Uma usebenza ku-VFX futhi unelukuluku lokuthatha ukugeleza kwesithombe sakho esikhiqizayo ukuyise ezingeni elilandelayo, i-ComfyUI kufanele ube nayo. Indlela yayo esekwe ku-node, i-modularity, namandla engeza i-FLUX Lokhu kuyenza ibe indawo ekahle yokuhlola yonke into kusukela ku-high-fidelity text2img kuya kumapayipi ayinkimbinkimbi ngokuqeqeshwa kwe-ControlNet, IP-Adapter, noma i-LoRA. I-ComfyUI ye-VFX: Ukufakwa nokusebenza kwe-FLUX kuchaziwe.
Kulo mhlahlandlela uzothola, ngendlela ehlelekile, konke okudingayo: ukuthi yikuphi ukuhamba komsebenzi ku-ComfyUI, indlela yokulondoloza nokwabelana ngayo, indlela yokufaka i-FLUX isinyathelo ngesinyathelo, yiziphi ihadiwe oyidingayo, nencazelo ecacile yokugeleza okusetshenziswe kakhulu (Txt2Img, Img2Img, Inpainting, LoRA, ControlNet, NF4, IPAdapter kanye ne-latent). Uzobona ezinye izindlela ezifana ne-FP8, NF4, noma i-GGUF ukuze unciphise i-VRAMIzinsiza zokusebenzisa i-Flux emafini, kanye nokufakwa kwe-ComfyUI okwenziwe nge-Windows okunamathiphu asebenzayo.
Kuyini ukuhamba komsebenzi ku-ComfyUI?
Ukugeleza komsebenzi ukumelela okubonakalayo kwenqubo yokukhiqiza oyakhayo ngokuhlanganisa amanodi. Ungacabanga njengebhodi lezicucu zohlobo lwe-LegoI-node ngayinye yenza umsebenzi (ukulayisha amamodeli, umbhalo wombhalo wekhodi, isampuli, ukuqopha, njll.) futhi izixhumanisi zichaza indlela elandelwa ulwazi kuze kutholakale isithombe sokugcina.
I-ecosystem inkulu kakhulu: isithombe esimile, ividiyo, umsindo kanye ne-3D Zihlala ndawonye ngenxa yomphakathi womthombo ovulekile. Okubi ukuthi kukhona ijika lokufunda, ngoba kubalulekile ukuqonda ukuthi ingxenye ngayinye yenzani nokuthi ungayihlanganisa kanjani ngempumelelo ukuze ugweme izingqinamba noma imiphumela engahambisani.
Ukuze uqalise ngokushesha, kungcono ukubheka ukugeleza okusemthethweni nokuyisisekelo (i-text2img ne-img2img) bese udlulela kumanodi ayinkimbinkimbi kakhulu. Imibhalo yomphakathi kanye nezibonelo ezisemthethweni ze-ComfyUI Zivame ukubuyekezwa, ngakho-ke kulula ukuqalisa kabusha noma ukubuyekeza ukugeleza komsebenzi okushintshayo ngezinguqulo ezintsha.
Imininingwane eyodwa ebalulekile: I-ComfyUI ingashumeka ukuhamba komsebenzi esithombeni sokugcina (i-PNG) ngokwayo njengemethadatha. Lokhu kukuvumela ukuthi wabelane ngesithombe kuphela futhi ubuyise yonke igrafu. ukuyihudula emuva kusixhumi esibonakalayo.
Ungafunda kanjani futhi uthuthuke kanjani ngokugeleza komsebenzi
Iseluleko sami ukuqala ngezibonelo ezilula ezishicilelwe kuzinsiza zohlobo lwe-wiki namagalari omphakathi. Ukuqhubeka okunengqondo yilokhu: Txt2Img, Img2Img, bese kuba ControlNet noma LoRANjengoba uqonda okokufaka, okuphumayo, nokuthi abahleli bakuthinta kanjani ukusampula, kuzoba ngokwemvelo ukufaka ama-node omhlahlandlela, imaski, nezimo ezengeziwe.
