Brandy loves cats.
Showcase images are at a weight of .9. It's pretty forgiving though.
Training info
There are 58 core images and 164 supporting images. Core images were used 5 times each epoch, and supporting images used once per epoch.
Captioning done with wd14.
I've included both the Kohya configuration json and all of the images used to generate this lora as zip that accompanies this model.
It is my hope that if you are making your own loras, you can use it see what I chose to select as core images vs supporting images and improve your own work.
Description
Brandy was the first model that I started using supporting images with. I wanted to revisit her and apply some changes to my protocol.
I went from a training batch size of 1 to 2. This should kind of mix the two images so they learn together. It lowers the total number of steps as well.
I also changed the captions. I'd used straight wd14 before, but this time I trimmed out all of the things that define her core features (aside from the hair, which I've left malleable).
I pushed this training to its limits, which is why it's 2.1. I only did 14 epochs comprised of 3178 steps, and wanted to continue training. I performed another 6 epochs for an additional 1386 steps to come to the epoch I selected for this version.
I used the same images for both versions. Instead of a 2 and 5 split, I went for 1 and 5.
There are 164 supporting images and 58 core images. So about 300 core iterations, and half that in supporting iterations.