Software Engineering Courses (SECourses)•3y ago

after update to the newest stable diffusion, when i use lora i trained before to generate images ,go

after update to the newest stable diffusion, when i use lora i trained before to generate images ,got this
RuntimeError: output with shape [60, 320, 1, 1] doesn't match the broadcast shape [60, 320, 3, 3]

sola_law•4/25/23, 4:03 AM

Does anyone know if its possible to blend two images in Dreambooth? I finally got it making pretty awesome images, but I'd like to put different hats and wigs on the realistic faces that Dreambooth is generating. I'm using CoLab.

crazygonuts•4/25/23, 4:46 AM

for autmatic1111 i had seperated my checkpoints into different directories and passed which one i wanted to use through different .bat file launchers. I'm trying vlad automatic now and it doesn't seem to have the --ckpt-dir flag anymore and its all saved in the config? is there a way to easily set up directory swapping like I had it before, does anyone know?

crazygonuts•4/25/23, 6:41 AM

basically like user profiles that point to different model dirs but without the need to duplicate all the rest of the data

Hhopeyan after update to the newest stable diffusion, when i use lora i trained before to...

Furkan Gözükara SECourses•4/25/23, 8:52 AM

ye you need additional networks extension

Ssola_law Does anyone know if its possible to blend two images in Dreambooth? I finally go...

Furkan Gözükara SECourses•4/25/23, 8:52 AM

you should be able to with prompting. or best way would be inpainting the object you want

Ccrazygonuts basically like user profiles that point to different model dirs but without the ...

Furkan Gözükara SECourses•4/25/23, 8:53 AM

havent used it yet but planning a tutorial

Ffelixsanz Click to see attachment

Furkan Gözükara SECourses•4/25/23, 9:01 AM

it is full res you need to open in browser

FFurkan Gözükara SECourses it is full res you need to open in browser

felixsanz•4/25/23, 11:03 AM

that's what i did

felixsanz•4/25/23, 11:03 AM

oh nevermind. i right clicked and opened in new tab

felixsanz•4/25/23, 11:03 AM

there is a "open in browser" button

felixsanz•4/25/23, 11:05 AM

3 looks better but doesn't seem flexible

Ffelixsanz that's what i did

Furkan Gözükara SECourses•4/25/23, 11:15 AM

i feel like 6 is best

Furkan Gözükara SECourses•4/25/23, 11:15 AM

but 3 is also decent

AlexP•4/25/23, 11:39 AM

Hi AI lovers, I'm new here, thanks for having me Dr. Furkan

AlexP•4/25/23, 11:39 AM

I want to start experimenting with training, I will do an avatar of my face for starters and have watched Dr. Furkans Tutorials which have been a great help.

AlexP•4/25/23, 11:39 AM

There are a few settings which are a massive source of confusion for me and I was hoping to get some advice here.

AlexP•4/25/23, 11:39 AM

One is the Epochs. I am trying to understand how many Epochs I should use. I've seen people use the words steps, samples, epochs, iterations, sometimes as if they are synonyms, sometimes not. So I read explanations and think I understand but the next thing I read just makes me confused again.

AlexP•4/25/23, 11:39 AM

I have seen the formula NUMBER OF IMAGES * 101 for determining epochs. I have 64 images, so that would be 6464. That seems a gigantic number, as in the tutorials I have seen suggested values of 150 and 200 Epochs.

AlexP•4/25/23, 11:40 AM

Please can someone clarify?

I am running Dreambooth on Runpod with 24GB VRAM.

AlexP•4/25/23, 11:44 AM

Another source of confusion for me is the classification images. Following the tutorial and suggested settings, when I got to the point to generate the classification images, Dreambooth created 3200 images. This seems a huge number and I have not understood how I can determine what the right amount is. Any pointers on the topic of classification images quantity?

