whatever the purpose you have, you can trian your style generate images with it
whatever the purpose you have, you can trian your style generate images with it
Ftrain object
Fitem
DThe purpose of training or finetuning is to add new concepts people or things that the model don't have, what other reason you can have?
F100% that is the aim
Cbut you never do anything but pose images. we tried yesterday to get you to do fun, unusual images.
Fnope i can do all poses i need
Fi can even surf
Flol
Ci want to see an apple with your face
Fi dont wanna see it
Fsimple as that
Csee? boring pose imges
Fnope
Fi got all the fun and images i need 
Fi dont see any fun in such image
Fif i can ride a dinasaur an eagle that is more fun
Fbut that is not the aim
Fas i said
Ftrain object style item
Fthese are more important
For get professional photo for linkedin and such
Fthe only lacking thing is stylization atm
Fi hope we will also solve it eventually
Fit is poor
Cnot if you keep trying to train an untrainable model, you won't
Fwell people trying distillation i cant say for sure will work
Fbut we will see
DI'm training it right now, many concepts at the same time, lets see if it bleeds
Dnyanko7/flux-dev-de-distill
DI'm trying right now with flux-dev-de-distill If it doesn't wok I will inform you and we will know if you are right
Dthe only problem is the sample images on kohya It needs real cfg and not distilled cfg, may be sampling is not implemented
Cthat could be a problem.
Dyes if so I will have to wait until the training is finished
D@Dr. Furkan Gözükara I think that flux-dev-de-distill is slower training 7s/it is too slow for 3090ti?
Drank 128
DApply T5 Attention Mask off
DTrain T5-XXL off
Fnot very slow but 6 second would be better
Fexpected
Dok thanks, to speed it up for the first test Im training at 512x512 2.77s/it
Dproblem with samples on kohya
Dit puts the real cfg on 1
Dor 0 adk
Dis still traying to use distilled cfg and it doesn't do anything
Chave you tried using onetrainer?
Dnot yet
DI always use kohya
Dno problem, I will train the model without samples and then I can try the lora on forge
OHas anyone tried merging a bunch of fine tunes/Loras together? Does this work to have multiple concepts in one checkpoint or is the quality impacted?
DThe problem with the samples is reported to @kohya_tech
