Ukuba ufuna ukuntywila kuqala kwi-AI-powered image kunye nokudalwa kwevidiyo ngaphandle kokuphulukana nolawulo oluhle, ufike kwindawo elungileyo: I-ComfyUI lujongano olusekwe kwindawo ekuvumela ukuba wakhe owakho umbhobho njengezitena ze-LEGO. Kwesi sikhokelo uya kufunda, ukusuka ekuqaleni kunye neenkcukacha, indlela yokuseta iziphumo ezibonakalayo (VFX) ukuhamba komsebenzi kwiComfyUI., kuzo zombini umfanekiso kunye nevidiyo, ngaphandle kokuphosa nantoni na ebalulekileyo.
Ukongeza kwizinto ezisisiseko, siza kujonga ukuhamba kwesicatshulwa-kuya-kumfanekiso, ukuhamba komfanekiso ukuya kumfanekiso, ukupeyinta, ukupeyinta, ukukala, i-ControlNet, i-SDXL, i-LoRA, kunye nokufakela. Siza kwenza umtsi ukuya kwividiyo ngeAnimateDiff, iHunyuanVideo, iVidiyo yeLTX kunye neWan 2.1Oku kubandakanya iimfuno, ukufakwa, iiparamitha eziphambili, kunye neengcebiso zokuvelisa kunye neendlela ezimfutshane kunye nabaphathi bee-node. Siza kugubungela iinketho zamafu ukuba ukhetha ukunqanda ufakelo oluntsonkothileyo.
Yintoni iComfyUI kwaye kutheni ifanelekile kwiVFX?
I-ComfyUI yi-GUI esekwe kwi-node ye-Stable Diffusion ekuvumela ukuba ujonge kwaye uguqule ukuhamba kwedatha ukusuka ekuqaleni ukuya ekugqibeleni. Indawo nganye yenza umsebenzi othile (ukulayisha imodeli, i-encoding text, isampuli, i-decoding VAE, njl. njl.) kwaye iqhagamshelwe ngeentambo. emele amagalelo kunye neziphumo. Le filosofi ifanelekile kwi-VFX: uyazi kakuhle apho umqondiso ungena khona, apho uguqulwa khona, kunye nendlela yokuchaphazela umphumo.
Xa kuthelekiswa nojongano lwe-monolithic, iComfyUI ibalasele ngokungafihlisi kunye nokuguquguquka. Ixabiso laloo nkululeko ligophe lokufunda kunye nokuphazamiseka okubonakalayo. (ukuhamba komsebenzi ngamnye kunokubekwa ngokwahlukileyo), kodwa umvuzo ukwazi ukwenza iprototype ngokukhawuleza, ukulungisa ingxaki ngokuchanekileyo, kunye nokwabelana ngokuqhutywa komsebenzi ngendlela ephinda iphindwe.
I-ComfyUI vs. AUTOMATIC1111
Abasebenzisi abaninzi bavela kwi-AUTOMATIC1111, iklasikhi ye-Stable Diffusion. IComfyUI izuza ekukhanyeni, elubala kunye nesakhono sokuchwethezaI-A1111 ivakalelwa ngakumbi kwaye iqonde, kodwa incinci. Ukuba ufuna ukuqonda ukusebenza kwangaphakathi kwaye ufumane okuninzi kwiVFX, iComfyUI lubhejo olukhuselekileyo.
Amanyathelo okuqala kunye nolawulo olusisiseko
Ukusebenzisana ne-canvas kulula: sondeza ngevili okanye i-pinch gesture, tsala ukuhamba, kwaye wenze uxhulumaniso ngokutsala ukusuka kwisiphumo se-node enye ukuya kwigalelo lenye. Uza kubona iibhloko (iindawo) njengeNqanawa yokuHlola, i-CLIP i-Encoud yokubhaliweyo, i-KSampler okanye i-VAE, kunye neentambo ezimele indlela yedatha.
Umbhalo kumfanekiso: ukuhamba kwesiseko kunye neendawo ezibalulekileyo
Umbhobho oqhelekileyo uquka ukulayisha indawo yokukhangela, ukufakwa kweekhowudi kwi-prompt, isampuli kwi-latents, kunye ne-decoding kwii-pixels. Lo ngumathambo ekwakhiwe kuwo phantse yonke into kwiComfyUI..
