Software Engineering Courses (SECourses)•3y ago

you dont have to use

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

if you dont have good ones

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

the logic is like this

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

we are leveraging a class info knowladge

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

so it depends on your class word

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

lets say you use style

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

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

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

Furkan Gözükara SECoursesOP•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 SECoursesOP•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 SECoursesOP•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 activateactivate then
pip install clip_interrogator, gradiopip 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?

SSpicy Chicken Sandwich Deluxe Would you say Dall-e is much better than other models?

Furkan Gözükara SECoursesOP•10/13/23, 10:58 AM

no i wouldnt

Furkan Gözükara SECoursesOP•10/13/23, 10:58 AM

but it can follow prompts really really good

MMetaphysix > Those who have such problem running on *windows* do the following: start **cmd...

Furkan Gözükara SECoursesOP•10/13/23, 10:59 AM

they are in installer file it is weird didnt install in your case

GGerdes of Earth Anyone have any experience merging LoRa's into models with Kohya? Having a hard ...

Furkan Gözükara SECoursesOP•10/13/23, 10:59 AM

it is on my research list

KKallamamran @Dr. Furkan Gözükara Since A1111 (and other SD's) build upon Python. What do yo...

Furkan Gözükara SECoursesOP•10/13/23, 10:59 AM

even 3.11 wont work so we still have far to get support for 3.12

Tt800se Okay! I left everything open and I woke up to find my screen like this: Does th...

Furkan Gözükara SECoursesOP•10/13/23, 11:00 AM

wow sad. i think best way is while it is generating checkpoints download them

Furkan Gözükara SECoursesOP•10/13/23, 11:00 AM

kaggle free is so limiting

BBenjamin in general kaggle doesnt want you to have long running tasks, that's why they di...

Furkan Gözükara SECoursesOP•10/13/23, 11:00 AM

yep

Furkan Gözükara SECoursesOP•10/13/23, 11:01 AM

they are so restrictive. best way is downloading them as they are generated

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

Furkan Gözükara SECoursesOP•10/13/23, 11:01 AM

which parameter

Furkan Gözükara SECoursesOP•10/13/23, 11:01 AM

check out this video : https://youtu.be/PNA9p94JmtY

YouTubeSECourses

Caption Images or Learn How To Prompt With Clip Vision of SDXL and ...

Download scripts from here : https://www.patreon.com/posts/sota-very-best-90744385

Full length Twitter video here : https://twitter.com/GozukaraFurkan/status/1712665890259447831

SOTA (The Very Best) Image Captioning Models Script For Stable Diffusion And More

1 Click install and use SOTA image captioning models on your computer. Supports 8 bi...

BBenjamin did you select that files will be persisted? at the right side under notebook op...

t800se•10/13/23, 11:03 AM

no mate, shall I choose persistance option?

Tt800se no mate, shall I choose persistance option?

Benjamin•10/13/23, 11:04 AM

it's what I do to be on the safe side. As @Dr. Furkan Gözükara mentioned or else you need to download the checkpoints right away.

If you persist the files you can resume afterwards if it got disconnected with a parameter in the training command

t800se•10/13/23, 11:04 AM

I started new instance without it tho