Software Engineering Courses (SECourses)•6mo ago

I am using your Kaggle Kohya ss file to start a kohya ss gradio ui, but it fails by saying no conda

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

SommmOP•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

SommmOP•7/23/25, 9:31 PM

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

SommmOP•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_10803•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_10803•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_10803•7/23/25, 10:45 PM

Oh, I see

minaricrew_10803•7/23/25, 10:45 PM

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

minaricrew_10803•7/23/25, 10:45 PM

Is this same speed?

minaricrew_10803•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_10803•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_10803•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_10803•7/24/25, 12:56 AM

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

minaricrew_10803•7/24/25, 2:02 AM

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

minaricrew_10803•7/24/25, 2:03 AM

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

Mminaricrew_10803 <@205854764540362752> What if we train with fp8? Have you tried that before?

Furkan Gözükara SECourses•7/24/25, 2:32 AM

not possible

Furkan Gözükara SECourses•7/24/25, 2:32 AM

even if you use fp8 weights the training happens at 16 bit precision

Furkan Gözükara SECourses•7/24/25, 2:32 AM

kohya doesnt support fp8 training

minaricrew_10803•7/24/25, 2:40 AM

Thank you so much for answering all the questions! I'll focus on facial accuracy with Kohya fisrt, and then think about speed later haha

FFurkan Gözükara SECourses but if classes are matching it can bleed

zadkhamaya•7/24/25, 7:17 AM

I see, thank you.
Where we can see the class type for this flux model?

Zzadkhamaya I see, thank you. Where we can see the class type for this flux model?

Furkan Gözükara SECourses•7/24/25, 9:48 AM

there is no class types

Furkan Gözükara SECourses•7/24/25, 9:48 AM

you need to understand yourself sadly

H A R B D U L•7/24/25, 6:42 PM

Yoo devs

underdog•7/24/25, 9:00 PM

@Furkan Gözükara SECourses Hello

I was wondering if you have any plans in the future to train a custom lora for wan 2.1/2.2 (when it releases) video generation?

Uunderdog <@205854764540362752> Hello 🙂 I was wondering if you have any plans in the futu...

Furkan Gözükara SECourses•7/24/25, 9:11 PM

yes

Furkan Gözükara SECourses•7/24/25, 9:12 PM

i saw today wan 2.2 coming haha

underdog•7/24/25, 9:12 PM

Cant wait for your video

underdog•7/24/25, 9:13 PM

When should I be looking out for it?

Uunderdog When should I be looking out for it? 😄

Furkan Gözükara SECourses•7/24/25, 9:20 PM

after this video upscaler i plan to research them :d

Furkan Gözükara SECourses•7/24/25, 9:20 PM

i need to fisnih this

underdog•7/24/25, 9:24 PM

Can't wait

underdog•7/27/25, 12:58 PM

Hey, I have been finding resources and documents for text to video for wan with images being dataset or image to video with videos being dataset, but I was wondering if there is a way to train lora with images and use first and last image as reference and generate a video based on the reference images? do you happen to know any models which lets you train on images and then use that lora with reference images?

Uunderdog Hey, I have been finding resources and documents for text to video for wan with ...

Furkan Gözükara SECourses•7/27/25, 5:10 PM

i dont know such model

Furkan Gözükara SECourses•7/27/25, 5:10 PM

but wan 2.1 can be trained from images to generate videos

007•7/27/25, 5:52 PM

Hey guys, so I'm really impressed with the fine tune accuracy with regards to my model's face. it looks great. but it still has a bit of AI look to it. and I know the model is trained on photorealism.

my goal is it make it look more candid, like it was shot on an iPhone or a digital cam. I've used loras with varying strengths and it completely changes the faces on the model. I've tried different resolutions but it doesnt help much.

I am using your Kaggle Kohya ss file to start a kohya ss gradio ui, but it fails by saying no conda

Similar Threads

Similar Threads

Similar Threads