Software Engineering Courses (SECourses)•3y ago

Hello @Dr. Furkan Gözükara , which Ai model of SD V2 I need to use for training 768x768 pixel images

Hello @Dr. Furkan Gözükara , which Ai model of SD V2 I need to use for training 768x768 pixel images ?

ashley•6/6/23, 12:44 PM

Kohya_ss dreambooth actually seems to train a person better for styling than the A1111 Dreambooth extension

ashley•6/6/23, 12:46 PM

I couldn't get the example data set from YouTube trained properly with the Dreambooth extension, Kohya_ss is doing a much better job

RReach Hello @Dr. Furkan Gözükara , which Ai model of SD V2 I need to use for training ...

Furkan Gözükara SECourses•6/6/23, 12:59 PM

sd 21 is really hard to train and get good results

Aashley I couldn't get the example data set from YouTube trained properly with the Dream...

Furkan Gözükara SECourses•6/6/23, 12:59 PM

interesting i think i should compare

Furkan Gözükara SECourses•6/6/23, 1:00 PM

by the way it turns out that 768x1024 loses styling ability significantly

Furkan Gözükara SECourses•6/6/23, 1:00 PM

good for real images but not for style

Furkan Gözükara SECourses•6/6/23, 1:00 PM

and my training dataset had so many repeating cloth and background

Furkan Gözükara SECourses•6/6/23, 1:00 PM

that degraded styling significantly

RRedman I am having so much fun with roop. I'll have to subscribe to your patreon next ...

Furkan Gözükara SECourses•6/6/23, 1:00 PM

awesome ty so much

teachtaha•6/6/23, 1:03 PM

@Dr. Furkan Gözükara in cloning, can I resume training Eg i left at 2550.gpt can I start again from that?

teachtaha•6/6/23, 1:04 PM

why is this picking random voices? Can i use one specific tone only? Please guide

Tteachtaha @Dr. Furkan Gözükara in cloning, can I resume training Eg i left at 2550.gpt can...

Furkan Gözükara SECourses•6/6/23, 1:04 PM

yes you can continue training from there

Furkan Gözükara SECourses•6/6/23, 1:05 PM

random voices could be related to your training progress and dataset

teachtaha•6/6/23, 1:07 PM

Does more training means better result? which parameter to select to resume training?

teachtaha•6/6/23, 1:08 PM

Dataset contains only my voice of 16 mins and trained it for 5 hours until 2155_gpt.pth

Tteachtaha Does more training means better result? which parameter to select to resume trai...

Furkan Gözükara SECourses•6/6/23, 1:14 PM

not at all

Furkan Gözükara SECourses•6/6/23, 1:14 PM

16 min voice could be low amount

Furkan Gözükara SECourses•6/6/23, 1:14 PM

usually more is better in training dataset as long as high quality

FFurkan Gözükara SECourses interesting i think i should compare

ashley•6/6/23, 1:15 PM

wand.ai helps to determine where styling gets lost as well

Furkan Gözükara SECourses•6/6/23, 1:15 PM

the model learned even the whites in my beard haha

Furkan Gözükara SECourses•6/6/23, 1:15 PM

ashley•6/6/23, 1:16 PM

Thats awesome

Furkan Gözükara SECourses•6/6/23, 1:16 PM

but there is too much memorization because of my training set

Furkan Gözükara SECourses•6/6/23, 1:16 PM

i should get good different bg and clothing pictures

ashley•6/6/23, 1:17 PM

Oh I thought you wanted it like that

Furkan Gözükara SECourses•6/6/23, 1:17 PM

it is good but cant stylze

FFurkan Gözükara SECourses usually more is better in training dataset as long as high quality

teachtaha•6/6/23, 1:23 PM

Whats the minimum amount of dataset? Also, will it resume if I start it like in video?

teachtaha•6/6/23, 1:29 PM

I mean will dataset training resume?

Zono50•6/6/23, 1:44 PM

they need to hurry up aand make it so us lower vram guys can dreambooth

Zono50•6/6/23, 1:44 PM

or should i just rent a runpod with someone with a beast card, train my models, then use them in auto

ZZono50 or should i just rent a runpod with someone with a beast card, train my models, ...

ashley•6/6/23, 1:45 PM

Yes

ashley•6/6/23, 1:46 PM

You don't need an 80GB beast though, 24GB is fine unless you want to use Adan D-Adapatation Optimizer, then you need at least 32GB

ashley•6/6/23, 1:47 PM

W&B can also send email or Slack notifications when your training completes

Tteachtaha Whats the minimum amount of dataset? Also, will it resume if I start it like in ...

Furkan Gözükara SECourses•6/6/23, 1:47 PM

i trained with 2 hours speech

Aashley W&B can also send email or Slack notifications when your training completes 😍

Furkan Gözükara SECourses•6/6/23, 1:48 PM

really this is 32 gb?

Furkan Gözükara SECourses•6/6/23, 1:48 PM

wow

FFurkan Gözükara SECourses wow

ashley•6/6/23, 1:49 PM

Only Adan D-Adaptation Optimizer, not the other D-Dadaptation optimizers, the others can all run on 24GB VRAM

ashley•6/6/23, 1:51 PM

Maybe its a bug in Dreambooth training for the DB extension, because someone is using the DB extension to train LoRA with Adan D-Adapt with just 8GB: https://civitai.com/articles/52/d8ahazard-dreambooth-extension-guide

Civitai | D8ahazard Dreambooth extension guide

FFurkan Gözükara SECourses sd 21 is really hard to train and get good results

ReachOP•6/6/23, 2:09 PM

So do you advice better to train with 512 x 512 pixel image and SD 1.5 model ?

RReach So do you advice better to train with 512 x 512 pixel image and SD 1.5 model ?

Furkan Gözükara SECourses•6/6/23, 2:12 PM

sd 1.5 much easier better

Furkan Gözükara SECourses•6/6/23, 2:12 PM

also using public models are even better

ashley•6/6/23, 2:12 PM

W&B also helps you compare 4 different trainings to determine the optimal average epoch so you don't overtrain

Furkan Gözükara SECourses•6/6/23, 2:12 PM

according to your aim

ashley•6/6/23, 2:13 PM

Well can be much more than 4, just I have only done 4 so far

Furkan Gözükara SECourses•6/6/23, 2:13 PM

W&B?

ashley•6/6/23, 2:13 PM

Weights & Balances (wandb.ai)

Furkan Gözükara SECourses•6/6/23, 2:14 PM

i see paid service

ashley•6/6/23, 2:14 PM

You can enter your API key into Kohya_ss and EveryDream2trainer, not yet supported by Dreambooth extension but I logged a feature request to add it

FFurkan Gözükara SECourses i see paid service

ashley•6/6/23, 2:14 PM

Free for personal and academic use.

FFurkan Gözükara SECourses also using public models are even better

ReachOP•6/6/23, 2:14 PM

okay so can I use your 512 x 512 Reg class man images of your Patreon for training ?

Hello @Dr. Furkan Gözükara , which Ai model of SD V2 I need to use for training 768x768 pixel images

Similar Threads