Software Engineering Courses (SECourses)•6mo ago

<@205854764540362752> When I change config to "cache_latents = false / cache_latents_to_disk = false

@Furkan Gözükara SECourses When I change config to "cache_latents = false / cache_latents_to_disk = false", GPU usage dropped to 36GB. Is there any quality drop if I change it like this?

Mminaricrew_10803 <@205854764540362752> When I change config to "cache_latents = false / cache_lat...

Furkan Gözükara SECourses•7/23/25, 12:53 AM

nope

zadkhamaya•7/23/25, 2:12 AM

Hi @Furkan Gözükara SECourses
Thank you for the dreambooth training tutorial, it really very clear and I successfully manage the training for a person and it performs really well

I have question for other case, if I want to train a person with specific clothing style and also in a specific illustration style, what will be the best approach? Shall I train first the illustration style, then the person with his specific clothing style? The aim is I want to generate this person in consistent clothing style and in consistent illustration style.

Thank you

Zzadkhamaya Hi <@205854764540362752> Thank you for the dreambooth training tutorial, it rea...

Furkan Gözükara SECourses•7/23/25, 2:22 AM

for illustration style use kontext model

Furkan Gözükara SECourses•7/23/25, 2:22 AM

it really does good job

Furkan Gözükara SECourses•7/23/25, 2:23 AM

for clothing if you have pictures of person with that clothing

Furkan Gözükara SECourses•7/23/25, 2:23 AM

that would be great to train

Furkan Gözükara SECourses•7/23/25, 2:23 AM

otherwise you can try to train with cloth pictures

Furkan Gözükara SECourses•7/23/25, 2:23 AM

or traing a clothing model and inpaint images of person

zadkhamaya•7/23/25, 2:52 AM

Ah I see. Previously I already prepare the character with correct clothing and the desired illustration style. But I have only this one image.

I plan to create other pose and other situation using Flux Kontext. Will it perform better if I do the train in one go, it has full body clothing style and also illustration style?

minaricrew_10803OP•7/23/25, 10:07 AM

Is there best config json file for batch 7? I can only find best config for batch 1 in the folder(Kohya_FLUX_DreamBooth_LoRA_v30)

Zzadkhamaya Ah I see. Previously I already prepare the character with correct clothing and t...

Furkan Gözükara SECourses•7/23/25, 4:18 PM

1 in go is best

Furkan Gözükara SECourses•7/23/25, 4:18 PM

but if classes are matching it can bleed

Mminaricrew_10803 Is there best config json file for batch 7? I can only find best config for batc...

Furkan Gözükara SECourses•7/23/25, 4:18 PM

for lora i recommend multiple gpu

Furkan Gözükara SECourses•7/23/25, 4:18 PM

to speed up

Furkan Gözükara SECourses•7/23/25, 4:18 PM

for lora i dont have batch config but you can use batch size 1

Furkan Gözükara SECourses•7/23/25, 4:18 PM

and set batch size 7

Furkan Gözükara SECourses•7/23/25, 4:18 PM

and multiply learnign rate with

Furkan Gözükara SECourses•7/23/25, 4:19 PM

square root 7 * base learning rate

Furkan Gözükara SECourses•7/23/25, 4:19 PM

however batch size 1 has the best quality

Sommm•7/23/25, 9:09 PM

I am using your Kaggle Kohya ss file to start a kohya ss gradio ui, but it fails by saying no conda environment found, and if i manually start i and try to load a json file forf the configuration, it says something like 'str' has not attribute 'get'? please help

Sommm•7/23/25, 9:15 PM

SSommm I am using your Kaggle Kohya ss file to start a kohya ss gradio ui, but it fails...

Furkan Gözükara SECourses•7/23/25, 9:26 PM

which versioın are you using?