Ukukhetha imodeli eneNqanawa yokuHlola
I-node ye-Load Checkpoint ihambisa iziqwenga ezintathu: i-MODEL (inethiwekhi yokubikezela ingxolo), i-CLIP (i-encoder yombhalo) kunye ne-VAE (ukusuka kwiipikseli ukuya kwi-latents kunye ne-vice versa). I-MODEL ityisa i-KSampler, i-CLIP iya kwiindawo zokubhaliweyo, kwaye i-VAE isetyenziselwa ukucacisa isiphumo sokugqibela.Ngaphandle kwendawo yokukhangela akukho mdlalo, ke khetha enye ehambelana nokuhamba komsebenzi wakho.
Iingcebiso ezilungileyo nezingalunganga nge-CLIP Text Encode
Sebenzisa iinodi ezimbini ze-CLIP Text Encode: eyona iphezulu ye-positive kunye nesezantsi ye-negative. Umbhalo uguqulwa ube yimixholo enesombululo ephezulu ekhokela usasazoUngalinganisa amagama ngesintaksi (ikota:1.2) ukunika ubunzima obungaphezulu okanye obungaphantsi kwiikhonsepthi.
Ukuveliswa kwe KSampler kunye neeparamitha
Isampulu iya kuqala xa uyibeke emgceni (i-Queue Prompt). I-KSampler ilawula imbewu, amanyathelo, isampulu, umcwangcisi, kunye namandla edenoiseImbewu esisigxina ibonelela ngokuphinda-phinda; amanyathelo amaninzi adla ngokuphucula iinkcukacha (ngexabiso lexesha); denoise=1 kwi-text2img isebenzisa inkqubo epheleleyo yokususa ingxolo.
Umfanekiso oLatent ongenanto: isisombululo kunye neebhetshi
Indawo yoMfanekiso eLatent engenanto yenza icanvas yokuqala efihlakeleyo. Ubude nobubanzi kufuneka buphindaphindwe ka-8Ubungakanani obuqhelekileyo: 512/768 ye-SD 1.5 kunye ne-1024 ye-SDXL. Lungisa ubungakanani bebhetshi ukuba ufuna imifanekiso emininzi ngokubaleka.
I-VAE: ukunyanzeliswa kunye nokwakhiwa kwakhona
I-VAE ifaka iikhowudi kwaye ichaza phakathi kwepikseli kunye nelatent. Ibonelela ngokusebenzayo kunye nendawo efihlakeleyo ephathwayo, ngokutshintshiselana nelahleko ethile okanye izinto zakudala.Kwi-text2img, uya kuyisebenzisa ikakhulu ekugqibeleni (VAE Decode) ukufumana umfanekiso kwiipikseli.
Umfanekiso kumfanekiso, i-SDXL kunye nokupeyinta/ukupeyinta
Umfanekiso ngomfanekiso
Lo msebenzi udibanisa ngokukhawuleza kunye nomfanekiso osisiseko. Khetha indawo yokukhangela, layisha umfanekiso, uphonononge imiyalelo, kwaye uhlengahlengise i-denoise kwi-KSampler ukwenza isigqibo sokuba utenxa kude kangakanani kweyoqobo (i-denoise encinci = ngakumbi njengomthombo).
SDXL kwiComfyUI
I-ComfyUI ixhasa i-SDXL kwangethuba nangempumelelo ngenxa yokumodareyitha kwayo. Lungiselela iingcebiso ezilungileyo/ezingalunganga kwaye uqalise inkqubo ngesampulu efanelekileyo; khumbula esona sisombululo sikwimeko efihlakeleyo (idla ngokuba ngu-1024).
Ukupeyinta
Ukuguqula iindawo ezithile, layisha umfanekiso, vula umhleli wemaski, kwaye ugcine imaski kwindawo. Le ndlela yokusebenza isebenzisa iimodeli eziqhelekileyo; ukuba usebenzisa i-"inpainting" checkpoint, sebenzisa i-VAE Encode (Inpaint). Esikhundleni somgangatho we-VAE Encode kunye ne-Set Noise Latent Mask nodes, ibeka ukukhawuleza ukuchaza utshintsho kunye namandla aqhelekileyo e-denoise njenge-0.6.
