that seems to be the cause with adafactor at least for me, adamw i dont get it
that seems to be the cause with adafactor at least for me, adamw i dont get it
Fbut this happens at inference. actually i should compare with my high res training and see if happens on that too
MI will check mine again, i tried training on 1280 and 1536 learns details but overfitted fast
Fi also didnt see much improvement with that resolution
Fbut lowering below 1024 reduces a lot
Ftorch 2.5 release candidate published
Fi updated installer
Myou could try multires
full body images at 1536
half body at 1280
and portrait at 1024
full body images at 1536
half body at 1280
and portrait at 1024
Fmakes sense
Mwill test and see
Fworking nice i am testing right now
Fnot fast as linux but maybe because i use gpu a lot of power 
Mxfromers works?
Fwithout training with all apps
Mmaybe a speed up with supir 
Fi havent tested yet maybe
Fkohya works great
Mtoo many tests for you 
Fyep
Mneed to figure out how to duplicate you 3x
Mit 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