ControlNet — 12 iintlobo zokulungiselela kwisixhobo esinye

Layisha phezulu umfanekiso wobhekiso, khetha uhlobo lolawulo, bhala umyalezo. I-AI igcina ukwakhiwa kobhekiso lwakho (imigca, ubeko, ubunzulu, njl.njl.) kwaye ibonisa imixholo entsha nangayiphi na indlela. Ixhaswe yi ControlNet- Union SDXL ProMax — Apache 2.0, ilungele ukusetyenziswa kwentengiso.

Canny / lineart yomsebenzi womgca ococekileyo. Pose yendawo yomzimba. Ubunzulu bobeko lwe 3D. Ushicilelo / umda ococekileyo we doodles. MLSD yoyilo lwe sakhiwo. Eqhelekileyo / uhlukaniso / iithayile zokuhamba komsebenzi okuphezulu.
Ulawulo luphuma kulo - imibala isuswa, kuphela iimpawu zokwakha (ngohlobo olukhethiweyo) zigcinwa.
I-Looser 0.7 Eqinileyo
~1,200 tokens (SDXL × 1.2 ControlNet)
I-Result

Indlela i ControlNet isebenza ngayo

ControlNet ikuvumela ukuba uqhube ukwenziwa komfanekiso ngokwakhiwa komfanekiso obhekisa kuwo endaweni yokungathembeki kumyalezo wombhalo kuphela. Umqhubi ophambili ufunda umyalezo wakho obhekisa kuwo kwaye ukhupha umqondiso wokungalingani ofanayo - imiphetho yayo, imaphu yobunzulu bayo, i-pose ejikelezayo yomntu, njalo njalo. Imodeli yosasazo iyatshixa kulo mqondiso ngelixa umyalezo obhekisa kuyo ugqiba uhlobo, imibala, ukukhanya, kunye nomxholo. Impumelelo igcina ukwenziwa ngokuchanekileyo okukhokelela kulo kodwa ibonakala njengento epheleleyo entsha.

Eli sixhobo sixhaswa yi ControlNet-Union SDXL ProMax (Apache 2.0) - imodeli enye eqonda zonke iintlobo ezili-12 zokumisela ezantsi, ngoko utshintshela phakathi kwazo ukusuka kumkhethi omnye ngaphandle kokufaka umxholo ohlukileyo ixesha ngalinye. Isebenza ngokufanelekileyo kwintengiso: gcina, thengisa, okanye guqula kancinane nantoni na oyenzayo.

Iinkqubo ezili-12 zokulungiselela

Canny
Ubhaqo lomda ocacileyo. Olungileyo ukugcina imizobo eqaqambileyo nomsebenzi womgca ococekileyo.
Ubunzulu
I-3D ye-deepness map. Igcina ubeko lwendawo - yintoni ekufuphi nentoni ekufuphi.
Iposi
I-OpenPose body skeleton. Itshixa indawo yobume besimo kunye nendawo yelungu.
I-Scribble
Ii-doodles ezizotyiweyo ngesandla ezilahlekileyo ziguqulwa zibe ngumzobo ogqityiweyo.
Ukwahlula-hlula
Iindawo ezikhowudiweyo ngemibala. Sebenzisa indawo nganye yendawo yokubonakala kuhlobo.
Eqhelekileyo
Imaphu yomhlaba-oqhelekileyo. Igcina ulungelelaniso oluhle lwe-3D lomhlaba kunye neengqungquthela.
Imifanekiso engumgca
Ukukhupha umgca omncinci - olungele ukushicilela, imanga, kunye nomfanekiso.
Umda oNceda
Ubophelelo lomda ococekileyo olandela iimpawu ngokulula kuneCanny.
MLSD
Iindawo ezithe nkqo. Ziye zakhiwa uyilo lwesinyithi, iinkalo zangaphakathi, kunye nemifanekiso yemveliso.
Iithayile
Ukugcina inkcukacha-ukulungelelanisa ukunyuka kwenqanaba kunye nomsebenzi ongenazingcingo wombala.
Umbala
Isigqubuthelo-esizimeleyo sokungalingani ukubuyisela kwakhona kuphela inxalenye yomfanekiso.
Umbala
Yandisa i-canvas okanye uphinde upeyinte imimandla ngelixa uhlonipha ukwakhiwa okujikelezayo.

Iinkqubo ezintathu

  1. Layisha phezulu umfanekiso obhekisa kuwo — umfanekiso, umzobo, iscreenshot, nantoni na enesakhiwo ofuna ukusigcina.
  2. Khetha uhlobo lokumisela olufana nolufunayo (ubeko lomzobo, ubunzulu bendawo, ubuhle okanye ubude bemigca ecocekileyo).
  3. Bhala umyalezo ochaza ukubonakala ofuna ukukwenza. Nciphisa ulawulo lomandla ukulandelela ubhekiso ngokuqinileyo, nyusa ukunciphisa ukufikelela kwinkululeko eninzi yobugcisa.

ControlNet — 12 iintlobo zokulungiselela kwisixhobo esinye - FAQ

Isixhobo esifanayo esibonisa zonke iintlobo ezili-12 zokumisela ukusuka kwimodeli ye ControlNet-Union SDXL ProMax - canny, pose, depth, scribble, lineart, anime-lineart, MLSD, HED, soft-edge, normal, segmentation, kunye ne tile. Khetha uhlobo lokumisela, shiya umfanekiso wobhekiso, bhala umyalezo, kwaye i-SDXL ibonisa umfanekiso omtsha olandela ukwakhiwa kobhekiso lakho.

img2img ipeyinta kwakhona ngaphezulu kwengeniso ngqo - imibala, imiphetho, NObunzulu bomhlaba odityanisiweyo ne prompt. I ControlNet isusa imibala kwaye igcina kuphela iimpawu zokwakha ezikhethiweyo (imigca, i-pose skeleton, imaphu yobunzulu, njl.njl.). Oku kukuvumela ukuba utshintshe ngokugqithisileyo imixholo ngelixa ugcina ubeko oluqinileyo. Ulawulo olunamandla ngakumbi lwesakhiwo kune img2img.

