Software Engineering Courses (SECourses)•3y ago

Your videos and results are great. The realistic portraits from Kohya you made are excellent.

TTheWhykiki Your videos and results are great. The realistic portraits from Kohya you made a...

Furkan Gözükara SECourses•5/1/23, 2:03 PM

thank you

Furkan Gözükara SECourses•5/1/23, 2:03 PM

i will show hopefully a way to

Furkan Gözükara SECourses•5/1/23, 2:04 PM

sort them with similarity

Furkan Gözükara SECourses•5/1/23, 2:04 PM

so you can generate thousands of images

Furkan Gözükara SECourses•5/1/23, 2:04 PM

and let AI the sort best ones for you

Furkan Gözükara SECourses•5/1/23, 2:04 PM

working on it

FFurkan Gözükara SECourses are you doing this on kohya ss? what is your repeating count?

Isaac Brock•5/1/23, 2:06 PM

i have ~20 images, so i have repeats on 30 for those and 1 for the regularization images.

IIsaac Brock i have ~20 images, so i have repeats on 30 for those and 1 for the regularizatio...

Furkan Gözükara SECourses•5/1/23, 2:06 PM

that means 1 epoch is actually 30 epochs of sd dreambooth extension

Furkan Gözükara SECourses•5/1/23, 2:06 PM

because it trains 30 times each image in 1 epoch of kohya lora script

Furkan Gözükara SECourses•5/1/23, 2:09 PM

1 epoch simply means every training data is trained 1 time. the step count for this also depends on your batch size

Furkan Gözükara SECourses•5/1/23, 2:10 PM

lets say i have 20 images

Furkan Gözükara SECourses•5/1/23, 2:10 PM

1 epoch is 20 steps

Furkan Gözükara SECourses•5/1/23, 2:10 PM

if i make batch size 4

Furkan Gözükara SECourses•5/1/23, 2:10 PM

then in 5 steps = 1 epoch

Furkan Gözükara SECourses•5/1/23, 2:10 PM

because each step it trains 4 images

Arvy•5/1/23, 2:10 PM

@Furkan Gözükara SECourses have you tried the Tile resample control net for upscaling images?

AArvy @Furkan Gözükara SECourses have you tried the Tile resample control net for upsc...

Furkan Gözükara SECourses•5/1/23, 2:10 PM

not yet. i was waiting controlnet 1.1 to become stable

Furkan Gözükara SECourses•5/1/23, 2:10 PM

for a new video

Arvy•5/1/23, 2:11 PM

Oh okay. Apparently it works like magic by adding a lot of details. I am not able to figure out how to properly use it. Will wait for your video

Furkan Gözükara SECourses•5/1/23, 2:12 PM

since automatic1111 published a new release i think time has come

FFurkan Gözükara SECourses that means 1 epoch is actually 30 epochs of sd dreambooth extension

BecauseReasons•5/1/23, 2:16 PM

Are there any best-practices on the settings/math for training Kohya_SS styles? ie) 100-300 images.

BBecauseReasons Are there any best-practices on the settings/math for training Kohya_SS styles? ...

Furkan Gözükara SECourses•5/1/23, 2:19 PM

well best way is saving frequent epoch and afterwards doing x/y/z comparison

Furkan Gözükara SECourses•5/1/23, 2:21 PM

i have to go the gym. will be afk for a while

rafraf•5/1/23, 2:27 PM

is there a batch processing tool in kohya_ss to process multiple json config settings consecutively, to test variations of LoRA's? @Furkan Gözükara SECourses

rafraf•5/1/23, 2:27 PM

I thought there was one

pat._.fenis•5/1/23, 2:43 PM

Is the new automatic111 stable and is dreambooth fully functional? What is the best method for upgrading it if so?

pat._.fenis•5/1/23, 2:48 PM

I am hoping the upgrade will fix the LDSR issue I have.

FFurkan Gözükara SECourses because each step it trains 4 images

Isaac Brock•5/1/23, 3:08 PM

thanks, makes sense now

rafraf•5/1/23, 4:42 PM

Getting 4it/s on Kohya_ss torch 1.13... isn't 2.0 faster? You said it wasn't working so well, is updating to 2.0 not recommended? @Furkan Gözükara SECourses thx

rafraf•5/1/23, 4:43 PM

Updating should in principle speed up the training 4x no (got 16it/s on webui generating but not sure how it carries over into kohya training)? Can you qualify what you meant?

ZZono50 biggest thing when doing captions is to just use the instance prompt, if you use...

rafraf•5/1/23, 5:26 PM

Oh I see. Slow thinking. You are using captions to add specific flexible details in the portraits. Sweet. Amazing how varied methods can be. Nothing is written on stone

Ppat._.fenis I am hoping the upgrade will fix the LDSR issue I have.

Furkan Gözükara SECourses•5/1/23, 6:47 PM

i saw they fixed LDSR issue

Furkan Gözükara SECourses•5/1/23, 6:47 PM

havent tested dreambooth yet

Rrafraf Getting 4it/s on Kohya_ss torch 1.13... isn't 2.0 faster? You said it wasn't wor...

Furkan Gözükara SECourses•5/1/23, 6:47 PM

torch 2 didnt do faster for me but you can test not sure why it wasnt

Furkan Gözükara SECourses•5/1/23, 6:47 PM

posted all our tutorials here : https://github.com/FurkanGozukara/Stable-Diffusion

GitHub

GitHub - FurkanGozukara/Stable-Diffusion: Best Stable Diffusion and...

Best Stable Diffusion and AI Tutorials, Guides, News, Tips and Tricks - GitHub - FurkanGozukara/Stable-Diffusion: Best Stable Diffusion and AI Tutorials, Guides, News, Tips and Tricks

Alamin H•5/1/23, 6:48 PM

whats the problem??

AAlamin H whats the problem??

Furkan Gözükara SECourses•5/1/23, 6:50 PM

it means space character error. replace that cell with below code

Furkan Gözükara SECourses•5/1/23, 6:52 PM

hf_token = None # <---- put hf_token here only in private notebooks;
if hf_token is None:
    from huggingface_hub import login; login()  # put your hf_token in output of this cell if you want to share notebook in public;

hf_token = None # <---- put hf_token here only in private notebooks;
if hf_token is None:
    from huggingface_hub import login; login()  # put your hf_token in output of this cell if you want to share notebook in public;

Furkan Gözükara SECourses•5/1/23, 6:52 PM

you see your from has space problem

Alamin H•5/1/23, 6:53 PM

lemme try again now

Alamin H•5/1/23, 7:00 PM

gc??