ControlNet — 12 noocyada conditioning ah oo ah qalab kaliya
Soo deji sawirka soo jeedinta, dooro nooc conditioning, qor fariin. AI haysta dhismaha soo jeedintaada (liiska, pose, qoto dheer, iwm) iyo content cusub ee qaab kasta.
Natiijo
Sida ControlNet u shaqeyso
ControlNet waxaad ku siin kartaa abuurista sawirka qaabka sawirka tilmaamaha, ka dibna ku tiirsanaanta qoraalka qoraalka oo keliya. Preprocessor ayaa akhriya tilmaamahaaga oo soo saara calaamad hal-abuur ah - jidadkeeda, khariidada qoto dheer, qaabka jirka ee qofka, iyo wixii la mid ah. Muuqaalka faafinta ayaa markaa loo xiray calaamaddaas, halka tilmaamaha uu go'aansado qaabka, midabada, iftiinka, iyo mawduuca. Natiijadu waxay haystaan waxyaabaha saxda ah ee aad ku siinayso laakiin waxay u egtahay wax aad u cusub.
Qalabkani waxaa taageera ControlNet-Union SDXL ProMax (Apache 2.0) — mid kaliya oo u fahma dhammaan 12 noocyada conditioning hoos, sidaas darteed waxaad u beddeli kartaa dhexdooda ka mid ah mid ka mid ah dooran oo aan la soo dejin shabakadda kala duwan mar kasta. Waa si buuxda u isticmaalka ganacsi saaxiibtinimo: haysta, iibsada, ama bedeli waxa aad abuurto.
12 noocyada conditioning
Canny
Ugu Fiican ee lagu ilaaliyo qaabab adag iyo linework nadiif ah.
Dhererka
3D dhumucda map. Ku haystaa qaabka spatial - waxa ku dhow iyo waxa ka fog.
Jid
OpenPose jirka skeleton. Ku xidha qaabka iyo meelaha xubinta jirka.
Qoraalo
Loose gacanta-sawir doodles noqday farshaxanka dhamaystiran.
Qaybta
Xarfaha gobollada ee la calaamadeeyay. U magacaabo meel kasta oo ka mid ah goobta fasal.
caadi ah
Map-saxaafadeed caadi ah. Ilaaliyaa 3D saxaafadeed iyo bumps.
Sawir-qaadista
Fine line soo saarka - ugu fiican inking, manga, iyo tusaale.
Jid-dheer
Ogaanshaha xuduudaha oo fudud oo ku raacaya qaabab ka sii fudud Canny.
MLSD
Qaybaha line-toosan. Made for dhismaha, interiors, iyo sawirro.
Qalooca
Detail-dhaqanka conditioning u upscaling iyo shaqada texture seamless.
Sawir
Mask-aware conditioning in ay soo celiyaan kaliya qayb ka mid ah sawirka.
Dib u xardho / outpaint
Kala qaybi canvas ama repaint gobolada iyadoo la xusayay dhismaha hareeraha.
Saddex tallaabo
- Soo deji sawirka tilmaame — sawir, a sketch, a screen, wax kasta oo la dhismaha aad rabto inaad sii.
- Dooro nooca conditioning oo la mid ah waxa aad ka walwasho (u muuji si aad u sawir, qoto dheer si ay u dhacdo, canny ama lineart si ay u nadiifiyaan outlines).
- Qoro farriin ku saabsan muuqaalka aad rabto iyo soo saaro. Kor u qaad xoogga xakamaynta in ay raacaan soo jeedinta si adag, hoos u dhigi si ay u xorriyadda abuurista badan.
ControlNet — 12 noocyada conditioning ah oo ah qalab kaliya - FAQ
A qalab kaliya oo muujinaya dhammaan 12 noocyada conditioning ka ControlNet-Union SDXL ProMax tusaale — canny, pose, dhererka, scribble, lineart, anime-lineart, MLSD, HED, soft-edge, caadiga ah, qaybinta, iyo tile. Dooro nooc conditioning, hoos sawir tixraacaya, qor fariin, iyo SDXL soo bandhigaa sawir cusub oo raacaya dhismaha aad tixraacaya.
img2img dib-qaab dhismeedka ka badan soo gudbinta si toos ah — midab, xagal, iyo geometry isku darka la prompt. ControlNet tuuro midab ka fog oo kaliya ka ilaaliya calaamadda dhisme ee la doortay (liiska, u muuqdaan skeleton, dhumucda map, iwm). Taasi waxay kuu ogolaanaysaa inaad si radically content kala wareegaan inta lagu hayo isku darka dhagax adag.
