Iri ibasa rakakosha uye rinodikanwa muchiratidzo chekombuta uye magiraidhi kugadzira mafirimu ekugadzira emhando yepamusoro.
Kunyangwe mamodheru akati wandei anoshanda eiyo portrait image toonification yakavakirwa pane ine simba StyleGAN yakatsanangurwa, aya maitiro anotarisana nemifananidzo ane zvipingamupinyi zvakajeka kana akashandiswa nemavhidhiyo, senge yakagadziriswa furemu saizi, chinodiwa chekumisikidza chiso, kusavapo kwezvinhu zvisiri zvechiso. , uye kusawirirana kwechinguvana.
Iyo shanduko yeVToonify chimiro inoshandiswa kubata iyo yakaoma inodzorwa yakakwirira-resolution mufananidzo wevhidhiyo kutamisa.
Isu tichaongorora yazvino ongororo paVToonify mune ino chinyorwa, kusanganisira mashandiro ayo, zvinokanganisa, uye zvimwe zvinhu.
Chii chinonzi Vtoonify?
VToonify framework inobvumira customizable yakakwirira-resolution mufananidzo wevhidhiyo kutapurirana.
VToonify inoshandisa StyleGAN yepakati-uye yepamusoro-resolution layers kugadzira emhando yepamusoro hunyanzvi mapikicha anoenderana neakawanda-mwero emukati maitiro anodzoserwa ne encoder kuchengetedza furemu ruzivo.
Mhedzisiro inokonzeresa dhizaini yekuvaka inotora zviso zvisina kurongeka mumabhaisikopo esaizi akasiyana-siyana sekuisa, zvichikonzera matunhu-akazara-ane mafambiro echokwadi mune zvinobuda.
Iyi dhizaini inoenderana neyazvino StyleGAN-yakavakirwa mufananidzo toonification modhi, ichivabvumidza kuti vawedzerwe kune vhidhiyo toonification, uye inogara nhaka inoyevedza hunhu senge inogadziriswa ruvara uye kusimba kugadzirisa.
ichi kudzidza inosuma zvirevo zviviri zveVToonify zvichibva paToonify uye DualStyleGAN yekuunganidza-yakavakirwa uye muenzaniso-yakavakirwa mufananidzo wevhidhiyo kutamisa, zvichiteerana.
Zvakawanda zviyedzo zvakawanikwa zvinoratidza kuti iyo VToonify yakarongwa inodarika nzira dziripo mukugadzira emhando yepamusoro, yenguva-inopindirana ehunyanzvi mabhaisikopo emifananidzo ane akasiyana masitaera paramita.
Vatsvakurudzi vanopa iyo Google Colab notebook, saka unogona kusvibisa maoko ako pairi.
Sei kushanda?
Kuzadzisa inogadzirika yakakwirira-resolution mufananidzo wedhidhiyo kutamisa, VToonify inosanganisa mabhenefiti eiyo mufananidzo wekushandura chimiro neiyo StyleGAN-yakavakirwa chimiro.
Kuti ive nehukuru hwakasiyana hwekuisa, shanduridzo yemifananidzo inoshandisa network inogoneka zvizere. Kudzidzira kubva mukutanga, kune rumwe rutivi, kunoita kuti kukwirira-kugadziriswa uye kudzorwa kwemaitiro kutapurirana kusagoneke.
Iyo isati yadzidziswa StyleGAN modhi inoshandiswa muStyleGAN-yakavakirwa chimiro chepamusoro-resolution uye inodzorwa dhizaini yekufambisa, kunyangwe ichiganhurwa kune yakagadziriswa saizi yemifananidzo uye kurasikirwa kweruzivo.
StyleGAN inogadziridzwa muiyo hybrid framework nekudzima chimiro chayo chakamisikidzwa-saizi yekupinza uye yakaderera-resolution maseru, zvichiita kuti inyatso convolutional encoder-jenareta architecture yakafanana neiyo yemufananidzo wekushandura chimiro.
Kuti uchengetedze ruzivo rwefuremu, dzidzisa encoder kuburitsa akawanda-scale emukati maficha eiyo furemu yekupinza sechimwe chekuwedzera chinodiwa kune jenareta. Vtoonify inogara nhaka yeStyleGAN modhi yemaitiro ekudzora kuchinjika nekuiisa mujenareta kuti ibvise zvese zvayo data uye modhi.
