InoteveraGaidhi rinotevera
Adversarial Diffusion Distillation uye Turbo Models
Tekinoroji
Nhungamiro yehunyanzvi
Diffusion mifananidzo modhi inowanzo gadzira mifananidzo mitsva, asi tsvakiridzo yakaratidza kuti dzimwe nguva vanogona kuburitsa chaiwo mifananidzo yekudzidziswa, kunyanya mifananidzo yaionekwa kakawanda mudhata rekudzidziswa.
Teerera kugwaro iriAI-yakarondedzerwa · ~3 min
Uku kunonzi kudzidzira nemusoro. Izvo zvine basa kune copyright, nekuti padhuze-kopi yebasa rakachengetedzwa inogona kubuda mumuenzaniso, uye yekuvanzika, nekuti foto yemunhu chaiye kana yekurapa scan inogona kudzokororwa.
Iyo yekuparadzira modhi inodzidza kudzosera kumashure maitiro eruzha: panguva yekudzidziswa inoona mifananidzo ine ruzha yakawedzerwa uye inodzidza kufanotaura nekubvisa ruzha irworwo. Kana iyo yakazara nemazvo, inodzidza huwandu hwemifananidzo muhuwandu uye inogadzira misanganiswa mitsva. Asi neural network inogonawo kukwana mienzaniso yega, uye kana mufananidzo mumwe ukaonekwa mazana kana zviuru zvenguva mudhata, nzira yakachipa yekudzikisa kurasikirwa kwekudzidziswa pairi ndeyekuichengeta. Humbowo hwakakosha hunobva kuCarlini uye vaanoshanda navo '2023 bepa 'Kubvisa Kudzidzisa Dhata kubva kuDiffusion Models'. Vakagadzira mifananidzo mizhinji yezvinyorwa zvemifananidzo yakadhindwa yekudzidziswa uye yakamisikidzwa mireza yaive pedyo zvakanyanya nemufananidzo wekudzidziswa. Vakatora anopfuura zana pedyo-akafanana makopi kubva kuStable Diffusion, kusanganisira mapikicha evanhu vanozivikanwa, uye vakawana mareti epamusoro ekurangarira muGoogle's Imagen. Zvakaparadzana, Somepalli nevamwe vaaishanda navo (2022-2023) vakaratidza kuti Stable Diffusion zvinobuda zvaiwanzo kukopa zvikamu kana marongero emifananidzo yekudzidzisa, uye yakabatana kukopa kune yakadhindwa mifananidzo uye yakanyatso tsanangurwa zvinyorwa. Pfungwa nhatu dzisiridzo dzakakodzera kugadziriswa. Chekutanga, kubata nemusoro haisiriyo maitiro emuenzaniso: makopi akabudiswa haawanzo hukama nezvizvarwa zvese. Chechipiri, kushomeka hakurevi kusakuvadza, nekuti mushandisi akatsunga anogona kunanga iyo mifananidzo ingangove inoyeukwa nemusoro. Chechitatu, kubata nemusoro hakusi kungotevedzera kwepixel chaiyo; mamodheru anogona kuburitsa mavara, marogo, watermark uye akasiyana maumbirwo. Humbowo uhu hwakabuda mumakakava emutemo, kusanganisira nyaya dzemumatare dzakaita seGetty Images v. Stability AI uye Andersen v. Stability AI, kunyangwe mabatiro anoita matare edzidziso yemuenzaniso uye zvinobuda zvichiri kuitwa uye zvichisiyana nenyika. Zvekuvanzika, kunetseka kwakanyanya mudiki-tuning datasets, senge yekurapa kana yemunhu mafoto, uko imwe neimwe muenzaniso ine simba rakawanda.
Vatsvaguri vanomhanyisa modhi ine tsananguro yemufananidzo wakadzokororwa wekudzidziswa uye vanodzosa yakabuda inenge pixel-yakafanana nemufananidzo wekutanga.
Mugadziri anokumbira chimiro chependi ine mukurumbira uye anogashira mufananidzo unokopa maumbirwo uye ruzivo rwechimwe chinhu chakadhirowa kwete chidimbu chitsva.
Iyo jenareta yemifananidzo yakadzidziswa pawebhusaiti nguva nenguva inoshandura vhezheni yakamonyaniswa yewatermark yestock agency, ichiratidza iyo yakatorwa mapatani kubva kune inodzokororwa kunobva mifananidzo.
Chikwata chechipatara chinonyatso gadzirisa modhi yekuparadzira pane diki seti yemurwere inoongorora bvunzo kuti modhi yacho inogona kuburitsa pedyo-makopi ema scan chaiwo kusati kwabvumidzwa kugovana.
Kudzora uye kuburitsa kufanana kwekutarisa kuri kuita kwakajairwa tsika yevagadziri vane mutoro, uye tsvakiridzo inoenderera pakuona zvinofungwa nemusoro panguva yekufungidzira. Kuvanzika kwakasiyana kunogona kuve kunoshanda kumatunhu ane hunyoro akadai sekufungidzira kwekurapa, kunyangwe zvemhando yekutengesa zvichiripo. Mufananidzo wepamutemo hauna kugadzikana: matare ari munyika dzakasiyana ari kufunga kuti kudzidziswa uye zvinobuda zvichityora kodzero here, uye mhedzisiro inogona kusiyana. Vagadziri vanogona kutarisana nekumanikidzwa kwakawanda kunyora madhata ekudzidziswa uye kuratidza maitiro avanoita uye nekudzikamisa kudzoreredza.
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
Diffusion mifananidzo modhi inowanzo gadzira mifananidzo mitsva, asi tsvakiridzo yakaratidza kuti dzimwe nguva vanogona kuburitsa chaiwo mifananidzo yekudzidziswa, kunyanya mifananidzo yaionekwa kakawanda mudhata rekudzidziswa. Uku kunonzi kudzidzira nemusoro. Izvo zvine basa kune copyright, nekuti padhuze-kopi yebasa rakachengetedzwa inogona kubuda mumuenzaniso, uye yekuvanzika, nekuti foto yemunhu chaiye kana yekurapa scan inogona kudzokororwa.
Kuchengeta nemusoro zvinoreva zvinobuda zviri pedyo-makopi emamwe mifananidzo yekudzidziswa, kwete mifananidzo mitsva.
Kudzokororwa kuratidzwa kumufananidzo mumwechete kunoita kuti kufanidza iwo chaiwo muenzaniso nzira inoshanda yekudzikisa kurasikirwa.
Vakatora anopfuura zana pedyo-makopi kubva kuStable Diffusion, kusanganisira mapikicha evanhu vanozivikanwa, uye vakawana mitengo yepamusoro muImagen.
Mufananidzo wakabatiswa nemusoro unoonekwa nguva dzese zvisinei nemhodzi, saka mapoka akasimba ezvinenge zvakafanana zvinoburitsa mureza.
Makopi akabudiswa haawanzo hukama kune zvese zvinobuda, asi vanorwisa vanogona kutarisa pamifananidzo ingangove yakabatwa nemusoro.
Ramba uchidzidza
Mamwe madhairekitori akasarudzirwa nyaya iyi
InoteveraGaidhi rinotevera
Adversarial Diffusion Distillation uye Turbo Models
Tekinoroji