Software Engineering Courses (SECourses)•13mo ago

training on different resolutions in the dataset seems to be working fine for me

DD4UNTLESS As most of my training images are below 1024x1024, can I do padding to bring all...

Furkan Gözükara SECourses•12/20/24, 10:51 AM

it may work yes

TTimson training on different resolutions in the dataset seems to be working fine for me

Furkan Gözükara SECourses•12/20/24, 10:51 AM

yes i also say compare both method and see working good or not

ArtChicken•12/20/24, 11:16 AM

I'm retraining some human models that I trained between May and June on onetrainer (with the presets available at that time), and I have to say that in terms of similarity, the new trainings (based on SDXL and some realistic PDXL models that gave me great results) can't even come close to the ones I got a few months ago. So I asked myself: has the workflow on onetrainer changed a lot in the last months? Are there any special or different precautions to take into account when using the updated presets? I'm using fast tier v4... and I almost felt like trying to use fast tier v2 again to see if I can get the excellent models of the past months again...

TTimson training on different resolutions in the dataset seems to be working fine for me

D4UNTLESS•12/20/24, 11:57 AM

Thanks, Timson! I'll give it a go then (without padding)

AArtChicken I'm retraining some human models that I trained between May and June on onetrain...

Furkan Gözükara SECourses•12/20/24, 12:43 PM

they made so many updates

Furkan Gözükara SECourses•12/20/24, 12:43 PM

maybe broken

FFurkan Gözükara SECourses they made so many updates

ArtChicken•12/20/24, 2:06 PM

Can you suggest me what is the best guide currently on Patreon for doing SDXL/PDXL training (primarily... and maybe also LORA Flux)?

AArtChicken Can you suggest me what is the best guide currently on Patreon for doing SDXL/PD...

Furkan Gözükara SECourses•12/20/24, 2:50 PM

lora flux : 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•12/20/24, 2:50 PM

Full Instructions and Links Written Post (the one used in the tutorial)

https://www.patreon.com/posts/click-to-open-post-used-in-tutorial-110879657

Furkan Gözükara SECourses•12/20/24, 2:51 PM

sdxl : https://youtu.be/0t5l6CP9eBg

YouTubeSECourses

Full Stable Diffusion SD & XL Fine Tuning Tutorial With OneTrainer ...

In this tutorial, I am going to show you how to install OneTrainer from scratch on your computer and do a Stable Diffusion SDXL (Full Fine-Tuning 10.3 GB VRAM) and SD 1.5 (Full Fine-Tuning 7GB VRAM) based models training on your computer and also do the same training on a very cheap cloud machine from MassedCompute if you don't have such compute...

Furkan Gözükara SECourses•12/20/24, 2:51 PM

SDXL Configs Updated to Better Version After 20 New Experiment

https://www.patreon.com/posts/96028218

FFurkan Gözükara SECourses sdxl : https://youtu.be/0t5l6CP9eBg

ArtChicken•12/20/24, 3:14 PM

and if I want to make comparison with dreambooth finetuning (SDXL), where is the best tutorial on Patreon?

AArtChicken and if I want to make comparison with dreambooth finetuning (SDXL), where is the...

Furkan Gözükara SECourses•12/20/24, 3:22 PM

can you eloborate more

FFurkan Gözükara SECourses can you eloborate more

ArtChicken•12/20/24, 3:39 PM

I would like to train a SDXL model of the same person, with same training images and same base model on Onetrainer (which gave me bad results lately) and Kohya/Dreambooth... and then compare them...

ArtChicken•12/20/24, 3:40 PM

So I was looking for the best presets for both training softwares...

QM•12/20/24, 4:18 PM

@Furkan Gözükara SECourses I have tried now 6 different times Finetuning with two people in the photos, but it is not getting both of the subjects 100% - i evaluated from Epoch 140 - 280 -
from 15 - 30 images

QM•12/20/24, 4:19 PM

Doing it on 1 person - works like a magic

QM•12/20/24, 4:19 PM

i left all other settings same as you have in your configuration file

QQM i left all other settings same as you have in your configuration file

ArtChicken•12/20/24, 4:22 PM

what software are you using? and what type of training (SDXL, Flux)?

QM•12/20/24, 4:23 PM

Flux - Finetuning with Kohya based on @Furkan Gözükara SECourses video guide -

AArtChicken what software are you using? and what type of training (SDXL, Flux)?

QM•12/20/24, 4:23 PM

have you had any success for 2 or more people?

QQM have you had any success for 2 or more people?

