Software Engineering Courses (SECourses)•9mo ago

I'm also wondering why "accelerate" isn't updated in kohya_ss scripts. As I understand, newer versio

I'm also wondering why "accelerate" isn't updated in kohya_ss scripts. As I understand, newer versions allow using fp8 weights during training. Or am I wrong?

Original message was deleted

Furkan Gözükara SECourses•4/16/25, 10:58 PM

well masked training causes body disproportions

KKonstantin I'm also wondering why "accelerate" isn't updated in kohya_ss scripts. As I unde...

Furkan Gözükara SECourses•4/16/25, 10:59 PM

accelerate working fine i think

Original message was deleted

Furkan Gözükara SECourses•4/16/25, 11:01 PM

well when i describe background good enough it usually works

Furkan Gözükara SECourses•4/16/25, 11:01 PM

during inference

Original message was deleted

Furkan Gözükara SECourses•4/16/25, 11:04 PM

i still prefer just ohwx man

Furkan Gözükara SECourses•4/16/25, 11:04 PM

or woman

FFurkan Gözükara SECourses yep

John Cena•4/17/25, 1:44 AM

Is there any way to get around that?

JJohn Cena Is there any way to get around that?

Furkan Gözükara SECourses•4/17/25, 1:59 AM

not that i know

KonstantinOP•4/17/25, 7:48 AM

Hmm, without regularization with the same settings it turned out very well, from the 10th epoch the features were slightly visible, by 40 it was already very similar. Now the 90th epoch is underway, I don't see any overfitting yet.
This raises the question of why regularization complicates the process so much... Or is it all because my batch size and step accumulation are set to 1? Does it learn on 1 gradient and roll itself back when it hits regularization?

KonstantinOP•4/17/25, 7:59 AM

So should de-distill flux models work better with regularization? Or should the gradient update always contain at least one target image when regularizing so as not to forget it?

KonstantinOP•4/17/25, 8:16 AM

I'll probably try to experiment with Flex.1-alpha and de-distill models in fine-grained learning.
And also I wonder if fine-grained learning with block freezing will work? That is, if I, for example, train a fine model on 4 doubles and 4 singles (or any other configuration), then later, when extracting Lora of rank 256, I can remove all layers except those that were trained, getting a more compact lora, right? Or are others that I did not train also important when extracting lora?

Original message was deleted

Furkan Gözükara SECourses•4/17/25, 3:46 PM

which guide did you watch?

Original message was deleted

Furkan Gözükara SECourses•4/17/25, 3:48 PM

ye this is for sdxl

Furkan Gözükara SECourses•4/17/25, 3:48 PM

what you want to train?

Original message was deleted

Furkan Gözükara SECourses•4/17/25, 3:49 PM

ok then use flux

Furkan Gözükara SECourses•4/17/25, 3:49 PM

https://youtu.be/FvpWy1x5etM

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...

Original message was deleted

Furkan Gözükara SECourses•4/17/25, 3:52 PM

for sdxl

Furkan Gözükara SECourses•4/17/25, 3:52 PM

we use reg images

Furkan Gözükara SECourses•4/17/25, 3:52 PM

to use max reg images you need to use repeating and do 1 epoch

Furkan Gözükara SECourses•4/17/25, 3:52 PM

onetrainer better for sdxl

Furkan Gözükara SECourses•4/17/25, 3:52 PM

watch its tutorial

Furkan Gözükara SECourses•4/17/25, 3:52 PM

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...

TrumpYouAll•4/18/25, 4:42 PM

@Furkan Gözükara SECourses IS there any guidance / video from you to teach us how to train Lora for Wan 2.1 ?

�

🍭🎀 𝒜𝓋𝒶 𝐹𝓇𝒾𝑔𝑔 🎀🍭•4/18/25, 8:36 PM

Guys, could a person with tattoo be trained and the Tattoos be reproduced correctly? Would sdxl be able to do that too? Fine tuning @Furkan Gözükara SECourses

�🍭🎀 𝒜𝓋𝒶 𝐹𝓇𝒾𝑔𝑔 🎀🍭Guys, could a person with tattoo be trained and the Tattoos be reproduced correc...

TrumpYouAll•4/18/25, 9:20 PM

I can tell u it's very hard for SDXL, maybe possible for Flux

TTrumpYouAll <@205854764540362752> IS there any guidance / video from you to teach us how to ...

Furkan Gözükara SECourses•4/18/25, 9:20 PM

hopefully will make but i still couldnt have time yet

�🍭🎀 𝒜𝓋𝒶 𝐹𝓇𝒾𝑔𝑔 🎀🍭Guys, could a person with tattoo be trained and the Tattoos be reproduced correc...

