Software Engineering Courses (SECourses)•3y ago

i must be doing something wrong 😄 it's RealisticVision5.1, Dreambooth trained. on do you guys know

i must be doing something wrong

it's RealisticVision5.1, Dreambooth trained. on do you guys know any good tutorials where I can learn how to get good looking realistic photos?

SuperTurboHeroOP•10/12/23, 9:25 PM

the best results so far I got from Automatic1111 v1.4.0 interface, because there are these options "restore faces" and "Hires.fix"

t800se•10/12/23, 9:26 PM

Hey @Dr. Furkan Gözükara , very silly question probably but I am new in all of these

I just started the training process in Kaggle Notebook by watching your videos. I didn't relise that it will take 7-9 hrs

Do I need to leave wifi on and my pc on while it trains? or because it's in the cloud, so it doesn't matter?

SSuperTurboHero i must be doing something wrong 😄 it's RealisticVision5.1, Dreambooth trained. ...

Furkan Gözükara SECourses•10/12/23, 11:21 PM

looks like training errors

Furkan Gözükara SECourses•10/12/23, 11:21 PM

what is your gpu 12gb?

Tt800se Hey @Dr. Furkan Gözükara , very silly question probably but I am new in all of t...

Furkan Gözükara SECourses•10/12/23, 11:22 PM

yes you need to leave browser open on kaggle

Furkan Gözükara SECourses•10/12/23, 11:22 PM

on runpod you dont have to

Jas•10/12/23, 11:40 PM

If i train a dreambooth SDXL model on a style, is it weird to use regulisation images?

Jas•10/12/23, 11:42 PM

i am making a model of images like this

Jas•10/12/23, 11:43 PM

so it would be strange to have regularisation images of portrait photos of people?

JJas If i train a dreambooth SDXL model on a style, is it weird to use regulisation i...

Furkan Gözükara SECourses•10/12/23, 11:51 PM

you dont have to use

Furkan Gözükara SECourses•10/12/23, 11:51 PM

if you dont have good ones

Furkan Gözükara SECourses•10/12/23, 11:51 PM

i would do 2 trainings and compare

FFurkan Gözükara SECourses you dont have to use

Jas•10/12/23, 11:59 PM

thanks for quick reply, from what I understand the regularisation images should be images that you would expect to see when using the class prompt?

So if i use the word "style" as a class prompt then should I have regularisation images of different styles?

Really the images will all be kind of macro photos of colourful water and such like this...

What kind of regularisation images would be best?

JJas thanks for quick reply, from what I understand the regularisation images should ...

Furkan Gözükara SECourses•10/13/23, 12:13 AM

the logic is like this

Furkan Gözükara SECourses•10/13/23, 12:13 AM

we are leveraging a class info knowladge

Furkan Gözükara SECourses•10/13/23, 12:13 AM

reg images should be as much as possibly like those class images used in the original training of model

Furkan Gözükara SECourses•10/13/23, 12:13 AM

so it depends on your class word

Furkan Gözükara SECourses•10/13/23, 12:14 AM

lets say you use style

Furkan Gözükara SECourses•10/13/23, 12:14 AM

then you need to collect images similar to original images used during initial training that contained style word

Furkan Gözükara SECourses•10/13/23, 12:14 AM

or use whatever the model generates easier

virginia•10/13/23, 12:17 AM

regularization images are reference of corrections? or what kind of function they offer?

Vvirginia regularization images are reference of corrections? or what kind of function the...

Furkan Gözükara SECourses•10/13/23, 1:08 AM

https://arxiv.org/pdf/2208.12242.pdf

Furkan Gözükara SECourses•10/13/23, 1:08 AM

Furkan Gözükara SECourses•10/13/23, 1:08 AM

read that paper

virginia•10/13/23, 1:09 AM

Hi Furkan, I have a question. for example, if I want to train panda dataset. could I use any panda image as regulation images?

Vvirginia Hi Furkan, I have a question. for example, if I want to train panda dataset. cou...

Furkan Gözükara SECourses•10/13/23, 1:20 AM

yes

virginia•10/13/23, 1:22 AM

dataset and regularization should look like the same, or for data sets have a big circle on the belly, but regularization have no circle or dataset and regularization images can be the same?

Furkan Gözükara SECourses•10/13/23, 3:06 AM

https://twitter.com/GozukaraFurkan/status/1712665890259447831

Spicy Chicken Sandwich Deluxe•10/13/23, 5:07 AM

Would you say Dall-e is much better than other models?

EElisoli Click to see attachment

Metaphysix•10/13/23, 5:54 AM

Those who have such problem running on windows do the following:

start cmd in venv\scripts
type

activate

activate

then

pip install clip_interrogator, gradio

pip install clip_interrogator, gradio

These packages are missing from the virtual environment

Gerdes of Earth•10/13/23, 6:58 AM

Anyone have any experience merging LoRa's into models with Kohya? Having a hard time with the LoRa overriding the model too much but requiring a higher % to make sure I preserve much of the LoRa's strengths.

Kallamamran•10/13/23, 8:16 AM

@Dr. Furkan Gözükara Since A1111 (and other SD's) build upon Python. What do you think about the new Python 3.12? It seem to have some serious performance enhancements

t800se•10/13/23, 8:55 AM

Okay! I left everything open and I woke up to find my screen like this:

Does this mean the job completed or session closed itself for some reason?

t800se•10/13/23, 8:56 AM

The connection was closed when I woke up, I connected to teh server again but scared to run the cell again, just in case if it start from the beginning

Benjamin•10/13/23, 9:19 AM

did you select that files will be persisted? at the right side under notebook options

Benjamin•10/13/23, 9:20 AM

you can also check if the files are still there with the data file browser at the right side

Benjamin•10/13/23, 9:20 AM

but usually everything is gone if you lost connection

Benjamin•10/13/23, 9:21 AM

in general kaggle doesnt want you to have long running tasks, that's why they disconnect you for example if your screensaver turns on and theres not activity for 40 minutes. Disabling the screensaver does the trick for me

Narva•10/13/23, 9:28 AM

Very amazing the new sdxl parameter, sadly tattos are not good, but the realism is amazing

NNarva Very amazing the new sdxl parameter, sadly tattos are not good, but the realism ...

Benjamin•10/13/23, 10:01 AM

nice, how many images did you use?

Elisoli•10/13/23, 10:49 AM

can i somehow take this model - https://civitai.com/models/139562?modelVersionId=169921 (real vision but for SDXL, still in training) - and use my trained dreambooth sdxl with it? will merge work?
Or should I train my whole dreambooth on ths as a base model?