Software Engineering Courses (SECourses)•3y ago

Hey guys, I've been trying to train lora SDXL but I seem to always run out of CUDA in runpod, despit

Hey guys, I've been trying to train lora SDXL but I seem to always run out of CUDA in runpod, despite not running anything.

My parameters are just like in SEcourse videos, i'm running in 3090 runpod with 24gb vram

ockhamrazorOP•12/14/23, 10:01 AM

ockhamrazorOP•12/14/23, 10:01 AM

I also got Untyped Storage error like this

mikemenders•12/14/23, 10:02 AM

After a couple of weeks I find that d_coef 0.75 is enough for SD 1.5 Lora prodigy training, but if it is strong, 0.6 helps. You don't need to make it stronger to train persons. This is true for 5 reps for 20 pictures, 4 reps for 25 pictures (20x5=25x4=100).

Oockhamrazor Hey guys, I've been trying to train lora SDXL but I seem to always run out of CU...

Bophedes•12/14/23, 1:18 PM

hey, where you get that cuda? i get an error saying i need it

BBophedes hey, where you get that cuda? i get an error saying i need it

ockhamrazorOP•12/14/23, 1:33 PM

what do you mean? i mean if you use nvidia gpu it already has cuda memory no?

Oockhamrazor what do you mean? i mean if you use nvidia gpu it already has cuda memory no?

Bophedes•12/14/23, 1:34 PM

yea nah i have an AMD Radeon RX 6600XT. if i had an nvidia id probably have less problems

Rrevoerom Hi I'm trying to train an SD1.5 model with my face on Kohya SS Dreambooth (check...

Pelolisu•12/14/23, 2:41 PM

I tried training on 1.5 in Kohya but it has been performing very slowly. If I discover the settings I will let you know.

Rrevoerom Hi I'm trying to train an SD1.5 model with my face on Kohya SS Dreambooth (check...

Benjamin•12/14/23, 3:00 PM

look up the messages from mikemenders in this channel here, he researched a lot and posted a lot of tipps and configurations for sd 1.5

JMPA•12/14/23, 3:23 PM

Hi guys, If I want to create a fictitious person (that does not exist as a base) how should I do? I want to create a model girl

JMPA•12/14/23, 3:24 PM

Because I learned about creating with a base person, but now I want to create from a fictitious person

Deleted User•12/14/23, 3:25 PM

Hello everyone, I have a quick question. I'm currently training a stable diffusion model using Kaggle, and I'm wondering about the required resolution for the training images. Do they need to be 1024x1024 pixels? I'm asking because my images are currently 512x512 pixels. I'm looking for clarification on this point,

as suggested by the information in this video: Stable Diffusion Training Guide on YouTube.

YouTubeSECourses

How To Do Stable Diffusion XL (SDXL) DreamBooth Training For Free -...

Master Stable Diffusion XL Training on Kaggle for Free!

Welcome to this comprehensive tutorial where I'll be guiding you through the exciting world of setting up and training Stable Diffusion XL (SDXL) with Kohya on a free Kaggle account. This video is your one-stop resource for learning everything from initiating a Kaggle session with dual ...

DDeleted User Hello everyone, I have a quick question. I'm currently training a stable diffusi...

Dr.Tekiti•12/14/23, 3:26 PM

if it a sdxl model, I think you need 1024x1024 images for training

JJMPA Hi guys, If I want to create a fictitious person (that does not exist as a base...

Dr.Tekiti•12/14/23, 3:29 PM

who do you mean by create ? you want to make one picture, a video, a instragram ? If it just one picture, you have multiple way, if you want to make a lot of pic, you probably need to create a lora for you fictional charater

DDr.Tekiti if it a sdxl model, I think you need 1024x1024 images for training

Deleted User•12/14/23, 3:29 PM

Thanks for the clarification.

DDr.Tekiti who do you mean by create ? you want to make one picture, a video, a instragram ...

JMPA•12/14/23, 3:30 PM

Multiple images, I want to create a real character, a woman.

JMPA•12/14/23, 3:30 PM

Normally, I put 40 real photos of the person as a base, in this case, I have not them because I want to create it, do you understand what I mean?

JJMPA Normally, I put 40 real photos of the person as a base, in this case, I have not...

Dr.Tekiti•12/14/23, 3:31 PM

yeah, you want to know how to create food for you future lora

DDr.Tekiti yeah, you want to know how to create food for you future lora

JMPA•12/14/23, 3:31 PM

Correct, what should I do?

