Software Engineering Courses (SECourses)•16mo ago

600 steps 10 Gradient Accumlation steps batch size 2, sort of fixes the bleed

lilwashuu10•9/14/24, 1:42 AM

So 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?

MMaxivi 600 steps 10 Gradient Accumlation steps batch size 2, sort of fixes the bleed

Furkan Gözükara SECourses•9/14/24, 1:44 AM

the problem is that

Furkan Gözükara SECourses•9/14/24, 1:44 AM

when i train longer the images become realistic

Furkan Gözükara SECourses•9/14/24, 1:44 AM

that plastic and blurred generations of model is gone

Furkan Gözükara SECourses•9/14/24, 1:44 AM

but when you do lesser training you cant get that realism

MaxiviOP•9/14/24, 1:44 AM

mine works fine, anime

MMaxivi mine works fine, anime

Furkan Gözükara SECourses•9/14/24, 1:44 AM

anime works

Furkan Gözükara SECourses•9/14/24, 1:44 AM

realism missing

Llilwashuu10 So is it possible to train style and characters in the same finetune? I guess I ...

Furkan Gözükara SECourses•9/14/24, 1:45 AM

you can try but results i cant tell for sure

nicearoni•9/14/24, 1:45 AM

@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?

Nnicearoni @Dr. Furkan Gözükara @Maxivi @dsienra think all of you have mentioned bleeding b...

MaxiviOP•9/14/24, 1:46 AM

tried that already

MaxiviOP•9/14/24, 1:46 AM

faces blend into each other

Nnicearoni @Dr. Furkan Gözükara @Maxivi @dsienra think all of you have mentioned bleeding b...

Furkan Gözükara SECourses•9/14/24, 1:46 AM

ye will likely to bleed

Furkan Gözükara SECourses•9/14/24, 1:46 AM

i wonder what will happen with fine tune

MaxiviOP•9/14/24, 1:46 AM

conspiracy theory is flux is poisoned pilled

lilwashuu10•9/14/24, 1:47 AM

Yeah face bleed sucks, you have one asian woman lora that turns all your guys korean

Furkan Gözükara SECourses•9/14/24, 1:47 AM

the problem is that it internally captions (effect of it)

Furkan Gözükara SECourses•9/14/24, 1:47 AM

so it bleeds a lot

MaxiviOP•9/14/24, 1:47 AM

everyone from simple tuner discord is suggesting that you don't need to train the text encoders for flux

MaxiviOP•9/14/24, 1:47 AM

not sure why

MMaxivi everyone from simple tuner discord is suggesting that you don't need to train th...

Furkan Gözükara SECourses•9/14/24, 1:48 AM

clip l improves i tested

Furkan Gözükara SECourses•9/14/24, 1:48 AM

T5 not much diff :d

MaxiviOP•9/14/24, 1:48 AM

t5 is good if you're training lots of texts other wise i see no use for it

Furkan Gözükara SECourses•9/14/24, 1:48 AM

i dont know lokr details but it could be training different layers for each concept to not bleed :d

MaxiviOP•9/14/24, 1:48 AM

just hogs vram

Furkan Gözükara SECourses•9/14/24, 1:48 AM

once individual layer training arrives to kohya this can be used

MaxiviOP•9/14/24, 1:52 AM

seems to decrease quality

MMaxivi seems to decrease quality

Furkan Gözükara SECourses•9/14/24, 2:09 AM

yep

Furkan Gözükara SECourses•9/14/24, 2:09 AM

trade-off

MaxiviOP•9/14/24, 2:38 AM

@Furkan Gözükara SECourses it seems your massed compute creator is down

MaxiviOP•9/14/24, 2:38 AM

others work except yours

MaxiviOP•9/14/24, 2:39 AM

MaxiviOP•9/14/24, 2:40 AM

@NicB

MaxiviOP•9/14/24, 3:01 AM

works now

AIGambino•9/14/24, 3:23 AM

dr do you have a one click installer for dreambooth models for SDXL

AIGambino•9/14/24, 3:23 AM

I only see the flux version for making loras

AIGambino•9/14/24, 3:27 AM

or is onetrainer better for sdxl

MMaxivi works now

MaxiviOP•9/14/24, 3:49 AM

actually nevermind its stuck on intializing for 1 hour now

nicearoni•9/14/24, 5:30 AM

has 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?)

nicearoni•9/14/24, 7:19 AM

dxqb•9/14/24, 7:48 AM

anyone 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

dxqb•9/14/24, 8:11 AM

the 8 step hyper Lora also doesn't seem to work with forge, even though other Loras do

dxqb•9/14/24, 8:11 AM

strange

Nnicearoni Click to see attachment

PixelMonkey•9/14/24, 8:19 AM

I just hade a oh F.. moment

thanks!

Luciani•9/14/24, 12:01 PM

Anyone 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

dxqb•9/14/24, 12:05 PM

there is a setting that enables saving of images

Luciani•9/14/24, 12:07 PM

Oh yes -- that was it -- I don´t know how this was not checked anymore! Thanks!

MMaxivi actually nevermind its stuck on intializing for 1 hour now

Furkan Gözükara SECourses•9/14/24, 12:27 PM

i am testing now

Furkan Gözükara SECourses•9/14/24, 12:27 PM

could be that image was getting updated

AAIGambino dr do you have a one click installer for dreambooth models for SDXL

Furkan Gözükara SECourses•9/14/24, 12:28 PM

which models you need?

600 steps 10 Gradient Accumlation steps batch size 2, sort of fixes the bleed

Similar Threads

Similar Threads

Similar Threads