Canny / lineart yengeniso ecocekileyo yemisebenzi yemigca. Anime-lineart yengeniso yemigca yohlobo lwe-anime. I-Scribble / soft-edge / HED ye-sketchs eziqinileyo kunye ne-doodles. Iposi yokukopa indawo yomzimba ukusuka kwifoto. Ubunzulu ukugcina i-geometry yendawo / ubeko lwe-3D. MLSD ukugcina imigca ethe nkqo (ubugcisa / ngaphakathi). Eqhelekileyo ukugcina ubeko lwesiphelo kunye nobukhulu. Ukwahlulahlula ukugcina imimandla. Iithayile ukugqiba okanye ukunyusa utshintsho lomfanekiso okhoyo.

ControlNet- Union SDXL ProMax (xinsir, Apache 2. 0) ipakisha onke amajelo okulungiselela 12 kwi 2. 5 GB bunzima. Unikezelo oludala lwakhuphela phantsi ubunzima obuhlukileyo ~2. 5 GB kuhlobo ngalunye - utshintshiselwano phakathi kwe canny ne pose kuthetha ukuqala okubandayo. Imodeli ye union ifaka kabini kwaye ihlala ishushu, ngoko ke udidi ngalunye lokulungiselela luyinxalenye yesibini emva koqhagamshelwano lokuqala.

Ewe. ~1,200 iitokeni nganye yokuvelisa (1,000 isiseko SDXL + 20% ControlNet iindleko zolawulo). Abasebenzisi ababhalisiweyo bafumana 30,000 iitokeni ezikhululekileyo ngosuku — malunga ne-25 eyenziweyo yokubonisa ngosuku ngaphandle kwexabiso. Anonimous: 2,500 iitokeni/ ngosuku (~2 ukubonisa).

Ewe - i Control strength slider (okwendalo 0.7) imisela ukuba imveliso ilandela njani ngokungqongqo ubhekiso lwakho. 1.0 = ingqongqo (imveliso ibonakala njenge-render ye-referensi yakho). 0.4 = ikhululekile (i-prompt inenkululeko eninzi). Nciphisa ukutshintsha okunobuchule, nyusa xa ukuthembeka kubaluleke.

512×512 okumiselweyo. Ii-SDXL eziqhelekileyo — 768×1024 umzobo, 1024×768 umthunzi, 1024×1024 isikwere — zonke zisebenza. Iimveliso ezinkulu zisebenzisa i-VRAM kunye nee-token ezininzi; i-H200 ixhasa ukuya kwi-1024×1024 ngokukhululekileyo.

Ewe. Uhlobo ControlNet lubiza ~1,000 ye-token nganye yomfanekiso, ifana nohlobo oluqhelekileyo SDXL. Abasebenzisi abangaziwayo bafumana i-6,000 ye-token/imini; i-akhawunti ekhululekileyo inika i-30,000/imini. Akukho khadi letyala, akukho phawu lwamanzi. Ubungakanani obuphezulu okanye utshintsho oluninzi lubiza ngokuhambelana ngakumbi kwi-pool.

I-ControlNet-Union SDXL ProMax ikhutshwa phantsi kwe-Apache 2.0 — ivumela ngokugcwele, kubandakanya ukusetyenziswa korhwebo. Isiseko se-SDXL si OpenRAIL++. Zonke zivumela ukusetyenziswa korhwebo; imifanekiso yakho eyenziweyo iyinikwe wena ukuyisebenzisa korhwebo ngaphandle kweerhafu.

Imodeli efanayo, umgangatho ofanayo, iimpawu ezifanayo zokulungiselela. I-ComfyUI ne-A1111 zifuna i-GPU yangaphakathi ene-12+ GB ye-VRAM kunye nokumisela. Siyiqhuba kwinkqubo yokwakha enikezelweyo ene-pool ekhululekileyo ebanzi — akukho fakelo, akukho GPU ifunekayo.

Ubizo lokuqala lukhuphela ezantsi ubunzima be-Union (~2.5 GB) kwi-GPU cache kwaye lufudumeza i-SDXL pipeline. Lindela imizuzwana engama-30-60 kwisicelo esiphambili emva kokufaka okanye ukuphuma kwe-LRU. Ubizo olulandelayo phantsi kofakelo oluqhelekileyo lubuyela kwimizuzu engama-4-7.

Ewe — INKXELO yenxalenye eninzi ku /v1/image/generate/ ngemodeli=sdxl (okanye imodeli=controlnet-union-sdxl-promax), umyalezo okhawulezayo, ulawulo_lomfanekiso (ifayili), ulawulo_lohlobo=<enye ye: canny, pose, ubunzulu, ukubhala, lineart, anime-lineart, mlsd, hed, umda othambileyo, oqhelekileyo, uhlukaniso, iithayile>, ulawulo_lomandla olukhethwayo (0.1-1.5). Ugunyaziso lomthwalo, 10K ii-token ezikhululekileyo/inyanga. /api/ inemizekelo yokujika.

Uthanda i-Free.ai? Nceda utshele abahlobo bakho!

Iphepha elilandelayo