ControlNet — 12 izinhlobo conditioning kuthuluzi elinye
Layisha phezulu isithombe esibhekiswe kuso, khetha uhlobo lokulungiselela, bhala isiphakamiso. I-AI igcina isakhiwo sesihloko sakho (ama-line, i-pose, ububanzi, njll.) futhi inikeza okuqukethwe okusha nganoma iyiphi indlela. Ixhaswe yi-ControlNet-Union SDXL ProMax — Apache 2.0, ngokugcwele isebenziseka ngokuhweba.
Imiphumela
Indlela ControlNet esebenza ngayo
ControlNet ikuvumela ukuthi uqondise ukukhiqizwa kwesithombe ngesakhiwo sesithombe esibhekiswe kuso endaweni yokuya ku-text prompt kuphela. Umphathi wokuqala ufunda isiqondiso sakho futhi ukhipha umlayezo wokusebenza ofanayo — ama-edge ayo, i-depth map yayo, i-pose skeleton yomntu, njll. Imodeli yokusakaza ivala kulowo mlayezo ngenkathi i-prompt ikhetha uhlobo, amamibala, ukukhangisa, nesihloko. Imiphumela igcina ukwakheka okulungile okunikeziwe kodwa ibukeka njengento entsha ngokuphelele.
Lezi zisetshenziswa zixhaswa yi ControlNet-Union SDXL ProMax (Apache 2.0) — imodeli eyodwa eqonda zonke izinhlobo ezingu-12 zokuphatha ezingezansi, ngakho ungashintsha phakathi kwazo kusuka ku-picker eyodwa ngaphandle kokufaka inethiwekhi ehlukile ngaso sonke isikhathi. Kulungile ngokuphelele ukusebenzisana kokuthengiswayo: gcina, thengisa, noma guqula noma yini oyenzayo.
Izinhlamvu ezingu-12 zokulungiselela
I-Canny
Ukuqapha okuqinile kwengxenye. Okungcono kakhulu ukulondoloza izibopho eziqinile kanye nomsebenzi we-line ohlanzekile.
Ububanzi
I-3D depth map. Igcina isitayela sendawo — yini ephakathi nokuthi yini ephakathi.
Iposi
I-OpenPose body skeleton. Ivala indawo yobude nendawo yengxenye yesithombe.
Ukubhala
Izandla ezizobe zidwetshwe ziguqulwe zibe umdwebi oqediwe.
Ukuhlukaniswa
I-maphu yesifunda eqokiwe ngombala. Linganisa indawo ngayinye yendawo yokubukela kwisigaba.
Okujwayelekile
I-surface-normal map. Igcina ukuqondiswa kwe-3D kwe-surface kanye ne-bumps.
Uhlobo lwesikhashana
Ukukhishwa komugqa oncane — ofanele ukuphrinta, i-manga, nesithombe.
Ingxenye ekhanyayo
Ukuqapha umkhawulo oqinile olandela ifomu ngokukhululekile kunaleyo yeCanny.
MLSD
Iziqephu ezihamba phambili. Ziyenziwe ngezifiso, izimo zangaphakathi, kanye nezithombe zemikhiqizo.
I-tile
Ukugcinwa kwemingwane-ukugcinwa kwe-conditioning ukuphakama nokusebenzela i-texture engangeni lutho.
Umbala
Ukulungiselela okuzizwayo kwe-mask ukuvuselela kabusha kuphela ingxenye yesithombe.
Phinda udwebe / udwebe ngaphandle
Yandisa i-canvas noma uphinde udwebe izindawo ngenkathi uhlonipha ukwakhiwa okubhekene nalo.
Izinhlamvu ezintathu
- Layisha phezulu isithombe esibhekiswe kuso — isithombe, isithonjana, isikrini, noma yini equkethe isakhiwo ofuna ukuyigcina.
- Khetha uhlobo lokulungiselela olufana nolufunayo (iposi lesithombe, ububanzi bendawo, i-canny noma i-lineart ye-outline ehlanzekile).
- Bhala umlayezo ochaza ukubonakala ofuna ukukwenza. Nciphisa amandla okuphatha ukuze ulandele isixhumanisi ngokuqinile, cindezela phansi ukuze ube nenkululeko eminingi yokudala.
ControlNet — 12 izinhlobo conditioning kuthuluzi elinye - FAQ
Ithuluzi elilodwa eliveza zonke izinhlobo ezingu-12 zokulungiselela kusuka ku-ControlNet-Union SDXL ProMax model - canny, pose, depth, scribble, lineart, anime-lineart, MLSD, HED, soft-edge, normal, segmentation, and tile. Khetha uhlobo lokulungiselela, khipha isithombe esibhekiswe kuso, bhala umlayezo, futhi i-SDXL inikeza isithombe esitsha esilandela ukwakhiwa kwesithombe sakho esibhekiswe kuso.
i-img2img iphinde ipeyinte phezu kwe-input ngokuqondile — amamibala, ama-edge, kanye ne-geometry mix nge-prompt. I-ControlNet isusa amamibala futhi igcina kuphela umlayezo wesakhiwo okhethiwe (ama-line, pose skeleton, ububanzi bama-map, njll). Lokhu kukuvumela ukuthi ushintshe ngokuqinile okuqukethwe ngenkathi ugcina ukumiswa kwe-rock solid. Ukulawula okuqinile kwesakhiwo kune-img2img.
