Software Engineering Courses (SECourses)•16mo ago

published a video as well haha 😄 https://www.youtube.com/watch?v=aVwu2Faw8Iw

published a video as well haha

https://www.youtube.com/watch?v=aVwu2Faw8Iw

YouTubeSECourses

FLUX LoRA: Free Hollywood-Quality Art at Home on Your PC - Unleash ...

These images are generated with FLUX LoRA training for free. Full workflow is as follows:

Medium article

https://medium.com/@furkangozukara/ultimate-flux-lora-training-tutorial-windows-and-cloud-deployment-abb72f21cbf8

Windows main tutorial

https://youtu.be/nySGu12Y05k

Cloud tutorial for GPU poor or scaling

https://youtu...

FFurkan Gözükara SECourses published a video as well haha 😄 https://www.youtube.com/watch?v=aVwu2Faw8Iw

Maxivi•9/14/24, 1:13 AM

wont it get copyrighted for the music haha

FFurkan Gözükara SECourses normally my experiments are targeting 150 epoch but i train up to 200 epochs wit...

Maxivi•9/14/24, 1:17 AM

have you tried multires training yet with flux shift timestep?

MMaxivi wont it get copyrighted for the music haha

Furkan Gözükara SECoursesOP•9/14/24, 1:18 AM

it is loyalty free

MMaxivi have you tried multires training yet with flux shift timestep?

Furkan Gözükara SECoursesOP•9/14/24, 1:18 AM

nope

Furkan Gözükara SECoursesOP•9/14/24, 1:18 AM

but when i increase inference resolution

FFurkan Gözükara SECourses it is loyalty free

Maxivi•9/14/24, 1:18 AM

nice video, quality has improved for you

Furkan Gözükara SECoursesOP•9/14/24, 1:18 AM

it adds vertical lines

FFurkan Gözükara SECourses it adds vertical lines

Maxivi•9/14/24, 1:19 AM

that seems to be the cause with adafactor at least for me, adamw i dont get it

MMaxivi that seems to be the cause with adafactor at least for me, adamw i dont get it

Furkan Gözükara SECoursesOP•9/14/24, 1:19 AM

but this happens at inference. actually i should compare with my high res training and see if happens on that too

Maxivi•9/14/24, 1:21 AM

I will check mine again, i tried training on 1280 and 1536 learns details but overfitted fast

MMaxivi I will check mine again, i tried training on 1280 and 1536 learns details but ov...

Furkan Gözükara SECoursesOP•9/14/24, 1:23 AM

i also didnt see much improvement with that resolution

Furkan Gözükara SECoursesOP•9/14/24, 1:23 AM

but lowering below 1024 reduces a lot

Furkan Gözükara SECoursesOP•9/14/24, 1:23 AM

torch 2.5 release candidate published

Furkan Gözükara SECoursesOP•9/14/24, 1:23 AM

i updated installer

Maxivi•9/14/24, 1:23 AM

you could try multires

full body images at 1536
half body at 1280
and portrait at 1024

Furkan Gözükara SECoursesOP•9/14/24, 1:24 AM

makes sense

Maxivi•9/14/24, 1:25 AM

will test and see

Furkan Gözükara SECoursesOP•9/14/24, 1:26 AM

working nice i am testing right now

Furkan Gözükara SECoursesOP•9/14/24, 1:27 AM

not fast as linux but maybe because i use gpu a lot of power

Maxivi•9/14/24, 1:27 AM

xfromers works?

Furkan Gözükara SECoursesOP•9/14/24, 1:27 AM

without training with all apps

Maxivi•9/14/24, 1:27 AM

maybe a speed up with supir

Furkan Gözükara SECoursesOP•9/14/24, 1:27 AM

i havent tested yet maybe

Furkan Gözükara SECoursesOP•9/14/24, 1:28 AM

kohya works great

Maxivi•9/14/24, 1:28 AM

too many tests for you

Furkan Gözükara SECoursesOP•9/14/24, 1:30 AM

yep

Maxivi•9/14/24, 1:31 AM

need to figure out how to duplicate you 3x

Maxivi•9/14/24, 1:32 AM

it seems very low learning rate helps with bleeding

MMaxivi it seems very low learning rate helps with bleeding

Furkan Gözükara SECoursesOP•9/14/24, 1:34 AM

possibly

Maxivi•9/14/24, 1:42 AM

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 SECoursesOP•9/14/24, 1:44 AM

the problem is that

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

when i train longer the images become realistic

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

that plastic and blurred generations of model is gone

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

but when you do lesser training you cant get that realism

Maxivi•9/14/24, 1:44 AM

mine works fine, anime

MMaxivi mine works fine, anime

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

anime works

Furkan Gözükara SECoursesOP•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 SECoursesOP•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?