Isithombe nguMbhali
I-ComfyUI ishintshe indlela abadali nabathuthukisi ababhekana ngayo nokukhiqiza isithombe esinamandla e-AI. Ngokungafani nezindawo zokusebenzelana ezivamile, i-architecture esekelwe endaweni ye-ComfyUI ikunikeza ukulawula okungakaze kubonwe ngaphambili kokugeleza komsebenzi wakho wokudala. Lesi sifundo sokuphahlazeka sizokuthatha kusukela ekuqaleni okuphelele kuye kumsebenzisi ozethembayo, sikuhambise kuwo wonke umqondo obalulekile, izici, nesibonelo esisebenzayo osidingayo ukuze ube yingcweti yaleli thuluzi elinamandla.
Isithombe nguMbhali
I-ComfyUI imahhala, umthombo ovulekile, isikhombimsebenzisi esisekelwe kuma-node kanye ne-backend ye Ukusabalalisa Okuzinzile namanye amamodeli akhiqizayo. Kucabange njengendawo yokuhlela ebonakalayo lapho uxhuma khona amabhulokhi wokwakha (okuthiwa “ama-node”) ukuze udale ukuhamba komsebenzi okuyinkimbinkimbi kokukhiqiza izithombe, amavidiyo, amamodeli e-3D, nomsindo.
Izinzuzo ezibalulekile ngaphezu kokusebenzelana kwendabuko:
- Unokulawula okugcwele kokwakha ukuhamba komsebenzi ngokubukeka ngaphandle kokubhala ikhodi, ngokulawula okuphelele phezu kwayo yonke ipharamitha.
- Ungakwazi ukulondoloza, wabelane, futhi uphinde usebenzise konke ukugeleza komsebenzi ngemethadatha eshumekwe kumafayela akhiqiziwe.
- Azikho izindleko ezifihliwe noma okubhaliselwe; kungenziwa ngokwezifiso ngokuphelele ngama-node wangokwezifiso, mahhala, nomthombo ovulekile.
- Isebenza endaweni emshinini wakho ukuze iphindaphinde ngokushesha kanye nezindleko eziphansi zokusebenza.
- Inokusebenza okunwetshiwe, okucishe kungapheli ngamanodi ngokwezifiso angahlangabezana nezidingo zakho ezithile.
# Ukukhetha Phakathi Kokufakwa Kwasendaweni Nokusekelwe Efu
Ngaphambi kokuhlola i-ComfyUI ngemininingwane eyengeziwe, kufanele unqume ukuthi uzoyiqhuba endaweni noma usebenzise inguqulo esekelwe emafini.
| Ukufakwa Kwasendaweni | Ukufakwa Okusekelwe Efwini |
|---|---|
| Isebenza ngokungaxhunyiwe ku-inthanethi uma isifakiwe | Idinga uxhumano lwe-inthanethi njalo |
| Azikho izinkokhelo zokubhalisa | Kungase kuhlanganise izindleko zokubhaliselwe |
| Gcwalisa ubumfihlo bedatha nokulawula | Ukulawula okuncane kudatha yakho |
| Idinga ihadiwe enamandla (ikakhulukazi enhle I-NVIDIA GPU) | Azikho izingxenyekazi zekhompuyutha ezinamandla ezidingekayo |
| Kudingeka ukufakwa mathupha nezibuyekezo | Izibuyekezo ezizenzakalelayo |
| Ikhawulelwe amandla okucubungula ekhompyutha yakho | Imikhawulo yesivinini engaba khona ngesikhathi sokusetshenziswa okuphezulu |
Uma usaqala, kunconywa ukuthi uqale ngesixazululo esisekelwe efwini ukuze ufunde isikhombimsebenzisi nemibono. Njengoba uthuthukisa amakhono akho, cabanga ukushintshela ekufakweni kwendawo ukuze uthole ukulawula okukhulu futhi wehlise izindleko zesikhathi eside.