Canny / lineart for clean linework input. Anime-lineart for anime-style line input. Scribble / soft-edge / HED for rough sketches and doodles. Pose to copy a body position from a photo. Depth to preserve scene geometry / 3D layout. MLSD to preserve straight lines (architecture / interiors). Normal to preserve surface orientation and volume. Segmentation to preserve regions. Tile to refine or upscale variations of an existing image.
ControlNet-Union SDXL ProMax (xinsir, Apache 2.0) ifaka zonke izixhumanisi ezingu-12 ezilungele ukufakwa ku-2.5 GB kuphela. Izisebenziso ezidala zikhuphula ubunzima obuhlukile ~2.5 GB kuhlobo ngalunye — ukushintsha phakathi kwe-canny ne-pose kufakazela ukuqala okubandayo. Imodeli ye-union ifaka futhi ihlala ipholile, ngakho-ke uhlobo ngalunye lokulungiselela luyingxenye yesibili ngemuva kocingo lokuqala.
Yebo. ~1,200 i-token ngayinye ye-render (1,000 base SDXL + 20% ControlNet conditioning surcharge). Abasebenzisi ababhalisiwe bathola i-30,000 ye-token yamahhala ngosuku — mayelana ne-25 conditioned renders ngosuku ngaphandle kwezindleko. Anonimous: 2,500 tokens / ngosuku (~2 renders).
Yebo — i-Control strength slider (iphutha 0.7) ichaza ukuthi i-output ilandela kanjani ngokuqinile ukubhekisa kwakho. 1.0 = iqinile (i-output ibukeka njenge-re-render ye-reference yakho). 0.4 = ikhululekile (i-prompt inenkululeko engaphezulu). Nciphisa ukuhluka okusha, phakamisa uma kubaluleke ukuthembela.
512×512 iphutha. SDXL isilinganiso esijwayelekile - 768×1024 isithombe, 1024×768 landscape, 1024×1024 square - zonke umsebenzi. Imiphumela enkulu usebenzise VRAM futhi tokens; H200 isekela kuze 1024×1024 ngokunethezeka.
Yebo. Uhlobo ControlNet lubiza ~1,000 amathokheni ngesithombe, ofanayo nehlobo SDXL elijwayelekile. Abasebenzisi abangaziwayo bathola amathokheni angama-6,000/izinsuku; i-akhawunti emahhala inikeza ama-30,000/izinsuku. Akukho ikhadi le-credit, akukho uphawu lwamanzi. Ukuxazululwa okuphezulu noma ukuhlukahluka okuningi kubiza ngokulingana kakhulu kusuka epulazini.
I-ControlNet-Union SDXL ProMax ikhishwa ngaphansi kwe-Apache 2.0 — ivunyelwe ngokuphelele, kufaka phakathi ukusetshenziswa kwezokuhweba. Isisekelo se-SDXL yi-OpenRAIL++. Zonke zivumela ukusetshenziswa kwezokuhweba; izithombe zakho ezikhiqizwe zikhona ukuze zisetshenziswa kwezokuhweba ngaphandle kwe-royalties.
Imodeli efanayo, umgangatho ofanayo, ama-signals afanayo okulungiselela. I-ComfyUI ne-A1111 zidinga i-GPU yasendaweni ene-12+ GB VRAM kanye nokumiswa. Siyiqhuba kusakhiwo esihlukaniswe nge-pool ekhululekile egcwele — akukho kufaka, akukho GPU edingekayo.
Ucingo lokuqala lulanda isisindo se-Union (~2.5 GB) kwi-GPU cache futhi lupholile i-SDXL pipeline. Lindela amasekondi angama-30-60 kusicelo esingukuqala ngemuva kokufaka noma ukuphuma kwe-LRU. Ucingo olulandelayo ngezansi kokufaka okujwayelekile lubuyela kumasekondi angama-4-7.
Yebo — POST ingxenye eminingi ku /v1/image/generate/ ngemodeli=sdxl (noma imodeli=controlnet-union-sdxl-promax), umlayezo, ukulawula_isithombe (ifayela), ukulawula_uhlobo=<enye ye: canny, pose, ububanzi, ukubhala, lineart, anime-lineart, mlsd, hed, soft-edge, ojwayelekile, ukuhlukaniswa, i-tile>, ukulawula_ubukhulu okukhethayo (0.1-1.5). Ukuqinisekisa umthwali, 10K ama-token amahhala/inyanga. /api/ unezibonelo zokujika.