Software Engineering Courses (SECourses)•3y ago

Is there any difference between sdxl finetuning and kohya's dreambooth? Since both methods use sdxl_

Is there any difference between sdxl finetuning and kohya's dreambooth? Since both methods use sdxl_train.py?
(except for the captioning method..)

__Captain_Is there any difference between sdxl finetuning and kohya's dreambooth? Since bo...

Furkan Gözükara SECourses•10/25/23, 11:34 PM

same

Furkan Gözükara SECourses•10/25/23, 11:34 PM

just dont use reg images

Furkan Gözükara SECourses•10/25/23, 11:34 PM

when no reg images used dreambooth = fine tuning

_Captain_OP•10/25/23, 11:35 PM

thanks a lot!! now it's little less confusing

_Captain_OP•10/25/23, 11:40 PM

oh, I have one more questions:
does it make a difference if i train image sets sequential (for instance 3000 steps, save checkpoint and use it as new base model to continue another 3000 steps) VS train all images in one run (eg 6000 steps)

__Captain_oh, I have one more questions: does it make a difference if i train image sets s...

Furkan Gözükara SECourses•10/25/23, 11:43 PM

in theory yes

Furkan Gözükara SECourses•10/25/23, 11:43 PM

but it depends on whether you are using dynamic LR

Furkan Gözükara SECourses•10/25/23, 11:43 PM

and if scripts are properly working

_Captain_OP•10/25/23, 11:44 PM

that makes sense

thank you!!!

_Captain_OP•10/26/23, 12:05 AM

https://github.com/kohya-ss/sd-scripts/issues/890
--train_text_encoder in sdxl_train.py does literally NOTHING

super exciting, curious how the training outcome will be after this is fixed.

GitHub

--train_text_encoder in sdxl_train.py does literally NOTHING · Issu...

Hi! I've been trying to perform Dreambooth training of the SDXL text encoders without affecting the unet at all. There doesn't seem to be an option in sdxl_train.py to specifically target o...

__Captain_https://github.com/kohya-ss/sd-scripts/issues/890 --train_text_encoder in sdxl_t...

Furkan Gözükara SECourses•10/26/23, 12:17 AM

yep i am waiting too

Furkan Gözükara SECourses•10/26/23, 12:17 AM

posted it on twitter

Furkan Gözükara SECourses•10/26/23, 12:17 AM

https://twitter.com/GozukaraFurkan/status/1717324781349974363

minxx•10/26/23, 3:34 AM

Hi all, i have been using runpod for A1111, since im using 4090 i want to fully maximize the gpu to generate image. can i know if i can divide the GPU process so that two people can generate the image at the same time without having to queue up with each other?

minxx•10/26/23, 3:36 AM

or perhaps just need to tweak the command line args in webui-user.sh?

Mminxx Hi all, i have been using runpod for A1111, since im using 4090 i want to fully...

Furkan Gözükara SECourses•10/26/23, 12:09 PM

start 2 instance of auto1111.

FFurkan Gözükara SECourses We have 16gb config

mrLLUU•10/26/23, 12:15 PM

had not found that yet, thanks

MmrLLUU had not found that yet, thanks 🙂

Furkan Gözükara SECourses•10/26/23, 12:22 PM

here https://www.patreon.com/posts/very-best-for-of-89213064

_Captain_OP•10/26/23, 2:17 PM

@Furkan Gözükara SECourses debugged the text encoder bug myself and reproduced the answer from arazela.
If you are curious too, here the debug output.

_Captain_OP•10/26/23, 2:18 PM

after implementing the fix from koyha (only local because it's not pushed yet):

_Captain_OP•10/26/23, 2:19 PM

the array is fixed

_Captain_OP•10/26/23, 2:20 PM

will test it to see if it makes a difference

__Captain_<@205854764540362752> debugged the text encoder bug myself and reproduced the an...

Furkan Gözükara SECourses•10/26/23, 2:24 PM

awesome

_Captain_OP•10/26/23, 3:49 PM

@Furkan Gözükara SECourses Wohooo! Big difference

can't share result because of private dataset, but man, looking forward to your next comparision

I think it will have more impact on things the text encoder doesn't already know.

