Software Engineering Courses (SECourses)•17mo ago

I'm training same parameters no class and regularization images, in the first samples 5 epochs nobod

I'm training same parameters no class and regularization images, in the first samples 5 epochs nobody looks like me, this is good I think, the resemblance must apear in the next epochs

FFurkan Gözükara SECourses and man prompt is still fully impacted

dsienraOP•8/29/24, 7:33 PM

yes but that happens with sd and sdxl too at a some extent, it generalizes if you train with your name only it bleeds to people even to other genders if you overtrain enough

Ddsienra yes but that happens with sd and sdxl too at a some extent, it generalizes if yo...

Furkan Gözükara SECourses•8/29/24, 7:35 PM

sd and sdxl doesnt have exact same mecahnism though

Furkan Gözükara SECourses•8/29/24, 7:35 PM

so if you dont caption it will make huge impact

Furkan Gözükara SECourses•8/29/24, 7:35 PM

for example even though i fully captioned, it reduced the likelineess very small. impact on sdxl way bigger

dsienraOP•8/29/24, 7:36 PM

that's true

dsienraOP•8/29/24, 7:38 PM

Lets see, it needs more research, It will be cool to know exactly how the base model was trained

Ddsienra Lets see, it needs more research, It will be cool to know exactly how the base m...

Furkan Gözükara SECourses•8/29/24, 7:41 PM

yep

Furkan Gözükara SECourses•8/29/24, 7:57 PM

i am training turbo model

Furkan Gözükara SECourses•8/29/24, 7:57 PM

lets see what comes out of it

FFurkan Gözükara SECourses i am training turbo model

nicearoni•8/29/24, 8:27 PM

what differences are you seeing\expecting? starting with same configs used on dev? how does speed compare?

Nnicearoni what differences are you seeing\expecting? starting with same configs used on de...

Furkan Gözükara SECourses•8/29/24, 8:38 PM

still training. i dont expect it to work :d

m3m3x•8/29/24, 8:55 PM

Hello, trying to train with 3080 10Go on Linux but even with the config Rank_9_7514MB, my GPU is always OOM at the "move text encoders to gpu" step

m3m3x•8/29/24, 8:56 PM

any idea ?

Mm3m3x Hello, trying to train with 3080 10Go on Linux but even with the config Rank_9_7...

Furkan Gözükara SECourses•8/29/24, 9:06 PM

yes

Furkan Gözükara SECourses•8/29/24, 9:06 PM

you loaded config into dreambooth tab at least once

Furkan Gözükara SECourses•8/29/24, 9:06 PM

so extract a new config

Furkan Gözükara SECourses•8/29/24, 9:06 PM

load into lora

Furkan Gözükara SECourses•8/29/24, 9:06 PM

also you are using accurate branch sd3 flux right

Furkan Gözükara SECourses•8/29/24, 9:07 PM

how did you install kohya?

FFurkan Gözükara SECourses you loaded config into dreambooth tab at least once

m3m3x•8/29/24, 9:07 PM

yes did it the first time and then use the lora tab so I was thinking it was still ok, I will retry

Mm3m3x yes did it the first time and then use the lora tab so I was thinking it was sti...

Furkan Gözükara SECourses•8/29/24, 9:07 PM

it corrupts

m3m3x•8/29/24, 9:08 PM

https://github.com/bmaltais/kohya_ss.git I'm on branch sd3-flux.1 , in the subdir sd-script I'm on brach sd3

GitHub

GitHub - bmaltais/kohya_ss

Contribute to bmaltais/kohya_ss development by creating an account on GitHub.

Mm3m3x https://github.com/bmaltais/kohya_ss.git I'm on branch sd3-flux.1 , in the subd...

Furkan Gözükara SECourses•8/29/24, 9:10 PM

using python 3.10.11?

