Ukuba usebenza kwi-VFX kwaye unomdla wokuthatha ukuhamba komfanekiso wakho ukuya kwinqanaba elilandelayo, iComfyUI kufuneka ube nayo. Indlela yayo esekwe kwi-node, ukumodareyitha, kunye namandla ayongezwayo yi-FLUX Oku kuyenza ibe yindawo efanelekileyo yokuhlola yonke into ukusuka kwi-high-fidelity text2img ukuya kwimibhobho enzima kunye ne-ControlNet, i-IP-Adapter, okanye uqeqesho lwe-LoRA. IComfyUI yeVFX: Ufakelo kunye nokuhamba komsebenzi kweFLUX kuchaziwe.
Kule khokelo uya kufumana, ngendlela ehlelekile, yonke into oyifunayo: yintoni i-workflow kwi-ComfyUI, indlela yokugcina nokwabelana ngayo, indlela yokufaka i-FLUX inyathelo ngesinyathelo, yiyiphi i-hardware oyifunayo, kunye nenkcazo ecacileyo yeendlela ezisetyenziswa kakhulu (Txt2Img, Img2Img, Inpainting, LoRA, ControlNet, NF4, IPAdapter kunye ne-latent). Uya kubona ezinye iindlela ezifana neFP8, NF4, okanye GGUF ukunciphisa iVRAMIzixhobo zokusebenzisa iFlux efini, kunye nofakelo lweComfyUI olulungiselelwe ngeWindows oluneengcebiso ezisebenzayo.
Yintoni ukuhamba komsebenzi kwiComfyUI?
Ukuhamba komsebenzi kukubonakaliswa okubonakalayo kwenkqubo yesizukulwana oyakhayo ngokujoyina iindawo. Unokuyicingela njengebhodi yohlobo lweLego yamaqhekezaI-node nganye yenza umsebenzi (imodeli yokulayisha, i-encoding text, isampuli, i-decoding, njl. njl.) kunye nodibaniso luchaza indlela elandelwa yinkcazelo de kufunyenwe umfanekiso wokugqibela.
I-ikhosistim inkulu kakhulu: umfanekiso, ividiyo, iaudio kunye ne-3D Zihlala ngokubulela kuluntu oluvulelekileyo. I-downside kukuba kukho i-curve yokufunda, kuba kubalulekile ukuqonda ukuba yintoni icandelo ngalinye lenzayo kunye nendlela yokudibanisa ngokufanelekileyo ukuphepha iibhotile okanye iziphumo ezingahambelani.
Ukuze uqalise ngokukhawuleza, kungcono ukujonga ukuhamba okusemthethweni kunye nesiseko (i-text2img kunye ne-img2img) kwaye emva koko uqhubele phambili kwiindawo ezinzima kakhulu. Amaxwebhu oluntu kunye nemizekelo esemthethweni yeComfyUI Zihlaziywa rhoqo, ngoko kulula ukuqalisa kwakhona okanye ukuphonononga ukuhamba komsebenzi okutshintsha ngeenguqulelo ezintsha.
Enye ingcaciso ebalulekileyo: I-ComfyUI inokushumeka ukuhamba komsebenzi kumfanekiso wokugqibela (PNG) ngokwawo njengemethadatha. Oku kukuvumela ukuba wabelane ngomfanekiso kuphela kwaye ufumane igrafu yonke. ukuyitsala umva kujongano.
Ufunda njani kwaye uqhubele phambili nokuhamba komsebenzi
Ingcebiso yam kukuqala ngemizekelo elula epapashwe kwimithombo yohlobo lwe-wiki kunye neegalari zoluntu. Ukuqhubela phambili okunengqiqo kukuba: Txt2Img, Img2Img, kwaye emva koko ControlNet okanye LoRANjengoko uqonda amagalelo, iziphumo, kunye nendlela abacwangcisi abayichaphazela ngayo iisampulu, kuya kuba yindalo ukubandakanya iindawo zesikhokelo, iimaski, kunye neemeko ezongezelelweyo.