_Captain_OP•10/26/23, 3:51 PM

used sdxl_train.py with .caption (full text descriptions) and .txt (with tags), no reg images

_Captain_OP•10/26/23, 3:53 PM

one line of code might improve sdxl a lot

crazy stuff.. always the same with coding xD

__Captain_<@205854764540362752> Wohooo! Big difference 😱 can't share result because of p...

Furkan Gözükara SECourses•10/26/23, 4:00 PM

awesome

GeorgeClintonGeorgeBush•10/26/23, 11:22 PM

What's the best way to fix weird eyes/pupils on your images?

GGeorgeClintonGeorgeBush What's the best way to fix weird eyes/pupils on your images?

Furkan Gözükara SECourses•10/26/23, 11:44 PM

inpainting

Furkan Gözükara SECourses•10/26/23, 11:44 PM

but first improve your training dataset

FFurkan Gözükara SECourses inpainting

GeorgeClintonGeorgeBush•10/26/23, 11:45 PM

Thanks. ADetailer right?

GGeorgeClintonGeorgeBush Thanks. ADetailer right?

Furkan Gözükara SECourses•10/26/23, 11:45 PM

yep

FFurkan Gözükara SECourses but first improve your training dataset

GeorgeClintonGeorgeBush•10/26/23, 11:47 PM

Do you know what kind of bad images causes weird eyes in your model? I don't know how to improve my dataset further

GGeorgeClintonGeorgeBush Do you know what kind of bad images causes weird eyes in your model? I don't kno...

Furkan Gözükara SECourses•10/26/23, 11:49 PM

image quality focus lightning distance

Furkan Gözükara SECourses•10/29/23, 9:37 PM

@helen i am checking your issue right now

Furkan Gözükara SECourses•10/29/23, 9:57 PM

@helen here images they are working. there must be some step you skip. did you make container size 20 gb before install?

Furkan Gözükara SECourses•10/29/23, 9:57 PM

Furkan Gözükara SECourses•10/29/23, 9:57 PM

Furkan Gözükara SECourses•10/29/23, 9:57 PM

i tested both templates

FFurkan Gözükara SECourses <@1161756280994676806> here images they are working. there must be some step you...

helen•10/29/23, 9:59 PM

OK, I'll try again and check container size

thank you

Hhelen OK, I'll try again and check container size😊 thank you

Furkan Gözükara SECourses•10/29/23, 10:01 PM

you are welcome.

Furkan Gözükara SECourses•10/29/23, 10:01 PM

also pick rtx 3090

David Ce•10/30/23, 2:22 AM

Hey @Furkan Gözükara SECourses
Please I've been having this issue and don't know how to fix it
Could you help?

David Ce•10/30/23, 2:28 AM

The error has changed after editing the webui_user.bat file to this

DDavid Ce Hey <@205854764540362752> Please I've been having this issue and don't know how...

Furkan Gözükara SECourses•10/30/23, 3:18 AM

just replied you in qwen-flux-wan-such

felixsanz•10/30/23, 7:51 PM

i can't make use of roop/insightface/whatever in runpod, does anyone know why? it's banned or something? CUDA execution provider doesn't appear

Ffelixsanz i can't make use of roop/insightface/whatever in runpod, does anyone know why? i...

Furkan Gözükara SECourses•10/30/23, 9:44 PM

try auto installer working i tested yesterday : https://www.patreon.com/posts/auto-installer-84511510

Patreon

Auto Installer Script For Roop (DeepFake) On RunPod & Example Image...

Get more from SECourses: Tutorials, Guides, Resources, Training, MidJourney, Voice Clone, TTS, ChatGPT, GPT, LLM, Scripts by Furkan Gözükara on Patreon

felixsanz•10/30/23, 9:47 PM

i'm not using roop. i created my own app using insightface

felixsanz•10/30/23, 9:47 PM

but mine uses more or less the same stuff

felixsanz•10/30/23, 9:48 PM

it's just that i can't get the cuda execution provider here, dunno why

Is there any difference between sdxl finetuning and kohya's dreambooth? Since both methods use sdxl_

Similar Threads