Software Engineering Courses (SECourses)•15mo ago

Hello community, I am having a 2gb lora on SD , it was like 147mo only for 40mn , what I am doi

Hello community, I am having a 2gb lora on SD , it was like 147mo only for 40mn ,
what I am doing wrong now ? the file is bigger and takes about 3h , the lora does not even look alike

heeeelp

SSiggiSmilez But the checkpoint size is around the same size as flux?

Furkan Gözükara SECourses•10/14/24, 2:08 PM

yes

Furkan Gözükara SECourses•10/14/24, 2:08 PM

23.8 gb

Mmayna Hello community, I am having a 2gb lora on SD , it was like 147mo only for 4...

Furkan Gözükara SECourses•10/14/24, 2:09 PM

lora network dimension

Furkan Gözükara SECourses•10/14/24, 2:09 PM

increases size

maynaOP•10/14/24, 2:10 PM

I used dreambooth training ( not lora only ) Tab 1 not 2 , what should I use ( for face only lora )

maynaOP•10/14/24, 2:15 PM

it worked you are the best @Dr. Furkan Gözükara

dsienra•10/14/24, 3:16 PM

@Dr. Furkan Gözükara Update on my https://huggingface.co/nyanko7/flux-dev-de-distill training tests. Some bad news and some good news, Good news: I can train many subjects at the same time without bleeding between each other 11 people in one LORA, simple captions "Name class" note: the token must be different a saw a little bleeding with similar names like "Diego man" "Dani man" works best with "NameLastname class" so they end up been very different. Bad news: there is still some class bleeding, may be my fault because I was using a higher lr than recommended to get faster results, Other thing I was using faster presets rather than quality presets "Apply T5 Attention Mask disabled" "Train T5-XXL disabled" now I'm testing with dose enabled, other bad news regularization images still reduces resemblance, I will try this again with "Apply T5 Attention Mask enabled" and "Train T5-XXL enabled" and with the recommended learning rate. The other option that I didn't try was "Guidance scale for Flux1" flux-dev-de-distill has a default cfg of 3.5, that option I leave it at 1 for now because I'm using regular Flux-Dev for inference is another thing I have to test. I will report my Updates.

nyanko7/flux-dev-de-distill · Hugging Face

Ddsienra @Dr. Furkan Gözükara Update on my https://huggingface.co/nyanko7/flux-dev-de-dis...

Furkan Gözükara SECourses•10/14/24, 3:18 PM

awesome

Furkan Gözükara SECourses•10/14/24, 3:18 PM

hopefully it will be my next research after i make tutorial

Furkan Gözükara SECourses•10/14/24, 3:48 PM

public article : https://www.patreon.com/posts/113970485

Patreon

Huge FLUX LoRA vs Fine Tuning / DreamBooth Experiments Completed, M...

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

Ddsienra @Dr. Furkan Gözükara Update on my https://huggingface.co/nyanko7/flux-dev-de-dis...

deepblue_1337•10/14/24, 4:03 PM

So you add the image captions, is that correct? Literally "NameLastname class"? so "JohnDoe class"?

Ddeepblue_1337 So you add the image captions, is that correct? Literally "NameLastname class"? ...

dsienra•10/14/24, 4:44 PM

yes thats it

dsienra•10/14/24, 4:46 PM

"NameLastname man" "NameLastname woman" etc.

SiggiSmilez•10/14/24, 5:16 PM

I have questions about Flux Lora training especially about prompting for characters:

SiggiSmilez•10/14/24, 5:16 PM

Do you prompt at all?
Do you use tags or natural language?

SiggiSmilez•10/14/24, 5:17 PM

What do you describe in the prompt?
Everything that is important about the character or everything that is not important about the character?
Or everything in the training picture?

SSiggiSmilez I have questions about Flux Lora training especially about prompting for charact...

dsienra•10/14/24, 5:21 PM

I use "Charactername class and the normal prompt" or "Charactername class, prompt"

dsienra•10/14/24, 5:22 PM

in the prompt put all you want, the clothing the setting location what is doing, etc

Ddsienra I use "Charactername class and the normal prompt" or "Charactername class, promp...

Sipriyani•10/14/24, 5:26 PM

are you using kohya? if so how do you feel these boxes:

Insance prompt: diego man,
and class prmpt: man .

dsienra•10/14/24, 5:27 PM

Insance prompt: Diego

dsienra•10/14/24, 5:27 PM

class prmpt: man

dsienra•10/14/24, 5:28 PM

that's for training

dsienra•10/14/24, 5:28 PM

then promt like I said

dsienra•10/14/24, 5:29 PM

Dr. Furkan have a great list of prompts

fadounav•10/14/24, 6:57 PM

Hey, @Dr. Furkan Gözükara I think 256 images is a lot, can I make 64 for a Dreamooth?

Ffadounav Hey, @Dr. Furkan Gözükara I think 256 images is a lot, can I make 64 for a Drea...

druhl•10/14/24, 7:18 PM

Yes, you can

fadounav•10/14/24, 7:25 PM

thanks

fadounav•10/14/24, 7:25 PM

@druhl

SSiggiSmilez Do you prompt at all? Do you use tags or natural language?

Furkan Gözükara SECourses•10/14/24, 7:27 PM

i shared test prompts check them out

Furkan Gözükara SECourses•10/14/24, 7:27 PM

and i dont caption

Furkan Gözükara SECourses•10/14/24, 7:27 PM

only

Furkan Gözükara SECourses•10/14/24, 7:27 PM

ohwx man

Ffadounav Hey, @Dr. Furkan Gözükara I think 256 images is a lot, can I make 64 for a Drea...

Furkan Gözükara SECourses•10/14/24, 7:27 PM

100%

Furkan Gözükara SECourses•10/14/24, 7:27 PM

you see even 15 images works :d

Furkan Gözükara SECourses•10/14/24, 7:44 PM

kohya is broken right now - newest version

Furkan Gözükara SECourses•10/14/24, 7:45 PM

https://github.com/kohya-ss/sd-scripts/issues/1696

GitHub

AttributeError: 'bool' object has no attribute 'num_processes' · Is...

Urgent fix needed for SD3 branch @kohya-ss this error is after latest @bmaltais update this config worked yesterday INFO prepare accelerator flux_train.py:177 accelerator device: cuda Saving traini...

ar.1•10/14/24, 7:51 PM

Hey guys, I'm looking to do a big finetune on FLUX1. Dev, not just a training for a single character, more like a big model finetune.

The aim is to get more realistic looking images of people. I know it was possible for SD 1.5 as I currently work with finetuned version of SD 1.5 that is making more realistic images than FLUX.

I have a few thousand high quality and well labelled images, and can get a lot more if that is required, so my main question is:

Is this possible to do with flux yet ? If so, can anyone point me in the direction of githubs or resources to find more info on how I can do this