for me is running 2x slow but I'm training the text encoder, maybe is it?
for me is running 2x slow but I'm training the text encoder, maybe is it?
Fyes text encoder would slow down
Fi am yet to test it
Ffirst i am gonna test new parameter
Fstarting text encoder trainings too 
Fye total 8 new trainings starting
Fnow need to record widnows tutorial
Ffor kohya training
Didk the learning rate, to start I'm using the same 0.0001 for unet and the TE
Fi am testing them :d
Ffirst tests
F5e05 unet and clip
Fsecond is
F5e05 unet and 1e05 text encoder
Fother 6 tests about flow shift
DI went more aggressive with 1e04
Fvram usage only increased 300 mb
Fimpressive
F16 bit training
Fspeed loss is also minimal
Fif works great
Fi will update all configs
Fneed to find best LR
DI hade a speed loss of 2x
Fwhat is your gpu?
D3090
Fit sounds inaccurate
Flet me test
Fwhat is your second / it?
Falso pay attention to shared vram
Dmay be is the update
Fi am testing right now my speed on rtx 3090
Fye i have no such speed loss when training text encoder
Frestart your pc and empty vram usage
Fcan we download the lora we've trained or do we have to stay on the Citivai platform to generate the images?
Fyou can download
Fi downloaded
Fbut civitai default config is not that great
D@Dr. Furkan Gözükara this my speed and memory usage with TE
DWSL was super fast yesterday
D3.5 to 4 it/s
F3.5 to 4 is amazing
Ffor rtx 3090
Fwsl makes huge diff but hard to use for average people
Fi still couldnt find this difference reason it doesnt make sense
Fi compared venv and it is same as linux
Fno diff
Frecording the video 
Fwow i am tired. tutorial recorded 90 minutes without introduction
F
VIt's fine you can record it later