Software Engineering Courses (SECourses)•3y ago

I upgraded already

LLorelai AI I upgraded already

FiveSAI•12/9/23, 8:50 PM

works on the new one too so no worries

FiveSAI•12/9/23, 8:50 PM

I just like to keep multiple versions of kohya

FiveSAI•12/9/23, 8:51 PM

but yes after changing transformers 
then do this 

Name your folder structure like this


I add dt to ohwx

dt is just the first and last name of the person

but yes after changing transformers 
then do this 

Name your folder structure like this


I add dt to ohwx

dt is just the first and last name of the person

FiveSAI•12/9/23, 8:52 PM

Then caption your images like this,

Change the number of beams to either 10 or 15

FiveSAI•12/9/23, 8:52 PM

With the new transformers you cant change the beams

FiveSAI•12/9/23, 8:52 PM

Then you should be done

Lorelai AIOP•12/9/23, 8:54 PM

Still here just setting it up

FiveSAI•12/9/23, 8:56 PM

now worries after your start the training lemme know

Lorelai AIOP•12/9/23, 8:57 PM

Do I need to do anything else than change the version of transformers in requirements to install it?

FiveSAI•12/9/23, 8:58 PM

Nope thats pretty much it, you can change it back to 4.30.2 after your done

FiveSAI•12/9/23, 8:58 PM

thats why i keep two versions of kohya one just for captioning

FiveSAI•12/9/23, 8:58 PM

so if you have time keep a second version of kohya just for the transformers

FiveSAI•12/9/23, 8:59 PM

blip captioning is best for realistic people

Lorelai AIOP•12/9/23, 8:59 PM

And can you clarify captions again I don’t understand the folder structure

FiveSAI•12/9/23, 9:00 PM

So for the folder structure

You have the base folder img which you input in kohya

FiveSAI•12/9/23, 9:00 PM

inside that folder you have a folder 20_ohwxdt woman

FiveSAI•12/9/23, 9:00 PM

20 is the repeats ohwxdt is the instance prompt, and woman is class prompt

FiveSAI•12/9/23, 9:01 PM

the dt in ohwxdt can be replaced by the first and last name of the person you're training

Lorelai AIOP•12/9/23, 9:02 PM

FiveSAI•12/9/23, 9:03 PM

if you're still getting bad results lemme know, will share my class images later, my upload speed is bad

FiveSAI•12/9/23, 9:07 PM

brb

Lorelai AIOP•12/9/23, 9:13 PM

Sure, just booting stuff up

Lorelai AIOP•12/9/23, 9:15 PM

3640 steps, one hour! You damn genius

Lorelai AIOP•12/9/23, 9:16 PM

This would have been 12 hours on my old config

LLorelai AI 3640 steps, one hour! You damn genius 🙂

FiveSAI•12/9/23, 9:19 PM

Ahhh im glad! All fixed now

FiveSAI•12/9/23, 9:20 PM

With min snr gama

these best settings are either 0 or 10

so if you get bad results change one

FFiveSAI With min snr gama these best settings are either 0 or 10 so if you get bad re...

Lorelai AIOP•12/9/23, 9:21 PM

I’ll play around. It’s easier to tweak when training isn’t 12-18 hours

LLorelai AI I’ll play around. It’s easier to tweak when training isn’t 12-18 hours 😄

FiveSAI•12/9/23, 9:22 PM

Yep keep testing and trying, some of my trainings i do 10+ with different settings hehe

FiveSAI•12/9/23, 9:23 PM

also wait did you disable hardware-accelerated gpu scheduling ?

Lorelai AIOP•12/9/23, 9:24 PM

I enabled gradient checkpointing, for advanced I don’t think I did anything else.

FiveSAI•12/9/23, 9:25 PM

FiveSAI•12/9/23, 9:25 PM

this one

Furkan Gözükara SECourses•12/9/23, 9:25 PM

I am preparing a very good tutorial can you check and let me know if anything else missing?

https://github.com/FurkanGozukara/Stable-Diffusion/blob/main/Tutorials/Install-Python-C-Plus-Plus-Tools-FFMPEG-CUDA-Install-Tutorial.md

GitHub

Stable-Diffusion/Tutorials/Install-Python-C-Plus-Plus-Tools-FFMPEG-...

Stable Diffusion, SDXL, LoRA Training, DreamBooth Training, Automatic1111 Web UI, DeepFake, Deep Fakes, TTS, Animation, Text To Video, Tutorials, Guides, Lectures, Courses, ComfyUI, Google Colab, R...