Software Engineering Courses (SECourses)•17mo ago

Is there a <@205854764540362752> in the house? I couldn't find any material (post or video) about t

Is there a @Furkan Gözükara SECourses in the house? I couldn't find any material (post or video) about the importance of captioning and regularization images for SDXL Lora/Dreambooth training. Can you point me in the right direction?

macbaOP•9/1/24, 9:43 AM

I got very good results with Flux (where you mentioned reg images are not useful, and a single trigger word is good for single persons), but my SDXL results where horrible…

EECD <@205854764540362752> Hello Doctor, I am having problems with the presets tier 1...

Furkan Gözükara SECourses•9/1/24, 8:44 PM

you probably loaded into dreambooth tab

Furkan Gözükara SECourses•9/1/24, 8:44 PM

and corrupted config

Furkan Gözükara SECourses•9/1/24, 8:45 PM

also hopefully i will update all configs today

Mmacba Is there a <@205854764540362752> in the house? I couldn't find any material (po...

Furkan Gözükara SECourses•9/1/24, 8:45 PM

it reduces likeliness

Furkan Gözükara SECourses•9/1/24, 8:45 PM

i tested it for person

Furkan Gözükara SECourses•9/1/24, 8:45 PM

https://medium.com/@furkangozukara/compared-effect-of-image-captioning-for-sdxl-fine-tuning-dreambooth-training-for-a-single-person-961087e42334

Medium

Compared Effect Of Image Captioning For SDXL Fine-tuning / DreamBoo...

Check all the images below to see used training dataset and used captionings.

Ddaroth tier 1 is too much for any consumer GPU I think, tier 3 is the best that fits in...

Furkan Gözükara SECourses•9/1/24, 8:45 PM

ye tier 1 is for only 40 GB and above

ECD•9/1/24, 8:46 PM

oooh, for 40 GB of VRAM?

EECD oooh, for 40 GB of VRAM?

Furkan Gözükara SECourses•9/1/24, 8:47 PM

yep

Furkan Gözükara SECourses•9/1/24, 8:47 PM

go tier 3

Furkan Gözükara SECourses•9/1/24, 8:47 PM

rank 3

Furkan Gözükara SECourses•9/1/24, 8:47 PM

you do flux right?

ECD•9/1/24, 8:48 PM

that's explain a lot... hehe, okey, thanks! no, Trying to train SDXL

ECD•9/1/24, 8:48 PM

the configs are only for flux?

EECD that's explain a lot... hehe, okey, thanks! no, Trying to train SDXL

Furkan Gözükara SECourses•9/1/24, 8:49 PM

i have different configs for each

Furkan Gözükara SECourses•9/1/24, 8:49 PM

for sdxl use onetrainer it is bestr

Furkan Gözükara SECourses•9/1/24, 8:50 PM

you using kohya or onetrainer?

Furkan Gözükara SECourses•9/1/24, 8:50 PM

i will link you accurate file

ECD•9/1/24, 8:50 PM

onetrainer

Furkan Gözükara SECourses•9/1/24, 8:50 PM

ECD•9/1/24, 8:50 PM

gotcha, thanks

Furkan Gözükara SECourses•9/1/24, 8:50 PM

https://www.patreon.com/posts/96028218

Patreon

OneTrainer Stable Diffusion XL (SDXL) Fine Tuning Best Presets | SE...

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

Furkan Gözükara SECourses•9/1/24, 8:50 PM

use tier1_15.4GB_fast_v4.json

Furkan Gözükara SECourses•9/1/24, 8:50 PM

this is for sdxl

Furkan Gözükara SECourses•9/1/24, 8:51 PM

use reg imags hugely improves for training a person

FFurkan Gözükara SECourses use reg imags hugely improves for training a person

macbaOP•9/1/24, 9:08 PM

There it is, that was my second question. So you actually do prefer OneTrainer, at least for SD* models…

Mmacba There it is, that was my second question. So you actually do prefer OneTrainer, ...

Furkan Gözükara SECourses•9/1/24, 9:23 PM

yep

Furkan Gözükara SECourses•9/1/24, 9:23 PM

sd trainer really does handle vram better

Furkan Gözükara SECourses•9/1/24, 9:23 PM

also it is all in one

Furkan Gözükara SECourses•9/1/24, 9:23 PM

with kohya we have gui and scripts

Fino•9/2/24, 7:22 AM

Hello together,
have anyone trained multiple concepts in one or more dreambooth training(s)?
I want to train 2 different characters with unique tokens into my model and face the issue that the one concept is always bleeding into the other.
I tried to remove the class token and made several tests with a combined training set and single trainings. Unfortunately no success. Each character has a versatile dataset of 35 high quality images.
Any idea to reduce the bleeding?

FFino Hello together, have anyone trained multiple concepts in one or more dreambooth ...

Furkan Gözükara SECourses•9/2/24, 10:49 AM

i just replied on youtube but it may not be possible since FLUX has internal text encoding. so it knows both are man. however you can try this. only ohwx and bbuk as rare token and use clip text encoder training that i added yesterday

Furkan Gözükara SECourses•9/2/24, 10:49 AM

still may bleed. fine tuning may solve though not sure

Fino•9/2/24, 10:53 AM

Sorry for the missing information. I talk about a SDXL Model

FFino Sorry for the missing information. I talk about a SDXL Model 😅

Furkan Gözükara SECourses•9/2/24, 11:01 AM

yes people said they do very good. are you doing lora or fine tuning?

Fino•9/2/24, 11:03 AM

I fine tune the model with character images.
Lora training works great with your guides.

FFino I fine tune the model with character images. Lora training works great with you...

Furkan Gözükara SECourses•9/2/24, 11:06 AM

are you using onetrainer?

Fino•9/2/24, 11:06 AM

I used OneTrainer (with your config) and dreambooth from kohya_ss also with your configuration

FFino I used OneTrainer (with your config) and dreambooth from kohya_ss also with your...

Furkan Gözükara SECourses•9/2/24, 11:07 AM

i see you tested on both

Furkan Gözükara SECourses•9/2/24, 11:07 AM

and still bleed?

Fino•9/2/24, 11:10 AM

Yes, onetrainer was much worse than kohya for some reason.
I made also dataset tagging with following schemas:
-> <unique identifier>
-> <unique identifier > <class token>
-> <unique identifier>, <describing image without character attitudes >

Fino•9/2/24, 11:10 AM

To test if the tagging is influencing common model tokens like men or women

FFino Yes, onetrainer was much worse than kohya for some reason. I made also dataset t...

Furkan Gözükara SECourses•9/2/24, 11:20 AM

are you using reg images?

Furkan Gözükara SECourses•9/2/24, 11:20 AM

reg makes huge impact on onetrainer

Furkan Gözükara SECourses•9/2/24, 11:20 AM

and which base model you are training sdxl base?

Fino•9/2/24, 11:22 AM

It’s a customized pony based model for a client.
I didn’t use reg images.

Fino•9/2/24, 11:22 AM