Software Engineering Courses (SECourses)•15mo ago

Hi, @Dr. Furkan Gözükara Did you try training OpenFLUX / Schnell with kohya's latest commits on sd3

Hi, @Dr. Furkan Gözükara Did you try training OpenFLUX / Schnell with kohya's latest commits on sd3-flux.1 branch ? I always get a NotImplemented error. tried on torch 2.4.1 and 2.5.0

DavsamaOP•10/7/24, 9:41 PM

tried a diffusers format and bfl too

DavsamaOP•10/7/24, 9:42 PM

NotImplementedError: Cannot copy out of meta tensor; no data! Please use torch.nn.Module.to_empty() instead of torch.nn.Module.to() when moving module from meta to a different device.

NotImplementedError: Cannot copy out of meta tensor; no data! Please use torch.nn.Module.to_empty() instead of torch.nn.Module.to() when moving module from meta to a different device.

anyway just wanted to know if its familiar to you

FFurkan Gözükara SECourses you need cuda 12.4, python 3.10 and cudnn 8.9.7 and c++ and visual studio

sergio_gonzo•10/7/24, 10:42 PM

Dr Furkan i hope everythong is ok, i just saw thta onetrainer support Flux , do you plan to do a tutorial in onetrainer sometime in the future?

pianogospel•10/7/24, 10:46 PM

Dr. how to set Kohya SS, in a DreamBooth FLUX Train, to save the train every "x" steps if I need to stop the train and restart it before exactly where I stopped?

xlads•10/7/24, 11:36 PM

got live portrait to work!

xlads•10/7/24, 11:36 PM

super good video going to try experimenting with other ai apps soon after watching the video i feel confident

Vexmachina•10/8/24, 2:16 AM

When fine tuning flux, any suggestions on lora_rank? The default on the replicate trainer is 16, and it says higher numbers capture more features but take longer to train. Any guidance on this?

xlads•10/8/24, 4:40 AM

https://hotshot.co

xlads•10/8/24, 4:41 AM

Haven’t tried it

IIWishIWereGood Hi everyone, should I use regularisation images when training dreambooth for sdx...

Furkan Gözükara SECourses•10/8/24, 6:50 AM

yep 100%

DDavsama Hi, @Dr. Furkan Gözükara Did you try training OpenFLUX / Schnell with kohya's l...

Furkan Gözükara SECourses•10/8/24, 6:50 AM

i tried official Schnell it was working

Furkan Gözükara SECourses•10/8/24, 6:50 AM

23.8 gb

Ssergio_gonzo Dr Furkan i hope everythong is ok, i just saw thta onetrainer support Flux , do ...

Furkan Gözükara SECourses•10/8/24, 6:51 AM

yes but they still dont have fp8

Furkan Gözükara SECourses•10/8/24, 6:51 AM

i am waiting upgrade of that

Furkan Gözükara SECourses•10/8/24, 6:51 AM

i already have some configs working

Furkan Gözükara SECourses•10/8/24, 6:53 AM

https://www.patreon.com/posts/onetrainer-fine-101606471

Patreon

OneTrainer Fine Tuning vs Kohya SS DreamBooth & Huge Research of On...

Get more from SECourses: FLUX, Tutorials, Guides, Resources, Training, Scripts on Patreon

Ppianogospel Dr. how to set Kohya SS, in a DreamBooth FLUX Train, to save the train every "x"...

Furkan Gözükara SECourses•10/8/24, 6:53 AM

you have to start a new training

Furkan Gözükara SECourses•10/8/24, 6:53 AM

it is same as lora

Furkan Gözükara SECourses•10/8/24, 6:53 AM

save every n steps

Xxlads super good video going to try experimenting with other ai apps soon after watchi...

Furkan Gözükara SECourses•10/8/24, 6:54 AM

awesome

VVexmachina When fine tuning flux, any suggestions on lora_rank? The default on the replicat...