Uma ufuna ukubona ama-topology ahlukahlukene, kukhona amakhosombe wokugeleza komphakathi namakhasi anezibonelo eziphindaphindekayo lapho ungalanda khona izithombe ngemethadatha noma amafayela we-.json. Kungumkhuba omuhle ukungenisa, ukugijima njengoba kwenzeka, bese uphindaphinda ngamapharamitha ukuqonda umthelela webhulokhi ngayinye ngaphandle kokuphula yonke.
Ezisekelweni zamafu uzothola nezindawo ezilungiselelwe ngaphambili ukuze usebenzise amapayipi ngaphandle kokuzabalaza nokuncika kwendawo. Inzuzo ukuthi balayisha kuqala ama-node asindayo namamodeliNokho, kuyatuseka ukuthi uhlole izinguqulo ezitholakalayo kanye ne-VRAM ukuze uqiniseke ukuthi imiphumela ihlangabezana nokulindelwe.
Ukulondoloza, ukungenisa, nokuthekelisa ukuhamba komsebenzi ku-ComfyUI
I-ComfyUI isekela izindlela ezimbili eziyinhloko zokulondoloza: isithombe esinemethadatha (PNG) noma ifayela le-JSON kwegrafu. Eyokuqala iyona elula kakhulu yokwabelana ezinkundleni; okwesibili kukunikeza ukulawula okusobala kwefayela, okuwusizo ekwenzeni inguqulo.
Ukuze ungenise, mane udonse ifayela le-PNG noma le-Json kusixhumi esibonakalayo, noma usebenzise isinqamuleli esithi Ctrl (Command) + O. Ukuze uthumele, ungagcina isithombe esikhiqiziwe noma usebenzise imenyu yokuthekelisa ye-JSON.Qaphela uma uminyanisa noma udlulisa izithombe: ezinye izindlela zokuminyanisa namashaneli athile asusa imethadatha, okungadala ukuthi ulahlekelwe ukuhamba komsebenzi okushumekiwe.
Ngenxa yokuvela njalo kwe-ComfyUI, Akuwona wonke amafayela e-JSON amadala asebenza ezinguqulweni ezintsha.Uma kukhona okungahambi kahle, vula ukugeleza, shintsha amanodi aphelelwe yisikhathi, noma faka kabusha okuncikile ngenguqulo ehambisanayo; ukusebenzisa i-ComfyUI-Manager kwenza kusheshe kakhulu ukuthola nokuxazulula izingxenye ezingekho.
I-FLUX ku-ComfyUI: ukuthi iyini nokuthi kungani ibalulekile
I-FLUX.1 iwumndeni wamamodeli avela ku-Black Forest Labs egxile ekuguquleleni umbhalo uye-sithombeni othembeke kakhulu. Isakhiwo sayo esiyingxubevange esinamapharamitha acishe abe yizigidi eziyizinkulungwane eziyi-12 Ilungiselelwe ukunamathela ngokushesha, ukuphatha izigcawu eziyinkimbinkimbi, futhi ikhiqize umbhalo ofundekayo ngaphakathi kwesithombe, umsebenzi lapho amanye amamodeli evamise ukuntengantenga.
Enye inzuzo: ukuguquguquka kwayo. Kusukela ku-photorealism kuya kuzitayela zobucikoI-FLUX.1 igqama ngenxa yokungaguquguquki kwayo okubonakalayo kanye nemininingwane, okuhlanganisa ukunikezwa kwezandla, indawo ebuthakathaka yakudala emidwebeni ekhiqizayo. Akumangazi ukuthi iqhathaniswa nezixazululo ezifana ne-Stable Diffusion noma iMidjourney, ephuma phambili mayelana nokusebenziseka kalula kanye nekhwalithi.
