Preparing a LoRa (low Rank) dataset

In my previous article, I asked to prepare a dataset of a few pictures, to train a LoRA model

This is actually paramount for successful results later, and therefor I will go deeper into that process

First gather about a dozen pictures of the same subject, this could be a character, a celebrity, or yourself (to create your avatar)

choose various background, various clothes, various face expression, angles etc ... the more divert the better

then resize the pictures to 512x512 pixels, and save it with no too much quality loss 

the name of the pictures should be 1.jpg, 2.jpg, 3.jpg etc ...

After this is done, open a text editor and create 1.txt, 2.txt, 3.txt for each pictures

In that file describes the pictures, using a keyword that will be the LoRa keyword (its name) make it simple and unique, something like Gdlf for Gandalf or Hptr for Harry Potter 

then add the variations made in the picture, the hair style, the background, the angle, the clothes etc ... separated by comma.

Here is an example:

1.jpg is

1.txt will contains:

cbpk, white hair, looking at viewer, leather jacket, yellow abstract background

The LoRa will be cbpk.safetensor (for Cyberpunk)