Kuganhurirwa kweStyleGAN & Yakarongwa Vtoonify
Mifananidzo yehunyanzvi yakajairika muhupenyu hwedu hwezuva nezuva pamwe nemumabhizinesi ekugadzira akadai sehunyanzvi, evanhu vezvenhau maavatars, mafirimu, kushambadzira kwevaraidzo, zvichingodaro.
Nekuvandudzwa kwe kudzidza zvakadzika tekinoroji, zvave kugoneka kugadzira emhando yepamusoro mifananidzo yemifananidzo kubva kuhupenyu chaihwo mapikicha uchishandisa otomatiki dhizaini yekufambisa.
Kune nzira dzakasiyana-siyana dzakabudirira dzakagadzirwa dzechifananidzo-based style kuendesa, mazhinji ayo anowanikwa zviri nyore kune vanotanga vashandisi nenzira yemafoni ekushandisa. Vhidhiyo zvinhu zvakakurumidza kuve musimboti wesocial media feeds mumakore akati wandei apfuura.
Kusimuka kwesocial media uye ephemeral mafirimu kwakawedzera kudiwa kwekuvandudza vhidhiyo editing, senge portrait vhidhiyo kutamisa maitiro, kugadzira akabudirira uye anonakidza mavhidhiyo.
Maitiro aripo ekuona mifananidzo ane zvakakomba zvakaipira kana akaiswa kumabhaisikopo, zvichidzikamisa kushanda kwawo mune otomatiki mufananidzo wevhidhiyo stylization.
StyleGAN yakajairika musana wekugadzira mufananidzo wemhando yekufambisa modhi nekuda kwekugona kwayo kugadzira zviso zvemhando yepamusoro zvine gadziriso manejimendi.
A StyleGAN-based system (inozivikanwawo sepicture toonification) inoisa chiso chaicho muStylGAN yakadzikama nzvimbo uye yobva yaisa kodhi yemasitaera inoguma kune imwe StyleGAN yakanatswa pane artic portrait dataset kugadzira shanduro ine stylized.
StyleGAN inogadzira mapikicha ane zviso zvakaenzanirana uye nehukuru hwakatarwa, izvo zvisingafarire zviso zvine simba mune chaiyo-yenyika tsoka. Kucheka kwechiso uye kurongeka muvhidhiyo dzimwe nguva kunokonzeresa kusarudzika kwechiso uye maitiro asina kunaka. Vatsvakurudzi vanodana nyaya iyi StyleGAN's 'fixed-crop restriction.'
Kune zviso zvisina kurongeka, StyleGAN3 yakatsanangurwa; zvisinei, inongotsigira seti yemufananidzo saizi.
Uyezve, ongororo ichangoburwa yakawana kuti encoding zviso zvisina kurongeka zvinonetsa pane zviso zvakatarisana. Kuisa kodhi kumeso kusiriyo kunokuvadza kutamisa chimiro, zvichikonzera nyaya dzakaita sekushandura chiziviso uye zvisipo mumafuremu akavakwa patsva nemaitiro.
Sezvakakurukurwa, dhizaini inoshanda yekutamisa vhidhiyo yekufambisa inofanirwa kubata zvinotevera:
- Kuchengetedza mafambiro echokwadi, iyo nzira inofanirwa kukwanisa kubata nezviso zvisina kurongeka uye akasiyana mavhidhiyo saizi. Vhidhiyo yakakura saizi, kana kona yakafara yekuona, inogona kutora rumwe ruzivo uchichengetedza chiso kubva kunze kwefuremu.
- Kukwikwidza neanhasi anowanzo kushandiswa HD gadget, yakakwirira-resolution vhidhiyo inodiwa.
- Flexible style control inofanirwa kupihwa kune vashandisi kuti vachinje uye vatore sarudzo yavo kana vachigadzira yechokwadi mushandisi yekudyidzana system.
Nechinangwa ichocho, vaongorori vanokurudzira VToonify, inoveli hybrid chimiro chevhidhiyo toonification. Kuti vakunde kusungirirwa kwechirimwa, vaongorori vanotanga kudzidza kuenzana kweshanduro muStyleGAN.
