Aisha (von Daphnel), the Violet Princess* from King's Raid
trained on NAI, strength varies from 0.7-1.0, recommended strength is 0.9, higher for slightly more accuracy, lower when doing other outfits.
training done locally with Prodigy, 10 epochs, 21 images, 5 steps for a total of 1050
trigger word is aishakr.
additionnal tags:
drill hair, twin drills - for her signature drills
black dress, frills, black gloves, crown, bare shoulders, detached sleeves, detached collar, purple flower, purple ribbon - her "usual outfit"
large breasts - for more accuracy though even without it SD makes them fairly large usually.
yellow eyes - if for some reason they're changed to blue or purple
negative tags:
two crowns, multiple crowns, greatly reduces the chances it fumbles and make it two
a few things:
her outfit seems to be an absolute nightmare for AI to comprehend, with a lot of elements sporting multiple colors, I went for a compromise between amount of tags and flexibility so it results in some inaccuracies and is unevitable, as she has very little fanarts, most being of very low quality.
the LoRA can fight you a bit due to the fairly small dataset, notably she loves holding things in one hand, and her outfit tends to creep back, best advice is lowering strength a bit and putting the tags that are creeping back in your negative prompt.
despite not being trained with shuffling left and right, it still tends to mess up with her crown.
images were not exactly cherry-picked, they were picked out of a 20 images batch for each prompt then upscaled and went through ADetailer.
==========================================================
This is the very first LoRA I upload, and the very first character I wanted to train (when I joined Civitai 1 year ago, so ig it's kind of a birthday celebration?).
There was no LoRA of her, and now there is one, which is something I guess.
dataset: 3 fanarts, 7 images from in-game illustrations(including crops), 11 images from in-game screenshots(including crop)
credits/mentions:
Civitai(duh)
@holostrawberry - whose excellent written guide got me into LoRA training
@novowels - for sharing their Mirianne dataset which helped me a whole bunch.
*even though she's actually a duchess, I don't get it either.