# Ukuqonda i-Core Architecture
Ngaphambi kokusebenza ngama-node, kubalulekile ukuqonda isisekelo sethiyori sokuthi i-ComfyUI isebenza kanjani. Kucabange njengokuhlukahluka phakathi kwamayunivesiti amabili: indawo yonke ebomvu, eluhlaza, eluhlaza okwesibhakabhaka (i-RGB) (esikubonayo) kanye nomkhathi ocashile womkhathi (lapho ukubala kwenzeka khona).
// Ama-Universes Amabili
Umkhathi we-RGB ngumhlaba wethu obonakalayo. Iqukethe izithombe ezivamile kanye nedatha esingayibona futhi siyiqonde ngamehlo ethu. Isikhala esicashile (indawo yonke ye-AI) yilapho kwenzeka khona “umlingo”. Kuwukumelela kwezibalo amamodeli angakuqonda futhi akulawule. Isiyaluyalu, igcwele umsindo, futhi iqukethe ukwakheka kwezibalo okufushanisa ukwenziwa kwesithombe.
// Ukusebenzisa i-Variational Autoencoder
I-variational autoencoder (VAE) isebenza njengengosi phakathi kwalawa mayunivesi.
- I-Encoding (RGB – Latent) ithatha isithombe esibonakalayo futhi isiguqule sibe isethulo esifihlekile esingabonakali.
- I-Decoding (Latent — RGB) ithatha isethulo esifihlekile esingabonakali futhi isiguqulele esithombeni esisibonayo.
Lo mqondo ubalulekile ngoba ama-node amaningi asebenza ngaphakathi kwendawo yonke eyodwa, futhi ukuyiqonda kuzokusiza ukuxhuma ama-node alungile ndawonye.
// Ukuchaza amaNodes
Ama-Node ayizisekelo zokwakha ze-ComfyUI. I-node ngayinye iwumsebenzi ozimele owenza umsebenzi othile. Amanodi ane:
- Okokufaka (uhlangothi lwesobunxele): Lapho idatha ingena khona
- Okuphumayo (uhlangothi lwesokudla): Lapho idatha ecutshunguliwe iphuma khona
- Amapharamitha: Izilungiselelo ozilungisayo ukuze ulawule ukuziphatha kwenodi
// Ukuhlonza Izinhlobo Zedatha Enekhodi Enemibala
I-ComfyUI isebenzisa isistimu yombala ukukhombisa ukuthi yiluphi uhlobo lwedatha olugeleza phakathi kwamanodi:
| Umbala | Uhlobo Lwedatha | Isibonelo |
|---|---|---|
| Okuluhlaza okwesibhakabhaka | Izithombe ze-RGB | Izithombe ezibonakalayo ezijwayelekile |
| Pink | Izithombe Ezifihlekile | Izithombe ekumeleleni okucashile |
| Okuphuzi | I-CLIP | Umbhalo uguqulelwe olimini lomshini |
| Okubomvu | I-VAE | Imodeli eguqula phakathi kwendawo yonke |
| iwolintshi | I-Conditioning | Ukwaziswa nokulawula imiyalelo |
| Okuhlaza | Umbhalo | Izinhlamvu zombhalo ezilula (izixwayiso, izindlela zamafayela) |
| Okunsomi | Amamodeli | Izindawo zokuhlola nezisindo zemodeli |
| I-Teal/Turquoise | ControlNets | Lawula idatha yokukhiqiza okuyisiqondiso |
Ukuqonda le mibala kubaluleke kakhulu. Bakutshela khona manjalo ukuthi amanodi angaxhumeka yini enye kwenye.
// Ukuhlola Izinhlobo Zezindawo Ezibalulekile
Amanodi okulayisha angenisa amamodeli nedatha ekuhambeni kwakho komsebenzi:
CheckPointLoader: Ilayisha imodeli (ngokuvamile equkethe izisindo zemodeli, Ukuqeqeshwa Kwangaphambi Kolimi-Isithombe Okuphikisanayo (CLIP), kanye ne-VAE kufayela elilodwa).Load Diffusion Model: Ilayisha izingxenye zemodeli ngokwehlukana (kumamodeli amasha njenge I-Flux ezingahlanganisi izingxenye).VAE Loader: Ilayisha idekhoda ye-VAE ngokuhlukana.CLIP Loader: Ilayisha umbhalo wekhodi ngokuhlukana.
