Using my collections in training Gen AI locally

[Kjetil Kjernsmo]

Hi all,

I'm wondering if anybody has done anything around taking a text-to-image GenAI 
model, fine-tune or train it with your own pictures, and generate images from 
that?

I have a Digikam managed collection of tens of thousands of images, most of 
which I've tagged and rated through the years in a neat taxonomy, so now, it 
would be interesting to use those tagged images as input to such a model. I 
wouldn't share my pictures beyond my firewall (for now), so I'd like to do it 
locally.

I have looked around, and seen similar stuff being done, but not quite what I 
was looking for. There is DreamBooth: https://dreambooth.github.io/  but 
that's a fine tuning over a small number of images for a specific purpose. 
My idea is to add to the model locally, which I suppose is more characterized 
as more training than fine-tuning, but I'm not too deep into these things yet.

The hardware I have is a 6-core Intel(R) Core(TM) i7-8700 CPU with a Intel® 
UHD Graphics 630 and 32GB of RAM, and I would be getting a video card with 10 
GB of VRAM too, which I realize is not a lot, but what I can currently afford.

It would be great to hear if anyone has done something similar!

Kind regards,

Kjetil