ArtChicken•12/20/24, 4:36 PM

never tried such thing... I think it's very hard to achieve something good... unless you try with some bomb-proof captions

QM•12/20/24, 4:36 PM

what wood make the captions bomb proof

QM•12/20/24, 4:37 PM

I tried using ohwx man, bbuk woman, i also made long identifiers for the subjects,
It gets the the subjects about 80-90% but when you look at the images generated you know something is off

ArtChicken•12/20/24, 5:07 PM

well... if it's only "something" off I think that you got already a good result... sometimes some models are "all" off even with a single person in it

macba•12/20/24, 5:49 PM

Hi everyone

macba•12/20/24, 5:53 PM

@Furkan Gözükara SECourses do you have experience with training multiple models at a time with Kohya on multi-GPU systems? I read what you wrote about the bug leading to excess memory utilization when multi-GPU finetuning, so you don't get linear increases in training speed with increases in GPU numbers.

I've been lucky enough to get a lot of credits from the great guys at TensorDock so I could try with multi-H100 SXM systems; I can confirm the bug but I can also tell you that you can run 2 instances of Kohya at the same time, and in that case you do get double the training speed.

Of course this is only useful if you have more than one model to train in the first place…

macba•12/20/24, 6:00 PM

Every GPU is taking about 4.3 seconds per iteration (using batch 4), don't know how that compares. However, it's really a shame to see more than half the VRAM lying there unused

AArtChicken I would like to train a SDXL model of the same person, with same training images...

Furkan Gözükara SECourses•12/20/24, 9:21 PM

we have kohya config too

Furkan Gözükara SECourses•12/20/24, 9:21 PM

so you can compare checkpoints of both in swarmui

QQM <@205854764540362752> I have tried now 6 different times Finetuning with two peo...

Furkan Gözükara SECourses•12/20/24, 9:21 PM

it is limitation of flux

Furkan Gözükara SECourses•12/20/24, 9:22 PM

sadly we couldnt solve it yet

Mmacba <@205854764540362752> do you have experience with training multiple models at a ...

Furkan Gözükara SECourses•12/20/24, 9:22 PM

yes i do this all the time

Furkan Gözükara SECourses•12/20/24, 9:22 PM

you can start as many as instances of kohya and make each one run on each gpu seperately

Mmacba <@205854764540362752> do you have experience with training multiple models at a ...

QM•12/20/24, 9:31 PM

I ran 6x Kohya - finetuning 6 different Flux models on 4090s

FFurkan Gözükara SECourses it is limitation of flux

QM•12/20/24, 9:32 PM

what would be the best way, if one has multiple concepts that need to be brought together -

QQM what would be the best way, if one has multiple concepts that need to be brought...

Furkan Gözükara SECourses•12/20/24, 10:16 PM

sadly no way atm

Furkan Gözükara SECourses•12/20/24, 10:16 PM

you need to have them in same image during training

TimsonOP•12/20/24, 10:56 PM

@Furkan Gözükara SECourses can you use FLUX.1 Fill and FLUX.1 Depth with lora / dreambooth somehow? DO you have to train it separately? I'm having problem with yolov face inpaint when something in the square area of the face gets corrupted during the in-painting. Is there a way to do a non rectangular more advanced mask?

TimsonOP•12/20/24, 10:59 PM

It's pretty much guaranteed to fail if you have another face or object partially obstructing face caught in the inpaint square

TTimson <@205854764540362752> can you use FLUX.1 Fill and FLUX.1 Depth with lora / dream...

�

🍭🎀 𝒜𝓋𝒶 𝐹𝓇𝒾𝑔𝑔 🎀🍭•12/21/24, 11:37 AM

yes, when using depth/canny you can't use segment in the prompt, you will have to inpaint face after, which can easily be done on all of them using image batch under tools. use init image, set creativity to 0, and only prompt the segment face part then run batch.

�

🍭🎀 𝒜𝓋𝒶 𝐹𝓇𝒾𝑔𝑔 🎀🍭•12/21/24, 11:41 AM

@Furkan Gözükara SECourses question about flux dev fine tune and reg. images. A thought occured to me that if you were training something which is completely unknown to flux dev, say i.e. a particular alien race which lives on planet X, and your fine tune is a particular character of that race. Is this a situation were regularisation images would make sense? Training imgs would be the character, and reg. images just images of many different kind of aliens in this race. tokens ohwx planetxalien

�🍭🎀 𝒜𝓋𝒶 𝐹𝓇𝒾𝑔𝑔 🎀🍭yes, when using depth/canny you can't use segment in the prompt, you will have t...

TimsonOP•12/21/24, 12:26 PM

the question was: do I have to train separate model based on FLUX.1 Fill checkpoint for that? Or I have to train on the base flux dev then extract lora and apply it to flux fill?