Glad you like it :D

The dataset contains images that I don't have permission to distribute, but here is my training settings https://files.catbox.moe/z5u6de.json

Most of the images can be found on danbooru however. One thing I've been doing is upscaling my dataset with automatic1111. So for example, any image with a side shorter than 2048 (arbitrarily chosen) I'll scale up to 2048x2048 (crop to fit = false). Under the extras tab you can batch it out by directory. RealESRGAN x4 Anime worked extremely well for this set. I feel that doing this at least helps keep the gens sharp when upscaling with hires fix or img2img.

For captioning, I just prefix with my trigger word and let kohya_ss auto caption.

And for this lora I did 10 repeats and 10 epochs