Software Engineering Courses (SECourses)•5mo ago

@Dr. Furkan Gözükara Dr. Furkan. Do you think that training on the flux krea model in Kohya GUI Drea

@Dr. Furkan Gözükara Dr. Furkan. Do you think that training on the flux krea model in Kohya GUI DreamBooth will be the same as in your tutorial from 22.10.2024:

https://youtu.be/FvpWy1x5etM?si=CI-dObKmDprZesoN

Should I just replace flux.dev with Krea?

Or maybe during this time, some significant optimizations have been introduced that would also be worth changing? Any other options?

Are you planning to do a comparison of flux.dev vs Krea in DreamBooth training?

YouTubeSECourses

FLUX Full Fine-Tuning / DreamBooth Training Master Tutorial for Win...

If you want to train FLUX with maximum possible quality, this is the tutorial looking for. In this comprehensive tutorial, you will learn how to install Kohya GUI and use it to fully Fine-Tune / DreamBooth FLUX model. After that how to use SwarmUI to compare generated checkpoints / models and find the very best one to generate most amazing image...

JonkoXL•8/3/25, 6:16 PM

Wan Training

dschulie•8/3/25, 9:40 PM

is it a bad idea when training a character lora (SDXL) to use two sets of the same images - one in 1024x1024 and one in 896x1152 to be more flexible depending on which aspect ration/resolution i want to create my images with? and if it is not a bad idea, should i create one concept per resolution? would i have to use 2 concepts with reg images?

dschulie•8/3/25, 9:42 PM

and is there a benefit or downside in using full resolution images (in the correct aspect ratio) or should i downscale everything to the exact resolution (1024x1024/896x1152) or does it downscale on its own when the source images are higher resolution

KK4bronsky @Dr. Furkan Gözükara i was wondering if doing a finetune on krea is a bit differ...

Furkan Gözükara SECourses•8/3/25, 10:20 PM

i will test hopefully

Ccyberbol @Dr. Furkan Gözükara Dr. Furkan. Do you think that training on the flux krea mod...

Furkan Gözükara SECourses•8/3/25, 10:20 PM

yes i will do detailed test

JJonkoXL Wan Training 😄

Furkan Gözükara SECourses•8/3/25, 10:20 PM

lots of shared vrma

Furkan Gözükara SECourses•8/3/25, 10:20 PM

it must be very slow

Furkan Gözükara SECourses•8/3/25, 10:21 PM

ostris still doesnt support block swap

Furkan Gözükara SECourses•8/3/25, 10:21 PM

interesting did he add block swap?

Furkan Gözükara SECourses•8/3/25, 10:21 PM

it is not as slow as much i would say

Furkan Gözükara SECourses•8/3/25, 10:21 PM

what is your training resolution?

Ddschulie and is there a benefit or downside in using full resolution images (in the corre...

Furkan Gözükara SECourses•8/3/25, 10:22 PM

i would say make all images exact resolution

Furkan Gözükara SECourses•8/3/25, 10:22 PM

you probably can train with 2 different aspect ratio

Furkan Gözükara SECourses•8/3/25, 10:22 PM

make sure to enable bucketing

dschulie•8/3/25, 10:23 PM

thank you. would i need both sets of reg images? so in the end 2 image concepts and 2 reg image concepts? or would you just throw them all in one folder (1 image folder, 1 reg folder)

Ddschulie thank you. would i need both sets of reg images? so in the end 2 image concepts ...

Furkan Gözükara SECourses•8/3/25, 10:36 PM

good question. probably having different aspect ratio would help

Furkan Gözükara SECourses•8/3/25, 10:36 PM

you training sdxl right?

dschulie•8/3/25, 10:36 PM

yes sdxl

Furkan Gözükara SECourses•8/3/25, 10:37 PM

yes it should work

dschulie•8/3/25, 10:37 PM

currently i have two folders, one with 1024x1024 and one with 896x1152. so the question is: should i create one concept in onetrainer with both resolutions in one folder, or separate them and have two concepts

dschulie•8/3/25, 10:37 PM

and would i add 2 reg image concepts for either resolution

dschulie•8/3/25, 10:38 PM

or one concept with both reg resolutions inside

Furkan Gözükara SECourses•8/3/25, 10:49 PM

in one trainer

Furkan Gözükara SECourses•8/3/25, 10:49 PM

just add them as 2 concepts

Furkan Gözükara SECourses•8/3/25, 10:49 PM

and properly set their repeat count

dschulie•8/3/25, 10:51 PM

so 2 concepts for the images (1024x1024 and 896x1152). i have 50 images each. since it's 2 concepts do i use 50 or 100?

Ddschulie so 2 concepts for the images (1024x1024 and 896x1152). i have 50 images each. si...

Furkan Gözükara SECourses•8/3/25, 10:52 PM

logic simple

Furkan Gözükara SECourses•8/3/25, 10:52 PM

at 1 epoch you have 100 train images

Furkan Gözükara SECourses•8/3/25, 10:52 PM

then you should use 100 reg images

Furkan Gözükara SECourses•8/3/25, 10:52 PM

50 from each in that case

FFurkan Gözükara SECourses lots of shared vrma

JonkoXL•8/4/25, 7:05 AM

due to block swap and vae_cache_cpu enabled.
interestingly enough it made no difference in my case for speed whether i turn it on or not for that 480z360 dataset , tested speeds with and without. (without it fits in dedicated vram without shared mem as well)
(though usualy vram bottleneck does limit speeds in other instances)

JonkoXL•8/4/25, 7:06 AM

i am still messing around with what barely fits in vram, apparently as long as there is about 1gb vram dedicated available it will run full speed (400 watt at 70% power limit)

JonkoXL•8/4/25, 9:57 AM

For Wan2.1 T2V using (25 videos) 480x360@61 and (23) images 960x720 fits just perfectly in an RTX 5090. Wondering about the results, but this is the most I was sable to cram into the VRAM without going overboard and not using any offloading / shared vram

JonkoXL•8/4/25, 9:57 AM

FFurkan Gözükara SECourses yes i will do detailed test

cyberbolOP•8/4/25, 12:03 PM

Yesterday, I extracted a LORA from my Dreambooth model (of my character) that I trained in Kohya GUI on the FLUX.dev model. I wanted to see how the LORA would perform with the Flux Krea model, but unfortunately, it doesn't look great.

I guess I'll have to train a new LORA specifically on the Flux Krea model, or train an entire Dreambooth model from scratch.

JonkoXL•8/4/25, 1:13 PM

Makes sense I guess

Ccyberbol Yesterday, I extracted a LORA from my Dreambooth model (of my character) that I ...

AiInfluence•8/4/25, 3:05 PM

I can confirm. Lora trained on Dev inferenced on krea is not that good

AiInfluence•8/4/25, 3:05 PM

But a checkpoint with krea Lora works

AAiInfluence But a checkpoint with krea Lora works

cyberbolOP•8/4/25, 3:12 PM

I'm currently training a test LORA on the wan22 low noise model, specifically a LORA for a person or character. From what I understand, it's best to train it on a low noise model. It's a fast training, only 30 images in diffusion pipe wsl. It should take about 1.5 hours.

Ccyberbol I'm currently training a test LORA on the wan22 low noise model, specifically a ...

AiInfluence•8/4/25, 3:13 PM

lmk how it went

AiInfluence•8/4/25, 3:13 PM

I tried wan 2.2 t2i and I was very disappointed

AiInfluence•8/4/25, 3:13 PM

text was all wrong and textures / quality same as krea

cyberbolOP•8/4/25, 3:13 PM

I will