ControlNet — 12 mefuta ea ho conditioning ka sesebelisoa se le seng
Kopitsa setšoantšo sa ho ngola, khetha mofuta oa ho ngola, ngola lengolo-tsoibila. AI e boloka sebopeho sa ho ngola (lihlooho, pose, bophahamo, jj.)'me e etsa litaba tse ncha ka mokhoa ofe kapa ofe. E tšehelitsoe ke ControlNet-Union SDXL ProMax — Apache 2.0, e lumellanang ka ho feletseng le ho sebelisoa ha khoebo.
Bo_lemo
ControlNet e sebetsa joang
ControlNet e u lumella ho tsamaisa ho hlaha ha setšoantšo ka sebopeho sa setšoantšo sa ho ngola ho fapana le ho itšetleha ka ho ngola feela. Preprocessor e bala ho ngola ha hau'me e tlosa lets'oao leng la ho ngola - likhutlo tsa eona, lethathamo la eona la bophahamo, lesela la pose la motho, le joalo-joalo. Mofuta oa ho khomarela o koaloa ka lets'oao lena ha lets'oao le bolela setaele, liaparo, ho khanna, le sehlooho. Phetoho e boloka kopano e nepahetseng eo u e tlisitseng empa e shebahala e le ntho e ncha ka ho feletseng.
E 'ngoe ea li-tools tse ka sehloohong tsa ProMax ke ControlNet-Union SDXL ProMax (Apache 2.0) - e' ngoe ea li-model tse ka sehloohong tse utloisisang mefuta e 12 ea ho hlophisa e ka tlase, ka hona u ka li fetola ho tloha ho mokhethoa o le mong ntle le ho kenya marang-rang a fapaneng ka nako e ngoe. E na le ts'ebeliso e ntle ea khoebo: boloka, rekisa, kapa fetola seo u se etsang.
Li-conditioning tse 12
Canny
Ho bona likhutlo tse otlolohileng. E loketseng ho boloka li-contours tse otlolohileng le ho boloka li-linework tse hloekileng.
Bophara
Map'a ea bophahamo ba 3D. E boloka sebopeho sa sebaka - seo se ka tlaase le seo se ka holimo.
_Posa
OpenPose skeleton ea'mele. E koala boemo ba motho le libaka tsa litho.
Scribble
Li-dots tse nyenyane li hlahisoa ka ho tjheseletsa.
Sebopeho sa leqephe
Map'a ea sebaka se nang le li-code tsa'mala. Tsebisa sebaka se seng le se seng sa sehlooho ho sehlopha.
Bophara
Map'a ea lefatše-e tloaelehileng. E boloka ho nepahala ha lefatše la 3D le ho bump.
Lineart
Li-invertebrate li na le li-anatomy tse fapaneng, li-anatomy tsa li-invertebrate, li-anatomy tsa likokoana-hloko le li-anatomy tsa likokoana-hloko.
Bophara bo bobebe
Ho bona bophahamo bo bobebe bo latela li-shapes tse bobebe ho feta Canny.
MLSD
Li-segments tsa li-line tse otlolohileng. Li entsoe bakeng sa meralo ea meralo, li-inner, le lifoto tsa lihlahisoa.
Tile
Ho na le li-conditioning tse khethehileng bakeng sa ho theola le ho etsa li-texture tse se nang sekoli.
Bo_ra
Ho hlophisoa ha mask ho etsa hore ho hlahe hape karolo ea setšoantšo feela.
Ho bopa ka ho hlaka
E tlisa canvas kapa e etsa li-regions tse ncha ka ho boloka sebopeho se ka morao.
Litsela tse tharo
- Kopitsa setšoantšo sa ho ngola — setšoantšo, sepheo, skrine, leha e le sefe se nang le sebopeho seo u batlang ho se boloka.
- Khetha mofuta oa ho hlophisa o lumellanang le seo u se batlang (pose bakeng sa setšoantšo, bophahamo bakeng sa sehlooho, canny kapa lineart bakeng sa li-outlines tse hloekileng).
- Ngola lengolo le hlalosang ho shebahala ha u batlang le ho e hlahisa. Ntšetsa pele matla a taolo ho latela ho ngolisoa ka thata, e theole bakeng sa ho ba le bokhoni bo eketsehileng ba ho etsa.
ControlNet — 12 mefuta ea ho conditioning ka sesebelisoa se le seng - FAQ
Setsebi se le seng se bonts'ang mefuta eohle ea 12 ea ho hlophisa ho tloha ho mofuta oa ControlNet-Union SDXL ProMax - canny, pose, depth, scribble, lineart, anime-lineart, MLSD, HED, soft-edge, normal, segmentation, le tile. Khetha mofuta oa ho hlophisa, o tlose setšoantšo sa ho ngola, o ngola lengolo-tsoibila,'me SDXL e etsa setšoantšo se secha se latelang sebopeho sa ho ngola.
img2img e tšoantša ka ho hlaka kahare ho tlhahiso ka ho toba — li-colors, li-edge, le ho kopanya li-geometry ka ho lekola. ControlNet e tlosa li-colors'me e boloka feela sesupo sa ho hlophisa se khethileng (li-line, pose skeleton, map ea bophahamo, jj.). Seo se u lumella ho fetola litaba ka ho feletseng ha u ntse u boloka kompositi e le thata. Ho laola ho hlophisa ho matla ho feta img2img.