Furkan Gözükara SECourses•8/29/24, 9:10 PM

i did the vram tests on linux

Furkan Gözükara SECourses•8/29/24, 9:10 PM

so they are pretty accurate

FFurkan Gözükara SECourses using python 3.10.11?

m3m3x•8/29/24, 9:14 PM

python 3.12.15

Mm3m3x python 3.12.15

Furkan Gözükara SECourses•8/29/24, 9:15 PM

wont work

m3m3x•8/29/24, 9:15 PM

I commented out the part of the code that test the version

Furkan Gözükara SECourses•8/29/24, 9:15 PM

python is most annoying programming language

m3m3x•8/29/24, 9:18 PM

I will try to switch version, the thing is everything seems to load well before "INFO move text encoders to gpu "

m3m3x•8/29/24, 9:18 PM

"CUDA out of memory. Tried to allocate 80.00 MiB. GPU 0 has a total capacity of 9.67 GiB of which 42.69 MiB is free. Including non-PyTorch memory, this process has 8.10 GiB memory in use. Of the allocated memory 7.87 GiB is allocated by PyTorch, and 14.33 MiB is reserved by PyTorch but unallocated"

Mm3m3x I will try to switch version, the thing is everything seems to load well before ...

Furkan Gözükara SECourses•8/29/24, 9:21 PM

really dont waste time

Furkan Gözükara SECourses•8/29/24, 9:21 PM

reinstall with python 3.10.11

m3m3x•8/29/24, 9:22 PM

Ok thank you ! I will try that

Sharinel•8/29/24, 9:24 PM

got a question on Adetailer - how do I stop it 'feminizing' my male character? If I have male and female in the same picture it always adds lipstick or something to the male character. I could set it so adetailer works on onely 1 char, but no guarantee that it will pick the right one, plus sometimes you need to do both. Am I missing something?

SSharinel got a question on Adetailer - how do I stop it 'feminizing' my male character? I...

Furkan Gözükara SECourses•8/29/24, 9:25 PM

you are not missing. it would be your luck to get it accurate char with accurate prompt

Furkan Gözükara SECourses•8/29/24, 9:25 PM

it has multi window option

Sharinel•8/29/24, 9:25 PM

Rae•8/29/24, 9:27 PM

Ssmoke Pieter talked to the team behind Flux and obtained a commercial license from the...

sexyboi•8/29/24, 9:31 PM

Does anybody have any ideas how much this commercial license costs?

FFurkan Gözükara SECourses reinstall with python 3.10.11

m3m3x•8/29/24, 9:49 PM

I've tried with 3.10.11 it goes a bit further now it OOM at "caching Text Encoder outputs..."

Poppy•8/29/24, 9:53 PM

@Dr. Furkan Gözükara hey

on runpod when I deploy comfy ui template, Jupiter lab doesn’t allow me to open the checkpoint folder to add safetensors. Is this normal? I’m able to open the the other folders such as Lora etc

I’ve also tried to train kohya flux Lora using lion scheduler but it fails. Should it be possible to train with the lion scheduler?

pixelscience•8/29/24, 10:15 PM

I'm watching the new video now and I noticed that the CFG is set to one in SwarmUI. If I'm using regular comfyUI with Flux, is the CFG set to the default 7 (i think) still? I was researching how to set the CFG for Flux, but didn't know if I should be manually setting this to 1 or not in comfy.

pixelscience•8/29/24, 10:17 PM

Judging by how crappy the images out of swarm are at CFG 7. I'm going to assume it's set to 1 in comfy for flux by default.

Ppixelscience I'm watching the new video now and I noticed that the CFG is set to one in Swarm...

nicearoni•8/29/24, 10:29 PM

you can probably check the metadata of one of your generated images?

pixelscience•8/29/24, 10:33 PM

AIGambino•8/29/24, 10:51 PM

How long will 102 images take to train on flux for a lora at 1k res? Seems like its been 8 hours already for me lol

KGTachi•8/29/24, 11:19 PM

@Dr. Furkan Gözükara would it be possible to add torch 2.5.1 for supir? Also I noticed I got better results with an old version of supir with v37 and using stage one and stage two

AAIGambino How long will 102 images take to train on flux for a lora at 1k res? Seems like ...

nicearoni•8/30/24, 12:03 AM

it's really more about steps\epochs than 'how long' since the speed of cards varies so much...

Nnicearoni it's really more about steps\epochs than 'how long' since the speed of cards var...

AIGambino•8/30/24, 12:58 AM

3000 steps bs 1

I'm training same parameters no class and regularization images, in the first samples 5 epochs nobod

Similar Threads