I-Black Forest Labs yasungulwa nguRobin Rombach, umuntu obalulekile owayesebenza emnyombweni we-Stability AI. Uma ufuna ukubheka kuqala, iwebhusayithi yayo esemthethweni iku blackforestlabs.ai.
I-FLUX.1 isatshalaliswa ngezinhlobonhlobo ezintathu: I-Pro, i-Dev ne-SchnellI-Pro inikeza ikhwalithi ephezulu kakhulu yezindawo ezichwepheshile; I-Dev yakhelwe ukusetshenziswa okungezona ezohwebo ngebhalansi enhle kakhulu; U-Schnell ugxile esivinini kanye nokukhanya, futhi ungumthombo ovulekile ngaphansi kwelayisensi ye-Apache 2.0.
Izidingo zezingxenyekazi zekhompuyutha ngenguqulo ye-FLUX
Nge-FLUX.1 Pro, iyanconywa I-NVIDIA RTX 4090 thayipha i-GPU eno-24 GB we-VRAM32 GB we-RAM kanye ne-SSD esheshayo. Isebenzisa i-FP16 ukugwema ukuphuma kwesibhakela, futhi kuhle kakhulu ukusebenzisa isifaki khodi sombhalo ku-FP16 ngekhwalithi ephezulu.
Ku-FLUX.1 Dev, a I-RTX 3080/3090 eno-16 GB we-VRAM Isebenza kahle nge-16 GB ye-RAM kanye ne-25 GB yesikhala sediski. Isekela i-FP16 ngisho ne-FP8 kwezinye izimo, kuya nge-GPU yakho.
Okwe-FLUX.1 Schnell, I-RTX 3060/4060 eno-12 GB we-VRAM I-8GB ye-RAM ne-15GB yesitoreji kwanele. Idizayinelwe isivinini, idela ezinye ze-headroom yokusebenza ngokuqhathaniswa nezinguqulo ze-Pro/Dev.
Uma unenkumbulo emfushane, umphakathi unikeza ezinye izindlela ezifana ne-FP8, NF4, noma i-GGUF leyo Banciphisa kakhulu i-VRAM edingekayo, ngokulungiselelwa kusukela ku-6 kuye ku-12 GB kuye ngokugeleza.
Ukufaka i-FLUX ku-ComfyUI: izinyathelo ezibalulekile
Okokuqala, qiniseka ukuthi uyayisebenzisa inguqulo yakamuva ye-ComfyUIUkuhlanganiswa kwe-FLUX kudinga ama-node nezici ezibuyekezwa njalo.
Landa umbhalo nezishumeki ze-CLIP: clip_l.safetensors kanye nelinye lamafayela e-T5 XXL, kungaba yi-t5xxl_fp16.safetensors (uma une-VRAM/RAM eningi) noma ama-t5xxl_fp8_e4m3fn.safetensors (uma unesabelomali esiqinile). Zibeke kufolda ye-ComfyUI/models/clip/. Uma usebenzise i-SD3 Medium, kungenzeka ukuthi usunawo lawa mafayela..
VAE: landa ae.ama-safetensors bese uyihambisa ku-ComfyUI/models/vae/. Iqambe kabusha ku-flux_ae.safetensors uma ufuna ukuyithola kalula. Le VAE ithuthukisa ukuqoshwa kokugcina futhi kuyisihluthulelo sekhwalithi.
UNET: khetha phakathi i-flux1-dev.safetensors noma i-flux1-schnell.safetensors ngokwenkumbulo yakho, bese uyibeka kokuthi ComfyUI/models/unet/. Ngalokhu, unesisekelo sokusebenzisa ukugeleza kwe-FLUX. wendawo.