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 a existing image.
ControlNet-Union SDXL ProMax (xinsir, Apache 2.0) waxay ku soo bandhigtaa dhammaan 12-ka shabakadaha ku habboon miisaanka 2.5 GB. Isku-darka hore wuxuu soo dejiyey miisaan ~ 2.5 GB oo kala duwan nooc kasta - isbeddelka u dhexeeya canny iyo pose wuxuu u muuqdaa bilow kulul.
Haa. ~1,200 tokens per render (1,000 salka SDXL + 20% ControlNet conditioning surcharge). Users diiwaan-in helaan 30,000 tokens bilaash ah maalin kasta — ku saabsan 25 conditioned renders maalin kasta lacag la'aan ah. Anonymous: 2,500 tokens / maalin (~ 2 renders).
Haa — the Control xoog slider (default 0.7) go'aamiya sida adag ee soo saarka raacaa soo jeedintaada. 1.0 = adag (soo saarka u eg tahay dib-soo bandhigid ee soo jeedintaada). 0.4 = loose (prompt leedahay xoriyad dheeraad ah). hoos u dhigo u kala duwanaansho abuurista, kor u qaado marka ay tahay waxyaabo run ah.
512 × 512 default. SDXL heerka saamiga - 768 × 1024 sawir, 1024 × 768 muuqaalka, 1024 × 1024 square - dhammaan shaqada. soo saarka weyn isticmaali VRAM iyo calaamadaha in ka badan; H200 taageertaa ilaa 1024 × 1024 raaxo leh.
Haa. ControlNet dhalasho waxay ku kacaysaa ~ 1,000 tokens sawir kasta, sida caadiga ah SDXL dhalasho. isticmaalayaasha aan la aqoonsanin waxay helayaan 6,000 tokens / maalin; xisaab bilaash ah wuxuu siiyaa 30,000 / maalin. Karti deynta ma jirto, ma jiro calaamad biyo ah. Go'aannada sare ama isbeddelada badan waxay ku kacayaan si isku mid ah oo ka badan.
ControlNet-Union SDXL ProMax waxaa lagu soo saaray hoos Apache 2.0 — si buuxda u oggolaansho, oo ay ku jiraan isticmaalka ganacsi. SDXL salka waa OpenRAIL ++. Labaduba oggolaadaan isticmaalka ganacsi; sawirka aad soo saartay waa adiga inaad isticmaali karto ganacsi la'aan.
Same, tayo isku mid ah, calaamadaha conditioning isku mid ah. ComfyUI iyo A1111 u baahan yihiin GPU deegaanka la 12+ GB VRAM iyo sidoo kale setup. Waxaan ku socdaan on dhismaha wadaagaan la jasiiradda xor ah oo deeqsi ah — ma soo dejisan, GPU ma baahan.
Codsiga koowaad wuxuu soo dejiyaa miisaanka Midowga (~ 2.5 GB) GPU cache-ka oo ku qaboojiyaa tuubada SDXL. Ka fiiri 30-60 ilbiriqsi ee codsiga ugu horreeyay ka dib markii la soo bandhigo ama la saaro LRU. Codsiyada ka dib waxaa hoos yimaada soo noqoshada culeyska caadiga ah 4-7 ilbiriqsi.
Haa — POST multipart in /v1/image/generate/ la model = sdxl (ama model = controlnet-union-sdxl-promax), soo jeedin, control_image (file), control_type = <mid ka mid ah: canny, u muuqdaan, qoto dheer, scribble, lineart, anime-lineart, mlsd, hed, soft-edge, caadiga ah, qaybinta, tile>, doorasho control_strength (0.1-1.5).