Ukupeyinta ngaphandle
Yandisa ngaphaya kwemida yemifanekiso ngePad Image for Outpainting: lawula ekhohlo/phezulu/ekunene/ezantsi kunye neentsiba zotshintsho olugudileyo. Kwi-VAE Encode (ye-Inpainting) lungisa i-grow_mask_by (ngcono > 10) ukufezekisa ukuzaliswa kwendalo kunye nokudityaniswa kakuhle.
Ukukala: pixel vs latent
Ukuphakama kwePixel
Iindlela ezimbini: nge-algorithm (i-bicubic, i-bilinear, ikufutshane-echanekileyo) kunye noMfanekiso oPhezulu ngo, okanye ngemodeli eneModeli yoMlayisho oPhezulu + uMfanekiso oPhezulu (usebenzisa iModeli). Ii-algorithms ziyakhawuleza kodwa zicocwa kancinci; iimodeli zithatha ixesha elide kwaye zihlala zinika iinkcukacha ezingconokwaye unokudibanisa ukuphuma kunye ne pakisha 50 iziphumo emva Effects.
Ukuphakama kwe-Latent
Oku kubizwa ngokuba yi-Hi-Res Latent Fix izikali ngokuthe ngqo kwindawo efihlakeleyo, iphucula iinkcukacha ngexesha lokwakhiwa kwakhona. Isenokutenxa kancinane kowokuqala kwaye icotha, kodwa yongeza ulwazi endaweni yokolula nje pixels.
Ukuthelekisa ngokukhawuleza
Ukunyuswa kwePixel: ngokukhawuleza, ngaphandle kokongeza ulwazi olutsha, ukugudisa okunokwenzeka. Ukunyuka okufihlakeleyo: kancinci, Yongeza iinkcukacha kodwa inokutshintsha umfanekiso osisiseko.Khetha ngokomxholo kunye nokunyaniseka okufunekayo.
ControlNet: ulawulo olulungileyo lwesakhiwo
I-ControlNet ibonelela ngezikhokelo ezinje ngemida, ukuma, ubunzulu, okanye ukwahlulahlula ukuqinisekisa ukuba imodeli ihlonipha ubume. Sisixhobo esinamandla kakhulu kwi-VFX kuba ihlala ilungisa ukwakheka kunye nentshukumo.Zama i-Lineart, Ubunzulu, okanye i-OpenPose kwaye ulungelelanise amandla ukuze ulungelelanise ukunyaniseka / ukuyila.
UMlawuli weComfyUI: IiNodi zesiNtu ukuya kuthi ga ngoku
Faka iindawo ezilahlekileyo
Ukuba i-workflow icela iinodi ongenazo, sebenzisa uMphathi: Umphathi weqhosha, "Faka iindawo ezilahlekileyo zesiko", qalisa kwakhona i-ComfyUI kwaye uphinde ulayishe isikhangeli. Oku kuqinisekisa ukuba uphindaphinda ukuhamba okwabelwana ngako kanye..
Hlaziya iindawo
Ukusuka kuMphathi, khangela uhlaziyo kwaye ucofe "Faka iinodi zesiko." Ukuba "Uhlaziyo" luvela ecaleni kwepakethe, lusebenzise, qala ngokutsha, kwaye uhlaziye. Ukugcina ii-nodes kusexesheni kuthintela iimpazamo kwaye kuphucula ukusebenza.
Khangela iindawo kwiseyile
Cofa kabini kwi-canvas engenanto ukuvula i-node finder kwaye ubongeze ngegama. Oku kukhawulezisa ukuhlanganiswa kwamatyathanga anzima. ngaphandle kokukhangela iimenyu.
Ufakelo (uguqulo lokubhaliweyo)
Ukwenza ufako lusebenze, chwetheza ufakelo:Igama kumyalelo olungileyo okanye ongalunganga. Beka ifayile kwi-ComfyUI/models/embeddings I-ComfyUI iya kuyisebenzisa ukuba ifumene umdlalo. Yindlela enamandla yokubandakanya izimbo okanye iikhonsepthi ezithile.
Ukuzinzisa ukugqibezela ngokuzenzekelayo
Fakela iComfyUI-Custom-Scripts package ukuze uzigqibezele ngokuzenzekelayo. Nje ukuba usebenze, ukuqalisa ukuchwetheza "ufakelo:" luza kubonisa usebenziso lwakho olukhoyoukukhawulezisa umsebenzi ngeengqokelela ezinkulu.
