Software Engineering Courses (SECourses)•3y ago

Is it possible to resume dreambooth training from a previously generated checkpoint? I have already

Is it possible to resume dreambooth training from a previously generated checkpoint? I have already generated one checkpoint for my model, but I don't know how to use it to train from there.

NNarva I did that but still the same

Furkan Gözükara SECourses•7/11/23, 10:46 AM

how many images you have and what is your per instance class count?

JJehex SDXL 1.0 has been released ?

Furkan Gözükara SECourses•7/11/23, 10:46 AM

not yet only in discord bot

FFurkan Gözükara SECourses not yet only in discord bot

Jehex•7/11/23, 10:46 AM

Ssquishycantaloupe E.g, in the tutorial screenshot I see this:

Furkan Gözükara SECourses•7/11/23, 10:47 AM

yes you can set from settings of auto111

Jehex•7/11/23, 10:47 AM

Guy's for Kohya_ss, images stabilization are really useful for your tests ?

LLT Hey 👋 I’m a musician/artist, been developing a new project with midjourney and...

Furkan Gözükara SECourses•7/11/23, 10:47 AM

hey welcome

Furkan Gözükara SECourses•7/11/23, 10:47 AM

unfortunately i dont have mac so cant produce tutorials for it

Furkan Gözükara SECourses•7/11/23, 10:47 AM

but you can use runpod

Furkan Gözükara SECourses•7/11/23, 10:47 AM

i have plenty tutorials

JJehex Guy's for Kohya_ss, images stabilization are really useful for your tests ?

Furkan Gözükara SECourses•7/11/23, 10:47 AM

wht you mean by images stabilization?

FFurkan Gözükara SECourses wht you mean by images stabilization?

Jehex•7/11/23, 10:48 AM

Regularisation sorry

Jehex•7/11/23, 10:49 AM

JJehex Regularisation sorry

Furkan Gözükara SECourses•7/11/23, 10:49 AM

well for dreambooth they are useful according to my experiments. lora is optimized way of dreambooth so should be

FFurkan Gözükara SECourses well for dreambooth they are useful according to my experiments. lora is optimiz...

Jehex•7/11/23, 10:50 AM

ok because when I use them its give me pretty bad results compare to not use them, I generate 1000 images of women with the same model I use for the training

JJehex ok because when I use them its give me pretty bad results compare to not use the...

Furkan Gözükara SECourses•7/11/23, 10:51 AM

well repeating logic of kohya is the issue

Furkan Gözükara SECourses•7/11/23, 10:51 AM

dreambooth extension much better for this

Jehex•7/11/23, 10:53 AM

sorry this is a noob question but what is the difference between kohya ss and dreambooth extension ? Both can create lora models ?

JJehex sorry this is a noob question but what is the difference between kohya ss and dr...

Furkan Gözükara SECourses•7/11/23, 10:55 AM

yes

Furkan Gözükara SECourses•7/11/23, 10:56 AM

difference is different scripts

Furkan Gözükara SECourses•7/11/23, 10:56 AM

so how they implemented the logic differs

Furkan Gözükara SECourses•7/11/23, 10:56 AM

which bugs they have differs :d

Jehex•7/11/23, 10:56 AM

thanks ! I will try both

PPawix Is it possible to resume dreambooth training from a previously generated checkpo...

PawixOP•7/11/23, 10:56 AM

do I just click "train" and it trains from latest epoch I trained before -- "10"?

PPawix do I just click "train" and it trains from latest epoch I trained before -- "10...

PawixOP•7/11/23, 10:57 AM

I don't have any "snapshot to resume" to select

PPawix do I just click "train" and it trains from latest epoch I trained before -- "10...

Furkan Gözükara SECourses•7/11/23, 11:10 AM

it will continue from 600 steps

Furkan Gözükara SECourses•7/11/23, 11:10 AM

epoch 10

FFurkan Gözükara SECourses it will continue from 600 steps

PawixOP•7/11/23, 11:11 AM

when continuing the training, on top left corner do I choose realistic vision as my base checkpoint or my newly created 600 steps endpoint?

PPawix when continuing the training, on top left corner do I choose realistic vision as...

Furkan Gözükara SECourses•7/11/23, 11:26 AM

you need to select your previous trained model

Furkan Gözükara SECourses•7/11/23, 11:26 AM

to continue

Furkan Gözükara SECourses•7/11/23, 11:26 AM

on top left

Furkan Gözükara SECourses•7/11/23, 11:26 AM

wont matter

Furkan Gözükara SECourses•7/11/23, 11:26 AM

it uses selected training model

Furkan Gözükara SECourses•7/11/23, 11:26 AM

this

Furkan Gözükara SECourses•7/11/23, 11:26 AM

PawixOP•7/11/23, 11:27 AM

okay, I see, source checkpoint, thanks!

PPawix okay, I see, source checkpoint, thanks!

Furkan Gözükara SECourses•7/11/23, 11:27 AM

you are welcome

hugojackson18•7/11/23, 12:01 PM

@Dr. Furkan Gözükara Love your Calibration images btw. Have you done the same for Women?

SC0RP10US•7/11/23, 12:37 PM

Hello everyone. I’m looking to buy 4070 12gb to upgrade from my 3060ti 8gb. The 4070 fits my budget and I’ll only be using for stable diffusion. My gaming rig has 6900xt, so is there any other options within the price range or is this the best choice? I tried running SD on 6900xt and it sucks lol

SSC0RP10US Hello everyone. I’m looking to buy 4070 12gb to upgrade from my 3060ti 8gb. The ...

ashley•7/11/23, 12:48 PM

If you want to train Dreambooth, 12GB is not enough, you can train LoRA with it. You might be able to train Dreambooth by enabling xformers and all kinds of things that don't give you the best settings.

Aashley If you want to train Dreambooth, 12GB is not enough, you can train LoRA with it....

SC0RP10US•7/11/23, 12:49 PM

At present I don't know the whole scope of training models myself and how will they be better than the models I can find on civit/hugging face?

SSC0RP10US At present I don't know the whole scope of training models myself and how will t...

ashley•7/11/23, 12:49 PM

If you aren't wanting to train custom models of people you know then its good enough

Aashley If you aren't wanting to train custom models of people you know then its good en...

SC0RP10US•7/11/23, 12:50 PM

I trained a model with 14 images on my on dreamlook.ai and I failed so for now I've given up on training loras/models lol

ashley•7/11/23, 12:51 PM

Training dataset quality is the most important thing when it comes to training. Must all be different clothing, background, lighting, close up shots, etc for good results.

SC0RP10US•7/11/23, 12:51 PM

I've trained lora on my 3060ti 14images 512x, it took 4ish hours

ashley•7/11/23, 12:52 PM

Wow thats very slow

Aashley Wow thats very slow 🙈

SC0RP10US•7/11/23, 12:53 PM

Painfully. But at the same time I have an nvidia gtx 1060 to scratch the itch while the lora was training. The only reason I chose 4070 as my buying option because that's whats available at bestbuy and I can't find 3080/90 24gb