Dr.Tekiti•12/14/23, 3:32 PM

I think you need first to define the look of you're character and you try to make one good face and one good body

Dr.Tekiti•12/14/23, 3:33 PM

then you use IPadator and reactor to put the face on other situation

Dr.Tekiti•12/14/23, 3:35 PM

and then you can create you're food, but you don't really need it, it just create a more simpler workflow (1 lora vs ipa + reactor)

DDr.Tekiti then you use IPadator and reactor to put the face on other situation

JMPA•12/14/23, 3:35 PM

But can I create the face from scratch with that or how? I never used it

JJMPA But can I create the face from scratch with that or how? I never used it

Dr.Tekiti•12/14/23, 3:39 PM

you can but I think to create you're first face, a good promting with a Sdxl model would be better, execpt if you want some facials features (not sure if it's the right term ^^') like a specific nose you can't get with prompting, then you can use IPa

Dr.Tekiti•12/14/23, 3:40 PM

all you really need is one good first face

JMPA•12/14/23, 3:40 PM

Ok perfect, then after generated, how many photos should I do to start generating the final ones? 100 200?

Dr.Tekiti•12/14/23, 3:44 PM

you use IPa and reactor to deepfake that face in many picture as possible if you want a lora

Dr.Tekiti•12/14/23, 3:45 PM

since you've got the face, you can juste deepfake it in many pic and said job done

JJMPA Ok perfect, then after generated, how many photos should I do to start generatin...

Dr.Tekiti•12/14/23, 3:48 PM

https://www.youtube.com/watch?v=7m9ZZFU3HWo

Dr.Tekiti•12/14/23, 3:48 PM

https://www.youtube.com/watch?v=shc83TaQmqA&t=338s&pp=ygUZaXBhZGFwdGVyIGxjcmVhdG9yIHVzZSBpdA%3D%3D

JMPA•12/14/23, 3:49 PM

Ok perfect brother thank you

JJMPA Ok perfect, then after generated, how many photos should I do to start generatin...

Dr.Tekiti•12/14/23, 3:49 PM

that should help you with IPa, reactor is really for the deepface part

JJMPA Ok perfect brother thank you

Dr.Tekiti•12/14/23, 3:49 PM

my pleasure, don't forget to ping to show the result ;p

CCharkel What is your favorite SDXL model? 🙂

mtnmecca•12/14/23, 7:34 PM

realvisXL 2.0 and juggernautXL_v7 - both excellent at human figure and details, no experience with generating other type of art.

Ssexyboi What are your opinions on identical training parameters but different GPUs and h...

Furkan Gözükara SECourses•12/14/23, 7:45 PM

only fp16 vs bf16 differs

Furkan Gözükara SECourses•12/14/23, 7:45 PM

between different gpus

SSarcasticest I don't think anything else is using my VRAM, unless it's starting simultaneousl...

Furkan Gözükara SECourses•12/14/23, 7:46 PM

yes 1 gb is good

SSarcasticest Click to see attachment

Furkan Gözükara SECourses•12/14/23, 7:46 PM

you need to start as half precision

Oockhamrazor Hey guys, I've been trying to train lora SDXL but I seem to always run out of CU...

Furkan Gözükara SECourses•12/14/23, 7:46 PM

hello welcome

Furkan Gözükara SECourses•12/14/23, 7:47 PM

it must be due to you are failing to kill automatically started SD web ui instance

DDeleted User Hello everyone, I have a quick question. I'm currently training a stable diffusi...

Furkan Gözükara SECourses•12/14/23, 7:48 PM

yes 1024x1024 for sdxl

Furkan Gözükara SECourses•12/14/23, 7:48 PM

if you do sd 1.5 training

Furkan Gözükara SECourses•12/14/23, 7:48 PM

they need to be 512x512 depending base model you use

Dobby•12/14/23, 8:03 PM

@Dr. Furkan Gözükara any idea why my latents are caching so slow on koyahss dreambooth? i have 4090

Dobby•12/14/23, 8:09 PM

im gonna restart env just in case

DDobby @Dr. Furkan Gözükara any idea why my latents are caching so slow on koyahss drea...

Furkan Gözükara SECourses•12/14/23, 8:10 PM

super slow

Furkan Gözükara SECourses•12/14/23, 8:10 PM

i think it is a bug of nvidia driver

Dobby•12/14/23, 8:11 PM

im going to resintall whole koya env lets see

FFurkan Gözükara SECourses you need to start as half precision

Sarcasticest•12/14/23, 10:01 PM

Wouldn't whatever is happening effect every process I try and run? If you can run

Clip_Interrogator.py

Clip_Interrogator.py

and use Clip_Interrogator Gradio Web UI with ViT-bigG-14/laion2b_s39b_b160k and blip2-2.7b with short generation time, and I have almost the same GPU, shouldn't I be able to do the same?

hozaggers•12/14/23, 10:17 PM

does anyone have any tips for further training on kaggle after uploading to hugging face? I have a model I am training but I will run out of qouta before it finishes, the 83% checkpout should save before my qouta runs out though. If I set the initial training for 100 steps, and save at 83% how do I re-upload it to train it the remaining 17% for the next checkpoint?

SSarcasticest Wouldn't whatever is happening effect every process I try and run? If you can ru...

Furkan Gözükara SECourses•12/14/23, 10:18 PM

you know full precision uses almost double of half

Hey guys, I've been trying to train lora SDXL but I seem to always run out of CUDA in runpod, despit

Similar Threads