Suggested LORA training settings

#3
by Hellomattieo - opened

What settings are you using when training these LORAs?

yes please answer.

for z-image it is between 0.85 and 1.1

you know that each sample that you see in the browser is a working comfyui workflow? so you can take a look and reuse it :)

cheers!

I think they were asking about the settings you were using to create the LoRA itself. I was coming to ask something similar. Would you consider a mini tutorial on how you format the dataset, as well as the AI Toolkit settings you are using, or if it's a different software entirely?

Here are my training settings (I use AI Toolkit) -> https://civitai.com/articles/23158/my-z-image-turbo-quick-training-guide

As for datasets, there is no dedicated articles but in articles regarding training tutorials I mention some stuff here and there so you could take a peak at some of them -> https://civitai.com/user/malcolmrey/articles

But in short, for Z Image. I use around 25 images and for that I pick 2500 steps (for 40 images I would set 4000 steps, and so on). I make sure that there is only one face visible in the image, the images are not blurry, the faces are clear and representing of the person. I set in AI Toolkit the 512 training res (it uses buckets, so the images do not have to be 512x512)

Here are my training settings (I use AI Toolkit) [snip]

Great writeup, thank you!

Sign up or log in to comment