Ukucubungula amanodi kuguqula idatha:
CLIP Text Encodeiguqulela ukwaziswa kombhalo kulimi lomshini (i-conditioning).KSampleriyinjini yokukhiqiza isithombe esiyinhloko.VAE Decodeiguqula izithombe ezicashile zibuyele ku-RGB.
Ama-Utility Node asekela ukuphathwa kokuhamba komsebenzi:
- I-Primitive Node: Ikuvumela ukuthi ufake amanani mathupha.
- I-Node yomzila kabusha: Ihlanza ukubonakala kokugeleza komsebenzi ngokuqondisa kabusha ukuxhumana.
- Layisha Isithombe: Ingenisa izithombe ekuhambeni kwakho komsebenzi.
- Londoloza Isithombe: Ithekelisa izithombe ezikhiqiziwe.
# Ukuqonda i-KSampler Node
I KSampler ngokungangabazeki iyindawo ebaluleke kakhulu ku-ComfyUI. “Ngumakhi werobhothi” okhiqiza izithombe zakho. Ukuqonda imingcele yayo kubalulekile ekudaleni izithombe ezisezingeni eliphezulu.
// Ibuyekeza amapharamitha e-KSampler
Imbewu (Okuzenzakalelayo: 0)
Imbewu yisimo sokuqala esingahleliwe esinquma ukuthi yimaphi amaphikseli angahleliwe abekwe ekuqaleni kwesizukulwane. Kucabange njengendawo yakho yokuqala yokwenza okungahleliwe.
- Imbewu Elungisiwe: Ukusebenzisa imbewu efanayo enezilungiselelo ezifanayo kuzohlale kukhiqiza isithombe esifanayo.
- Imbewu Engahleliwe: Isizukulwane ngasinye sithola imbewu entsha engahleliwe, ekhiqiza izithombe ezihlukene.
- Ibanga Lenani: 0 kuya ku-18,446,744,073,709,551,615.
Izinyathelo (Okuzenzakalelayo: 20)
Izinyathelo zichaza inani lokuphindaphinda kwe-denoising okwenziwe. Isinyathelo ngasinye siphucula kancane kancane isithombe sisuka emsindweni omsulwa siye kokuphumayo osifunayo.
- Izinyathelo Eziphansi (10-15): Isizukulwane esisheshayo, imiphumela engacolisisiwe.
- Izinyathelo Ezimaphakathi (20-30): Ibhalansi enhle phakathi kwekhwalithi nesivinini.
- Izinyathelo Eziphakeme (50+): Ikhwalithi engcono kodwa ihamba kancane kakhulu.
Isikali se-CFG (Okuzenzakalelayo: 8.0, Ibanga: 0.0-100.0)
Isikali se-classifier-free guide (CFG) silawula ukuthi i-AI iwulandela ngokuqinile ukwaziswa kwakho.
Isifaniso – Cabanga nje unikeza umakhi ipulani:
- I-CFG Ephansi (3-5): Umakhi ubheka ipulani bese enza into yakhe — ubuciko kodwa angase azibe imiyalelo.
- I-CFG Ephezulu (12+): Umakhi ulandela ngokucophelela yonke imininingwane yepulani — enembile kodwa ingase ibukeke iqinile noma icutshungulwe ngokweqile.
- I-Balanced CFG (7-8 yokusabalalisa okuzinzile, 1-2 ye-Flux): Umakhi ulandela kakhulu ipulani ngenkathi engeza ukuhluka kwemvelo.
Igama lesampula
Isampula i-algorithm esetshenziselwa inqubo yokukhipha umsindo. Amasampula ajwayelekile afaka Euler, DPM++ 2Mfuthi UniPC.
Isihleli
Ilawula ukuthi umsindo uhlelwe kanjani ezinyathelweni zokukhipha umsindo. Abahleli banquma ijika lokunciphisa umsindo.
