Software Engineering Courses (SECourses)•3y ago

just curious if anyone prefers training loras on comfyUI (https://github.com/johnman3032/simple-lora

just curious if anyone prefers training loras on comfyUI (https://github.com/johnman3032/simple-lora-dreambooth-trainer/) or is kohya still the better choice / whats the difference?

GitHub

GitHub - johnman3032/simple-lora-dreambooth-trainer: A collection o...

A collection of simple training GUIs for SD1.5 and SDXL. - GitHub - johnman3032/simple-lora-dreambooth-trainer: A collection of simple training GUIs for SD1.5 and SDXL.

Max•12/26/23, 9:58 AM

Hi guys good morning! I was just trying to train a girl on kaggle (with koyha ss), and was wondering how many steps were needed for 30 photos. What is the perfect number for you?

Max•12/26/23, 9:58 AM

PS: keep in mind that on Koyha there are 2 GPUs

DDigitalArt_Lover Actually, my concern is, that my GPU and CPU is more capable of what it's curren...

Furkan Gözükara SECourses•12/26/23, 10:10 AM

ye the trainer of sdxl or sd is not fully optimized

DDigitalArt_Lover Additional info, this is how much time it is showing after i run the training, j...

Furkan Gözükara SECourses•12/26/23, 10:10 AM

you have shared vram problem

Furkan Gözükara SECourses•12/26/23, 10:10 AM

it is using shared vram thus many times slower like 10x 20x more

RRen Does anyone used this one with sdxl? https://github.com/huggingface/diffusers/bl...

Furkan Gözükara SECourses•12/26/23, 10:10 AM

i know but didnt try yet

MMrJoozy just curious if anyone prefers training loras on comfyUI (https://github.com/joh...

Furkan Gözükara SECourses•12/26/23, 10:11 AM

i would use kohya. even kohya fixing bugs at today

Furkan Gözükara SECourses•12/26/23, 10:11 AM

it takes huge time for such program to mature

MMax Hi guys good morning! I was just trying to train a girl on kaggle (with koyha ss...

Furkan Gözükara SECourses•12/26/23, 10:12 AM

i usually go with 4000 5000 steps and take checkpoint

Furkan Gözükara SECourses•12/26/23, 10:12 AM

when reg images are also used

Furkan Gözükara SECourses•12/26/23, 10:12 AM

so you should 2000-2500

Furkan Gözükara SECourses•12/26/23, 10:12 AM

since batch size will be 2x

Max•12/26/23, 10:20 AM

Ok, done. Thank you!

Deleted User•12/26/23, 10:36 AM

30 images with 100 repeats and regularisation images repeat 1. Max steps 6000 then how much save_every_n_steps="" is good for best quality checkpoints

FFurkan Gözükara SECourses it is using shared vram thus many times slower like 10x 20x more

DigitalArt_Lover•12/26/23, 10:39 AM

Is there any solution to this!! i got it, that it's running slower.

DDeleted User 30 images with 100 repeats and regularisation images repeat 1. Max steps 6000 th...

Furkan Gözükara SECourses•12/26/23, 11:00 AM

try to get checkpoints every 15 20 epochs

DDigitalArt_Lover Is there any solution to this!! i got it, that it's running slower.

Furkan Gözükara SECourses•12/26/23, 11:00 AM

try to reduce VRAM usage before starting training

FFurkan Gözükara SECourses try to get checkpoints every 15 20 epochs

Deleted User•12/26/23, 11:01 AM

You mean 1500 -2000 steps

DDeleted User You mean 1500 -2000 steps

Furkan Gözükara SECourses•12/26/23, 1:00 PM

it depends

Furkan Gözükara SECourses•12/26/23, 1:00 PM

how many train images

Furkan Gözükara SECourses•12/26/23, 1:00 PM

and reg used or not

Ren•12/26/23, 3:56 PM

@Dr. Furkan Gözükara I wanna create sdxl Lora based on 15 product images.
What is the best training options?

