Sekhmet of Death (Base SDXL)

Basic attempt at creating a lora from a small dataset.

Trained on the base SDXL 1.0 model, using ClearVAE.

Generates decently with be base model, generates EXTREMELY well with most anime tuned checkpoints. bluePencilXL in particular produces some really nice results, some of them even almost SFW!

(Base model WILL NOT generate the pasties, so she is always topless. More training will fix it, but I kinda can't atm. Well, I can, but at a super slow speed. Like 24 hours to fully train a lora with only 18 images.)

I am... stunned how well it works, considering it was trained in less than an hour in 4 epochs on a free google colab T4 (Until they realized I was training SDXL and politely asked me to stop T_T)

Trigger is sekhmet_of_death (with the underscores) This works to get a general look, but to get the full character, the flowing tags must be used as well: dark skin, hime cut, huge breasts, egyptian clothes, revealing clothes, pasties, single thighhigh <-- Not really needed, but can help. Huge breasts is only needed half the time. Seems the only detail not trained was the hime cut. (Though it always adds the gild trim properly, so it was at least partway there.)