Ukuba ufuna ukubona ii-topology ezahlukeneyo, kukho iindawo zokugcina zokuqukuqela kukawonke-wonke kunye namaphepha anemizekelo ephinda-phindayo apho unokukhuphela imifanekiso ngemetadata okanye iifayile ze-json. Kuluqhelo olulungileyo ukurhweba ngaphandle, ukuqhuba njengoko kunjalo, kwaye emva koko uphindaphinda iiparamitha ukuqonda impembelelo yebhloko nganye ngaphandle kokuphula yonke.
Kwiiplatifomu zamafu uya kufumana kwakhona iimeko-bume ezicwangcisiweyo zokuqhuba imibhobho ngaphandle kokuzabalaza nokuxhomekeka kwendawo. Inzuzo kukuba balayisha kwangaphambili ii-node ezinzima kunye neemodeliNangona kunjalo, kuyacetyiswa ukuba ujonge iinguqulelo ezikhoyo kunye ne-VRAM ukuqinisekisa ukuba iziphumo ziyahlangabezana nokulindelweyo.
Ukugcina, ukungenisa, kunye nokuthumela ngaphandle ukuhamba komsebenzi kwiComfyUI
IComfyUI ixhasa iindlela ezimbini eziphambili zokugcina: umfanekiso onemetadata (PNG) okanye ifayile yeJSON yegrafu. Eyokuqala yeyona ilungele ukwabelana ngamaqonga; eyesibini ikunika ulawulo olucacileyo lwefayile, iluncedo kuguqulelo.
Ukungenisa, tsala ngokulula iPNG okanye ifayile yeJson kujongano, okanye sebenzisa indlela emfutshane Ctrl (Command) + O. Ukuthumela ngaphandle, ungagcina umfanekiso owenziweyo okanye usebenzise imenyu yokuThumela ngaphandle kweJSON.Lumka xa ucinezela okanye ugqithisela imifanekiso: ezinye iindlela zokucinezela kunye netshaneli ezithile zisusa imetadata, ezinokubangela ukuba ulahlekelwe kukuhamba komsebenzi okuzinzisiweyo.
Ngenxa yokuzivelela rhoqo kweComfyUI, Ayizizo zonke iifayile zeJSON ezindala ezisebenza kwiinguqulelo ezintsha.Ukuba kukho into engahambi kakuhle, vula ukuhamba, buyisela ii-nodes eziphelelwe lixesha, okanye ubuyisele ukuxhomekeka kunye noguqulelo oluhambelanayo; usebenzisa iComfyUI-Manejala yenza ukuba kukhawuleze kakhulu ukubona nokusombulula izinto ezingekhoyo.
I-FLUX kwi-ComfyUI: yintoni kwaye kutheni ibalulekile
I-FLUX.1 yintsapho yeemodeli ezivela kwi-Black Forest Labs egxile kwi-high-fidelity text-to-image conversion. Ubume bayo obuxubeneyo obumalunga ne-12 yeebhiliyoni zeeparamitha Yenzelwe ubambelelo olukhawulezileyo, ukuphatha imiboniso entsonkothileyo, kunye nokuvelisa isicatshulwa esifundekayo ngaphakathi komfanekiso, umsebenzi apho ezinye iimodeli zihlala zihexa.
Enye inzuzo: ukuguquguquka kwayo. Ukusuka kwi-photorealism ukuya kwizitayile zobugcisaI-FLUX.1 igqamile ngokuhambelana kwayo okubonakalayo kunye neenkcukacha, kubandakanya ukunikezelwa kwezandla, indawo ebuthathaka yakudala kwimizobo evelisayo. Akumangalisi ukuba kuthelekiswa nezisombululo ezifana ne-Stable Diffusion okanye i-Midjourney, ephuma phezulu malunga nokulula kokusetyenziswa kunye nomgangatho.