Ubunzima bokufaka
Ungawulinganisela njengamagama: (ufakelo:Igama:1.2) kwandisa impembelelo kwaye (ufakelo:Igama:0.8) kuyayinciphisa. Ukulungelelanisa iintsimbi kukunika ulawulo olulungileyo kwimpembelelo yembonakalo.
I-LoRA: ihlengahlengisa isitayile ngaphandle kokuchukumisa i-VAE
I-LoRA ilungisa iMODELI kunye ne-CLIP yendawo yokukhangela esisiseko ukwazisa izimbo, iimpawu, okanye izinto, ishiya i-VAE ingaguquki. Ukuhamba okusisiseko: khetha indawo yokukhangela, yongeza i-LoRAs enye okanye ngaphezulu, uphonononge izindululo, kwaye uqalise umgca.
Ii-LoRA ezininzi ezihlayo
Unokufaka ii-LoRA ezininzi ekuhambeni okufanayo; zidityaniswe ngokulandelelanayo. Linga ngocwangco kunye neentsimbi zokudibanisa ngokuyilayo izimbo de umlinganiselo ofunekayo ufezekiswe.
Iindlela ezimfutshane kunye namaqhinga agcina iiyure
Khuphela/uncamathisele: Ctrl+C, Ctrl+V kunye no-Ctrl+Shift+V ukuze uncamathelise ngelixa ugcina amangeniso. Khetha iindawo ezininzi ngeCtrl, yenza iibhokisi zokukhetha, kwaye uzihambise ngeShift yoyilo olukhawulezayo.
Ukuthulisa indawo nge Ctrl+M kuyayishiya okwexeshana; Nciphisa i-node ngokucinezela ichaphaza kwikona yalo ephezulu ngasekhohlo ukucima i-canvas kwiiprojekthi ezinkulu.
Umgca wesizukulwana: Ctrl+Enter. I-ComfyUI iphinda iphumeze iindawo kuphela ukuba amagalelo ayatshintsha; ilungisa imbewu ukunqanda ukubala kwakhona amatyathanga amade kwaye wonge ixesha.
Ukuqukuqela okuzinzisiweyo kwe-PNG: Tsala umfanekiso owenziweyo kwi-ComfyUI ukuze ufumane kwakhona ukuqhutywa komsebenzi kwimetadata yawo. Yeyona ndlela ibalaseleyo yokwabelana kunye nemibhobho yoguqulelo ngaphandle kokulahlekelwa ngamaqhekeza. Ukuba ufunda ngcono ngevidiyo, jonga Izifundo zevidiyo ezili-10 ezisisiseko.
IComfyUI yevidiyo: iAnimateDiff inyathelo ngenyathelo
I-AnimateDiff ikuvumela ukuba uvelise ulandelelwano ngokubhaliweyo, imifanekiso, okanye ividiyo. Kwi-Windows ene-NVIDIA, i-10 GB ye-VRAM ilungile (i-8 GB ubuncinane enezigqibo ezisezantsi okanye i-Txt2Vid); kwiiprojekthi ezifunayo ungalindela malunga ne-10 GB kunye ne-2 ControlNets.
Ukufakela kunye nokuxhomekeka
Faka i-Git kwi-clone nodes kunye ne-7-Zip ukukhupha i-ComfyUI ephathekayo. I-FFmpeg ayinyanzelekanga (yokupakisha i-GIF/MP4 ukusuka kwiindawo zokuhlanganisa)Ukuba ayikho kuMENDO, imijelo iyaqhubeka nokuvelisa izakhelo ezixeliweyo.
Khuphela iComfyUI ephathwayo kwaye usebenzise run_nvidia_gpu okokuqala ukuze uqalise. Kwincwadi eneenkcukacha yesiko, fanisa I-ComfyUI-AnimateDiff-Evolved, iComfyUI-Manejala, iComfyUI-Advanced-ControlNet kunye neComfyUI-VideoHelperSuite.
Ukusuka kuMphathi, faka "i-ControlNet Auxiliary Preprocessors" kunye ne "FizzNodes". Qala kwakhona i-ComfyUI ukuze ulayishe yonke into ngokuchanekileyo kwaye uphephe iimpazamo ezivela ngaphandle.
