InoteveraGaidhi rinotevera
Kuyeuka muDiffusion Models
Tekinoroji
Nhungamiro yehunyanzvi
Adversarial Diffusion Distillation (ADD) inzira yekudzidzisa inoshandura modhi inononoka, yakawanda-nhanho yekuparadzira kuita mudzidzi inoburitsa mifananidzo munhanho imwe kusvika ina.
Teerera kugwaro iriAI-yakarondedzerwa · ~3 min
Inobatanidza kurasikirwa kwemhandu kubva kune anosarura nehutungamiri kubva kumuenzaniso wemudzidzisi wekutanga. ADD yakagadzira SDXL Turbo, uye yakasarudzika yakasarudzika yakashandiswa kune inokurumidza mhando seSD3 Turbo uye, maererano neBlack Forest Labs, Flux.1 schnell. Mhedzisiro yacho iri pedyo-chaiyo-chaiyo-chaiyo mufananidzo chizvarwa pane zvakajairika Hardware, pane imwe mutengo mukusiyana uye kutonga kwakanaka.
Iyo yakajairwa diffusion modhi inogadzira mufananidzo nekubvisa ruzha pamatanho mazhinji, kazhinji makumi maviri kusvika makumi mashanu, uye nhanho imwe neimwe inopfuura yakazara kuburikidza netiweki hombe. Distillation inodzidzisa network yevadzidzi kuti isvike pamhedzisiro yakafanana nematanho mashoma. Stability AI yakaunza Adversarial Diffusion Distillation mune raNovember 2023 bepa naAxel Sauer nevamwe vaaishanda navo, yakaburitswa padivi peSDXL Turbo. Mudzidzi anotanga kubva kune akadzidziswa SDXL huremu uye anodzidziswa nekurasikirwa kuviri. Kurasikirwa kweanopikisa kunobva kune munhu anosarura anoedza kutaura zvinobuda nemudzidzi kubva pamifananidzo chaiyo. MuADD, rusarura urwu rwakavakwa pamusoro peganda rakaomeswa rekuona (DINOv2) rine misoro midiki inodzidziswa. Kurasikirwa kwe distillation kunoshandisa iyo yekutanga SDXL semudzidzisi. Kubuda kwomudzidzi kunonzwikazve, mudzidzisi obva aita ruzha, uye mudzidzi anosundirwa kune zvakafanotaurwa nomudzidzisi, imwe nzira yokuderedza zvibodzwa. Izwi remhanduri rinochengeta nhanho-nhanho mifananidzo yakapinza, uye temu yemudzidzisi inovachengeta vakatendeka kune izvo mhando huru inoziva. Musiyano wakazotevera, Latent Adversarial Diffusion Distillation (LADD), inomhanyisa isaruro munzvimbo yakavanzika ichishandisa zvimiro zvemudzidzisi, izvo zvinodzivirira kudhura kunodhura kumapikisi. Yakashandiswa SD3 Turbo, uye Black Forest Labs yati Flux.1 schnell yakadzidziswa nenzira iyi. Kutengeserana kwechokwadi. Distilled modhi dzinowanzorasikirwa nemhando dzakasiyana siyana, saka mbeu dzakasiyana dzekukurumidza kumwechete dzinogona kutaridzika zvakafanana. Classifier-yemahara gwara rinowanzo vakwa mukati, saka chiyero chekutungamira uye zvakaipa zvinokurudzira zvinobata zvakasiyana kana kusashanda. Tsanangudzo yakanaka, manyorerwo ezvinyorwa uye zvinyorwa zvisina kujairika zvinogona kusarira kumashure kwemudzidzisi. Pfungwa isiriyo yakajairika ndeyekuti turbo modhi ingori yakazara modhi inomhanya nematanho mashoma. Iyo yakajairika modhi inomhanya pane imwe nhanho inopa blurry mess, uye kumhanya kunobva pakudzidzirazve. Dzimwe nzira dzekukurumidza sampling dzinosanganisira kufanana modhi, LCM, inofambira mberi distillation uye SDXL Mheni.
Iyo mhenyu yekudhirowa app inovandudza mufananidzo pese apo mushandisi anonyora kana kugadzirisa sketch, achishandisa SDXL Turbo pane imwe nhanho kuitira kuti mhedzisiro ioneke ipapo ipapo.
A game studio prototyping concept art inogadzira mazana ekukurumidza kusiyana neFlux.1 schnell pamatanho mana, yozokwenenzvera vakasarudzwa nemodhi izere zvishoma.
