Same result with fresh .json... and it's way over
Same result with fresh .json... and it's way over
Fplease message me your json file
Ffrom private message
Falso show your vram before starting training
F@Zet your config is dreambooth corrupted config
Fwhy you dont follow what i say
DFresh install of Kohya_GUI_Flux_Installer_v40.zip and Kohya_FLUX_DreamBooth_v7.zip both yield:
ERROR: Cannot install -r requirements.txt (line 1), -r requirements.txt (line 10), diffusers[torch]==0.25.0 and huggingface-hub==0.24.5 because these package versions have conflicting dependencies.
ERROR: Cannot install -r requirements.txt (line 1), -r requirements.txt (line 10), diffusers[torch]==0.25.0 and huggingface-hub==0.24.5 because these package versions have conflicting dependencies.
Zchange requirements.txt huggingface-hub==0.24.5 to huggingface-hub==0.25.2
ZI am, or at least I am trying my very best
Fthis looks like bmaltasi error
Fwe need to report this
ZYes, some of the newer packages require huggingface-hub>=0.25. And careful, 0.26 removes cached_download
Zthat will need attention as well
Fi am making a fresh install to test
D@Dr. Furkan Gözükara New test Flux-De-Distilled with this settings the lora at 160 epochs looks severely undertrained, may be Flux-De-Distilled requieres higher LR
DI test it on the de-distilled model and on regular Flux-Dev
DI'm going to resume the training with LR 0.0001
Dand see if it converges
Ftotally could be
Fi will research it later hopefully
FGitHub
SD3 flux branch is broken when doing fresh install @kohya-ss @bmaltais Collecting wcwidth>=0.2.5 Using cached wcwidth-0.2.13-py2.py3-none-any.whl (34 kB) INFO: pip is looking at multiple version...
DI gonna resume from 160 epochs at 0.001 for 10 epochs and test if get signs of converging, I will report
Fguys dont use 24gb massed compute instances - or get at least 2x GPU for bigger RAM
F
VHow to know what the vram is when selecting a gpu? Its not shown for most gpus.
Fbest way is searching google
D@Dr. Furkan Gözükara now with lr at 0.0001 and 10 more epochs can see some resemblance with lora weight at 1.5, so is learning
EGeneral question on dreambooth finetuning: When switching from 15 img to 150 img, do you still train as many steps as with 15 img or is it 10x more?
Fsadly no formula
Ffor example 160*15 steps is best for 15 images
F80 * 256 best for 256 images
Fbatch size 1
EThanks a lot
FHey, is it now possible to train a style?
Fit was always
Fhere full info
Fyou guys tested these 2 against fp16?
Fe4m3fn vs e5m2
Fi guess i will grid test now
F
F32 GB RAM sucks so bad
Fok because you told me that (For stylized formations the flow is still not great) thank you very much
FThat's a good stuff to compare
Fwell if you want fully stylized person generation
Fyou should train on stylized images
Ffirst do fine tuning
Fgenerate stylized images
Fretrain
Fby the way fine tuning yields way better stylization
Fthan lora