Cishe uke waphawula ukuthi i-inthanethi igcwele izithombe ezikhiqizwe yi-AI. Abaningi bayazibuza ukuthi kungenzeka yini ukuthola imiphumela efanayo ekhaya, ngaphandle kokukhokha ukubhalisa kwanyanga zonke futhi ngenkathi begcina ubumfihlo obuphelele bezicelo zabo. Impendulo inguyebo ozwakalayo, futhi ithuluzi elinamandla kakhulu lokwenza kanjalo njengamanje yi-ComfyUI, isikhombimsebenzisi, nakuba ekuqaleni sesabisa, esikunikeza ukulawula okuphelele ekudalweni okubonakalayo.
Ngokungafani nezinketho ezilula, i-ComfyUI isebenza ngokusebenzisa uhlelo lwegrafu lapho uxhuma khona amabhokisi noma ama-node ahlukene. Lokhu kusho ukuthi awunqunyelwe yifomu eqinile; ungaklama ipayipi lakho lokwakha . Kungakhathaliseki ukuthi uneseva enamandla ye-Linux noma i-laptop ye-Windows, ukusetha le ecosystem kuzokuvumela ukuthi uzame amamodeli wakamuva, njenge-FLUX noma i-Stable Diffusion, ngaphandle kokuthi othile ahlole lokho okudalayo.
Izidingo Zehadiwe: Umhlaba we-VRAM
Ngaphambi kokuthi uqale ukufaka, sidinga ukukhuluma ngalokho okubaluleke ngempela: imemori yevidiyo (i-VRAM). Akunandaba ukuthi iprosesa yakho inamandla kangakanani noma ukuthi une-RAM engakanani; uma ikhadi lehluzo lingasekeli i-VRAM, uhlelo lwakho luzosebenza ngesivinini esikhulu noma lumane luphahlazeke. Ukuze uhlole izidingo zakho ze-VRAM ku-Linux, ungasebenzisa umyalo othi `nvidia-smi` we-NVIDIA noma `rocm-smi` we-AMD.
Uma une-4GB ye-VRAM, lungela ukusebenza okuhamba kancane futhi kuzodingeka usebenzise izinqumo eziphansi. Nge-8GB, ungakwazi ukusingatha imisebenzi eminingi kahle, yize kumamodeli adinga amandla afana ne-FLUX.2 kuzodingeka usebenzise i-GGUF quantization (njenge-Q4_K_M) , enciphisa usayizi wemodeli ngaphandle kokulahlekelwa ikhwalithi. Uma unenhlanhla yokuba ne-16GB noma ngaphezulu, njengakwi-RTX 3090 noma i-4090, ungalayisha amamodeli e-FP8 noma ngisho ne-FP16 ngaphandle kokujuluka.
Ukufakwa Kwesinyathelo Ngesinyathelo Ezinhlelweni Ezihlukene
Kuye ngesistimu yakho yokusebenza, inqubo iyahlukahluka. Ku-Windows, indlela elula ukulanda iphakheji ephathekayo esemthethweni , uyikhiphe, bese usebenzisa ifayela le-.bat elihambisana ne-GPU yakho. Lena yindlela eshesha kakhulu ngoba isivele ifaka i-Python kanye nokuncika okudingekayo, igwema ukungqubuzana neziguquguquko zemvelo.
Uma ukhetha i-Linux noma i-macOS, inqubo ilula kakhulu kodwa ihlanzekile. Okokuqala, udinga ukuhlanganisa indawo yokugcina idatha kusuka ku-GitHub bese udala i- Python virtual environment (venv) . Lokhu kubalulekile ukuze ugweme ukuphula amalabhulali esistimu yokusebenza. Uma imvelo isisebenza, faka ukuncika okubhalwe kufayela le-requirements.txt kanye nenjini yezibalo ye-PyTorch. Abasebenzisi be-NVIDIA bayelulekwa ukuthi basebenzise inguqulo ye-CUDA 12.1 , kuyilapho abasebenzisi be-AMD kufanele bakhethe i-ROCm.