- Okuvamile: Ukuhlela umsindo okuvamile.
- I-Karras: Ngokuvamile inikeza imiphumela engcono ekubalweni kwezinyathelo eziphansi.
I-Denoise (Okuzenzakalelayo: 1.0, Ibanga: 0.0-1.0)
Lesi esinye sezilawuli zakho ezibaluleke kakhulu zokugeleza komsebenzi kwesithombe kuya esithombeni. I-Denoise inquma ukuthi yiliphi iphesenti lesithombe esifakiwe esizofakwa esikhundleni sokuqukethwe okusha:
- 0.0: Ungashintshi lutho — okukhiphayo kuzofana nokokufaka
- 0.5: Gcina u-50% wesithombe sangempela, ukhiqize kabusha u-50% njengesisha
- 1.0: Khiqiza kabusha ngokuphelele — unganaki isithombe sokufaka bese uqala ngomsindo omsulwa
# Isibonelo: Ukukhiqiza Ukuma Ngomlingiswa
Ukwaziswa: “I-cyberpunk android enamehlo e-neon aluhlaza, izingxenye zomshini ezinemininingwane, ukukhanya okumangalisayo.”
Izilungiselelo:
- Imodeli: Flux
- Izinyathelo: 20
- I-CFG: 2.0
- Isampula: Okuzenzakalelayo
- Ukulungiswa: 1024×1024
- Imbewu: Yenza ngokungahleliwe
Ukwaziswa okungalungile: “ikhwalithi ephansi, ukufiphala, ukugcwala ngokweqile, okungelona iqiniso.”
// Ihlola ukugeleza kokusebenza kwesithombe kuya kwesithombe
Ukugeleza kokusebenza kwesithombe kuya kwesithombe kwakhela phezu kwesisekelo sombhalo kuya esithombeni, okwengeza isithombe sokufaka ukuze siqondise inqubo yokwenza.
Isimo: Unesithombe sokwakheka kwezwe futhi usifuna ngesitayela sokudweba uwoyela.
- Layisha isithombe sakho sokwakheka kwezwe
- I-Positive Prompt: “umdwebo kawoyela, isitayela se-impressionist, imibala egqamile, imivimbo yebhulashi”
- Isilinganiso: 0.7
// Ukuqhuba i-Pose-Guided Character Generation
Isimo: Ukhiqize umlingiswa omthandayo kodwa ufuna ukuma okuhlukile.
- Layisha isithombe sakho sokuqala sohlamvu
- Ukwaziswa Okuhle: “Incazelo efanayo yomlingiswa, ukuma, izingalo eceleni”
- Isilinganiso: 0.3
# Ukufaka nokusetha i-ComfyUI
I-Cloud-based (Elula Kwabaqalayo)
Vakashela RunComfy.com bese uchofoza ukuqalisa i-Comfy Cloud ngakwesokudla phezulu. Kungenjalo, ungakwazi kalula bhalisela esipheqululini sakho.
Isithombe nguMbhali
Isithombe nguMbhali
// Ukusebenzisa iWindows Portable
- Ngaphambi kokuthi ulande, kufanele ube nokusethwa kwezingxenyekazi zekhompuyutha okuhlanganisa i-NVIDIA GPU enosekelo lwe-CUDA noma i-macOS (Apple Silicon).
- Landa ukwakhiwa kweWindows okuphathekayo kusuka ekhasini lokukhishwa kwe-ComfyUI GitHub.
- Khiphela endaweni yakho oyifunayo.
- Gijima
run_nvidia_gpu.bat(uma une-NVIDIA GPU) nomarun_cpu.bat. - Vula isiphequluli sakho ukuze
// Ukwenza Ukufaka Mathupha
- Faka I-Python: Landa inguqulo 3.12 noma 3.13.
- I-Clone Repository:
git clone - Faka I-PyTorch: Landela imiyalelo eqondene nenkundla ye-GPU yakho.
