@Dr. Furkan Gözükara I see in the supir app you can upload a dreambooth model for face correction,
@Dr. Furkan Gözükara I see in the supir app you can upload a dreambooth model for face correction, can we use a LORA instead for low vram?
Fno platform for free
Fit requires bf16
Ffree platofrms dont provide such gpu
Fit uses a base model
Zah ok any recommendations for improving images of self on those trained on lora?
Zusing SUPIR?
V@Dr. Furkan Gözükara how big (mb wise) is the LoRA you did with the 15 image dataset?
Fyep
Fit depends several factors
Fbut starting from 2 gb to 3.5 gb
Fsingle layer training arrived to kohya
KKohya gets stuck here
KNo errors
KWell I find this scrolling up
VI trained a 200 epoch LoRA (81 images dataset), with the following settings:
VVGot a 18mb LoRA out of it
VIt works, but is not incredible
Fi explained in video the factors
Fset accelerator and gpu id 0
Fit is using cpu that is the error
KOh shit
KMay I ask where? command line?
Fread
K
Fok then set accelerator
Fyou have multiple gpus?
KI do
KI shitty internal GPU that gets listed
KAha
K
Kaccelerator device: cpu :/
KWell have to eat. Hopefully I can solve this later
KThanks!
Fok try gpu id 1
KTo test a Lora which guidance value is best? Like 1 or 3.5 or even higher?
KI have a pretty hard time to find the sweet spot. Sometimes it changes the image quite a bit like on epoch 200 strenght 1.0 and on epoch 150 at strenght 1 .1
Kok this image isn't very useful as discord resizes it
Fflux guidance 4
Fi got a better config for fine tuning
Fit is nice
Kdefault shift (base 0.5 / max 1.15) ?
Fah from your image i dont know what are they 
Kif discord doesn't resize the image maybe this is more useful. Seem pretty spot on with epoch 200 and lora_strenght 1.0 at 1.1 if starts mutilate hands.
Fye i use lora str 1
FPatreon
Get more from SECourses: Tutorials, Guides, Resources, Training, FLUX, MidJourney, Voice Clone, TTS, ChatGPT, GPT, LLM, Scripts by Furkan Gözü on Patreon
KOkey, but that is not for lora. check that out another time as it was a 48hr run with 100 images in a data set. for this lora.
Kon a RTX 3090 ti