Software Engineering Courses (SECourses)•17mo ago

https://youtu.be/HKX8_F1Er_w

Zero to Hero Stable Diffusion 3 Tutorial with Amazing SwarmUI SD We...

Do not skip any part of this tutorial to master how to use Stable Diffusion 3 (SD3) with the most advanced generative AI open source APP SwarmUI. Automatic1111 SD Web UI or Fooocus are not supporting the #SD3 yet. Therefore, I am starting to make tutorials for SwarmUI as well. #StableSwarmUI is officially developed by the StabilityAI and your m...

Furkan Gözükara SECoursesOP•8/27/24, 2:03 AM

put first tab loras

Furkan Gözükara SECoursesOP•8/27/24, 2:03 AM

second tab prompts

Furkan Gözükara SECoursesOP•8/27/24, 2:03 AM

that is what i do

�

𝔸𝖨ΞӾ•8/27/24, 2:03 AM

I get it that this one is more powerful and advanced but holy shit it took me an hour just to get something close enough to what I was looking for. At first I just selected first field for Loras and autofilled that section and then it said it was generating a million images lol

FFurkan Gözükara SECourses https://youtu.be/HKX8_F1Er_w

�

𝔸𝖨ΞӾ•8/27/24, 2:04 AM

Yes I watched that video but you were showing the different upscalers there. I'll look again

�𝔸𝖨ΞӾ Yes I watched that video but you were showing the different upscalers there. I'l...

Furkan Gözükara SECoursesOP•8/27/24, 2:06 AM

just do small test

Furkan Gözükara SECoursesOP•8/27/24, 2:06 AM

like 2 lora

Furkan Gözükara SECoursesOP•8/27/24, 2:06 AM

2 prompt

Furkan Gözükara SECoursesOP•8/27/24, 2:06 AM

and see results

Furkan Gözükara SECoursesOP•8/27/24, 2:06 AM

that will give you idea

Furkan Gözükara SECoursesOP•8/27/24, 2:06 AM

also play with very top options

�

𝔸𝖨ΞӾ•8/27/24, 2:09 AM

Ok thank you @Dr. Furkan Gözükara I think I'm figuring out now. At first I was just using prompt settings from top left only but now I have the prompt on second tab and it's getting much more closer to what I was going for

�

𝔸𝖨ΞӾ•8/27/24, 2:11 AM

Sorry I get so easily frustrated, I wanted to figure it out on my own and not have to bother asking you until I had no other option. Very much appreciated, Dr!

Furkan Gözükara SECoursesOP•8/27/24, 2:59 AM

i got to sleep hoepfully see you tomorrow

�

𝔸𝖨ΞӾ•8/27/24, 3:48 AM

Goodnight Dr. I just ended up using this workflow instead in the Comfy tab: https://civitai.com/models/652483/flux-lora-tester

FLUX Lora Tester - v1.01 | Stable Diffusion Workflows | Civitai

Hello

for whatever reason you want to show me appreciation, you can:

Ko-Fi

With this workflow you just save time. A lot of time. Have fun. To...

�

𝔸𝖨ΞӾ•8/27/24, 3:50 AM

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 SECoursesOP•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 SECoursesOP•8/27/24, 10:46 AM

multi gpu?

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

you training 880 images or what?

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

ok you have setup error

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

make folder repeat 1

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

train up to like 100 epochs at least

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

so it would be 5000 steps

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

your folder setup inaccurate

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

make repeat 1

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

i, plan to start recording tuıtorial video today hopefully

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

yes

Poppy•8/27/24, 11:20 AM

PPoppy Click to see attachment

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

yep

Furkan Gözükara SECoursesOP•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 :/

https://youtu.be/HKX8_F1Er_w

Similar Threads