ComfyUI LoRA: C谩ch D霉ng, K岷縯 H峄 v脿 S峄璦 L峄梚

LoRA l脿 g矛?

LoRA (Low-Rank Adaptation) l脿 k峄?thu岷璽 fine-tune m么 h矛nh AI m脿 kh么ng c岷 hu岷 luy峄噉 l岷 t峄?膽岷. Thay v矛 ch峄塶h s峄璦 t岷 c岷?tham s峄?c峄 m么 h矛nh, LoRA ch峄?膽i峄乽 ch峄塶h m峄檛 t岷璸 con nh峄?鈥?t岷 ra c谩c file nh峄?g峄峮 (th瓢峄漬g 10鈥?00 MB) c贸 th峄?t岷 l锚n b岷 k峄?m么 h矛nh base t瓢啤ng th铆ch n脿o.

H茫y ngh末 m么 h矛nh base nh瓢 m峄檛 h峄峚 s末 膽a n膬ng, v脿 LoRA nh瓢 c谩c bu峄昳 膽脿o t岷 chuy锚n bi峄噒 鈥?m峄檛 bu峄昳 d岷 thi岷縯 k岷?nh芒n v岷璽 c峄?th峄? bu峄昳 kh谩c d岷 phong c谩ch ngh峄?thu岷璽, v脿 bu峄昳 n峄痑 d岷 k峄?thu岷璽 谩nh s谩ng.

T岷 sao n锚n d霉ng LoRA?

B岷痶 膽岷

1. T岷 LoRA

T矛m LoRA tr锚n c谩c trang chia s岷?model. 膼岷 b岷 LoRA t瓢啤ng th铆ch v峄沬 m么 h矛nh base c峄 b岷 (v铆 d峄? LoRA SD1.5 d霉ng v峄沬 checkpoint SD1.5, LoRA SDXL d霉ng v峄沬 checkpoint SDXL).

2. C脿i 膽岷穞 file

膼岷穞 file .safetensors 膽茫 t岷 v脿o:

ComfyUI/models/loras/

3. Th锚m node

Trong workflow, th锚m node Load LoRA gi峄痑 Load Checkpoint v脿 ph岷 c貌n l岷 c峄 pipeline. K岷縯 n峄慽:

• 膼岷 v脿o model 鈫?t峄?膽岷 ra MODEL c峄 Load Checkpoint
• 膼岷 v脿o clip 鈫?t峄?膽岷 ra CLIP c峄 Load Checkpoint

Node Load LoRA xu岷 ra phi锚n b岷 膽茫 ch峄塶h s峄璦 c峄 c岷?hai, sau 膽贸 b岷 k岷縯 n峄慽 v峄沬 c谩c node CLIP Text Encode v脿 KSampler.

Th么ng s峄?Load LoRA

Ch峄塶h Strength

• 0.5鈥?.8 l脿 kho岷g kh峄焛 膽岷 t峄憈 cho h岷 h岷縯 LoRA
• Qu谩 cao (>1.0) c贸 th峄?g芒y artifact ho岷穋 b茫o h貌a qu谩 m峄ヽ
• Qu谩 th岷 (<0.3) c贸 th峄?kh么ng c贸 hi峄噓 峄﹏g r玫 r峄噒
• M峄梚 LoRA c贸 膽i峄僲 t峄慽 瓢u ri锚ng 鈥?h茫y th峄?nghi峄噈 膽峄?t矛m ra

X岷縫 ch峄搉g nhi峄乽 LoRA

B岷 c贸 th峄?n峄慽 nhi峄乽 node Load LoRA theo chu峄梚. 膼岷 ra c峄 node n脿y n峄慽 v脿o 膽岷 v脿o c峄 node k岷?ti岷縫:

Load Checkpoint 鈫?Load LoRA (style) 鈫?Load LoRA (character) 鈫?CLIP Text Encode 鈫?KSampler

Khi x岷縫 ch峄搉g:

• Gi岷 strength t峄玭g c谩i 鈥?b岷痶 膽岷 峄?0.4鈥?.6 m峄梚 c谩i 膽峄?tr谩nh xung 膽峄檛
• Th峄?t峄?c贸 th峄?岷h h瓢峄焠g 鈥?th峄?s岷痯 x岷縫 kh谩c nhau n岷縰 k岷縯 qu岷?ch瓢a l媒 t瓢峄焠g
• Ch煤 媒 xung 膽峄檛 鈥?hai LoRA 膽瓢峄 hu岷 luy峄噉 tr锚n kh谩i ni峄噈 t瓢啤ng t峄?c贸 th峄?g芒y nhi峄卽 l岷玭 nhau

C谩c lo岷 LoRA ph峄?bi岷縩

M岷筼 膽峄?c贸 k岷縯 qu岷?t峄憈 nh岷

• D霉ng trigger word 鈥?nhi峄乽 LoRA y锚u c岷 t峄?kh贸a c峄?th峄?trong prompt. Ki峄僲 tra trang m么 t岷?c峄 LoRA 膽峄?bi岷縯
• Kh峄沺 m么 h矛nh base 鈥?LoRA SD1.5 kh么ng ho岷 膽峄檔g v峄沬 SDXL v脿 ng瓢峄 l岷
• Th峄?v峄沬 prompt 膽啤n gi岷 tr瓢峄沜 鈥?c么 l岷璸 hi峄噓 峄﹏g c峄 LoRA tr瓢峄沜 khi k岷縯 h峄 v峄沬 prompt ph峄ヽ t岷
• L脿m m峄沬 ComfyUI sau khi th锚m file LoRA m峄沬 n岷縰 ch煤ng kh么ng xu岷 hi峄噉 trong dropdown

L峄梚 th瓢峄漬g g岷穚 v脿 c谩ch kh岷痗 ph峄

LoRA kh么ng xu岷 hi峄噉 trong dropdown

• 膼岷 b岷 file n岷眒 trong ComfyUI/models/loras/ (kh么ng ph岷 th瓢 m峄 con, tr峄?khi ComfyUI 膽瓢峄 c岷 h矛nh qu茅t 膽峄?quy)
• L脿m m峄沬 ComfyUI (F5) ho岷穋 kh峄焛 膽峄檔g l岷 sau khi th锚m file m峄沬
• Ki峄僲 tra ph岷 m峄?r峄檔g file 鈥?ph岷 l脿 .safetensors (kh么ng ph岷 .ckpt hay .pt)

LoRA kh么ng c贸 hi峄噓 峄﹏g r玫 r峄噒

• Strength qu谩 th岷 鈥?t膬ng strength_model l锚n 0.7鈥?.0
• Thi岷縰 trigger word 鈥?nhi峄乽 LoRA y锚u c岷 t峄?kh贸a c峄?th峄?trong prompt. Ki峄僲 tra trang model tr锚n Civitai 膽峄?bi岷縯 trigger word
• Sai m么 h矛nh base 鈥?LoRA SD1.5 kh么ng ho岷 膽峄檔g tr锚n checkpoint SDXL v脿 ng瓢峄 l岷

K岷縯 qu岷?b峄?"ch谩y" ho岷穋 b茫o h貌a qu谩 m峄ヽ

• Strength qu谩 cao 鈥?gi岷 strength_model xu峄憂g 0.5鈥?.7
• Xung 膽峄檛 LoRA 鈥?n岷縰 x岷縫 ch峄搉g nhi峄乽 LoRA, gi岷 strength t峄玭g c谩i
• cfg qu谩 cao 鈥?gi岷 gi谩 tr峄?cfg trong KSampler xu峄憂g 6鈥?

L峄梚: "LoRA key not found" ho岷穋 t瓢啤ng t峄?

• Kh么ng kh峄沺 phi锚n b岷 鈥?LoRA 膽瓢峄 hu岷 luy峄噉 cho ki岷縩 tr煤c model kh谩c
• T岷 xu峄憂g b峄?l峄梚 鈥?t岷 l岷 file LoRA

B瓢峄沜 ti岷縫 theo

• Text to Image 鈥?Th脿nh th岷 workflow c啤 b岷 tr瓢峄沜 khi th锚m LoRA
• Image to Image 鈥?K岷縯 h峄 LoRA v峄沬 岷h tham chi岷縰
• Upscale Guide 鈥?N芒ng cao ch岷 l瓢峄g 岷h t岷 b岷眓g LoRA