Furkan Gözükara SECourses•7/23/25, 9:26 PM

file name

Sommm•7/23/25, 9:31 PM

i guess its v25, i just ran it using the py file let me try

Sommm•7/23/25, 9:33 PM

its 2 am now, i will let you know later what happens

SSommm i guess its v25, i just ran it using the py file let me try

Furkan Gözükara SECourses•7/23/25, 9:44 PM

latest is v28

Furkan Gözükara SECourses•7/23/25, 9:44 PM

use it

Furkan Gözükara SECourses•7/23/25, 9:44 PM

https://www.patreon.com/posts/88397937

Patreon

A Kaggle Notebook to Train Stable Diffusion 1.5 and XL (SDXL) on a ...

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

FFurkan Gözükara SECourses for lora i recommend multiple gpu

minaricrew_10803OP•7/23/25, 10:42 PM

Is it even possible to set batch size 7 with VRAM 48GB? Because even with batch size 1, VRAM usage is around 36GB

minaricrew_10803OP•7/23/25, 10:43 PM

FYI, I want best quality, but I should also consider training speed

Mminaricrew_10803 Is it even possible to set batch size 7 with VRAM 48GB? Because even with batch ...

Furkan Gözükara SECourses•7/23/25, 10:43 PM

i think not with lora

Furkan Gözükara SECourses•7/23/25, 10:43 PM

it is for dreambooth at 1024x1024

minaricrew_10803OP•7/23/25, 10:45 PM

Oh, I see

minaricrew_10803OP•7/23/25, 10:45 PM

For dreambooth, batch size 1 & 7 gpu vs batch size 7 & 1 gpu

minaricrew_10803OP•7/23/25, 10:45 PM

Is this same speed?

minaricrew_10803OP•7/23/25, 10:46 PM

I assume quality would be better with batch size 1

Mminaricrew_10803 For dreambooth, **batch size 1 & 7 gpu** vs **batch size 7 & 1 gpu**

Furkan Gözükara SECourses•7/23/25, 11:13 PM

of course not

Furkan Gözükara SECourses•7/23/25, 11:14 PM

normally 7 gpu means it scales

Furkan Gözükara SECourses•7/23/25, 11:14 PM

with lora almost 7x speed

Furkan Gözükara SECourses•7/23/25, 11:14 PM

but for dreambooth you need 80 gb gpus and i didnt test

Furkan Gözükara SECourses•7/23/25, 11:14 PM

batch size 7 is like 1.3x speed

minaricrew_10803OP•7/23/25, 11:27 PM

Then, best way to speed up the training is

lora
batch size 1
multiple gpus(for linear speed up)

minaricrew_10803OP•7/23/25, 11:30 PM

Additional question, is there any way that I can improve training speed itself?

I used best Tier 1 config for 48GB vram for lora training
- But, set "cache_latents = false / cache_latents_to_disk = false"
With RTX 6000 Ada, it took 24 mins for 20 epoch
With A100 SXM, it took 16 mins for 20 epoch

Mminaricrew_10803 Then, best way to speed up the training is - lora - batch size 1 - multiple gpus...

Furkan Gözükara SECourses•7/23/25, 11:48 PM

yes

Mminaricrew_10803 **Additional question, is there any way that I can improve training speed itself...

Furkan Gözükara SECourses•7/23/25, 11:48 PM

sadly not much else

Furkan Gözükara SECourses•7/23/25, 11:48 PM

graident checkpointing slows down

Furkan Gözükara SECourses•7/23/25, 11:48 PM

but turning it of requires huge vram

minaricrew_10803OP•7/24/25, 12:56 AM

hmmm... in that case, I need to think about how to optimize gpu cost

minaricrew_10803OP•7/24/25, 2:02 AM

@Furkan Gözükara SECourses What if we train with fp8? Have you tried that before?

minaricrew_10803OP•7/24/25, 2:03 AM

I'm curious how much quality difference between fp8 and fp16

<@205854764540362752> When I change config to "cache_latents = false / cache_latents_to_disk = false

Similar Threads