Software Engineering Courses (SECourses)•2y ago

https://www.youtube.com/watch?v=0t5l6CP9eBg&t=1886s is this still the best training video?

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...

SSandstormleader https://www.youtube.com/watch?v=0t5l6CP9eBg&t=1886s is this still the best train...

Furkan Gözükara SECourses•7/28/24, 11:34 PM

yep

Furkan Gözükara SECourses•7/28/24, 11:34 PM

it is up to date

SandstormleaderOP•7/28/24, 11:34 PM

I lost my job so i have some extra time on my hands

SandstormleaderOP•7/28/24, 11:35 PM

time to learn now i guess

SSandstormleader I lost my job so i have some extra time on my hands

Furkan Gözükara SECourses•7/28/24, 11:35 PM

i hope you can find asap

SandstormleaderOP•7/28/24, 11:35 PM

going to try and just do freelance

SandstormleaderOP•7/28/24, 11:42 PM

what is the it/s supposed to be?

SSandstormleader what is the it/s supposed to be?

Furkan Gözükara SECourses•7/28/24, 11:45 PM

depends on gpu

Furkan Gözükara SECourses•7/28/24, 11:45 PM

what is your gpu

SandstormleaderOP•7/28/24, 11:45 PM

rtx 4090

SandstormleaderOP•7/28/24, 11:45 PM

i did the 15gb fast preset

Furkan Gözükara SECourses•7/28/24, 11:45 PM

1-1.2 range

SandstormleaderOP•7/28/24, 11:46 PM

okay that's what i'm getting was hoping i could get it faster haha

SSandstormleader okay that's what i'm getting was hoping i could get it faster haha

Furkan Gözükara SECourses•7/28/24, 11:46 PM

youy need to sacrifice quality with xformers

Furkan Gözükara SECourses•7/28/24, 11:46 PM

or batch size + xformers

SandstormleaderOP•7/28/24, 11:47 PM

oh i can use xformers? i think i need to rewatch and look for that part

SSandstormleader oh i can use xformers? i think i need to rewatch and look for that part

Furkan Gözükara SECourses•7/28/24, 11:49 PM

yes but that can reduce quality

Furkan Gözükara SECourses•7/28/24, 11:49 PM

xformers + turn off gradient checkpoint

Furkan Gözükara SECourses•7/28/24, 11:49 PM

will speed up

SandstormleaderOP•7/28/24, 11:52 PM

gradient is under learning rate scaler?

SandstormleaderOP•7/29/24, 12:01 AM

hmm xformers only took it up to 1.4

SSandstormleader hmm xformers only took it up to 1.4

Furkan Gözükara SECourses•7/29/24, 12:09 AM

ye you cant get much faster

Furkan Gözükara SECourses•7/29/24, 12:09 AM

did you disable gradient checkpoint?

SandstormleaderOP•7/29/24, 12:09 AM

i did

Furkan Gözükara SECourses•7/29/24, 12:10 AM

ok then that is what you can get

Furkan Gözükara SECourses•7/29/24, 12:10 AM

try batch size 2

Furkan Gözükara SECourses•7/29/24, 12:10 AM

but it may give out of vram error likely

highjourney•7/29/24, 8:52 AM

I'm currently trying to learn how to train a model like realvisXL (https://civitai.com/models/139562/realvisxl-v40).
I've been training LoRA models, but have no experience with checkpoint training.
To train a target-focused (realism) and also flexible model like realvisXL, do I just need a lot more images followed by detailed captions? Or is there something more to it?
If you have any suggestions or links to share, I'd be really grateful.

Hhighjourney I'm currently trying to learn how to train a model like realvisXL (https://civit...

Furkan Gözükara SECourses•7/29/24, 9:19 AM

you need a big dataset

Furkan Gözükara SECourses•7/29/24, 9:19 AM

proper captioning

Furkan Gözükara SECourses•7/29/24, 9:19 AM

and training

FFurkan Gözükara SECourses you need a big dataset

highjourney•7/29/24, 9:33 AM

How big to be exact?

Do I need 100+ image? or 1000+?
And for the captions, is it ok to use captioning tools such as blip or wd1.4? It seems impossible to manually caption thousands of images. Thanks.

Hhighjourney How big to be exact? Do I need 100+ image? or 1000+? And for the captions, is ...

Furkan Gözükara SECourses•7/29/24, 9:37 AM

you want to do a model like real vis XL 4?

Furkan Gözükara SECourses•7/29/24, 9:38 AM

you need 10s of thousands high quality image to fine to SDXL base into such model

Furkan Gözükara SECourses•7/29/24, 9:38 AM

you can use CogVLM for captioning

Furkan Gözükara SECourses•7/29/24, 9:38 AM

especially CogVLM v2

highjourney•7/29/24, 9:51 AM

I'm wondering if the Dreambooth method would be effective for training with 10,000 high-quality images. Based on the publicly available SDXL base model, there are various fine-tuning methods such as Dreambooth, LoRA, HyperNetwork, and Textual Inversion. From what I know, Dreambooth seems to be the most appropriate. I would appreciate it if you could share any specifically recommended methods for the final training process you emphasized, or if there are any noteworthy precedents for training with such a large dataset.
(P.S. I've watched all of your fine-tuning tutorial videos on YouTube multiple times. I'm a big fan! Thank you!)

Hhighjourney I'm wondering if the Dreambooth method would be effective for training with 10,0...

Furkan Gözükara SECourses•7/29/24, 10:31 AM

dreambooth without class images = fine tuning

Furkan Gözükara SECourses•7/29/24, 10:31 AM

it is best method

highjourney•7/29/24, 10:32 AM

by without class images you mean without regularization images right?

Hhighjourney by without class images you mean without regularization images right?

Furkan Gözükara SECourses•7/29/24, 10:34 AM

yep

Furkan Gözükara SECourses•7/29/24, 10:34 AM

class = reg same meaning

highjourney•7/29/24, 10:44 AM

Okay thanks a lot! Lastly, I'm curious if it would be appropriate to use a dynamic optimizer in this case as well. I've consistently used the Prodigy optimizer and cosine LR scheduler with LoRA, and I'm wondering if they would be effective when training with a large number of images.
Of course, frequent experimentation should be prioritized, but it would be greatly helpful if there's any guidance on where to set the starting point. Thank you!

Hhighjourney Okay thanks a lot! Lastly, I'm curious if it would be appropriate to use a dynam...

Furkan Gözükara SECourses•7/29/24, 10:54 AM

i tested them and i found adafactor works better

Furkan Gözükara SECourses•7/29/24, 10:54 AM

but it is a choice

Furkan Gözükara SECourses•7/29/24, 10:54 AM

before doing a big training

Furkan Gözükara SECourses•7/29/24, 10:54 AM

make smaller ones and see if works

highjourney•7/29/24, 11:40 AM

Thanks bunch. I'll try smaller version first and share the results if it turns out to be ok.

FFurkan Gözükara SECourses yep