Trained on Tsubaki, 14 images.
Because I trained her LoRA on images that had both human and "elf" ears, it can be sometimes inconsistent, but you can prompt it, if you desire either human or elf ears.
Nevermind, I didn't check if I used her LoRA :) She generates completely fine.
I use Euler A. 28 steps. Because the results are pretty good, I think you can generate 3 images in parallel, and you will still be fine. Unlike SDXL, DiT is pretty consistent from my initial experience. Of course, if you demand a lot of detail, then maybe go cheap first.
I use 3, 4.5 and 6CFG. 0.7 Rescale CFG, same as every DiT V1 LoRAs I use.
Thumbnail model: Tsubaki