Ukucushwa Okuthuthukisiwe kanye namaNode Abalulekile
Uma usuqalise i-ComfyUI okokuqala ku-http://127.0.0.1:8188, uzobona ideshibhodi egcwele amakhebula. Ukuze ugweme ukukhungatheka, into yokuqala okufanele uyifake yi- ComfyUI Manager . Le plugin iyigugu langempela, njengoba ikuvumela ukuthi useshe futhi ufake amanye ama-node ngokwezifiso, futhi okubaluleke kakhulu, ufake ngokuzenzakalelayo noma yimaphi ama-node angekho uma ungenisa umsebenzi womunye umuntu.
Ukuze usebenze namamodeli esimanje afana ne-FLUX, uzodinga i -ComfyUI-GGUF node . Ngokungafani ne-Stable Diffusion, lapho konke kuvame ukugcinwa khona kufolda yezindawo zokuhlola, i-FLUX idinga ukuhlukanisa izingxenye zayo: imodeli yokusakaza (i-UNet) iya ku- models/unet/ , i-text encoder (CLIP) ku-models/clip/, kanye ne-VAE ku-models/vae/. Uma uzibeka endaweni engafanele, i-loader ngeke izithole, futhi uzogcina unesikrini esingenalutho.
Ukuqonda kahle ukuhamba komsebenzi kanye ne-API
Umsebenzi oyisisekelo uhilela ukulayisha imodeli, ukubhala i-prompt, ukudlula ku-sampler (lapho kwenzeka khona umlingo wokusabalalisa), kanye nokuhumusha isithombe. Icebo elibalulekile le-FLUX ukugcina i- CFG iku-1.0 bese usebenzisa izinyathelo ezimbalwa (phakathi kuka-4 no-8) nge-sampler ye-Euler; uma ukhulisa i-CFG, izithombe zizobonakala kakhulu futhi zibe nemibala engenangqondo.
Kulabo abanehadiwe kuseva abafuna ukusebenza besuka kwenye ikhompyutha, i-ComfyUI ivumela ukufinyelela okukude kusetshenziswa ipharamitha ethi `--listen 0.0.0.0` . Lokhu kuvula i-HTTP API ekuvumela ukuthi uthumele imisebenzi ngefomethi ye-JSON futhi uthole isithombe esicutshunguliwe. Kuyindlela ekahle yokuguqula i-PC yakho ene-GPU ibe isevisi yokukhiqiza izithombe yangasese, okuzenzakalelayo ukudalwa kwamakhulu ezithombe zokumaketha noma ze-e-commerce kusetshenziswa izikripthi ze-Python.
Ukuxazulula izinkinga Nokuthuthukisa
Kuvamile ukuhlangana nephutha le -CUDA Out of Memory . Uma wenza kanjalo, ungaqala kabusha i-ComfyUI ukuze ususe ukuhlukana kwememori noma uqale uhlelo ngomyalo othi `--lowvram` , ohambisa idatha phakathi kwe-GPU ne-RAM yesistimu. Kumasistimu e-AMD, uma ikhadi lingatholakali, inkinga ivame ukuxazululwa ngokwengeza umsebenzisi kumaqembu okunikeza kanye namavidiyo noma ngokusebenzisa i-HSA_OVERRIDE_GFX_VERSION environment variable.
Uma ubona ukusebenza okungekuhle, hlola ukuthi awusebenzisi i-CPU ngephutha. Futhi, cabanga ngokuthuthukisa i-RAM yesistimu yakho ibe ngu-32GB noma u-64GB, njengoba i-ComfyUI iyisebenzisa njengebhuloho lapho i-VRAM inganele. Kwisitoreji, idrayivu ye-NVMe esheshayo ibalulekile, njengoba ukulayisha amamodeli angu-10GB njalo uma ushintsha imisebenzi kungathatha isikhathi eside kudrayivu enzima yomshini.
Ukusetha indawo yakho yokusebenza yezithombe yendawo kukukhulula emikhawulweni yamafu kanye nezindleko zesikweletu. Kusukela ekuphathweni kwe-VRAM kanye nokusethwa kwemvelo ebonakalayo kuya e-API ezenzakalelayo kanye nokusetshenziswa kwamamodeli e-GGUF quantized, manje usunawo wonke amathuluzi okuguqula i-RTX ibe yifektri yobuciko bedijithali, uthuthukise isinyathelo ngasinye ngokusebenzisa ama-node futhi uqinisekise ukuthi ihadiwe isebenza kahle kakhulu.