Furkan Gözükara
patreon
Best Stable Diffusion 1.5 Based Models Training Config - Parameters With Kohya Scripts GUI
🕑 Added 2023-10-04 14:32:41 +0000 UTCJoin discord and tell me your discord username to get a special rank : SECourses Discord
I have done over 40 trainings so far. But today I found out that since version v21.8.6 of Kohya SS GUI the script was broken for Stable Diffusion SD 1.5 based models training.
It was only training U-NET thus huge resemblance and quality loss.
So I started hyper parameter experimentation again sadly :(
Once I got best hyper parameters hopefully I will update this thread.
So far best working config : me_sd_7e7_v2.json
If you get error or if it starts using shared VRAM enable gradient checkpointing to reduce VRAM usage. If your VRAM still be insufficient, reduce resolution to 512x512px. Currently it is 768x768px.
So how to use latest working v21.8.4 of Kohya SS?
First git clone the repo as below
https://github.com/bmaltais/kohya_ss
Then open a CMD inside clone repo. If you have already existing installation you don't have to reinstall.
And type git checkout v21.8.4
After that restart GUI and you will have a working version.
How did I know it was only training U-NET?
Here I have opened an issue thread : https://github.com/kohya-ss/sd-scripts/issues/855
Comments
Furkan Gözükara
9 hours is so wrong for 4090. if you are getting all same picture that means the model is cooked. i need more info to give you better picture
محمد الذوادي
I'm getting the same picture as a result no matter what prompt I use, do I have to change the repeats or lower the epoch? , I tried the sdxl config with stock setting and I'm getting better results but 9 hours of training on 4090
Furkan Gözükara
with SDXL best settings are still valid.
Meito
hi with the new updated kohya ss, do we change the text encoder learning rate as well?
Furkan Gözükara
For style as many as possible with each one having different context. logic explained in this video : https://youtu.be/m-UVVY_syP0
George Gostyshev
What's your best practices recommendation oh now much dataset should be for training best possible style lora/best possible finetune? Thanks in advance)
Furkan Gözükara
haha good catch i will remove it. mistake.
Zsédely Ruben
I don't know if it's intentional or not, but you shared your WANDB API Key in the json.