Iimodeli ezifunekayo
Beka iindawo zokuhlola ezihambelanayo ze-SD 1.5 kwifolda efanelekileyo kunye ne-VAE eqhelekileyo ukuba kuyimfuneko. Khuphela iimodyuli ezishukumayo (umzekelo, ezoqobo ezivela kwi-AnimateDiff, TemporalDiff, okanye i-AD Stabilized Motion) kwaye uzikhuphele kwindlela yakho. Kwi-ControlNet, yongeza i-Lineart, Ubunzulu, kunye ne-OpenPose (pth/yaml).
Ukuhamba okungundoqo: Vid2Vid kunye neTxt2Vid
I-Vid2Vid: ilayisha uluhlu lwezakhelo ezinomfanekiso/ividiyo yokufakwayo kwenode, ilawula umfanekiso_load_cap, skip_first_images kwaye ukhethe_every_nth kangangexesha kunye nesampulu. Ukhetho Lwemeko Efanayo lubalulekile: ubude bomxholo ~16, ziphilelana ngokuqhubekayo, kunye nelophu evaliweyo kuphela kwi Txt2Vid.
I-Txt2Vid: isebenzisa i-node yesakhelo esisiseko (ngaphandle komlayishi womfanekiso) kwaye yenza ngokuthe ngqo kwi-prompt. NgeDenoise=1 kwi-KSampler uya kuba nesiphumo sokuvelisa ngokupheleleyo, ilungele iiklip zentelekelelo.
Ukucwangcisa ibhetshi ngokukhawuleza
I-BatchPromptSchedule ye-FizzNodes ikuvumela ukuba utshintshe ukwaziswa ngesakhelo ngasinye. Isebenzisa i-pre_text kunye ne-app_text kwiiheader eziqhelekileyo kunye nokuvalwa, kwaye ichaza "isakhelo: ngokukhawuleza" ngababini. Qaphela isiphumlisi sokugqibela kwisiqalelo sokugqibela, siya kubangela impazamo; iphinda-phinda umyalelo ukuba ufuna ukuwugcina phakathi kwezithuba.
Isampuli kunye nokudibanisa izicwangciso
I-KSampler yevidiyo ifuna amanyathelo amaninzi (ama-25 ubuncinci kwaye ngcono ukuwanyusa). Zama i-Euler_a sampler kwaye ulungelelanise i-CFG ngokuthanda kwakho.Kwi-Vid2Vid, yehlisa i-denoise ukuze usondele kwikliphu yemvelaphi. I-Combine node ithumela ngaphandle i-GIF/MP4: chaza ifreyim_rate, loop_count, ifomathi, kunye nokuba uyafuna na i-ping-pong.
Iingcebiso ezisebenzayo: yehlisa amandla e-ControlNet kwimifanekiso emileyo, zama i-OpenPose, sebenzisa i-KSampler yesibini yolungiso "lokuqesha". Zama i-Motion LoRA ukutyebisa iintshukumo ezithile kwaye idibanisa iControlNets kancinci.
Ezinye iinjini zevidiyo kwiComfyUI
I-HunyuanVideo (img2vid ikhokelwa yimibhalo engezantsi)
Lungisa umfanekiso wakho kwi-512x512 kwaye uvelise isihlokwana ngeFlorence2Run. Buyisela amagama afana “nomfanekiso/ifoto/umzobo” nge “vidiyo” usebenzisa iStringReplace Ukulungelelanisa noqeqesho lwemodeli. Guqulela kwindawo efihlakeleyo usebenzisa iHunyuanVideo Sampler + wrappers, faka iLora Khetha kwaye uthumele ngaphandle ngesihlanganisi.
Ividiyo ye-LTX (umbhobho we-LTX node)
Faka iComfyUI-LTXVideo nodes kunye neemodeli (ibandakanya i-PixArt-XL encoder). Bhala ngokukhawuleza kwi-CLIP, yenza ividiyo efihlakeleyo nge-EmptyLTXVLatentVideo kwaye ulungise i-LTXVSchedulerInani lamanyathelo linyusa umgangatho, kunye neeparamitha ezifana ne-max_shift, i-base_shift, yolula, kunye ne-terminal imilo ye-clip's dynamics. Gcina nge-SaveAnimatedWEBP (inyani engalahlekiyo, umgangatho we-100) okanye uthumele ngaphandle kwezinye iifomati.
