Theoretical Question for learning and creating images

Beginners
1

Hello,

I am using kohya_ss (LORA) - stabilityai/stable-diffusion-xl-base-1.0.

I have a dataset of 300 images made from say 1000 sub images which are stitched together to give original images.

My question is:

  1. Do I train on 300 images and generate the output image directly?

  2. Or do I train on 1000 sub images, generate the output sub images which are then stitched together to give final output image by a tool like Adobe Photoshop?

Please advise. Thank You.

1 Like
2

If you are trying to reproduce abstract features such as the pattern (e.g. the touch or the color) of an image, then a larger number of images is more reliable, but if you are trying to reproduce the shape, then 100 images is more than enough for LoRA learning.
Even 10 images will produce a reasonably similar result.

So I think option 1 is fine.

1 Like