Software Engineering Courses (SECourses)•2y ago

mobile RTX4090, so 16GB of VRAM

DDJRich mobile RTX4090, so 16GB of VRAM

Furkan Gözükara SECourses•12/29/23, 11:44 PM

i see

Furkan Gözükara SECourses•12/29/23, 11:44 PM

I would also do SDXL DreamBooth training with 16 GB config we have

Furkan Gözükara SECourses•12/29/23, 11:44 PM

and compare results

Furkan Gözükara SECourses•12/29/23, 11:44 PM

it could be better than LoRA

Furkan Gözükara SECourses•12/29/23, 11:44 PM

also training dataset matters a lot in terms of training quality

Furkan Gözükara SECourses•12/29/23, 11:44 PM

that could be difference reason

FFurkan Gözükara SECourses and compare results

DJRichOP•12/29/23, 11:46 PM

DreamBooth is definitely better, but I like the flexibility of using other models as well as the disk space savings. The image quality results are good from all four Loras, it is just that two of them essentially diminish the subject's likeness as more tokens are added.

DDJRich DreamBooth is definitely better, but I like the flexibility of using other model...

Furkan Gözükara SECourses•12/29/23, 11:47 PM

ok do this

Furkan Gözükara SECourses•12/29/23, 11:47 PM

dreambooth train

Furkan Gözükara SECourses•12/29/23, 11:47 PM

extract lora

Furkan Gözükara SECourses•12/29/23, 11:47 PM

it works great too

FFurkan Gözükara SECourses extract lora

DJRichOP•12/29/23, 11:48 PM

I thought I had read that the Lora extraction was deprecated in the backend code, but still in the UI. Are there any special settings that should be used for a successful extraction?

DDJRich I thought I had read that the Lora extraction was deprecated in the backend code...

Furkan Gözükara SECourses•12/29/23, 11:49 PM

it works great in kohya gui

Furkan Gözükara SECourses•12/29/23, 11:49 PM

i just tested

Furkan Gözükara SECourses•12/29/23, 11:49 PM

try different ranks

Furkan Gözükara SECourses•12/29/23, 11:49 PM

such as

Furkan Gözükara SECourses•12/29/23, 11:49 PM

128x128

Furkan Gözükara SECourses•12/29/23, 11:49 PM

192x192

Furkan Gözükara SECourses•12/29/23, 11:49 PM

256x256

Furkan Gözükara SECourses•12/29/23, 11:49 PM

whichever you like most

DJRichOP•12/29/23, 11:50 PM

Is there a json on Patreon for 16GB VRAM local Dreambooth training?

FFurkan Gözükara SECourses whichever you like most

DJRichOP•12/29/23, 11:55 PM

Or should I just use the Lora preset in the DreamBooth tab?

DDJRich Is there a json on Patreon for 16GB VRAM local Dreambooth training?

Furkan Gözükara SECourses•12/29/23, 11:56 PM

yes

Furkan Gözükara SECourses•12/29/23, 11:56 PM

we have here

Furkan Gözükara SECourses•12/29/23, 11:56 PM

https://www.patreon.com/posts/very-best-for-of-89213064

Furkan Gözükara SECourses•12/29/23, 11:56 PM

16GB_TextEncoder.json

Furkan Gözükara SECourses•12/29/23, 11:57 PM

i did 100+ testing and i believe my findings much better than those presets

FFurkan Gözükara SECourses have you tested my auto installer?

Digital [Starburst]•12/30/23, 12:58 AM

Yes I used the autoinstaller. My GPU is a 4070 ti

FFurkan Gözükara SECourses i did 100+ testing and i believe my findings much better than those presets

DJRichOP•12/30/23, 1:04 AM

Thank you. I ended up using your settings from the Lora training video you made. The issue is that the video is five months old and I know how quickly the technology is advancing. I will try the DreamBooth preset and let you know how it goes.

DDigital [Starburst]Yes I used the autoinstaller. My GPU is a 4070 ti

Furkan Gözükara SECourses•12/30/23, 1:21 AM

how much vram it has?

DDJRich Thank you. I ended up using your settings from the Lora training video you made....

Furkan Gözükara SECourses•12/30/23, 1:21 AM

i did very extensive serach for dreambooth

Furkan Gözükara SECourses•12/30/23, 1:21 AM

lora was not that much extensive

Furkan Gözükara SECourses•12/30/23, 1:21 AM

also dreambooth far superior than lora i compared

FFurkan Gözükara SECourses i did 100+ testing and i believe my findings much better than those presets

DJRichOP•12/30/23, 1:23 AM

By the way, I think I figured out the issue. I was using too many regularization images. I am doing a dreambooth test now using your text encoder 16GB preset with no modifications. I'll post my results when it completes. Thanks again

FFurkan Gözükara SECourses also dreambooth far superior than lora i compared

Haashim•12/30/23, 2:26 AM

I need to test this. Will check out recent vids

FFurkan Gözükara SECourses hello. are you installing on activated venv?

banfi6885•12/30/23, 5:55 AM

Hello, I finally got it working. Also, have you trained a style on the "dreamshaper xl turbo" model? If yes, did you use colab or runpod with kohya? Thank you

Deleted User•12/30/23, 8:08 AM

Hi @Dr. Furkan Gözükara My DB test finished a short time ago and I am still seeing similar results. I ran x,y,z plot tests and while facial features were similar, some prominent characteristics such as the subject's long hair and blue eyes were not present. One example prompt I used had instagram selfie at the start of the prompt and even the last epoch's model turned the subject into a young, muscular man when the subject is in his 60's. I was able to see slightly better results by using (Instance prompt Class prompt:2), but I feel that shouldn't be necessary if I am training correctly.

I've included a screenshot of my dataset preparation page if it helps. I am also using the 16GB_TextEncoder.json provided with no changes at all other than modifying the output model name.

Deleted User•12/30/23, 10:06 AM

DDeleted User Click to see attachment

Deleted User•12/30/23, 10:08 AM

Please guide me i have 15 images repeat 150 + regularisation images batch size 2 total steps is 4500, on how much steps i need to save every n steps for best quality checkpoints

Deleted User•12/30/23, 10:28 AM

15 * 30 * 2 + 1=901 is this ok for checkpoints to save

DDeleted User Hi @Dr. Furkan Gözükara My DB test finished a short time ago and I am still see...

DJRichOP•12/30/23, 12:42 PM

I ran one more test overnight. The only thing I changed is that I pointed the regularization image folder to the one that has the 5200 images you compiled whereas before I was calculating the images needed for the number of repeats I was doing for the subject images and creating a folder with that total number. I'm not sure if that would make a difference, but it sure seems to. The likeness is much better now. The only challenge which I don't know the resolution for yet is that when putting additional names in the prompt such as the artist's style, it does cause issues with the likeness. I am not sure if perhaps doing more epochs is the answer or something else.