Umhlahlandlela osebenzayo wokugeleza komsebenzi we-FLUX ku-ComfyUI
I-Txt2Img ene-FLUX
Qala ngokulayisha izingxenye: I-UNETLoader, i-DualCLIPLoader ne-VAELaderInodi ye-CLIPextEncode ifaka ukwaziswa kwakho; I-EmptyLatentImage idala isithombe esicashile sokuqala; I-BasicGuider iqondisa inqubo ngokuhlanganisa ingqondo enemibandela ne-UNET ye-FLUX.
Khetha isampula nge KSamplerKhethaIdala umsindo nge-RandomNoise futhi ichaza irempu ye-sigma nge-BasicScheduler. I-SamplerCustomAdvanced ihlanganisa yonke into: umsindo, umhlahlandlela, isampula, ama-sigmas, kanye ne-latent. Ekugcineni, i-VAEDecode iguqula isignali efihlekile ibe isithombe. futhi nge-SaveImage ulondoloza umphumela.
I-Img2Img nge-FLUX
Ipayipi lengeza isithombe sokuqala: LayishaIsithombe + Isikali Sesithombe Usayizi uyalungiswa, futhi i-VAEEncode iyibeka kokuthi ifihlekile. Ukwaziswa kufakwe ikhodi nge-CLIPextEncode futhi amandla aso alungiswa nge-FluxGuidance. I-ModelSamplingFlux ilawula ukushintshwa kwezikhala kanye nobukhuluI-KSamplerSelect, RandomNoise, kanye ne-BasicScheduler iphatha amasampula. I-SamplerCustomAdvanced ihlanganisa okunemibandela nokokufaka okufihlekile, bese i-VAEDecode ikhiqiza okukhiphayo.
I-LoRA ene-FLUX
Ukuze ucwenge isitayela noma izici, engeza I-LoraLoaderModelOnly kanye ne-UNETLoader, i-DualCLIPLoader, ne-VAELader. Ngemva kokufaka ikhodi yombhalo nokusebenzisa i-FluxGuidance, udala isithombe esicashile nge-EmptyLatentImage, uchaze ukusampula nge-ModelSamplingFlux, bese usebenzisa i-SamplerCustomAdvanced. Nge-VAEDecode uthola isithombe esivele sithonywe yi-LoRA. Isibonelo esijwayelekile: realism_lora.safetensors ku-flux1-dev.
ControlNet nge FLUX
Amacala amabili awusizo kakhulu e-VFX: ukujula kanye nemiphetho ye-canny. Ukuze uthole ukujula, qhubeka nge I-MiDaS-DepthMapPreprocessorLayisha ukujula kwe-ControlNet futhi ukusebenzise nge-ApplyFluxControlNet. Sebenzisa i-XlabsSampler ukuze ukhiqize i-wave wave form enemibandela, bese I-VAEDecode ikhiqiza isithombe.
Ukuze Canny, sebenzisa I-CannyEdgePreprocessor, layisha i-Canny ControlNet bese uphinda uhlelo: ApplyFluxControlNet → XlabsSampler → VAEDecode. Lokhu kulawula okwengeziwe kunikeza ukunemba phezu komumo nokwakheka..
Ukudweba nge-FLUX
Layisha i-UNET, i-VAE ne-CLIP, futhi ulungiselele ukwaziswa okuhle nokubi. I-LoadAndResizeImage iletha isithombe kanye nemaskiUshintsho lwenziwe bushelelezi nge-ImpactGaussianBlurMask. I-InpaintModelConditioning ihlanganisa isimo, isithombe, nemaski. Ngemva kokumisa isampula, umsindo, kanye nama-sigmas, i-SamplerCustomAdvanced yakha kabusha isifunda esifihliwe. I-VAEDecode ihlanganisa isichibi ngokungaguquki nabanye.
I-FLUX NF4
Nge-NF4 quantization, inkumbulo incishisiwe. Layisha izingxenye nge I-CheckpointLoaderNF4 futhi ichaza ubude/ububanzi ngamanodi akudala. I-ModelSamplingFlux ibeka imingcele; I-EmptySD3LatentImage idala isithombe esicashile; I-BasicScheduler ne-RandomNoise zihlela umsindo. I-SamplerCustomAdvanced ikhiqiza isignali ecashile futhi i-VAEDecode iyihumushela esithombeni.Ngokukala, i-UltimateSDUpscale, kanye ne-UpscaleModelLoader kanye nokwaziswa okwengeziwe okuhle, kwenza umehluko.
