Closed SlavyanDesu closed 1 year ago
Thanks for the reply @noriwaru. Yes, I already put the prompt/tag on Dreambooth Config, so do the caption file for each image.
For your info, here are my token settings:
instance_token="mika"
class_token="1girl"
Here are my example of 1.txt
:
1girl, mika, halo, solo, wings, smile, hair bun
The original sd-script names the learning folder "number of iterations, instance_token and class_token" (e.g. 10_john girl). However, if there is a text file in the folder, the folder name is ignored and the tags entered in the text are used for training. kohya-trainer Tested. 5.2 Dataset Config Dreambooth Config train_repeats:10 reg_repeats:10 instance_token: car class_token: dog Text tags - train image= cpc, reg image= scs train image is circular, regularized image is rectangular
The generated results were drawn as a circle in cpc and a rectangle in scs. In other words, when using text files, I think that instance_token: car instance_token: dog is ignored when using text files.
I see, so that means it's OK to leave instance_token
and class_token
empty and stick to .txt
files?
That's what I mean.
Thanks for your answer! @noriwaru
Is
instance_token
same asinstance_prompt
? If so, let's say I want to create LoRA for Misono Mika and the activator tag ismika (blue archive)
. It's the same tag as Booru does. Do I still need underscore as a space or just keep it? Thanks in advance.