Software Engineering Courses (SECourses)•15mo ago

Thanks this is really encouraging to see. People have been telling me you only need 2000 steps or it

Thanks this is really encouraging to see. People have been telling me you only need 2000 steps or it is overtrained, but most of my attempts look like your examples. 2000 steps PS3 style and closer to 6000+ start to look more photographic. BTW could you make a guide on how to tell if you have over or undertrained if you haven't already?

Ddsienra @Dr. Furkan Gözükara what is the best speed I can get full finetuning with my 30...

Furkan Gözükara SECourses•10/10/24, 12:21 PM

10 second / it

Furkan Gözükara SECourses•10/10/24, 12:21 PM

it is decent believe me worths it

EError-404-unknown Thanks this is really encouraging to see. People have been telling me you only n...

Furkan Gözükara SECourses•10/10/24, 12:21 PM

yes i will make a guide

Furkan Gözükara SECourses•10/10/24, 12:21 PM

actually preparing grids right now

Furkan Gözükara SECourses•10/10/24, 12:21 PM

will share all grids on patreon post

Furkan Gözükara SECourses•10/10/24, 12:21 PM

with explanation

FFurkan Gözükara SECourses 256 images batch size 1

OinkSauce•10/10/24, 12:34 PM

Curious what your conclusions will be on batch sizes, you seemed to be getting good results with 7

hamuz•10/10/24, 12:35 PM

How long would it take to train a Flux LoRA with 105 images? I'm using an RTX 3090

hamuz•10/10/24, 12:35 PM

with Dr. Furkan's best preset

FFurkan Gözükara SECourses 10 second / it

dsienra•10/10/24, 12:37 PM

extracting the lora then gives you a similar results?

Ddsienra extracting the lora then gives you a similar results?

Furkan Gözükara SECourses•10/10/24, 12:50 PM

extracted lora should give better than lora itself

Furkan Gözükara SECourses•10/10/24, 12:50 PM

lora training

daroth•10/10/24, 1:03 PM

I swear I'm getting better likeness, much quicker with 2 repeats. no overcooking

Papst•10/10/24, 1:10 PM

it seems like my W11 machine keeps going into engery savings mode while computing, any good idea on how to stop that, besides turning off the energy savings mechanisms?

cyberbol•10/10/24, 1:12 PM

Probably noon question. Please don't be mad. How I can add grid prompts to SwarmUI and use it?.

Or for example. I have trained my own face. Have 9 different loras. Now I want choose best. I'm doing all manually. Is any way to generate one Pic automatic using all loras and create a grid pictures?

FFurkan Gözükara SECourses i think i will pick 70 epoch

Dazzastrous [4090]•10/10/24, 1:52 PM

or 65

Ccyberbol Probably noon question. Please don't be mad. How I can add grid prompts to Swar...

dxqb•10/10/24, 2:28 PM

cannot answer, but that's the reason I'm using forge

PPapst it seems like my W11 machine keeps going into engery savings mode while computin...

Furkan Gözükara SECourses•10/10/24, 2:47 PM

yes did you set to performance?

Ccyberbol Probably noon question. Please don't be mad. How I can add grid prompts to Swar...

Furkan Gözükara SECourses•10/10/24, 2:48 PM

i have shown this

Furkan Gözükara SECourses•10/10/24, 2:48 PM

exactly

Furkan Gözükara SECourses•10/10/24, 2:48 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•10/10/24, 2:48 PM

49:46 Overfitting in FLUX training and training image quality
51:59 Finding the best checkpoint using the Grid Generator tool in Swarm UI
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
57:56 Finding the best LoRA checkpoint based on grid results

Furkan Gözükara SECourses•10/10/24, 3:17 PM

i am gonna share so much information even full logs :d

dsienra•10/10/24, 3:20 PM

@Dr. Furkan Gözükara Now flux lora training is much slower, before I achieved 4.8 - 4.5 it/s, now I'm around 8.5. why is that?

Ddsienra @Dr. Furkan Gözükara Now flux lora training is much slower, before I achieved 4....