IPadaptha ene-FLUX
Uma ufuna ukubeka isimo usebenzisa isithombe esiyisethenjwa, sebenzisa LoadFluxIPAdapter kanye ne-ApplyFluxIPAdapter kanye ne-clip_vision_l.safetensors. Kala isithombe esiyisethenjwa nge-ImageScale, lungiselela imiyalo, bese usebenzisa i-XlabsSampler. Nge-VAEDecode uzobona okukhiphayo okuthonywa ubuhle noma izici kusukela esithombeni somhlahlandlela.
Umqeqeshi we-LoRA we-FLUX
Ukuqeqesha i-LoRA ngqo ku-ComfyUI, ukuhamba komsebenzi kuhlanganisa: I-FluxTrainModelSelect, i-OptimizerConfig kanye ne-TrainDatasetGeneralConfigI-InitFluxLoRATraining iyaqalisa, i-FluxTrainLoop yenza izinyathelo, bese i-FluxTrainValidate ikhiqiza ukuqinisekiswa ngezikhathi ezithile.
Nge-VisualizeLoss ulandelela ukulahlekelwa; I-ImageBatchMulti kanye ne-ImageConcatFromBatch Bahlanganisa ukuqinisekiswa; I-FluxTrainSave ilondoloza izindawo zokuhlola, futhi i-FluxTrainEnd ivala inqubo. Uma ufuna, layisha umphumela ku-Hugging Face nge-UploadToHuggingFace futhi wabelane ngayo.
I-Flux Latent Upscaler
Ukuze ukale ngokuningiliziwe, chaza usayizi nge SDXLEmptyLatentSizePicker+ kanye namaketanga i-LatentUpscale ne-LatentCrop. Ngamaski adalwe i-SolidMask kanye ne-FeatherMask, i-LatentCompositeMasked ihlanganisa okucashile okukaliwe nokwangempela. I-InjectionLatentNoise+ ithuthukisa imininingwane ngaphambi kwe-VAEDecodeFuthi ukuthinta nge-ImageSmartSharpen+ kuqeda inqubo. Amanodi wokubala afana ne-SimpleMath+ asiza ukuvumelanisa izilinganiso.
Ezinye izinguqulo: FP8, NF4 kanye ne-GGUF ukwehlisa i-VRAM
Uma ushoda ngezinsiza, unezinketho. Izindawo zokuhlola ze-FP8 ezivela ku-Comfy.org nababhali abafana no-Kijai. ikuvumela ukuthi usebenzise i-FLUX ngefayela elilodwa ku-ComfyUI/models/checkpoints/. Kunconywa ukuthi uqambe kabusha noma uhlukanise kube amafolda ukuze uhlukanise phakathi kwezinhlobonhlobo ze-dev ne-schnell.
Nge-NF4 (bitsandbytes), faka i-plugin ComfyUI_bitsandbytes_NF4 futhi usebenzise i-flux1-dev-bnb-nf4-v2 kumamodeli/izindawo zokuhlola. Le nguqulo ithuthukisa imininingwane uma iqhathaniswa nokuphindaphinda kokuqala.
Ukulinganisa kwe-GGUF kweCity96, kanye ne-plugin I-ComfyUI-GGUFyehlisa ibha nakakhulu: Landa imodeli ye-FLUX GGUF, isifaki khodi se-t5-v1_1-xxl-encoder-ggufAmafayela okuthi `clip_l.safetensors` kanye nokuthi `ae.safetensors` kufanele abekwe kumafolda awo ngokulandelana kwawo. Kunamacala okusebenzisa alula ane-6 GB ye-VRAM.
