Software Engineering Courses (SECourses)•17mo ago

I completely had forgotten how much VRAM it was going to take to load up 40 Loras at once 😂 but thi

I completely had forgotten how much VRAM it was going to take to load up 40 Loras at once

but this works good for now

Kallamamran•8/27/24, 9:06 AM

Testing loras must be easier in Forge

Poppy•8/27/24, 9:06 AM

@Dr. Furkan Gözükara for kohya flux Lora training, should it save a safetensor at each epoch? As it hasn’t been saving them in my model folder for some reason. It says “epoch is incremented”, which I haven’t seen before. Pretty sure in the past with Sdxl Lora training, it says “epoch saved”

Poppy•8/27/24, 9:06 AM

Poppy•8/27/24, 9:36 AM

Also it’s just finished training and has provided just one safetensor

�𝔸𝖨ΞӾ I completely had forgotten how much VRAM it was going to take to load up 40 Lora...

Furkan Gözükara SECourses•8/27/24, 10:46 AM

i find swarmui best

PPoppy @Dr. Furkan Gözükara for kohya flux Lora training, should it save a safetensor a...

Furkan Gözükara SECourses•8/27/24, 10:46 AM

multi gpu?

Furkan Gözükara SECourses•8/27/24, 10:46 AM

you training 880 images or what?

Furkan Gözükara SECourses•8/27/24, 11:10 AM

clip l arrived i am gonna train

FFurkan Gözükara SECourses multi gpu?

Poppy•8/27/24, 11:12 AM

1 GPU

Poppy•8/27/24, 11:13 AM

Training 50 images with captions

PPoppy Training 50 images with captions

Furkan Gözükara SECourses•8/27/24, 11:13 AM

ok you have setup error

Furkan Gözükara SECourses•8/27/24, 11:13 AM

make folder repeat 1

Furkan Gözükara SECourses•8/27/24, 11:13 AM

train up to like 100 epochs at least

Furkan Gözükara SECourses•8/27/24, 11:13 AM

so it would be 5000 steps

Furkan Gözükara SECourses•8/27/24, 11:13 AM

save every 10 epoch

FFurkan Gözükara SECourses ok you have setup error

Poppy•8/27/24, 11:17 AM

Hmm, no I haven’t received an error.
I’m using your rank 4 config at the moment

Poppy•8/27/24, 11:18 AM

It’s strange as I tried the tank 4 config on 20 training images last night, and it made around 170,000 steps which surely isn’t correct?

PPoppy Hmm, no I haven’t received an error. I’m using your rank 4 config at the moment

Furkan Gözükara SECourses•8/27/24, 11:18 AM

not error folder setup

Poppy•8/27/24, 11:18 AM

It reached only 8% training in 12 hours

PPoppy It’s strange as I tried the tank 4 config on 20 training images last night, and ...

Furkan Gözükara SECourses•8/27/24, 11:18 AM

your folder setup inaccurate

Furkan Gözükara SECourses•8/27/24, 11:18 AM

make repeat 1

Furkan Gözükara SECourses•8/27/24, 11:18 AM

i, plan to start recording tuıtorial video today hopefully

Furkan Gözükara SECourses•8/27/24, 11:19 AM

pay attention to folder names

Poppy•8/27/24, 11:19 AM

Ah so here should be 1, not 40?

Furkan Gözükara SECourses•8/27/24, 11:19 AM

as a folder path i give this : R:\mass_compute\training_imgs

PPoppy Ah so here should be 1, not 40?

Furkan Gözükara SECourses•8/27/24, 11:20 AM

yes

Poppy•8/27/24, 11:20 AM

PPoppy Click to see attachment

Furkan Gözükara SECourses•8/27/24, 11:20 AM

make 1

Poppy•8/27/24, 11:20 AM

Is this the reason it was making 170,000 steps?

PPoppy Is this the reason it was making 170,000 steps?

Furkan Gözükara SECourses•8/27/24, 11:21 AM

yep

Furkan Gözükara SECourses•8/27/24, 11:21 AM

it makes 40x more

FFurkan Gözükara SECourses it makes 40x more

Poppy•8/27/24, 11:25 AM

Thank you, I’ll adjust this

Poppy•8/27/24, 11:25 AM

I’ve noticed training flux on Renaissance painting details is polishing the result (giving a slightly plastic feel), not picking up the granular textures of paint, crack etc. Do you know why is this?
Is there parameters such as learning rate which will effect this or perhaps it’s the architecture of flux itself?
When training the same images on kohya Sdxl, the trained model will clearly be show the texture of painting and canvas surface etc.

Kallamamran•8/27/24, 11:53 AM

I still think Kohya with it's folder-preparation and such feels like a stone-age tool :/

PPoppy I’ve noticed training flux on Renaissance painting details is polishing the resu...

Furkan Gözükara SECourses•8/27/24, 11:54 AM

we already use lower LR

Furkan Gözükara SECourses•8/27/24, 11:54 AM

what will impact is

Furkan Gözükara SECourses•8/27/24, 11:54 AM

training clip l text encoder - which arrived today

Furkan Gözükara SECourses•8/27/24, 11:54 AM

and full fine tuning

KKallamamran I still think Kohya with it's folder-preparation and such feels like a stone-age...

Furkan Gözükara SECourses•8/27/24, 11:54 AM

so true

Furkan Gözükara SECourses•8/27/24, 11:54 AM

onetrainer way better

Furkan Gözükara SECourses•8/27/24, 12:05 PM

currently with flux we are literally cooking the u-net

Furkan Gözükara SECourses•8/27/24, 12:05 PM

i trained with only ohwx word

Furkan Gözükara SECourses•8/27/24, 12:05 PM

didnt use in prompt only man

Furkan Gözükara SECourses•8/27/24, 12:05 PM

it still generates my image :d

Furkan Gözükara SECourses•8/27/24, 12:24 PM

new test results published

Furkan Gözükara SECourses•8/27/24, 12:24 PM

Furkan Gözükara SECourses•8/27/24, 12:24 PM

https://www.patreon.com/posts/110293257

Patreon

Kohya FLUX LoRA and Fine Tuning Training Full Tutorial For Local Wi...

Get more from SECourses: Tutorials, Guides, Resources, Training, FLUX, MidJourney, Voice Clone, TTS, ChatGPT, GPT, LLM, Scripts by Furkan Gözü on Patreon

smoke•8/27/24, 12:35 PM

Is this with the new clip_l?

smoke•8/27/24, 12:35 PM

Or unet only

Ssmoke Or unet only

Furkan Gözükara SECourses•8/27/24, 12:37 PM

still unet only

I completely had forgotten how much VRAM it was going to take to load up 40 Loras at once 😂 but thi

Similar Threads