I was using the one from BlackLabs but it wasnt encoding the image correctly for caching
I was using the one from BlackLabs but it wasnt encoding the image correctly for caching
SI hope it’s even effective with flux. I’ve read mixed things about it, curious to your findings
Fusing downloader is best :d
Fi plan to train today
Nwith latest build, didn't work eather
Fwhich config you tried
Nepochs 50, repeats 1, size 512,512. Tried v5 slow
Fit is impossible
Fit should use lesser than 10 gb
Fwhat is your batch size
N
Fsend me your json file as prrivatre message
Fyou are not using lora tab
Fnow extract config from zip file again
Ffresh
Fload into lora tab fresh
Fthen set your files
Fit will work
Fi looked your config file
Noh damn
Nkohya is new for me, my bad
F
Fi am making video hopefully today
Nyou can add a first advice " Change Tabs to LORA!")))
F i am gonna do nowFbut recording video and will show all
Nyey! but speed is lower than expected. I guess 4060ti is not the best one for training
Fare you using shared vram
Fspeed will get better
Fwait like 150 steps to see final speed
NI guess not
Nbut my ram is full I guess with model and clips
Fye not using
Fwait 150 steps
Fand show me speed
Ndidn't change much
N@Furkan Gözükara SECourses is there any method to train lora by parts, because not always there is possibility to have pc turned on?
N200 steps same
Fok
Fso to speed up you have 2 options
Fturn off t5 text encoder masking attention
Fvery little quality loss
Fspeed up signficiant
Fif still slow reduce training resolut
Flike 768x768
Fthis will hve more impact on quality
Ngot it. Will be trying new parameters tomorrow. Will lyk results
Nthanks
Fyou are welcome
XI was wondering, do warmup steps damage a lora?
Fwarmup steps means lower or higher LR
