I tried Lora training yesterday and garbage results on sd 2.1
I tried Lora training yesterday and garbage results on sd 2.1
HInteresting. Was it locally?
乂yes, locally
乂today I installed kohya_ss for further training with Lora + Dreambooth and it ran in to problem
Fyou need full precision for 2.1
F--no-half
M@Dr. Furkan Gözükara can It be cache? Or something ele?
MSorry for the spam
MIf I run: import torch
print(torch.version)
It says my version is 1.13.1+cu116
print(torch.version)
It says my version is 1.13.1+cu116
乂that means, to use lora we need to run it with argss --no-half
Fyou are not in the correct folder
Fuse cmd as i have shown in video please
Ffor sd 2.1 yes
乂understood
Falso using fp16 could be causing problem but i dont have enugh memory to test
Fp16 actually half precision
MOh, my fault. Thank you! It worked
MNow I'll train a model in Dreambooth and I give u a feedback
FYou are welcome
MThanks very much for your attention
HOk. I need a style model for a project this week. I will try a dreambooth for this time.
Fyes please let us know
Fi am also planning experiments on runpod
F24 gb
MNow I'm getting some real results and not a bunch of random shapes and colors lol
Fyep that fixes. people skipping that part and not working :d
Fleave a comment to the video really helpful
MOf course
Freddit
0 votes and 0 comments so far on Reddit
DNow there is one comment. 
DWhen you are doing img2img restore faces what are the prompt/steps/cfg/demonizing strength to get the same picture with just face restore? Or you have to do it iwth inpainting?
!Merhaba everyone, ben yapmak hablar turkche
Fwelcome. you can talk english
乂thanks to @thirdman for this 
CYou are welcome nephew
乂huehue uncle
CI am running ahmed 3
Cwhich should be 400, 800, 1600, 2200 cycles or so
乂woa
Clet me fire up the 4090
乂let's see the peek of 4090
Chow do I switch port?
乂port?
C乂ohhh
乂there would be configuration in web gui
乂let me check
乂--port 7870
乂arugument
C