Wan 2.1 (umbhalo kwividiyo, umfanekiso kwividiyo, ividiyo ukuya kwividiyo)
I-ComfyUI ikwadibanisa ukuhamba kweWan 2.1. Ukusetyenziswa kwayo kubandakanya iTxt2Vid, Img2Vid kunye neVid2Vid, kunye nolawulo lweparameter olufana nemibhobho yangaphambili kunye neenzuzo kwisigxina sexesha ngokuxhomekeke kwimeko yokusetyenziswa.
Imizobo eshukumayo: ulwahlulo, ubunzulu, kunye nokudibanisa
Kwimifanekiso eshukumayo yemifanekiso eshukumayo evela kwividiyo, qala ngeLoadVideoInput, lawula uTyiba iifreyimu zokuqala kwaye ukhethe yonke isakhelo se-Nth, kwaye ulinganise nge-ImageScaleToMegapixels ukuya kuthi ga kwi ~1MP. Oku kulungiswa kwangaphambili kulungelelanisa umthwalo we-VRAM kunye nesantya sokusebenza kwisizukulwana. Ungajonga kwakhona ukuba njani Yenza izihloko kwiPremiere ukudibanisa imizobo kunye neekhredithi.
Yahlula umxholo usebenzisa i-GroundingDINO kunye ne-SAM kunye ne-GroundingDinoSAMSsegment kwisicatshulwa. Yandisa imaski nge-GrowMaskWithBlur kwaye uyiguqulele kumfanekiso ngeMaskToImage ngolwandlalo olomelele ngakumbi.
Yenza umqondiso wexesha nge-TimeFeatureNode kwaye uyimodareyithe nge-FeatureScaler (linear, logarithmic, exponential). Oku kuya kukuvumela ukuba ulawule ubunzulu (Z) ukufuduswa okanye iindawo zemaski ecaleni kwekliphu ngeziphumo ezininzi ze-cinematic.
Yenza imaski yokupeyinta exhomekeke kubunzulu ngeFlexMaskDepthChamber, ukudibanisa imaski yomxholo, umqondiso wexesha, kunye nemephu enzulu yekliphu. Lungisa iZ Front/Z Emva ukuchaza indawo esebenzayo nangaliphi na ixesha. kunye nokufezekisa umphumo oqinisekileyo we-3D.
Kwinqanaba lokuvelisa, i-checkpoint yomthwalo, sebenzisa i-LoRAs, cwangcisa i-prompts, kwaye ungeze i-ControlNet ukuba ifanelekile. I-AnimateDiff iya kukunika iifreyimu; emva koko udibanise nge-RIFE VFI ukuphinda kabini ukuguda kunye neenguqu ezigudileyo.
Ukuba ufuna ukuxuba amapasi: velisa iinguqulelo ezininzi ezinezalathisi ezahlukeneyo, khetha amacandelo nge ImageIntervalSelectPercentage, xuba iinguqu nge-ImageBlend kwaye udibanise ne-ImageBatchMulti. Ukudlula kokugqibela nge-RIFE VFI kushiya upopayi ugudile kwaye ilungele ukuthunyelwa ngaphandle.
ComfyUI kwi-intanethi kunye nezinye iindlela zamafu
Ukuba awufuni kufakela nantoni na, kukho iinkonzo zelifu ezine-ComfyUI esele ziqwalaselwe, kunye namakhulu eenodi / iimodeli kunye nenani lemisebenzi esele yenziwe. Ziluncedo kuvavanyo olukhawulezayo okanye kumaqela abelana ngeetemplates ngaphandle kokujongana nee-arhente zendawo. Njengenye indlela ekhawulezayo nelula, kukwakho nezibonelelo kwi oopopayi kunye neziphumo ezibonakalayo kwiCapCut.
Enye inketho kukusebenzisa iijenereyitha zevidiyo ezisekwe kwilifu njengeDreamina: ujongano olulula, akukho VRAM yendawo, kwaye iziphumo kwimizuzwana engama-20 ukuya kwengama-60. Inika ezongezelelweyo ezifana ne-HD upscale, i-frame interpolation, kunye nokuveliswa kwesandi.Ngeekhredithi zemihla ngemihla zasimahla zokuqalisa, yindlela elungelelanisiweyo xa isantya sibaluleke kakhulu kunolawulo lwegranular.