Mugadziri anoona kuti kusimudzira kwakashata uye hukuru hwekutungamira hunoita zvishoma pane yeTurbo modhi, nekuti nhungamiro yakavakirwa mukati panguva yedistillation. Vanogadzirisa mazwi anokurumidza pane kudaro.
Iyo diki isina purofiti inomhanyisa yemuno mufananidzo chishandiso pane yepakati-renji GPU uye inotora yakanyungudutswa modhi nekuti mana-nhanho chizvarwa chinoramba chakamirira ipfupi zvakakwana kune mhenyu mashopu.
Vashoma-matanho chizvarwa chave sarudzo inotarisirwa yemhando nyowani dzemifananidzo, uye mamwe mazano edistillation ari kuiswa kuvhidhiyo, uko nhanho imwe neimwe yakachengetwa inonyanya kukosha. Tsvagiridzo inoenderera mberi nekuvhara mikaha mukusiyana uye kutonga nekukurumidza, semuenzaniso nekubatanidza kurasikirwa kweanopikisa nezvinangwa zvekugovera. Kuti nhanho imwe-nhanho modhi inogona kunyatsoenderana nevadzidzisi vavo ichiri mubvunzo wakavhurika, uye mhinduro inotsamira pakuyerwa kwemhando. Kune vashandisi, mhedzisiro inoshanda isarudzo yakawanda: inokurumidza distilled modhi yemadhizaini, previews uye maturusi anodyidzana, uye anononoka akazara modhi kana ruzivo uye kutonga zvinonyanya kukosha.
Tsamba yemahara
Nhatu dzakasimbiswa dzeAI nyaya mangwanani ega ega pakati pevhiki, dzakanyorwa neChirungu chakajeka. Yemahara zvachose, hapana ads.
Imwe email zuva rega rega revhiki. Bvisa kunyoresa kamwe chete. Hatimbotengese kana kugovera kero yako.
Zviedze iwe pachako
Mhinduro yepakarepo pamhinduro yega yega, uye chitupa chinogovaniswa chine ID inobatika kana wapedza kosi.
Tsigira dzidzo yemahara yeAI. AI Understanding ndeye 501(c)(3) isingabatsiri - hapana ads, hapana paywall, zvachose. Ita chipo
Adversarial Diffusion Distillation (ADD) inzira yekudzidzisa inoshandura modhi inononoka, yakawanda-nhanho yekuparadzira kuita mudzidzi inoburitsa mifananidzo munhanho imwe kusvika ina. Inobatanidza kurasikirwa kwemhandu kubva kune anosarura nehutungamiri kubva kumuenzaniso wemudzidzisi wekutanga. ADD yakagadzira SDXL Turbo, uye yakasarudzika yakasarudzika yakashandiswa kune inokurumidza mhando seSD3 Turbo uye, maererano neBlack Forest Labs, Flux.1 schnell. Mhedzisiro yacho iri pedyo-chaiyo-chaiyo-chaiyo mufananidzo chizvarwa pane zvakajairika Hardware, pane imwe mutengo mukusiyana uye kutonga kwakanaka.
ADD inobatanidza kurasikirwa kweanosarura, izvo zvinoita kuti mifananidzo irambe yakapinza, ine zvibodzwa-distillation kurasikirwa kubva kumudzidzisi, izvo zvinovaita kuti varambe vakatendeka kumuenzaniso wekutanga.
Stability AI yakaburitsa SDXL Turbo neiyo Mbudzi 2023 ADD bepa naSauer nevamwe vaanoshanda navo.
Anosarura anoshandisa chando pretrained DINOv2 maficha uye anodzidzisa chete misoro midiki pamusoro payo.
LADD inofambisa rusarura munzvimbo yakavanzika ichishandisa zvimiro zvemudzidzisi, izvo zvinodzivirira kudhura kunodhura kumapikiseli.
Negative zvirevo zvinoshanda kuburikidza neclassifier-yemahara nhungamiro. Distilled modhi dzinowanzotora kutungamira panguva yekudzidziswa uye kumhanya pasina iyo, saka kusimudzira kwakashata kunorasikirwa nekuita.
Ramba uchidzidza
Mamwe madhairekitori akasarudzirwa nyaya iyi
InoteveraGaidhi rinotevera
Kuyeuka muDiffusion Models
Tekinoroji