Furkan Gözükara SECourses•4/18/25, 9:20 PM

sdxl hard

Furkan Gözükara SECourses•4/18/25, 9:20 PM

but flux is doing i trained dwayne johnson

Furkan Gözükara SECourses•4/18/25, 9:20 PM

https://civitai.com/models/911087/dwayne-johnson-aka-the-rock-flux-dev-fine-tuning-dreambooth-model-for-educational-and-research-purposes-dwayne-johnson-aka-the-rock-flux-dev-lora-model-for-educational-and-research-purposes-full-tutorial

FFurkan Gözükara SECourses hopefully will make but i still couldnt have time yet 😦

TrumpYouAll•4/18/25, 9:21 PM

Can your video to educate us on Wan 2.1 Lora involving both Local Computer Training and Massed Compute Training ? For now, can u tell me if I can train Wan 2.1 on kohya ss for Lora ?

TTrumpYouAll Can your video to educate us on Wan 2.1 Lora involving both Local Computer Train...

Furkan Gözükara SECourses•4/18/25, 9:25 PM

yes will do hopefully

Furkan Gözükara SECourses•4/18/25, 9:25 PM

sorry for delay so many stuff

Furkan Gözükara SECourses•4/18/25, 9:26 PM

wan 2.1 full quality lora is my priority task

�

🍭🎀 𝒜𝓋𝒶 𝐹𝓇𝒾𝑔𝑔 🎀🍭•4/18/25, 9:33 PM

@TrumpYouAll I'm curious if including photos zoomed in on tattoos with correct tagging would improve details and consistency

FFurkan Gözükara SECourses yes will do hopefully

TrumpYouAll•4/19/25, 1:33 AM

thanks Dr, I would liek to use Massed Compute as well as local computer to train Wan 2.1 Lora. Thanks Dr ! @Furkan Gözükara SECourses +

�🍭🎀 𝒜𝓋𝒶 𝐹𝓇𝒾𝑔𝑔 🎀🍭<@393809848883740672> I'm curious if including photos zoomed in on tattoos with...

TrumpYouAll•4/19/25, 1:34 AM

Sorry, no knowledge on tattoos but can't be done with SDXL

cat cafe•4/19/25, 1:39 AM

How should you determine which base model to train your dream booth/lora on? Is there a good comparison sheet somewhere

Ccat cafe How should you determine which base model to train your dream booth/lora on? Is ...

TrumpYouAll•4/19/25, 11:22 AM

The best model to start is always the 'base' model , not any 'trained' model

TTrumpYouAll thanks Dr, I would liek to use Massed Compute as well as local computer to train...

Furkan Gözükara SECourses•4/19/25, 11:24 AM

yes will cover both hopefully

Ccat cafe How should you determine which base model to train your dream booth/lora on? Is ...

Furkan Gözükara SECourses•4/19/25, 11:24 AM

flux base

Furkan Gözükara SECourses•4/19/25, 11:24 AM

best

FFurkan Gözükara SECourses flux base

cat cafe•4/19/25, 3:54 PM

If it's objectively best why are so many people still training on other models nowadays?

Ccat cafe If it's objectively best why are so many people still training on other models n...

Furkan Gözükara SECourses•4/19/25, 3:55 PM

testing :d

Furkan Gözükara SECourses•4/19/25, 3:56 PM

i tested these for example : https://medium.com/@furkangozukara/thoroughly-experimented-with-fine-tuning-dreambooth-training-of-flux-dev-de-distill-pixelwave-13fab8c625bf

Medium

Thoroughly experimented with Fine-Tuning / DreamBooth training of F...

Thoroughly experimented with Fine-Tuning / DreamBooth training of Flux-dev-de-distill, PixelWave v03, Verus Vision and base FLUX Dev model…

cat cafe•4/19/25, 3:56 PM

oh cool, thanks!!

Ccat cafe If it's objectively best why are so many people still training on other models n...

N7-Blackbird•4/20/25, 11:46 PM

I did one training on Flux DeDistilled, because negative prompts were important to me, and I am very satisfied with result. I didn't see any objective benefits from base model over the de-distilled one - and negative prompt is a huge benefit on the other hand.
that said, I guess it all depends on your personal goal and needs.

�🍭🎀 𝒜𝓋𝒶 𝐹𝓇𝒾𝑔𝑔 🎀🍭<@393809848883740672> I'm curious if including photos zoomed in on tattoos with...

Aikage•4/21/25, 8:18 PM

Yes a little, but in the end it will not work

Aikage•4/21/25, 8:18 PM

I have tried extensively. Same for piercings

Aikage•4/21/25, 8:20 PM

I had some small luck with a piercing https://civitai.com/user/Aikage you can see my Sanne character has black ear clips but they're not always consistent.

Aikage•4/21/25, 8:23 PM

You are correct the best way I have found is to take a lot of pictures of JUST the tattoo and tag it something, then take pictures of the character with the tattoo and without and when the tattoo is present tag it with the tattoo tag which will link it to the close up pictures. Still it mostly just gets the rough shape correct

I'm also wondering why "accelerate" isn't updated in kohya_ss scripts. As I understand, newer versio

Similar Threads

Similar Threads

Similar Threads