Ren•12/26/23, 3:57 PM

I am using learning_rate=1e-4 and max_train_steps=500
After I train the model, it makes some bad images

Ren•12/26/23, 3:58 PM

it often mixing other colors even I only trained black ones

RRen @Dr. Furkan Gözükara I wanna create sdxl Lora based on 15 product images. What i...

Furkan Gözükara SECourses•12/26/23, 4:16 PM

what is your gpu?

Furkan Gözükara SECourses•12/26/23, 4:16 PM

SDXL dreambooth much better than lora

Ren•12/26/23, 4:23 PM

I am using A100

Max•12/26/23, 4:42 PM

Good afternoon guys. I was trying to ally a potential AI influencer. The problem is the photos to be taken as an example. If I take them from Instagram there is the problem of privacy/copyright but above all of quality. Solutions?

Max•12/26/23, 5:30 PM

Guys, why from this: https://civitai.com/images/2411073, on my stable diffusion became this : ? Did I do something wrong?

MMax Guys, why from this: https://civitai.com/images/2411073, on my stable diffusion ...

shortdwarf•12/26/23, 7:28 PM

did you use the same seed? same prompt+different seed=different picture

Max•12/26/23, 7:30 PM

I copied the generation data and pasted on prompt clicking the blue arrow to import the data on sd

AIGambino•12/26/23, 7:36 PM

AIGambino•12/26/23, 7:37 PM

gives me this error for some reason I dont know why

RRen I am using A100

Furkan Gözükara SECourses•12/26/23, 7:37 PM

you can use our dreambooth config

Furkan Gözükara SECourses•12/26/23, 7:37 PM

however for product training the dataset matters a lot

AAIGambino gives me this error for some reason I dont know why

Furkan Gözükara SECourses•12/26/23, 7:37 PM

i would say reinstall

Furkan Gözükara SECourses•12/26/23, 7:38 PM

do a fresh one

Furkan Gözükara SECourses•12/26/23, 7:38 PM

library missing

Furkan Gözükara SECourses•12/26/23, 7:38 PM

or you need to activate venv and install missing library manually

AIGambino•12/26/23, 7:47 PM

tried manually installing bitsandbytes no luck

FFurkan Gözükara SECourses try to reduce VRAM usage before starting training

DigitalArt_Lover•12/26/23, 8:48 PM

I didn't run anything else, while starting the training.

I've a core i9 cpu and a 4080 (16gb vram gpu).

I think in terms of my power, this 3000 steps should take a lot less time to train??

Again, I've used regularization images and also using a concept folder as well.

The settings are below:

Train Batch Size: 1
Epoch: 1
repeats for training images: 20.
repeats for regularization images: 2.
And for the concept folder: Training steps per concept per epoch is 20.
Mixed precision: fp16
Save precision: fp16

DDigitalArt_Lover I didn't run anything else, while starting the training. I've a core i9 cpu an...

Furkan Gözükara SECourses•12/26/23, 9:09 PM

it is using shared vram

Furkan Gözükara SECourses•12/26/23, 9:09 PM

you need to use our 16gb config

Furkan Gözükara SECourses•12/26/23, 9:09 PM

enable xformers and gradient checkpointing

FFurkan Gözükara SECourses i would say reinstall

AIGambino•12/26/23, 9:35 PM

Reinstall stable diffusion? I tried activating venv and that didn't do anything

AAIGambino Reinstall stable diffusion? I tried activating venv and that didn't do anything

Furkan Gözükara SECourses•12/26/23, 9:39 PM

it is kohya

FFurkan Gözükara SECourses it is kohya

AIGambino•12/26/23, 10:04 PM

AAIGambino Click to see attachment

Furkan Gözükara SECourses•12/26/23, 10:05 PM

restart

Furkan Gözükara SECourses•12/26/23, 10:05 PM

also your antivirus could be problem

Zax•12/26/23, 10:10 PM

Hello! New here. Dreambooth training question. If I'm trying to train a model on images of a woman who doesn't have a thin body type, is having regularization images of thin women influencing the training towards a thin body type? I've been banging my head on this training for 3 days now with no real success.

just curious if anyone prefers training loras on comfyUI (https://github.com/johnman3032/simple-lora

Similar Threads