Software Engineering Courses (SECourses)•17mo ago

@Dr. Furkan Gözükara new test with high TE learning rate, now with unique tokens no class and person

@Dr. Furkan Gözükara new test with high TE learning rate, now with unique tokens no class and person class regularization, I see that using person class avoid class mixing

dsienraOP•8/31/24, 2:10 PM

these are my regularizations

dsienraOP•8/31/24, 2:11 PM

works much better if I use man class it overtrains the class man, en mix them all

dsienraOP•8/31/24, 2:12 PM

is a couple of thousands of regularization pictures

Ddsienra @Dr. Furkan Gözükara new test with high TE learning rate, now with unique tokens...

Furkan Gözükara SECourses•8/31/24, 2:13 PM

do you see improved results?

Furkan Gözükara SECourses•8/31/24, 2:14 PM

i am about to compare 800 images generation

Furkan Gözükara SECourses•8/31/24, 2:14 PM

waiting grid to finish

Furkan Gözükara SECourses•8/31/24, 2:14 PM

780 - each one 48 steps FLUX image

FFurkan Gözükara SECourses do you see improved results?

dsienraOP•8/31/24, 2:15 PM

yes but is a preliminary test, I stopped and restarted the training with different settings a couple of times

dsienraOP•8/31/24, 2:15 PM

now I started the real test

dsienraOP•8/31/24, 2:16 PM

I will report in a few hours

dsienraOP•8/31/24, 2:17 PM

remember my goal is to train multiple people no concept or class bleading

dsienraOP•8/31/24, 2:17 PM

lest see if I succeed

Furkan Gözükara SECourses•8/31/24, 2:18 PM

great

dsienraOP•8/31/24, 2:19 PM

In my last test I wasn't using unique tokens and I changed tokens and removed class in the middle of a training resuming..

dsienraOP•8/31/24, 2:19 PM

It was a mess

dsienraOP•8/31/24, 2:22 PM

other note, all my regularization are captioned

dsienraOP•8/31/24, 2:33 PM

27hs left

Ddsienra 27hs left

Furkan Gözükara SECourses•8/31/24, 2:34 PM

nice. is this single gpu?

dsienraOP•8/31/24, 2:35 PM

rtx 3090

dsienraOP•8/31/24, 2:35 PM

multiple gpu can be used on kohya?

Furkan Gözükara SECourses•8/31/24, 2:35 PM

ok it is really good speed

Ddsienra multiple gpu can be used on kohya?

Furkan Gözükara SECourses•8/31/24, 2:36 PM

of course

Furkan Gözükara SECourses•8/31/24, 2:36 PM

i recorded tutorial

Furkan Gözükara SECourses•8/31/24, 2:36 PM

you can even use 8 gpus :d

Furkan Gözükara SECourses•8/31/24, 2:36 PM

editing the tutorial

dsienraOP•8/31/24, 2:36 PM

I have just one

dsienraOP•8/31/24, 2:36 PM

Furkan Gözükara SECourses•8/31/24, 2:39 PM

Furkan Gözükara SECourses•8/31/24, 2:40 PM

you can rent on massed compute

Furkan Gözükara SECourses•8/31/24, 2:40 PM

1.25$ per hour 4x A6000 GPU

Furkan Gözükara SECourses•8/31/24, 2:42 PM

ok i am gonna update configs with clip l

Furkan Gözükara SECourses•8/31/24, 2:42 PM

there is slight improvement

Furkan Gözükara SECourses•8/31/24, 2:43 PM

i will make it seperate folder

FFurkan Gözükara SECourses you can even use 8 gpus :d

dsienraOP•8/31/24, 2:45 PM

how it works you increment the batch size, it proceses 1 step per gpu or what?

FFurkan Gözükara SECourses you can rent on massed compute

dsienraOP•8/31/24, 2:46 PM

yes I know but I'm just doing it for fun on my own computer

Ddsienra how it works you increment the batch size, it proceses 1 step per gpu or what?

Furkan Gözükara SECourses•8/31/24, 2:48 PM

explained in video please wait it :d

dsienraOP•8/31/24, 2:49 PM

sure thanks

umuthasanoglu•8/31/24, 4:41 PM

Hey guys, how can I use .gguf format models in SwarmUI?

umuthasanoglu•8/31/24, 4:41 PM

I wanted to try this: https://civitai.com/models/705681?modelVersionId=789392

GGUF: HyperFlux 16-Steps (Flux.1 Dev + ByteDance HyperSD LoRA) - Q8...

Warning: This is the 16-step model. The faster 8-step model can be found HERE [Note: Unzip the download to get the GGUF. Civit doesn't support it n...

Uumuthasanoglu Hey guys, how can I use .gguf format models in SwarmUI?

Furkan Gözükara SECourses•8/31/24, 5:04 PM

i am gonna ask this now

umuthasanoglu•8/31/24, 5:05 PM

I just figured it out. For those who want to try out: In order to work, you need to manually edit metadata of the model and select architecture type. All gguf models goes into unet folder and after restart it automatically downloads Comfy GGUF Workflow.

Uumuthasanoglu I wanted to try this: https://civitai.com/models/705681?modelVersionId=789392

umuthasanoglu•8/31/24, 5:06 PM

I'm testing this right now but so far it looks great for 12-16 gb VRAM GPU's.

druhl•8/31/24, 5:19 PM

Is there like a ComfyUI XY Plot workflow to test/ compare multiple Flux Lora models?

Ddruhl Is there like a ComfyUI XY Plot workflow to test/ compare multiple Flux Lora mod...

Furkan Gözükara SECourses•8/31/24, 5:20 PM

why not use swarm?

FFurkan Gözükara SECourses why not use swarm?

druhl•8/31/24, 5:21 PM

Would it mean relearning a lot of stuff?

Ddruhl Would it mean relearning a lot of stuff?

Furkan Gözükara SECourses•8/31/24, 5:21 PM

nope

Furkan Gözükara SECourses•8/31/24, 5:21 PM

i shown how to make

Furkan Gözükara SECourses•8/31/24, 5:22 PM

https://youtu.be/nySGu12Y05k

YouTubeSECourses

FLUX LoRA Training Simplified: From Zero to Hero with Kohya SS GUI ...

Ultimate Kohya GUI FLUX LoRA training tutorial. This tutorial is product of non-stop 9 days research and training. I have trained over 73 FLUX LoRA models and analyzed all to prepare this tutorial video. The research still going on and hopefully the results will be significantly improved and latest configs and findings will be shared. Please wat...

Furkan Gözükara SECourses•8/31/24, 5:22 PM

52:55 Grid Generator tool for selecting LoRAs and prompts
53:59 Generating the grid and expected results
56:57 Analyzing grid results in Swarm UI

Furkan Gözükara SECourses•8/31/24, 5:22 PM

way easier than comfyui

@Dr. Furkan Gözükara new test with high TE learning rate, now with unique tokens no class and person

Similar Threads