Ukusebenzisa i-FLUX.1 efwini nakwezinye izinsiza
Uma ukhetha ukungafaki lutho, ungazama i-FLUX in Izikhala Zobuso Abangana: I-FLUX.1-dev y I-FLUX.1-schnell. Futhi ku- Phindaphinda, Mystic.ai o fal.ai. Lezi izinketho eziwusizo zokuqinisekisa imiyalo nezilungiselelo ngaphambi kokulanda amamodeli endawo.
Ukuze uthole ugqozi nokugeleza komsebenzi osekwenziwe, hlola izibonelo ezisemthethweni ze-ComfyUI kanye namagalari okuhamba komsebenzi afana I-OpenArt. Khumbula ukuthi izithombe eziningi ziqukethe imethadathaukuze ukwazi ukuwahudulela ku-ComfyUI ukuze ubuyise igrafu.
Izinto ezengeziwe: amaqoqo we I-LoRA ye-FLUX njengoba OngokoqoboLora noma ukuhlanganiswa ku XLabs-AII-ControlNet ye-FLUX njenge amaqoqo y Union; I-adaptha ye-IP en XLabs-AIUkuze uqeqeshe i-LoRA nge-VRAM ephansi, zama i-fluxgym noma umqeqeshi wokuphindaphinda ovela ku-Ostris; Kukhona umhlahlandlela we-DreamBooth we-FLUX.1 dev ku-diffusers repo.
Faka i-ComfyUI ku-Windows 11 ngokusebenza okuqinile
Uma uthanda ukufakwa okuhlanzekile, lena indlela efakazelwe. Isebenza kahle kakhulu ngamakhadi ochungechunge lwe-NVIDIA 40/50 futhi ikusiza ukuthi ugweme amaphutha ajwayelekile.
1) Faka uhlelo lokusebenza lwe-NVIDIA kanye nomshayeli weSitudiyo kusuka nvidia.com. Qala kabusha. 2) CUDA Toolkit kusuka developer.nvidia.com (kuyakhetheka kodwa kuwusizo ukugwema izexwayiso nge-Triton uma ungasebenzisi i-venv). 3) ffmpeg Ukusuka endaweni yokugcina ye-BtbN, engeza u-C:\ffmpeg\bin ku-Path. 4) I-Git yeWindows kusuka git-scm.com5) Python 3.12 x64 kusuka python.org, ngesiqalisi se-py sabo bonke abasebenzisi futhi wengeze kokuguquguqukayo kwemvelo.
5.5) Uma ukhetha i-venv, yidale nge python -m venv CUVenv futhi uyivule nge-CUVenv\Scripts\activate.bat. Ukusuka lapho, sebenzisa noma yimiphi imiyalo ye-pip noma ye-git ngaphakathi kwe-venv. Isikripthi sakho sokuqalisa singenza imvelo isebenze futhi siqalise i-ComfyUI ekuhambeni okukodwa.
6) Clone ComfyUI nge git clone https://github.com/comfyanonymous/ComfyUI.git D:\CU7) Iya ku-D:\CU bese usebenzisa i-pip install -r needs.txt. 8) Uma i-pip ixwayisa ngemibhalo engaphandle kweNdlela, engeza indlela ye-Python Scripts kuziguquko zesistimu bese uqala kabusha. 9) Faka i-PyTorch CUDA 12.8 Faka ithoshi usebenzisa `pip install torchvision torchaudio --index-url https://download.pytorch.org/whl/cu128`. Uma kukhona okubonakala kucishiwe, khipha ithoshi bese ulifaka kabusha usebenzisa umyalo ofanayo.
9 bis) Yethula i-ComfyUI nge python main.py futhi uvule 127.0.0.1:8188. 10) Faka i-Triton yeWindows nge-pip install -U triton-windows. 11) Sheshisa ukunakwa nge-Sage Attention 2.2: Landa isondo elihambisanayo i-CU128/Torch2.8 le-CP312, yifake ngepayipi bese uqala i-ComfyUI ngefulegi -use-sage-attention.
