Software Engineering Courses (SECourses)•16mo ago

we are #1 in the sub 😄

we are #1 in the sub

Pelolisu•9/14/24, 12:37 AM

this nitro features is cool

FFurkan Gözükara SECourses we are #1 in the sub 😄

Pelolisu•9/14/24, 12:38 AM

Bravo, I think few people are doing training like this in Flux.

PPelolisu Bravo, I think few people are doing training like this in Flux.

Furkan Gözükara SECoursesOP•9/14/24, 12:38 AM

yep

Maxivi•9/14/24, 1:08 AM

@Dr. Furkan Gözükara how many steps do you normally train for and how long does it take you?

MMaxivi @Dr. Furkan Gözükara how many steps do you normally train for and how long does ...

Furkan Gözükara SECoursesOP•9/14/24, 1:09 AM

normally my experiments are targeting 150 epoch but i train up to 200 epochs with my 15 images dataset

Furkan Gözükara SECoursesOP•9/14/24, 1:10 AM

for last 256 images i trained up to 200 epoch again but 80 epoch was best

Furkan Gözükara SECoursesOP•9/14/24, 1:10 AM

used 8x gpu

Furkan Gözükara SECoursesOP•9/14/24, 1:10 AM

duration totally depends on config and gpu

Furkan Gözükara SECoursesOP•9/14/24, 1:10 AM

so normally 2250 steps - 15 images

Furkan Gözükara SECoursesOP•9/14/24, 1:10 AM

256 images were 256 * 80 / 8 steps for best checkpoint

Furkan Gözükara SECoursesOP•9/14/24, 1:11 AM

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

we are #1 in the sub 😄

Similar Threads