Ukusebenza, iimfuno kunye namaxesha
Ekuhlaleni, iComfyUI yomfanekiso kwividiyo ihlala ifuna phakathi kwe-8 kunye ne-24 GB ye-VRAM ngokuxhomekeke kwimodeli (AnimateDiff, HunyuanVideo, LTX Video) kunye nesisombululo. NakwiiGPU ezinamandla, isizukulwana sinokuthatha imizuzu eyi-10 ukuya kwengama-30 Ukuba ikliphu inde okanye ukuba usebenzisa ii-ControlNets ezininzi kunye nezihluzi zokupasa okuphezulu, umthwalo utshintshelwa kumboneleli efini.
Isoftware yeComfyUI isimahla, kodwa ixabiso likwihardware kunye nombane ukuba usebenza ixesha elide. Ilifu lizinqanda ezo ndleko, ngokutshintshiselana ngokuxhomekeka kwinkonzo kunye nomrhumo okanye iikhredithi.Vavanya ukuba zeziphi iingenelo zokuqhutywa komsebenzi wakho kakhulu.
Ukusombulula ingxaki eqhelekileyo
Ukuba ubona iimpazamo ezingekhoyo okanye iindawo "ezingekho", mhlawumbi ulahlekile imifuziselo kwiincwadi zakho ezineenkcukacha okanye ukuxhomekeka okungafakwanga. Qinisekisa ukuba i-node nganye inemodeli ehambelana nayo kwaye usebenzise uMphathi ukufaka naziphi na iipakethe ezingekhoyo.Ziphephe ezingqubanayo zokugcina ukuba sele usebenzisa iComfyUI kweminye imisebenzi.
Iindlela ezigqwesileyo zeVFX engaguqukiyo
Tshixa imbewu ukuze ikwazi ukuvelisa kwakhona ngelixa uhlengahlengisa amacandelo ekhonkco. Gcina imifanekiso ngemetadata yokuhamba komsebenzi kunye ne-annotate node kunye neenguqulelo zemodeliKwividiyo, chaza ubude bomxholo kunye nokugqithisa ngononophelo, kwaye ugcine ulandelelwano olucacileyo lwe-ControlNets kunye ne-LoRAs.
Tshintshela ngobuchule phakathi kwe-pixel ephezulu kunye ne-latent ngokuxhomekeke kuhlobo lokudubula kunye nenqanaba leenkcukacha. Kwi-Vid2Vid, nciphisa i-denoise ukuhlonipha intshukumo yesiseko.Kwi-Txt2Vid, tyhala amanyathelo kunye nesampulu ukufumana uzinzo olubonakalayo.
Dibanisa i-ControlNet preprocessors (i-canny, ubunzulu, indawo evulekileyo…) ukusuka kuMphathi ukwandisa ibhokisi yakho yesixhobo. Kwaye khumbula: amandla amancinci e-ControlNet ahlala enza ngcono kwividiyo.ukuphepha inkangeleko ehluziweyo kunye nokugcina inkangeleko yendalo.
Qiniseka ukuba uphonononga i-HunyuanVideo kunye nevidiyo ye-LTX ukuba imeko yakho ifuna izikhokelo zemibhalo engezantsi, ulawulo lwamaxesha ahlukeneyo, okanye enye imibhobho. IWan 2.1 yongeza iinketho eziqinileyo zeTxt2Vid, Img2Vid, kunye neVid2Vid ngeeparamitha ezihambelanayo kunye neziphumo ezikhuphisanayo.
Abo bafuna isantya kunye ne-zero friction banokuthembela kwiinkonzo ze-intanethi, ngelixa ezo zifuna ulawulo lotyando kunye nokuveliswa kwakhona ngokupheleleyo kuya kuqaqamba ngeComfyUI yasekhaya. Ngamaqhekeza owabonileyo-iindawo, iiparamitha, iindlela ezimfutshane, kunye nokuhamba-ngoku unemephu yendlela yokuvelisa i-VFX ephezulu. zombini kwimifanekiso emileyo kunye nolandelelwano lwevidiyo, ngendlela eguquguqukayo kunye neyokunyuka.
