Ibisobanuro

Iyandikwa-ku-ishusho ni urwego rukomeye mubyerekezo bya mudasobwa bigezweho kandi byageze ku iterambere ryinshi binyuze mu bwihindurize bwubaka. Muri ibyo, hari ikwirakwizwa rishingiye ku gukwirakwiza kwerekana ireme ryiza. Izi moderi muri rusange zigabanyijemo ibyiciro bibiri: pigiseli-urwego nuburyo bwihishe-urwego. Turerekana Kandinsky [1], ubushakashatsi bushya bwububiko bwihishwa bwo gukwirakwiza, duhuza amahame yishusho yicyitegererezo cyambere hamwe nubuhanga bwo gukwirakwiza bwihishwa. Ishusho yabanjirije icyitegererezo yatojwe ukwayo gushushanya ikarita yanditswemo amashusho ya CLIP. Ikindi kintu cyihariye kiranga icyitegererezo cyatanzwe ni MoVQ yahinduwe ishyirwa mubikorwa, ikora nkibishusho autoencoder. Muri rusange, icyitegererezo cyateguwe kirimo ibipimo 3.3B. Twashyizeho kandi sisitemu-yerekana-sisitemu yerekana uburyo butandukanye butanga umusaruro nkumwandiko-ku-shusho, guhuza amashusho, inyandiko hamwe no guhuza amashusho, ibisekuru bitandukanya amashusho, hamwe no kuyobora-gushushanya / gushushanya. Twongeyeho, twasohoye inkomoko yinkomoko hamwe na bariyeri ya moderi ya Kandinsky. Isuzuma ryikigereranyo ryerekana amanota FID ya 8.03 kuri dataset ya COCO-30K, bikerekana urugero rwacu nkumukinnyi wambere ufungura isoko-nkomoko yubuziranenge bwibishusho.

1 Intangiriro

Mugihe gito cyane, ubushobozi bwo kubyara imiterere-yinyandiko-y-ishusho byateye imbere cyane, biha abakoresha ubwiza bwamafoto, hafi yigihe-nyacyo cyo kwihuta, umubare munini wibisabwa nibiranga, harimo byoroshye gukoresha-urubuga -bishingiye kumurongo hamwe nubuhanga buhanitse bwa AI.

Uru rupapuro rugaragaza iperereza ryihariye ryubushakashatsi bwihishe bwububiko, butanga icyerekezo gishya kandi gishya kuri ubu buryo bwo kwiga. Ubwa mbere, turasobanura imyubakire mishya ya Kandinsky nibisobanuro byayo. Sisitemu ya demo hamwe nibikorwa byashyizwe mubikorwa nabyo byasobanuwe. Icya kabiri, twerekana igeragezwa, ryakozwe mubijyanye nubwiza bwibishusho kandi tukazana amanota menshi ya FID muburyo busanzwe bwo gufungura isoko. Byongeye kandi, turerekana ubushakashatsi bukomeye bwo gukuraho ibyashizweho mbere twakoze, bidushoboza gusesengura neza no gusuzuma ibishushanyo bitandukanye kugirango tugere ku gishushanyo mbonera cyiza kandi cyiza.

Intererano zacu nizi zikurikira:

• Turerekana inyandiko yambere-y-ishusho yububiko bwateguwe dukoresheje guhuza amashusho mbere kandi ikwirakwizwa ryihishwa.

• Turerekana ibisubizo byubushakashatsi bigereranywa na moderi igezweho (SotA) nka Stable Diffusion, NIBA, na DALL-E 2, ukurikije ibipimo bya FID kandi tugera ku manota ya SotA mubyitegererezo byose bifungura isoko.

• Dutanga porogaramu yo gushyira mubikorwa uburyo bugezweho buteganijwe bwo kwerekana inyandiko-shusho, no gusohora ibyitegererezo byabanje gutozwa, bidasanzwe muburyo bwo hejuru. Uruhushya rwa Apache 2.0 rutuma bishoboka gukoresha icyitegererezo haba mubucuruzi butari ubucuruzi nubucuruzi.2 3

• Dushiraho urubuga rwerekana amashusho yurubuga rushobora gukoreshwa muguhuza amashusho kumashusho ukoresheje inyandiko (Icyongereza nu Burusiya zirashyigikirwa) dushingiye kuburyo bwateganijwe, kandi itanga imikorere yo gushushanya / gusohora.4 Kwerekana amashusho birahari kuri YouTube.5

Sisitemu yitiriwe Wassily Kandinsky, umunyabugeni uzwi cyane akaba n'umuhanga mu buhanzi.

[2] https://github.com/ai-imbere / Kandinsky-2

[3] https://huggingface.co/kandinsky-umuryango

[4] https://fusionbrain.ai/en/editor

[5] https://www.youtube.com/watch?v=c7zHPc59cWU