Software Engineering Courses (SECourses)•13mo ago

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)?

QMOP•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)?

QMOP•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

QMOP•12/20/24, 4:36 PM

what wood make the captions bomb proof

QMOP•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 ...

QMOP•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

QMOP•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

Timson•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?

Timson•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...

Timson•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?