Software Engineering Courses (SECourses)•3y ago

Hello, I am ML engineer and currently developing image generation engine with sd-1.5 dreambooth lora

Hello,
I am ML engineer and currently developing image generation engine with sd-1.5 dreambooth lora.
I am trying to train a model with 1 product photo and have some issues to do that.

SStarBuffer Hello, I am ML engineer and currently developing image generation engine with sd...

Benjamin•12/7/23, 11:47 AM

I never heard someone was successful doing that with only 1 photo

StarBufferOP•12/7/23, 11:55 AM

This is the example product image.

StarBufferOP•12/7/23, 11:55 AM

StarBufferOP•12/7/23, 11:56 AM

What I am doing is augement background and use them as training data

StarBufferOP•12/7/23, 11:56 AM

And I am using diffusers to train dreambooth lora

StarBufferOP•12/7/23, 11:57 AM

Could you tell me hyperparameters for dreambooth lora sd-1.5?

SStarBuffer Hello, I am ML engineer and currently developing image generation engine with sd...

Dungeon_Mst•12/7/23, 11:59 AM

Have you checked the videos from Furkan he covers many points on Lora and Dreambooth training, following his settings and having a good variety of of you dataset 20images approx and captioning them well then training it on the model that you want to use should give you good results to start with. Also suggest doing via run pod so that the training is quick to test various iterations

StarBufferOP•12/7/23, 11:59 AM

I am using diffusers currently

StarBufferOP•12/7/23, 12:00 PM

How can I use captions for images when I use diffusers?

StarBufferOP•12/7/23, 12:00 PM

https://github.com/huggingface/diffusers/blob/main/examples/dreambooth/train_dreambooth_lora.py

GitHub

diffusers/examples/dreambooth/train_dreambooth_lora.py at main · hu...

Diffusers: State-of-the-art diffusion models for image and audio generation in PyTorch - huggingface/diffusers

StarBufferOP•12/7/23, 12:00 PM

I am currently using this code to train lora model

Zono50•12/7/23, 3:05 PM

hey boss, any eta on when the video to densepose runpod scripts will bea vailable? @Dr. Furkan Gözükara

Mishroom•12/7/23, 4:28 PM

Hello everyone! How many images do I have have to put in dataset to train dreambooth kohya as and what length of prompts is optimal to describe images in dataset, what model is better to use to describe images in dataset?

thewingag•12/7/23, 5:43 PM

ValueError: Cannot load <class 'magicanimate.models.appearance_encoder.AppearanceEncoderModel'> from F:\AI\ComfyUI_windows_portable\ComfyUI\models\MagicAnimate\densepose_controlnet because the following keys are missing:
....
Please make sure to pass low_cpu_mem_usage=Falselow_cpu_mem_usage=False and device_map=Nonedevice_map=None if you want to randomly initialize those weights or else make sure your checkpoint file is correct.
can u help me. thanks

Zono50•12/7/23, 5:46 PM

what's the best cudatoolkit version to use now that is compatible with most of the ai software out there now?

ZZono50 what's the best cudatoolkit version to use now that is compatible with most of t...

J.•12/7/23, 6:27 PM

12.1 or above i think

J.•12/7/23, 6:27 PM

i think most things use 11.8

J.•12/7/23, 6:27 PM

and i have seen some use 12.0

J.•12/7/23, 6:27 PM

@Dr. Furkan Gözükara Spam from Hinga.inc

Zono50•12/7/23, 6:29 PM

i'm a install 11.8

Furkan Gözükara SECourses•12/7/23, 6:43 PM

will check messages

Furkan Gözükara SECourses•12/7/23, 6:43 PM

preparing kaggle foocus notebook

liber vir•12/7/23, 6:43 PM

what is the fastest way to train reliable checkpoint or lore? i have heared about instant lora and how long does it take?

FiveSAI•12/7/23, 6:51 PM

Getting super fast speed on sdxl on 4090

Lliber vir what is the fastest way to train reliable checkpoint or lore? i have heared abo...