12) Faka i-ComfyUI-Manager: ku-ComfyUI/custom_nodes run git clone https://github.com/ltdrdata/ComfyUI-Manager comfyui-manager13) Dala ifayela le-boot .bat elinemigqa: cd D:\CU kanye ne-python main.py -use-sage-attention. Lapho uqalisa, Umphathi uzothatha isikhashana okokuqala.; hlola ukuthi ithebhu yoMphathi iyavela kusixhumi esibonakalayo.
14) Beka izifanekiso kumafolda alungile (izindawo zokuhlola, isiqeshana, i-unet, vae) bese uvula ukuhamba kwakho. Uma ukugeleza komsebenzi kuletha inodi yayo ye-Sage, ungayiyeka uma usuvele uqala ngefulegiAmathiphu: Gwema ukuba nezinhlelo ezidinga izinsiza ezivulekile, lungiselela inkumbulo ebonakalayo ye-Windows uma uphelelwa amandla, futhi uhlole izingxoxo zokusebenza endaweni yokugcina ye-ComfyUI. Uma uthola izexwayiso zenkumbulo, cabanga ukusebenzisa okuhlukile kwefayela elilodwa le-FP8.
Uma ukhetha imanuwali elandekayo, kune-PDF ewusizo lesi sixhumanisi. Eminye imihlahlandlela ibuyekeziwe ukuze kube lula ukufaka nokuphakamisa i-venvUma usebenzisa i-venv, khumbula ukusebenzisa i-pip ne-git ngaphakathi kwemvelo.
Amathiphu okusebenzisa, ukuhambisana nemikhuba engcono kakhulu
Ukungenisa kugeleza ngokuhudula ama-PNG anemethadatha noma i-JSON futhi uhlole izinguqulo zenodi ngoMphathi. Lapho wabelana ngezithombe, gwema ukucindezela okususa imethadatha.Uma i-JSON iphuka enguqulweni entsha, buyisela amanodi angasasebenzi noma faka izinguqulo ezihambisanayo.
Ukuze usebenze nama-FLUX LoRA amaningi, kunemibiko yokusebenzisa amandla aphezulu ekwakhiweni kwakamuva kwe-ComfyUI; Hlola i-GGUF noma izilayishi ezithile Ukuze unciphise i-VRAM. Ku-ControlNet, qala ngokujula noma i-Canny ukuze usungule i-rig yokuhlanganisa ezinzile.
Lapho uthumela efwini, hlola i-VRAM nolayini bokwenza. Endaweni, i-SSD esheshayo kanye nabashayeli bakamuva benza wonke umehluko.Bhala ipayipi lakho ngezigaba: ukulayishwa kwemodeli, ukubeka isimo, ukusampula, ukuqopha, kanye nokucubungula ngemva kokucubungula. Lokhu kuzokwenza ukulungisa iphutha kube lula uma okuthile kuphuka.
Ngakho konke okungenhla, manje ungakha ipayipi le-VFX eliqinile ngempela nge-ComfyUI: Ingabe uyakuqonda ukuthi ukugeleza komsebenzi kuyini nokuthi ungakulondoloza kanjani ngaphandle kokulahlekelwa imethadatha?Ungakwazi ukufaka i-FLUX nezinhlobonhlobo zayo (i-Dev, i-Schnell ne-FP8, i-NF4, izinketho ze-GGUF), uyazi ukuthi uqhuba kanjani ukuhamba komsebenzi okubalulekile kokuzijwayeza kwansuku zonke (Txt2Img, Img2Img, Inpainting, ControlNet, LoRA, IPAdapter kanye ne-Upscale) futhi unokufakwa okuthuthukisiwe kweWindows nge-Triton, i-Comfyger toUsItable kanye ne-Finent Atten System.