Furkan Gözükara SECourses•10/10/24, 3:21 PM

who knows need full debug on computer :d

dsienra•10/10/24, 3:22 PM

but you experienced the same right?

dsienra•10/10/24, 3:22 PM

is something on kohya?

dsienra•10/10/24, 3:25 PM

It happened and then was fixed with torch 2.5

Ddsienra but you experienced the same right?

Furkan Gözükara SECourses•10/10/24, 3:40 PM

nope i didnt

Furkan Gözükara SECourses•10/10/24, 3:40 PM

torch 2.5 speeds up

Furkan Gözükara SECourses•10/10/24, 3:40 PM

you need to run properly to no return back 2.4.1

FFurkan Gözükara SECourses nope i didnt

dsienra•10/10/24, 3:44 PM

so I should have the same speed 4.8 - 4.5 it/s on lora training with the current kohya version?

dsienra•10/10/24, 3:54 PM

dsienra•10/10/24, 3:54 PM

I'm updating torch

dsienra•10/10/24, 3:55 PM

may be i was on an older version again

dsienra•10/10/24, 4:04 PM

your speed up script have pip uninstall xformers --yes

dsienra•10/10/24, 4:04 PM

why?

dsienra•10/10/24, 4:05 PM

xformers don't work with torch 2.5?

Vexmachina•10/10/24, 4:18 PM

Anyone else notice their flux fine tunes drop in likeness/quality really fast the second the subject gets far away from the camera/small? Like close up headshots look great, but a picture of the subject 20 feet away their face not only stops looking like them but starts being glitchy

Vexmachina•10/10/24, 4:18 PM

Any way to fix this?

dsienra•10/10/24, 4:57 PM

@Dr. Furkan Gözükara I updated torch to 2.5 but kohya installs 2.4.1 again when starting

dsienra•10/10/24, 6:14 PM

I fixed it --noverify

Error-404-unknownOP•10/10/24, 7:38 PM

Hi everyone. Maybe this isn't news to many of you but just wanted to share my experience. I finally got fed up with windows using about 1.5gb of vram. So yesterday I bit the bullet and installed Linux, first I tried Mint but just couldn't get the Nvidia drivers to work so I switched to Pop_OS. It was a battle and so many bash commands later finally got comfyui and kohya working. And OMG! In win11 for Flux Dev Q8 gguf 50 steps at 1024 it took about 1:30 per image generation, that's after loading the models another minute+. But in Linux I'm now getting image gen in 32 seconds and models also load in just a few seconds. I'm using a 3090, 12700k, 64gb . I just wanted to share because I can't belive the difference. If you haven't tried SD with Linux and have the patience to get things working it's definitely something worth looking into.

EError-404-unknown Hi everyone. Maybe this isn't news to many of you but just wanted to share my ex...

dsienra•10/10/24, 8:07 PM

that speed is slow for windows, the speed should be similar

Vigilence•10/10/24, 8:08 PM

@Dr. Furkan Gözükara I’m doing a flux Dora with the same dataset as the lora I showed you, quality is 1000x better.

VVigilence @Dr. Furkan Gözükara I’m doing a flux Dora with the same dataset as the lora I s...

dsienra•10/10/24, 8:10 PM

Dora is suported in kohya?

Vigilence•10/10/24, 8:19 PM

Onetrainer is what I’m using but I read that kohya scripts supports Dora, I can’t remember if that was a fork or implemented. I read it on either one of the kohya GitHubs.

dsienra•10/10/24, 8:23 PM

thanks

dsienra•10/10/24, 8:37 PM

@Dr. Furkan Gözükara do you know that parameter function? I'm training flux-dev-de-distill and this model have a optimal cfg scale of 3.5, I was testing leaving it at 1 but it worked just fine.

Ddsienra so I should have the same speed 4.8 - 4.5 it/s on lora training with the curren...

Furkan Gözükara SECourses•10/10/24, 9:09 PM

yep

Thanks this is really encouraging to see. People have been telling me you only need 2000 steps or it

Similar Threads