Canny / lineart bakeng sa ho kenya linework e hloekileng. Anime-lineart bakeng sa ho kenya line-style anime. Scribble / soft-edge / HED bakeng sa li-sketch tse thata le doodles. Pose ho kopitsa sebaka sa'mele ho tloha ho foto. Bophahamo ho boloka li-geometry tsa sehlooho / 3D layout. MLSD ho boloka li-line tse teteaneng (li-architecture / interior). Bophara ho boloka ho otloloha ha lefatše le boholo. Segmentation ho boloka li-regions. Tile ho ntlafatsa kapa ho eketsa liphetoho tsa setšoantšo se teng.
ControlNet-Union SDXL ProMax (xinsir, Apache 2.0) e paka marang-rang a 12 a ho conditioning ho boima ba 2.5 GB. Li-deployments tse fetileng li arolelana boima ba 2.5 GB ka mofuta - ho fetola lipakeng tsa canny le pose ho bolela ho qala ka thata. Mofuta oa k'hamphani o arolelana nako le ho lula o le mofuthu, ka hona mofuta o mong le o mong oa ho conditioning ke sub-second kamora ho ngola ka pele.
E. ~ 1,200 tokens ka render (1,000 base SDXL + 20% ControlNet conditioning surcharge). Basebelisi ba nang le ak'haonte ba fumana 30,000 tokens tsa mahala ka letsatsi - ka 25 conditioned renders ka letsatsi ntle le tefo. Anonymous: 2,500 tokens / letsatsi (~ 2 renders).
Ee — slider ea matla a taolo (ho tloaelehile 0.7) e hlalosa hore na tlhahiso e latela bopaki ba hau joang. 1.0 = bopaki (e shebahala e le ho tlisa bopaki ba hau). 0.4 = bongata (prompt e na le boikarabello bo boholo). E fokotsa bakeng sa ho fetola ka ho etsa, e phahamisa ha ho hlokahala ho etsa.
512 × 512 default. SDXL standard ratios - 768 × 1024 portrait, 1024 × 768 landscape, 1024 × 1024 square - tsohle li sebetsa. Lihlahisoa tse kholo li sebelisa VRAM le li-token tse ngata; H200 e tšehetsa ho fihlela ho 1024 × 1024 ka mokhoa o sireletsehileng.
Ee. Ho thehoa ha ControlNet ho theko e boima ho ~1,000 tokens ka setšoantšo, e tšoanang le ho thehoa ha SDXL e tloaelehileng. Basebelisi ba sa tsejoeng ba fumana 6,000 tokens / letsatsi; akhaonto e mahala e fana ka 30,000 / letsatsi. Ha ho karete ea mokitlane, ha ho na letšoao la metsi. Resolution e phahameng kapa liphetoho tse ngata li theko e boima ho feta ka ho lekana ho tsoa ho pool.
ControlNet-Union SDXL ProMax e lokollotsoe ka tlase ho Apache 2.0 — e lumellanang ka ho feletseng, ho kenyeletsoa le ho sebelisoa ka khoebo. Setsi sa SDXL ke OpenRAIL++. Bothata bo lumella ho sebelisoa ka khoebo; litšoantšo tseo u li etsang ke tsa hau ho li sebelisa ka khoebo ntle le litokelo.
Mofuta o ts'oanang, boleng bo ts'oanang, li-signals tsa ho conditioning tse ts'oanang. ComfyUI le A1111 li hloka GPU ea naha le 12+ GB VRAM le ho kenya. Re e tsamaisa ka merero e tšehelitsoeng ka pool e ntle ea mahala - ha ho na ho kenya, ha ho hlokahale GPU.
Ho ngola ka pele ho kenya boima ba Union (~ 2.5 GB) ho GPU cache le ho pholile SDXL pipeline. E-ea ka metsotsoana ea 30-60 ka kopo ea pele ka mor'a ho kenya kapa ho tlohela LRU. Li-calls tse latelang ka tlase ho ho lokisa ho khutlela ka metsotsoana ea 4-7.
E-ea - POST multipart ho /v1/image/generate/ le model=sdxl (kapa model=controlnet-union-sdxl-promax), ho botsa, control_image (faele), control_type=<e mong oa: canny, pose, bophahamo, scribble, lineart, anime-lineart, mlsd, hed, soft-edge, bo tloaelehileng, ho aroloa, tile>, control_strength e sa tšoaneng (0.1-1.5). Bearer auth, 10K tokens / khoeli e mahala. /api/ e na le litlhaloso tsa curl.