VToonify inosanganisa mabhenefiti eiyo StyleGAN-yakavakirwa dhizaini uye chimiro cheshanduro yemufananidzo kuti uwane inogadzirika yakakwirira-resolution mufananidzo wevhidhiyo kutamisa.
Izvi zvinotevera mipiro mikuru:
- Vatsvaguri vanoongorora kudzvanyirirwa kweStyleGAN uye vopa mhinduro yakavakirwa pashanduro yakaenzana.
- Vatsvagiri vanopa yakasarudzika yakasarudzika yakazara VToonify sisitimu yeinodzorwa yakakwira-resolution mufananidzo wevhidhiyo kutamisa inotsigira isina kurongeka zviso uye akasiyana mavhidhiyo saizi.
- Vatsvaguri vanovaka VToonify pamusana weToonify uye DualStyleGAN uye kudzoreredza musana maererano neese data uye modhi kugonesa kuunganidzwa-kwakavakirwa uye muenzaniso-yakavakirwa mufananidzo wevhidhiyo kutamisa.
Kuenzanisa Vtoonify nemamwe marudzi-e-the-art mhando
Toonify
Iyo inoshanda senheyo yekuunganidza-yakavakirwa dhizaini yekuchinjisa pane dzakaenderana zviso uchishandisa StyleGAN. Kuti utorezve makodhi emaitiro, vaongorori vanofanirwa kuwiriranisa zviso uye kudyara 256256 mafoto ePSP. Toonify inoshandiswa kugadzira mhedzisiro ine 1024 * 1024 makodhi makodhi.
Pakupedzisira, vanogadzirisa zvakare mhedzisiro muvhidhiyo kunzvimbo yayo yepakutanga. Nzvimbo isina kugadzirwa yaiswa kuita nhema.
DualStyleGAN
Iyo musana weiyo exemplar-yakavakirwa dhizaini yekufambisa yakavakirwa paStyleGAN. Ivo vanoshandisa iyo yakafanana data pre- uye post-processing matekiniki seToonify.
Pix2pixHD
Imhando yeshanduro yemufananidzo-kune-mufananidzo unowanzo shandiswa kupfupisa mamodheru akadzidzira kugadziridzwa kwepamusoro-soro. Inodzidziswa uchishandisa paired data.
Vatsvagiri vanoshandisa pix2pixHD seyekuwedzera mamepu emepu sezvo vachishandisa yakadhindwa mepu.
First Order Motion
FOM ndiyo yakajairika mufananidzo animation modhi. Yakadzidziswa pamifananidzo ye256256 uye inoita zvisina kunaka nemamwe saizi yemifananidzo. Nekuda kweizvozvo, vaongorori vanotanga kuyera mafiremu evhidhiyo kusvika pa256 * 256 kuti FOM iite animation vobva vagadzirisa saizi yemhando yavo yepakutanga.
Kuenzanisa kwakaringana, FOM inoshandisa yekutanga stylized furemu yemaitiro ayo seyereferensi chimiro chemufananidzo.
DaGAN
Iyo 3D face animation modhi. Ivo vanoshandisa iyo yakafanana kugadzirira data uye postprocessing nzira seFOM.
Advantages
- Inogona kushandiswa muhunyanzvi, social media avatar, mafirimu, kushambadzira kwevaraidzo, zvichingodaro.
- Vtoonify inogona zvakare kushandiswa mune metaverse.
Nokuremara
- Iyi nzira inobvisa zvese data uye modhi kubva kuStyleGAN-based backbones, zvichikonzera data uye muenzaniso kusarura.
- Izvo zvigadzirwa zvinokonzereswa zvakanyanya nehukuru mutsauko pakati peiyo stylized face dunhu uye zvimwe zvikamu.
- Iri zano harina kubudirira pakubata nezvinhu zviri munzvimbo yechiso.
mhedziso
Chekupedzisira, VToonify chimiro chemaitiro-anodzorwa akakwira-resolution vhidhiyo toonification.
Iyi dhizaini inogonesa kuita kukuru mukubata mavhidhiyo uye inogonesa kutonga kwakafara pamusoro pechimiro chechimiro, chimiro chemavara, uye chimiro dhigirii nekudzoreredza StyleGAN-based image toonification modhi maererano nezvose zvavo. synthetic data uye network zvimiro.
Leave a Reply