There is now a
, trained on all the original 90 images used in training, but also an additional 90 synthetic images (see post above, which has a sample of some of these).
It's more of a complementary LoRA than an 'upgrade', since real data is always better. Nonetheless, the V2 Leia can do some things that V1 can't, even if V1 may sometimes be better for inpainting faces (though not always).
I haven't done that much with it yet. Most of the initial examples below are simply the work of others who used the original LoRA I made, but not with Carrie Fisher's face. I have modified these to resemble Carrie Fisher. Some are pure gen with inpainting.
At this stage I don't feel there is any further I can go with SD V1.5 Leia LoRAs. But I hesitate to spend so much time this year trying to master SDXL. I will likely delay that until the tools really mature (Kohya defaults suck) or until something better comes after.