Software Engineering Courses (SECourses)•16mo ago

@Dr. Furkan Gözükara Have you tried training a LORA for an object? I'm finding when i go to generate

@Dr. Furkan Gözükara Have you tried training a LORA for an object? I'm finding when i go to generate the image using the LORA i have to describe the object in detail for it to retain most of the objects details

ZZuck @Dr. Furkan Gözükara Have you tried training a LORA for an object? I'm finding w...

Furkan Gözükara SECourses•9/16/24, 11:49 PM

i only tried a style yet

Furkan Gözükara SECourses•9/16/24, 11:49 PM

it worked very well

Furkan Gözükara SECourses•9/16/24, 11:49 PM

did you caption object during training?

ZuckOP•9/16/24, 11:51 PM

Ya i captioned really well, assigned trigger word, and described everything around the object i was training but didn't describe the objects details.

I wonder if i have to increase the LORA's strength when generating images...

ZuckOP•9/16/24, 11:52 PM

to be fair ive been using ai-toolkit, going to watch your kohya video, does kohya have a more up to date training method?

ZuckOP•9/16/24, 11:54 PM

I've done it succesfully with SD, but flux is a whole other beast

Furkan Gözükara SECourses•9/16/24, 11:55 PM

red color blue beard almost made it fine tuning amazing

ZZuck Ya i captioned really well, assigned trigger word, and described everything arou...

Furkan Gözükara SECourses•9/16/24, 11:55 PM

captioning your issue

Furkan Gözükara SECourses•9/16/24, 11:55 PM

dont caption

Dazzastrous [4090]•9/16/24, 11:56 PM

heheh

Furkan Gözükara SECourses•9/16/24, 11:56 PM

also my config is very best

Furkan Gözükara SECourses•9/16/24, 11:56 PM

i did 104 trainings to prepare it

ZuckOP•9/16/24, 11:56 PM

Okay gotchu!! Wont caption, and will try your config thank you!

Furkan Gözükara SECourses•9/16/24, 11:56 PM

orange hair white beard

Dazzastrous [4090]•9/16/24, 11:57 PM

what about mustache

Furkan Gözükara SECourses•9/16/24, 11:57 PM

i didnt set it :d

Furkan Gözükara SECourses•9/16/24, 11:57 PM

man deciding quality really hard between different timestep samplers

Dazzastrous [4090]•9/16/24, 11:58 PM

pink mohican

Furkan Gözükara SECourses•9/17/24, 12:32 AM

haha added mustache of french i guess

Furkan Gözükara SECourses•9/17/24, 12:32 AM

Dazzastrous [4090]•9/17/24, 12:32 AM

very nice

Dazzastrous [4090]•9/17/24, 12:32 AM

hipster look

Furkan Gözükara SECourses•9/17/24, 12:46 AM

FFurkan Gözükara SECourses Click to see attachment

druhl•9/17/24, 2:04 AM

Now, you've become Doc Phineas from Pawn Stars

druhl•9/17/24, 2:05 AM

Use that guy as inspiration for fine tuning testing lol.

Furkan Gözükara SECourses•9/17/24, 2:12 AM

starting single layer training experiments

Furkan Gözükara SECourses•9/17/24, 2:12 AM

lora

Furkan Gözükara SECourses•9/17/24, 2:21 AM

single block training speed

Furkan Gözükara SECourses•9/17/24, 2:21 AM

really fast

AstraBat•9/17/24, 2:34 AM

When you have Quality 1 and 2 in your Flux_Fine_Tune_v2.zip file, does '1' indicate the better quality?

FFurkan Gözükara SECourses Click to see attachment

thisisrahul.•9/17/24, 3:09 AM

Wow amazing, did you use your classical 15 image dataset or the new one with different expressions

Tthisisrahul.Wow amazing, did you use your classical 15 image dataset or the new one with dif...

Furkan Gözükara SECourses•9/17/24, 3:16 AM

classical 15 images

Furkan Gözükara SECourses•9/17/24, 3:16 AM

i do experiment on it

AAstraBat When you have Quality 1 and 2 in your Flux_Fine_Tune_v2.zip file, does '1' indic...

Furkan Gözükara SECourses•9/17/24, 3:17 AM

yes quality 1 better

AstraBat•9/17/24, 3:18 AM

I'm about to run dreambooth for flux - I presume that we still are using regularization images, as for SDXL?

Furkan Gözükara SECourses•9/17/24, 3:18 AM

doing 8x single layer training

AAstraBat I'm about to run dreambooth for flux - I presume that we still are using regular...

Furkan Gözükara SECourses•9/17/24, 3:18 AM

no dont use reg images

Furkan Gözükara SECourses•9/17/24, 3:18 AM

it doesnt work

AstraBat•9/17/24, 3:18 AM

I'll leave the reg folder location blank in kohya.

AstraBat•9/17/24, 3:19 AM

I'm using your flux_fine_tune_v2 config.

AAstraBat I'm using your flux_fine_tune_v2 config.

Furkan Gözükara SECourses•9/17/24, 3:20 AM

yep good one

Furkan Gözükara SECourses•9/17/24, 3:20 AM

yes leave blank

ZZuck Okay gotchu!! Wont caption, and will try your config thank you! 🙂

Neo•9/17/24, 3:53 AM

let us know how it goes

jvmbo•9/17/24, 3:57 AM

I got this warning while doing install Step 2 for Kohya. The bat files you've provided automatically install xformers correct? I just ignore this warning?

Luiz xd•9/17/24, 4:03 AM

@Dr. Furkan Gözükara do you know anything about Fluxgym, have you tried it or is it in your plans? they say that it works with kohya sd-scripts... the most powerful training script for Stable Diffusion, Flux, etc.

I just saw that information that's why my query

Darth Teethius•9/17/24, 4:56 AM

has anyone tried reg images with flux yet?

FFurkan Gözükara SECourses starting single layer training experiments

druhl•9/17/24, 5:43 AM

I thought Kohya did not have that function, is it released now?

DDarth Teethius has anyone tried reg images with flux yet?

druhl•9/17/24, 5:43 AM

Does not work. Doc has tested and posted his experiments.

LLuiz xd @Dr. Furkan Gözükara do you know anything about Fluxgym, have you tried it or i...

druhl•9/17/24, 5:44 AM

Is this the ComfyUI one or the standalone website one?

@Dr. Furkan Gözükara Have you tried training a LORA for an object? I'm finding when i go to generate

Similar Threads