I-Black Forest Labs yasungulwa nguRobin Rombach, umntu oyintloko owayesebenza kwi-core of Stability AI. Ukuba ufuna ukuthatha ujongo lokuqala, Iwebhusayithi yayo esemthethweni ikhona blackforestlabs.ai.
I-FLUX.1 isasazwe ngeendlela ezintathu: UPro, uDev kunye noSchnellI-Pro inikezela ngowona mgangatho uphezulu weendawo zobuchwephesha; I-Dev yenzelwe ukusetyenziswa okungeyontengiso kunye nebhalansi ebalaseleyo; U-Schnell ugxile kwisantya kunye nokukhanya, kwaye ungumthombo ovulekileyo phantsi kwelayisensi ye-Apache 2.0.
Iimfuno zehardware ngoguqulelo lweFLUX
Kwi-FLUX.1 Pro, kuyacetyiswa I-NVIDIA RTX 4090 chwetheza i-GPU ene-24 GB ye-VRAMI-32 GB ye-RAM kunye ne-SSD ekhawulezayo. Isebenzisa iFP16 ukunqanda ukuphuma kwebhokisi, kwaye kungcono ukusebenzisa i-encoder yombhalo kwiFP16 kumgangatho ophezulu.
Kwi-FLUX.1 Dev, a I-RTX 3080/3090 ene-16 GB ye-VRAM Isebenza kakuhle kunye ne-16 GB ye-RAM kunye ne-25 GB yendawo yediski. Ixhasa iFP16 kunye neFP8 kwezinye iimeko, kuxhomekeke kwiGPU yakho.
Ye-FLUX.1 Schnell, I-RTX 3060/4060 ene-12 GB ye-VRAM I-8GB ye-RAM kunye ne-15GB yokugcina yanele. Yenzelwe isantya, ukuncama ezinye zentloko yokusebenza xa kuthelekiswa neenguqulelo zePro/Dev.
Ukuba umfutshane kwinkumbulo, uluntu lubonelela ngeendlela ezizezinye ezifana neFP8, NF4, okanye GGUF leyo Bayinciphisa kakhulu iVRAM efunekayo, kunye noqwalaselo oluqala kwi-6 ukuya kwi-12 GB kuxhomekeke ekuhambeni.
Ukufaka iFLUX kwiComfyUI: amanyathelo abalulekileyo
Okokuqala, qiniseka ukuba usebenzisa inguqulelo yamva nje yeComfyUIUkudityaniswa kwe-FLUX kufuna ii-nodes kunye neempawu ezihlaziywa rhoqo.
Khuphela okubhaliweyo kunye nezifaki khowudi ze-CLIP: ikliphu_l.izikhuselo kunye nenye yeefayile ze-T5 XXL, nokuba yi-t5xxl_fp16.safetensors (ukuba une-VRAM/RAM eninzi) okanye t5xxl_fp8_e4m3fn.safetensors (ukuba unohlahlo lwabiwo-mali olungqongqo). Zibeke kwiComfyUI/models/clip/ folder. Ukuba usebenzise i-SD3 Medium, usenokuba sele unazo ezi fayile..
VAE: Khuphela ae.izikhuselo kwaye uyihambise kwiComfyUI/models/vae/. Yithiye kwakhona ku-flux_ae.safetensors ukuba ufuna ukuyifumana ngokulula. Le VAE iphucula i-decoding yokugqibela kwaye ingundoqo kumgangatho.
UNET: khetha phakathi i-flux1-dev.safetensors okanye i-flux1-schnell.safetensors ngokwenkumbulo yakho, kwaye uyibeke kwi ComfyUI/models/unet/. Ngale nto, unesiseko sokuqhuba ukuhamba kweFLUX. yendawo.
Isikhokelo esisebenzayo sokuhamba komsebenzi kwe-FLUX kwi-ComfyUI
Txt2Img kunye FLUX
Qala ngokulayisha amacandelo: UNETLoader, DualCLIPLoader kunye neVAELaderI-CLIPTextEncode node ifaka iikhowudi kwisaziso sakho; I-EmptyLatentImage yenza umfanekiso ofihlakeleyo wokuqala; I-BasicGuider ikhokela inkqubo ngokudibanisa ingqiqo enemiqathango kunye ne-UNET ye-FLUX.