- Faka Okuncikile:
pip install -r requirements.txt - Engeza Amamodeli: Beka izindawo zokuhlola imodeli phakathi
models/checkpoints. - Gijima:
python main.py
# Ukusebenza Ngezinhlobo ezahlukene ze-AI
I-ComfyUI isekela amamodeli amaningi asezingeni eliphezulu. Nawa amamodeli aphezulu amanje:
| I-Flux (Inconyelwe Okwangempela) | Ukusabalalisa Okuzinzile 3.5 | Amamodeli Amadala (SD 1.5, SDXL) |
|---|---|---|
| Kuhle kakhulu ezithombeni ze-photorealistic | Ikhwalithi nesivinini esinokulinganisela | Icushwe kahle kakhulu umphakathi |
| Isizukulwane esisheshayo | Isekela izitayela ezihlukahlukene | I-Massive low-rank adaptation (LoRA) ecosystem |
| CFG: 1-3 ububanzi | CFG: 4-7 ububanzi | Isasebenza kahle kakhulu ekusebenzeni okuthile |
# Ukuthuthukisa Ukugeleza Kokusebenza Ngokuzijwayeza Kwezinga Eliphansi
Ukujwayela kwezinga eliphansi (ama-LoRA) amafayela e-adaptha amancane ashuna kahle amamodeli ezitayela ezithile, izifundo, noma ubuhle ngaphandle kokulungisa imodeli eyisisekelo. Ukusetshenziswa okuvamile kuhlanganisa ukungaguquguquki kwezinhlamvu, izitayela zobuciko, nemiqondo yangokwezifiso. Ukuze usebenzise eyodwa, engeza inodi ethi “Layisha i-LoRA”, khetha ifayela lakho, bese ulixhuma ekuhambeni kwakho komsebenzi.
// Ukukhiqiza Isithombe Esiqondisayo ngama-ControlNets
I-ControlNets ihlinzeka ngokulawula kwendawo phezu kokukhiqiza, okuphoqa imodeli ukuthi ihloniphe ukuma, amamephu onqenqema, noma ukujula:
- Phoqa ukuma okuqondile ezithombeni eziyizethenjwa
- Gcina ukwakheka kwento ngenkathi ushintsha isitayela
- Ukwakhiwa komhlahlandlela okususelwe kumamephu onqenqema
- Hlonipha ulwazi olujulile
// Ukwenza Ukuhlela Okukhethiwe Kwesithombe Nge-Inpainting
I-Inpainting ikuvumela ukuthi ukhiqize kabusha izifunda ezithile kuphela zesithombe ngenkathi ulondoloza okunye.
Ukuhamba komsebenzi: Layisha isithombe — Umdwebo wemaski — Inpainting KSampler — Umphumela
// Ukwandisa Ukulungiswa Ngokuphakama
Sebenzisa amanodi aphezulu emva kwesizukulwane ukuze ukhuphule ukulungiswa ngaphandle kokuvuselela sonke isithombe. Ama-upscaler adumile afaka I-RealESRGAN futhi I-SwinIR.
# Isiphetho
I-ComfyUI imele ushintsho olubaluleke kakhulu ekudaleni okuqukethwe. Isakhiwo sayo esisekelwe ku-node sikunikeza amandla abegcinelwe onjiniyela be-software ngenkathi etholakala kwabaqalayo. Ijika lokufunda lingokoqobo, kodwa wonke umqondo owufundayo uvula amathuba amasha okudala.
Qala ngokudala ukuhamba komsebenzi okulula kombhalo uye esithombeni, ukhiqize ezinye izithombe, futhi ulungise amapharamitha. Emavikini ambalwa, uzobe udala ukuhamba komsebenzi okuyinkimbinkimbi. Ezinyangeni ezimbalwa, uzobe ucindezela imingcele yalokho okungenzeka endaweni yokukhiqiza.
Shithu Olumide ungunjiniyela wesofthiwe nombhali wezobuchwepheshe othanda ukusebenzisa ubuchwepheshe obuphambili ekwenzeni izindaba ezithokozisayo, oneso elibukhali lemininingwane kanye nekhono lokwenza imiqondo eyinkimbinkimbi ibe lula. Ungathola futhi i-Shittu Twitter.