FiveSAI•12/7/23, 6:55 PM

Depends SDXL or SD1.5? for me around 10-15 mins for sd.15 and 30-50 mins on sdxl

Lliber vir what is the fastest way to train reliable checkpoint or lore? i have heared abo...

lucy•12/7/23, 7:09 PM

instant lora just uses ipadapter, it's not actually lora related at all and doesnt yield nearly as good results

FFiveSAI Depends SDXL or SD1.5? for me around 10-15 mins for sd.15 and 30-50 mins on sdx...

liber vir•12/7/23, 7:15 PM

thats too much

Llucy instant lora just uses ipadapter, it's not actually lora related at all and does...

liber vir•12/7/23, 7:15 PM

yeah i know i saw another video with semilia name , i hoped there would be something faster

liber vir•12/7/23, 7:16 PM

u know anything faster than that?

FiveSAI•12/7/23, 7:16 PM

best bet is roop or one of those face swappers but quality isnt the best

liber vir•12/7/23, 7:20 PM

yeah i also want it for animals, pets

Tthewingag thanks. error python.

Furkan Gözükara SECourses•12/7/23, 7:28 PM

when do you get this error?

BBaran has anyone used a different base model for sdxl dreambooth training than sdxl ba...

Furkan Gözükara SECourses•12/7/23, 7:28 PM

i am yet to test

SStarBuffer This is the example product image.

Furkan Gözükara SECourses•12/7/23, 7:29 PM

it is doable. but procudes same pose. i am teaching this with private consultation

ZZono50 hey boss, any eta on when the video to densepose runpod scripts will bea vailabl...

Furkan Gözükara SECourses•12/7/23, 7:29 PM

hopefully today

Furkan Gözükara SECourses•12/7/23, 7:29 PM

right now making a notebook for foocus on kaggle

Furkan Gözükara SECourses•12/7/23, 7:29 PM

working great

ZZono50 what's the best cudatoolkit version to use now that is compatible with most of t...

Furkan Gözükara SECourses•12/7/23, 7:30 PM

mostly cuda 11.8

Lliber vir what is the fastest way to train reliable checkpoint or lore? i have heared abo...

Furkan Gözükara SECourses•12/7/23, 7:30 PM

checkpoint best

Furkan Gözükara SECourses•12/7/23, 7:30 PM

if you need lora extract it

FFiveSAI Getting super fast speed on sdxl on 4090

Furkan Gözükara SECourses•12/7/23, 7:31 PM

wow nice. on windows?

FFurkan Gözükara SECourses wow nice. on windows?

FiveSAI•12/7/23, 7:33 PM

yep windows, testing on linux soon

Furkan Gözükara SECourses•12/7/23, 8:38 PM

new free Kaggle notebook for Fooocus added : https://www.patreon.com/collection/258108?view=expanded

FFurkan Gözükara SECourses it will continue caching

hozaggers•12/7/23, 8:42 PM

that means if I keep clicking 'continue training' it will try to continue caching the training dataset from where it left off and eventually complete? (or does it cache from the beggining again?)

is there a way to continue training a checkpoint in Kaggle with more input data after it has been trained once?

Hhozaggers that means if I keep clicking 'continue training' it will try to continue cachin...

Furkan Gözükara SECourses•12/7/23, 8:42 PM

it starts where it left

Furkan Gözükara SECourses•12/7/23, 8:43 PM

yes you can upload that checkpoint to kaggle temp and give its path

Furkan Gözükara SECourses•12/7/23, 8:43 PM

to continue

mtnmecca•12/7/23, 10:06 PM

@Dr. Furkan Gözükarain kohya version 22.2.1 they made a change to include the unet and text encoders so you didn't have to put it in the extra parameters area. In the json file it shows four training parameters now Do you know what the second text encoder "te" is for: "learning_rate": 1e-05,
"learning_rate_te": 1e-05,
"learning_rate_te1": 3e-06,
"learning_rate_te2": 0.0,

Mmtnmecca @Dr. Furkan Gözükarain kohya version 22.2.1 they made a change to include the un...

Furkan Gözükara SECourses•12/7/23, 10:08 PM

you can use gui to set each one

Furkan Gözükara SECourses•12/7/23, 10:08 PM

latest config updated for that

Hello, I am ML engineer and currently developing image generation engine with sd-1.5 dreambooth lora

Similar Threads