Khetha isampuli nge KSamplerKhethaYenza ingxolo ngeRandomNoise kwaye ichaza irempu ye-sigma nge-BasicScheduler. ISamplerCustomAdvanced imanyanisa yonke into: ingxolo, isikhokelo, isampulu, i-sigmas, kunye ne-latent. Okokugqibela, i-VAEDecode iguqula umqondiso ofihlakeleyo ube ngumfanekiso. kwaye ngeSaveImage ugcina isiphumo.
Img2Img kunye neFLUX
Umbhobho wongeza umfanekiso wokuqala: LayishaUmfanekiso + Isikali somfanekiso Ubungakanani bulungelelanisiwe, kwaye i-VAEEcode iyiseta kwi-latent. I-prompt ifakwe ngekhowudi kunye ne-CLIPexteEncode kwaye amandla ayo ahlengahlengiswa kunye ne-FluxGuidance. I-ModelSamplingFlux ilawula ukutshintshwa kunye nobukhuluKSamplerSelect, RandomNoise, kunye noSicwangcisi esiSisiseko siphatha isampulu. I-SamplerCustomAdvanced idibanisa i-conditional kunye ne-input latent, kwaye i-VAEDecode ivelisa imveliso.
LoRA kunye FLUX
Ukucokisa isitayile okanye iimpawu, yongeza LoraLoaderModelOnly kunye ne-UNETLoader, i-DualCLIPLoader, kunye ne-VAELader. Emva kokufaka ikhowudi yokubhaliweyo kunye nokusebenzisa i-FluxGuidance, udala umfanekiso ofihlakeleyo nge-EmptyLatentImage, chaza isampuli nge-ModelSamplingFlux, kwaye usebenzise i-SamplerCustomAdvanced. Nge-VAEDecode ufumana umfanekiso osele uphenjelelwe yi-LoRA. Umzekelo oqhelekileyo: realism_lora.safetensors kwi-flux1-dev.
ControlNet nge FLUX
Iimeko ezimbini eziluncedo kakhulu ze-VFX: ubunzulu kunye nemiphetho ye-canny. Ubunzulu, preprocess nge MiDaS-DepthMapPreprocessorLayisha ubunzulu be-ControlNet kwaye uyisebenzise nge-ApplyFluxControlNet. Sebenzisa i-XlabsSampler ukwenza i-conditional latent waveform, kwaye emva koko I-VAEDecode ivelisa umfanekiso.
KwiCanny, sebenzisa CannyEdgePreprocessor, layisha iCanny ControlNet kwaye uphinde inkqubo: Faka isicelo seFluxControlNet → XlabsSampler → VAEDecode. Olu lawulo olongezelelweyo lubonelela ngokuchanekileyo phezu kwemilo kunye nokwakheka..
Ukupeyinta ngeFLUX
Layisha i-UNET, i-VAE kunye ne-CLIP, kwaye ulungiselele iimpembelelo ezilungileyo nezingalunganga. LoadAndResizeImage izisa umfanekiso kunye nemaskiUtshintsho lugudiswe nge-ImpactGaussianBlurMask. I-InpaintModelConditioning idibanisa ukulungiswa, umfanekiso, kunye nemaski. Emva kokuqwalasela isampulu, ingxolo, kunye neesigmas, iSamplerCustomAdvanced iphinda yakha kwakhona ummandla ofihliweyo. I-VAEDecode idibanisa i-patch ngokuqhubekayo kunye nabanye.