Furkan Gözükara SECourses•10/8/24, 6:54 AM

we use 128

Furkan Gözükara SECourses•10/8/24, 6:54 AM

i compared lora rank impact

Furkan Gözükara SECourses•10/8/24, 6:54 AM

multiplem times

FFurkan Gözükara SECourses actually it is easy once you understand logic

Kallamamran•10/8/24, 6:58 AM

I'm ways from understanding the logic of Kohya training. To many parameters. I can't for the life of me understand why there is no better way to do it than the folder preparation

And I have watched your video several times. It's a looooong video

Kallamamran•10/8/24, 6:59 AM

In the video you say "You need to select LORA. Because we are currently training LORA"... Which we aren't since now we're talking finetuning, so it's a little bit confusing

KKallamamran In the video you say "You need to select LORA. Because we are currently training...

Furkan Gözükara SECourses•10/8/24, 6:59 AM

yes now select dreambooth :d

Furkan Gözükara SECourses•10/8/24, 6:59 AM

so there are 2 trainings we do

Furkan Gözükara SECourses•10/8/24, 6:59 AM

1 lora

Furkan Gözükara SECourses•10/8/24, 6:59 AM

1 dreambooth

Furkan Gözükara SECourses•10/8/24, 7:00 AM

i will also hopefully make a dreambooth tutorial

Furkan Gözükara SECourses•10/8/24, 7:00 AM

perfecting it

Furkan Gözükara SECourses•10/8/24, 7:00 AM

testing 15 imgs batch size 7 atm

Furkan Gözükara SECourses•10/8/24, 7:01 AM

Ddaroth Its impossible to set up a shared checkpoint and lora folder for different apps,...

Kallamamran•10/8/24, 7:05 AM

No it's not. Use symbolic links! This makes it easier -> https://schinagl.priv.at/nt/hardlinkshellext/linkshellextension.html

FFurkan Gözükara SECourses perfecting it

Kallamamran•10/8/24, 7:09 AM

BS 7 is way better

KKallamamran BS 7 is way better

Furkan Gözükara SECourses•10/8/24, 7:09 AM

i will compare it with bs 1

Furkan Gözükara SECourses•10/8/24, 7:09 AM

both 15 images

Furkan Gözükara SECourses•10/8/24, 7:10 AM

both 256 images

Furkan Gözükara SECourses•10/8/24, 7:10 AM

by the way i researched a unique LR for bs7 :d

Kallamamran•10/8/24, 7:13 AM

Looking at the right column it's way better than the left column

Kallamamran•10/8/24, 7:13 AM

I need to get into the "Kaggle" thing I guess. I can't keep locking up my computer

KKallamamran Looking at the right column it's way better than the left column

Furkan Gözükara SECourses•10/8/24, 7:14 AM

these are all same bs7

Furkan Gözükara SECourses•10/8/24, 7:14 AM

different checkpoints

Furkan Gözükara SECourses•10/8/24, 7:14 AM

finding best epopch

Kallamamran•10/8/24, 7:14 AM

Ran 2 trainings without a problem with AI-Toolkit over night though.. Unfortunately bad ones (As expected)

FFurkan Gözükara SECourses these are all same bs7

Kallamamran•10/8/24, 7:15 AM

Sorry read it wrong! Read BS 1 for the left column

Furkan Gözükara SECourses•10/8/24, 7:27 AM

15 images bs7 really good

Furkan Gözükara SECourses•10/8/24, 7:27 AM

not even cherry pick

Furkan Gözükara SECourses•10/8/24, 7:27 AM

FFurkan Gözükara SECourses 15 images bs7 really good

Error-404-unknown•10/8/24, 10:08 AM

Just out of curiosity, approximately how many steps was the checkpoint used for this image trained for? I'm just curious because for me even at around 2000 steps I'm still getting PlayStation2 CGI type looks even though I'm using real photos as a dataset (50 images). It starts to improve at much higher steps.

Hi, @Dr. Furkan Gözükara Did you try training OpenFLUX / Schnell with kohya's latest commits on sd3

Similar Threads