interesting comparison the left is much better
interesting comparison the left is much better
DI'd be curious to see 32. thats what I used in the few experiments I did on civitai
Fyes lora lower rank reduced quality
F256 overfit though
F128 looks best
Fi have 32 too
Fpreparing grid
DI wander if rank 8 could be enough, must be sweet spot for such small dataset
F128 vs 32 here
Fmy dataset already small
Fi will share full grid on patreon
Dalso training the TE could make the diference, is not the same training the entire man class that is was doing right now than training a unique person, currently training a subject is overwriting the entire class weights
D32 is good enough probably not perfect but very good
Fyes TE should improve
Fyep 32 and above good
F256 overkills
Fall ranks vs
Fi am doing massive grids tests right now
F3 more on training
DI bet that with TE training enabled lower ranks will improve a lot, just a theory!! hehe
FI get this error message when I run training on Umbubu
Fyet I gave the right path for the images
�That has to with the folder training directory having to be the main directory above the folder of the images and not the image folder itself, if that makes sense
F,could be but not supported yet :d
Fyou are missing training dataset
�So, Home/Ubuntu/Desktop/ instead of Home/Ubuntu/Desktop/trainingimages/
Fshow screenshots how did you give folder path and what is in folder
Dinteresting how the glasses are missed in all but the very high rank training
Fflux is very powerful
Fit omits eyeglasses it knows it is not me :d
Fi will retest with adding eyeglasses word
Dfor the test prompt?
Fye for test prompt
Fwhen eyeglasses added it perfectly generates
Falso 3 more final trainings are running
Fwith full captions of
Fjoycaption
�Is there any disadvantages to training at 128 rank @Dr. Furkan Gözükara ? I'm assuming longer training time and more VRAM usage perhaps, but what about with inference, same thing?
Fvram usage slightly more 19 mb per rank for 16 bit
Ffile size bigger
Fbut it also leads more overfit. that is why full fine tuning is the best way
�Ok but quality wise, 128 seems to be the sweet spot so it seems kinda worth it
F128 best
�Ok cool thank you Dr!
Finstead of reducing lora rank i will reduce resolution for lower VRAM gpus
Fwill prepare every GPU config
F9.5 GB, 11.5 GB, 15 GB, 18 GB 
FMy images are in the training_imgs folder, isn't that good?
�Image Folder directory should be: home/Ubuntu/Downloads
Fshow inside of it