I-FLUX NF4
Ngomlinganiselo we-NF4, imemori iyancitshiswa. Layisha amacandelo nge I-CheckpointLoaderNF4 kwaye ichaza ubude/ububanzi ngeenodi zamandulo. I-ModelSamplingFlux ibeka iiparamitha; I-EmptySD3LatentImage yenza umfanekiso ofihlakeleyo; I-BasicScheduler kunye neRandomNoise ziququzelela i-denoise. ISamplerCustomAdvanced yenza isiginali efihlakeleyo kwaye iVAEDecode iyiguqulela kumfanekiso.Ukukala, i-UltimateSDUpscale, kunye ne-UpscaleModelLoader kunye ne-prompt eyongezelelweyo elungileyo, yenza wonke umahluko.
IPadapter eneFLUX
Xa ufuna ukulungisa usebenzisa umfanekiso wereferensi, sebenzisa LoadFluxIPAdapter kunye ne-ApplyFluxIPAdapter kunye neclip_vision_l.safetensors. Linganisa umfanekiso wereferensi nge-ImageScale, lungiselela izivuseleli, kwaye usebenzise i-XlabsSampler. Nge-VAEDecode uya kubona imveliso ephenjelelwa bubuhle okanye iimpawu ukusuka kumfanekiso wesikhokhelo.
LoRA umqeqeshi for FLUX
Ukuqeqesha i-LoRA ngokuthe ngqo kwi-ComfyUI, ukuhamba komsebenzi kubandakanya: FluxTrainModelSelect, OptimizerConfig and TrainDatasetGeneralConfigI-InitFluxLoRATraining iqalisa, iFluxTrainLoop iphumeza amanyathelo, kwaye iFluxTrainValidate ivelisa ukuqinisekiswa kwamaxesha athile.
NgeVisualizeLoss ulandelela ilahleko; ImageBatchMulti kunye ImageConcatFromBatch Baqokelela ukuqinisekiswa; I-FluxTrainSave igcina iindawo zokukhangela, kwaye iFluxTrainEnd ivala inkqubo. Ukuba uyafuna, layisha isiphumo kuBuso obukwangayo nge-UploadToHuggingFace kwaye wabelane ngayo.
Flux Latent Upscaler
Ukukala ngokweenkcukacha, chaza ubungakanani nge SDXLEmptyLatentSizePicker+ kunye namatyathanga eLatentUpscale kunye neLatentCrop. Ngeemaski ezenziwe yiSolidMask kunye neFeatherMask, iLatentCompositeMasked idibanisa i-letent yesikali kunye neyokuqala. I-InjectLatentNoise+ yongeza iinkcukacha phambi kwe-VAEDecodeKwaye ucofe nge-ImageSmartSharpen+ igqibezela inkqubo. Iinowudi zokubala njengeSimpleMath + zinceda ukulungelelanisa imilinganiselo.
Iinguqulelo ezizezinye: FP8, NF4 kunye neGGUF ukuthoba iVRAM
Ukuba umfutshane kwizibonelelo, unokhetho. Iindawo zokukhangela zeFP8 ezivela kwiComfy.org kunye nababhali abafana noKijai. ikuvumela ukuba usebenzise iFLUX ngefayile enye kwi-ComfyUI/iimodeli/iindawo zokukhangela/. Kuyacetyiswa ukuthiya ngokutsha okanye ukwahlula kwiifolda ukwahlula phakathi kwe dev kunye ne schnell ezahlukeneyo.
Nge-NF4 (bitsandbytes), faka iplagin ComfyUI_bitsandbytes_NF4 kunye nokusetyenziswa flux1-dev-bnb-nf4-v2 kwiimodeli / iindawo zokukhangela. Le nguqulo iphucula iinkcukacha xa kuthelekiswa nokuphindaphindwa kokuqala.
Ubungakanani be-GGUF yeCity96, kunye neplagin I-ComfyUI-GGUFyehlisa ibar nangakumbi: Khuphela imodeli yeFLUX GGUF, i-t5-v1_1-xxl-encoder-gguder encoderIifayile `clip_l.safetensors` kunye `ae.safetensors` kufuneka zibekwe kwiifolda zazo. Kukho iimeko zokusetyenziswa ezifanelekileyo kunye ne-6 GB ye-VRAM.
