it seems very low learning rate helps with bleeding
it seems very low learning rate helps with bleeding
Fpossibly
M600 steps 10 Gradient Accumlation steps batch size 2, sort of fixes the bleed
LSo is it possible to train style and characters in the same finetune? I guess I would have a folders 1_60sbar style and then 1_ohwx woman, if I wanted to train a woman from that time period in the same finetune?
Fthe problem is that
Fwhen i train longer the images become realistic
Fthat plastic and blurred generations of model is gone
Fbut when you do lesser training you cant get that realism
Mmine works fine, anime
Fanime works
Frealism missing
Fyou can try but results i cant tell for sure
N@Dr. Furkan Gözükara @Maxivi @dsienra think all of you have mentioned bleeding before... had a random idea thought I'd throw out lol - what about training separate loras for each subject with different token and class, then merge them one after another into the flux checkpoint? should be able to reference everything you trained on more separately?
Mtried that already
Mfaces blend into each other
Fye will likely to bleed
Fi wonder what will happen with fine tune
Mconspiracy theory is flux is poisoned pilled
LYeah face bleed sucks, you have one asian woman lora that turns all your guys korean
Fthe problem is that it internally captions (effect of it)
Fso it bleeds a lot
Meveryone from simple tuner discord is suggesting that you don't need to train the text encoders for flux
Mnot sure why
Fclip l improves i tested
FT5 not much diff :d
Mt5 is good if you're training lots of texts other wise i see no use for it
Fi dont know lokr details but it could be training different layers for each concept to not bleed :d
Mjust hogs vram
Fonce individual layer training arrives to kohya this can be used
Mseems to decrease quality
Fyep
Ftrade-off
M@Furkan Gözükara SECourses it seems your massed compute creator is down
Mothers work except yours
M
M@NicB
Mworks now
Adr do you have a one click installer for dreambooth models for SDXL
AI only see the flux version for making loras
Aor is onetrainer better for sdxl
Nhas anyone trained a flux lora on a logo? any advice? treat it more like a character or a style? captions or no? (if yes, really simple captions or joy captions?)
N
Danyone trying to use Forge UI for flux? Swarm UI is inefficient with Loras, it apparently combines and uploads them to GPU for each generation. So I was trying forge but I am getting memory leaks I think
Dthe 8 step hyper Lora also doesn't seem to work with forge, even though other Loras do
Dstrange
PI just hade a oh F.. moment 
thanks!
thanks!
LAnyone any idea why my Forge doesn´t save images anymore on the output folder I configured? I can´t find the generated images anywhere... The console seems normal and does´t bring any errors
Dthere is a setting that enables saving of images
LOh yes -- that was it -- I don´t know how this was not checked anymore! Thanks!
Fi am testing now
