How are you training with 5s/it??? On A6000 I'm getting like 30s/it for training
How are you training with 5s/it??? On A6000 I'm getting like 30s/it for training
JAnd how are you running multiple trainings at the same time?
Fit is batch size 1 speed
Falso this is different optimizer not adafactor
Jah okay, I thought the quality was not worth the extra speed?
Fthis is actually more demanding
Fyou see it uses entire vram
Jsomething you'll show us in the next tutorial?
Fadafactor uses 28 gb at batrch size 1 :d
Fi will update configs
Fgenerating massive grid to fianlize
Fi did 22+ trainings
Joh yea sorry i meant the other way around. I thought the extra quality was not worth the wait
F
Fit is a choice
Fthis will be premium quality
Fmore demanding
Ffor wohever wants
Jdope!
Fi got to sleep see you tomorrow hopefully
Jexcited to see results
Jyea same!
X@Dr. Furkan Gözükara Hope all is well i was wondering if you were going to make a tutorial on browser use and how to install and run locally
XGitHub
Make websites accessible for AI agents. Contribute to browser-use/browser-use development by creating an account on GitHub.
F@Dr. Furkan Gözükara I got a new PC and tried to install Kohya using your setup file. I'm getting this error. line 35 on requirements.txt has scipy module.
�That would be very helpfull
�Can kohya now train flux fill?
GI was referring to the GPU RAM which is 8GB, my main PC RAM has 64GB, so what configuration would you recommend? i was following your tutorial and it takes several hours for single percentage in training progress... is there anything I could tweak to run it faster?
Fi saw this model
Flooks like they are utilizing openai api to make it do things
Fhello do you have python 3.10 installed?
Flet me make a fresh install and see
Fkohya cant
Fyes it will work
Fi recommend this config
F8GB_GPU_Quality_Tier_4_FP8_6850MB_13.2_Second_IT.json
FThe grid test I am analyzing to find newer FLUX DEV training. This is not even including the first 10 different trainings i made 
1176 images. I think we gonna have a better workflow than what we have with using more VRAM and more training time.
1176 images. I think we gonna have a better workflow than what we have with using more VRAM and more training time.
�How do you even analyze so many images? thanks for the hard work
Fmanually no easier way
�yes that's very hard and how do you comapre if the previous models were better? are you uisng same seed, same prompt, etc?
Fi generate image of previous model as well in same grid
�okay. I trained a dreambooth yesterday on that outfit I asked you about but I think the learning rate you set is very low should I increase that to 1e-4 ?
Fyou can increase if it is undertrained
P@Dr. Furkan Gözükara I have few questions: 1)The hands especially the hands and fingers isn't good on my flux+dreambooth trained model ,how to fix that ?2)as we had trained the model for 1024x1024, what if I want to generate image of high resolutions like full HD or 4k 1920x1080, 3840x2160?
Fupscale images it really fixes a lot
F2x latent upscale
FPAnd what about increasing resolution will it do that too along with fixing the hands?
Ayes also for the faces too
AI've tested it
Abut it will be slower