Ukusebenzisa i-FLUX.1 kwifu kunye nezinye izixhobo
Ukuba ukhetha ukungafaki nantoni na, ungazama i-FLUX ngaphakathi Izithuba zoBuso obukwangayo: FLUX.1-dev y FLUX.1-schnell. Kwakhona kwi Phindaphinda, Mystic.ai o fal.ai. Ezi ziinketho eziluncedo zokuqinisekisa izivuseleli kunye nolungelelwaniso phambi kokukhuphela iimodeli zasekhaya.
Ukuphefumlelwa kunye nokuhamba komsebenzi okulungeleyo, jonga imizekelo esemthethweni yeComfyUI kunye neegalari zokuhamba komsebenzi ezifana I-OpenArt. Khumbula ukuba imifanekiso emininzi inemetadatangoko ungazitsalela kwiComfyUI ukufumana igrafu.
Izinto ezininzi: ingqokelela ye I-LoRA ye-FLUX njengaye UkwenyaniLora okanye ukuhlanganiswa kwi XLabs-AIControlNet for FLUX njengoko ingqokelela y umanyano; IP Adapter en XLabs-AIUkuqeqesha i-LoRA ngeVRAM ephantsi, zama Fluxgym okanye uReplicate umqeqeshi osuka kuOstris; Kukho i-DreamBooth isikhokelo se-FLUX.1 dev kwi-diffusers repo.
Faka iComfyUI Windows 11 ngokusebenza okuqinileyo
Ukuba uthanda ukufakela okucocekileyo, le yindlela eqinisekisiweyo. Isebenza kakuhle kakhulu ngamakhadi e-NVIDIA 40/50 kwaye ikunceda uphephe iimpazamo eziqhelekileyo.
1) Faka i-NVIDIA app kunye nomqhubi weStudio ukusuka nvidia.com. Phinda Uqalele. 2) CUDA Toolkit ukusuka developer.nvidia.com (ukhetho kodwa luluncedo ukunqanda izilumkiso ngeTriton ukuba awusebenzisi i-venv). 3) ffmpeg Ukusuka kwindawo yokugcina yeBtbN, yongeza iC:\ffmpeg\bin kwiNdlela. 4) IGit yeeWindows ukusuka git-scm.com5) Python 3.12 x64 ukusuka python.org, kunye nesiqalisi se-py kubo bonke abasebenzisi kwaye ungeze kwizinto eziguquguqukayo zokusingqongileyo.
5.5) Ukuba ukhetha i-venv, yenze nge ipithon -m venv CUVenv kwaye uyivule ngeCUVenv\Scripts\activate.bat. Ukusuka apho, sebenzisa nayiphi na i-pip efanelekileyo okanye imiyalelo ye-git ngaphakathi kwe-venv. Iskripthi sakho sokuqalisa sinokuvula okusingqongileyo kwaye siqhube iComfyUI ekuhambeni kunye.
6) Clone ComfyUI nge git clone https://github.com/comfyanonymous/ComfyUI.git D:\CU7) Yiya ku-D:\CU kwaye usebenzise ipip install -r requirements.txt. 8) Ukuba i-pip ilumkisa ngezikripthi ngaphandle kweNdlela, yongeza indlela ye-Python Scripts kwindlela eguquguqukayo kwaye uqalise kwakhona. 9) Faka iPyTorch CUDA 12.8 Faka itotshi usebenzisa `ipip yokufaka itorchvision torchaudio --index-url https://download.pytorch.org/whl/cu128`. Ukuba kukho nantoni na ebonakala ivaliwe, yikhuphe itotshi kwaye uyibuyisele kwakhona usebenzisa umyalelo ofanayo.
9 bis) Qalisa iComfyUI nge python main.py kwaye uvule 127.0.0.1:8188. I-10) Faka i-Triton ye-Windows kunye ne-pip install -U triton-windows. 11) Khawulezisa ingqalelo ngeSage Ingqalelo 2.2: Khuphela ivili ehambelanayo CU128/Torch2.8 for CP312, yifake ngombhobho kwaye uqalise iComfyUI ngeflegi -sebenzisa-sage-ingqwalasela.
