starting single layer training experiments
starting single layer training experiments
Flora
Fsingle block training speed
Freally fast
AWhen you have Quality 1 and 2 in your Flux_Fine_Tune_v2.zip file, does '1' indicate the better quality?
TWow amazing, did you use your classical 15 image dataset or the new one with different expressions
Fclassical 15 images
Fi do experiment on it
Fyes quality 1 better
AI'm about to run dreambooth for flux - I presume that we still are using regularization images, as for SDXL?
Fdoing 8x single layer training 
Fno dont use reg images
Fit doesnt work
Aok
AI'll leave the reg folder location blank in kohya.
AI'm using your flux_fine_tune_v2 config.
Fyep good one
Fyes leave blank
Nlet us know how it goes
JI got this warning while doing install Step 2 for Kohya. The bat files you've provided automatically install xformers correct? I just ignore this warning?
L@Dr. Furkan Gözükara do you know anything about Fluxgym, have you tried it or is it in your plans? they say that it works with kohya sd-scripts... the most powerful training script for Stable Diffusion, Flux, etc.
I just saw that information that's why my query
I just saw that information that's why my query
Dhas anyone tried reg images with flux yet?
DI thought Kohya did not have that function, is it released now?
DDoes not work. Doc has tested and posted his experiments.
DIs this the ComfyUI one or the standalone website one?
AMy flux fine tuning is chugging away on a 24gb card. I notice that the s/its value gets a little longer with each epoch. It was at 9 s/it and has crept to 20 s/it by epoch 20. My vram sits at 24 gb with 0.6 gb in regular ram.
KThat flexibility!
KThat's dreambooth finetune?
KDoesn't work
Syou think 15 scar images would be enough to generate lora to draw realistic big cut scar on my face?
Fyep latest one will fix the issues dont worryu
Fi dont plan kohya gui working fine
Fi tried doesnt work
Fdid like 6-7 different trainings
Fpublished yesterday
Ftesting check points right now
Fyes
Fmore usually better
Fi dont see any reason to invest in this 
Fkohya gui working good and the guy is fixing errors
Fnone of the single layers looks like me i assume it will require a new learning rate
�You can use the layers he mentions here - https://www.reddit.com/r/StableDiffusion/comments/1f523bd/good_flux_loras_can_be_less_than_45mb_128_dim/?utm_source=share&utm_medium=web3x&utm_name=web3xcss&utm_term=1&utm_content=share_button
Reddit
Explore this post and more from the StableDiffusion community
Fi tested 0 to 7 none looks a like
Fi think new LR is mandatory
�might be the case then
Ttraining on layer 7 and layer 20 is not too bad. Using noise_scheduler: flowmatch, optimizer: adamw8bit and lr: 0.0005
Fye your LR is 10x of mine
Fi am doing more trainings
Fatm
Z@Dr. Furkan Gözükara so supir tool was built using cudatoolkit 11.8? but xformers is looking for pytorch 2.4.0 +cu121, i'm confused