Ffelixsanz the uploaded image quality is low

DeadMan1999•4/25/23, 12:35 PM

open it in browser

FFurkan Gözükara SECourses ok which one looking best?

DeadMan1999•4/25/23, 12:39 PM

x3, y6 looks good. what do you think?

DeadMan1999•4/25/23, 12:42 PM

x3, y3 follows the reference image/prompt and body posture more precisely in my opinion, while keeping face details.

DeadMan1999•4/25/23, 12:53 PM

are you going to put out a new video for Lora training? @Furkan Gözükara SECourses

Extemnemoy•4/25/23, 1:33 PM

selamlar stable diffision kullanmaya çalışıyorumda dreambooth ile yüz taraması yapmaya çalışyorum 4060 laptop versiyonu 8 g vram li ekran kartım var 4 5 gündür birçok yöntem denedim bir türlü beceremiyorum anladığım kadarıyla xformers kurmam gerekiyo ve ekran kartı mimarisinden dolayı otomatik kurulum yapamıyorum internetteki bi çok metodu denedim defalarca yeniden stable diffusion kurmak zorudna kaldım Memory Attention bölümünde xformersı göremedim dün gece bir yöntemle xformerı görmeyi başardım bu seferde "Exception training model: 'import of xformers halted; None in sys.modules'." hatası aldım birçok yöntem denediğim için kafam biraz çorba oldu açıkcsı sanırım çok hızlı güncellendiği için sistemler bulduğum yönlendirmeler düzgün çalışmıyor sanırım xformersmı kurmalıyım başka yöntemler var mıdır yardımcı olma imkanınız var mı acaba

AAlexP Hi AI lovers, I'm new here, thanks for having me Dr. Furkan 👋

Furkan Gözükara SECourses•4/25/23, 1:45 PM

hi welcome

Furkan Gözükara SECourses•4/25/23, 1:45 PM

number of epochs 100% depends on your training dataset

Furkan Gözükara SECourses•4/25/23, 1:46 PM

therefore best approach is saving checkpoints and comparing after training ended

DDeadMan1999 are you going to put out a new video for Lora training? @Furkan Gözükara SECours...

Furkan Gözükara SECourses•4/25/23, 1:46 PM

yep

Furkan Gözükara SECourses•4/25/23, 1:46 PM

i have done 15+ trainings so far to find best settings

DeadMan1999•4/25/23, 1:49 PM

that's great to hear, ill be waiting for your next video. I have learned a lot from your SD videos.

Zono50•4/25/23, 4:49 PM

getting this error while trying to run 4x ultrasharp

Zono50•4/25/23, 4:50 PM

Error completing request
_pickle.UnpicklingError: invalid load key, '\x08'.\

Zono50•4/25/23, 5:39 PM

trying to get 4x ultrasharp to work in my automatic111, i don't even have the ultrasharp in the esrgan folder anymore, yet they still show up here -

FFurkan Gözükara SECourses therefore best approach is saving checkpoints and comparing after training ended

AlexP•4/25/23, 6:06 PM

Thanks a lot for your response. So I guess I will try 200 Epochs and go from there. I guess the same applies to the amount of classification images, or is there a reccomended amout or formula for those?

Zono50•4/25/23, 8:11 PM

i can't even count how many times i've broken automatic1111 already.

Zono50•4/25/23, 8:11 PM

between compatibilty issues, different UI's, upscalers and extras not working right, dreambooth errors

Zono50•4/25/23, 8:12 PM

but because of that, i can setup an auto1111 installation in linux and be traiing a mode in about 10-15 minutes

AAlexP Thanks a lot for your response. So I guess I will try 200 Epochs and go from the...

Furkan Gözükara SECourses•4/25/23, 9:08 PM

well formula depends on your training set so best way is checkpoint testing with x/y/z

Furkan Gözükara SECourses•4/25/23, 10:37 PM

i am doing another experiment

Furkan Gözükara SECourses•4/25/23, 10:38 PM