12) Faka i-ComfyUI-Manejala: kwi-ComfyUI/custom_nodes run git clone https://github.com/ltdrdata/ComfyUI-Manager comfyui-manager13) Yenza ifayile ye-boot .bat enemigca: cd D:\CU kunye nepython main.py -use-sage-attention. Xa uqalisa, uMphathi uya kuthatha ixeshana okokuqala.; khangela ukuba uMphathi wesithuba uyavela kujongano.
14) Beka iitemplates kwiifolda ezichanekileyo (i-checkpoints, ikliphu, i-unet, i-vae) kwaye uvule ukuhamba kwakho. Ukuba ukuhamba komsebenzi kuzisa indawo yayo yeSage, ungayishiya ukuba sele uqala ngeflegiIingcebiso: Kuphephe ukuba neenkqubo ezinzulu zobutyebi ezivulekileyo, qwalasela inkumbulo yeWindows yenyani ukuba usebenza ephantsi, kwaye ujonge iingxoxo zokusebenza kwindawo yokugcina iComfyUI. Ukuba ufumana izilumkiso zememori, cinga ukusebenzisa uhlobo lwefayile enye FP8.
Ukuba ukhetha incwadi ekhutshelweyo, kukho iPDF eluncedo kuyo esi sixhobo. Ezinye izikhokelo ziye zahlaziywa ukwenza lula ufakelo kunye nokuphakamisa ivenvXa usebenzisa i-venv, khumbula ukuhlala uqhuba ipip kunye ne-git ngaphakathi kokusingqongileyo.
Iingcebiso zokusetyenziswa, ukuhambelana kunye nezenzo ezilungileyo
Ngenisa uhamba ngokutsala ii-PNG ezinemethadatha okanye i-JSON kwaye ujonge iinguqulelo zenodi kunye noMphathi. Xa usabelana ngemifanekiso, kunqanda ucinezelo olucima imetadata.Ukuba i-JSON iphula kuguqulelo olutsha, buyisela iindawo eziphelelwe lixesha okanye ufake iinguqulelo ezihambelanayo.
Ukusebenza kunye ne-FLUX LoRAs ezininzi, kukho iingxelo zokusetyenziswa kwamandla aphezulu kwi-ComfyUI yakutshanje yakha; Vavanya i-GGUF okanye izilayishi ezithile Ukunciphisa iVRAM. Kwi-ControlNet, qala ngobunzulu okanye i-Canny ukuseka i-rig ezinzileyo yokuqamba.
Xa uthumela efini, jonga i-VRAM kunye nemigca yophumezo. Ekuhlaleni, i-SSD ekhawulezayo kunye nabaqhubi abahlaziyiweyo benza wonke umahluko.Bhala umbhobho wakho ngokwamacandelo: ukulayishwa kwemodeli, ukulungiswa, iisampulu, ukuguqula iikhowudi, kunye nokulungiswa kwasemva. Oku kuya kwenza ukulungisa iimpazamo kube lula xa kukho into eqhekekileyo.
Ngako konke oku kungasentla, ngoku unokwakha umbhobho weVFX owomeleleyo ngeComfyUI: Ngaba uyakuqonda ukuba yintoni ukuhamba komsebenzi kunye nendlela yokuyigcina ngaphandle kokulahlekelwa imethadatha?Unako ukufaka i-FLUX kunye neentlobo zayo (i-Dev, i-Schnell kunye ne-FP8, i-NF4, i-GGUF iinketho), uyazi ukuba uqhuba njani ukuhamba komsebenzi ongundoqo wemihla ngemihla (Txt2Img, Img2Img, Inpainting, ControlNet, LoRA, IPAdapter kunye ne-Upscale) kwaye une-Windows ephuculweyo yofakelo kunye ne-Triton, i-Sage UsEtAttention kunye ne-FinentAtten system.