V2 UPDATE: Bigger and better model with a few tags fixed and updates in the dataset as well as tweaked training settings to improve quality. Updated the name and category to reflect the fact that the model is fully capable of producing the style, (and a few other characters to some degree) not just one character.
It is now more consistent in producing everything, including the other tagged characters from the show and just the style without any characters. ALL EXAMPLE IMAGES ARE WITHOUT ADETAILER/FIXERS/ETC so you can judge the lora output for yourself.
NOTES FOR V2:
1)SOME TRIGGER WORDS HAVE CHANGED! CHECK EXAMPLE IMAGES!
main ones are that the main clothes traits tags got spaces removed (ex: orange dress => orangedress, etc) and main style trigger is now "sastyle" can use "fanart" in positives for more style variability or stick it in the negative for more consistency with show screencaps
2)it seems like the model sometimes produces unprompted slit pupils for some reason now on small/medium faces. This is usually fixable with "slit pupils" in the negative, but that might reduce overall eye quality so only use it if necessary.
3)as the strength of all keywords has increased, it may be advisable to place the ones you do NOT want to show up in the negative if you are noticing them bleeding in or strange behavior in general.
Hope you like it!
Trained on base pony so it should be compatible with most merges of it.
Can produce Tarte in several outfits and kind of knows what her spear is. There was a bit of lugh, maha, and dia in the training data too, along with some concepts from the show, but seems like it was not enough to produce them consistently.
Main triggers for V1:
saTarte, (twintails OR long hair OR hair rings), blonde hair, (orange dress OR maid, mob cap, maid apron etc, OR nightgown OR pink bikini OR servant dress, brown sweater, turtleneck, brown skirt etc OR assassinSuit, black clothes, black bodysuit, elbow gloves, thigh boots etc), satSpear (for her spear, recommended to use "holding satSpear" for her holding it more consistently)
also kind of understands her younger version:
saTarteY, blonde hair, short hair, optional: blue dress, green jacket, dirty
might want to put ahoge in the negative as it likes to give her one for some reason.
You can try to also use saDia, saDiaY, saLugh, saMaha, saMahaY, along with character appearance descriptions, but the training on these is very weak and unlikely to work too well.
There are other tags I used in training that did not get properly recognized, but you can experiment with if you care: windCowl, hiddenSpear, manaAura, etc
This is my first try making a lora, so I just randomly picked a character from the list of things I watched.
General format is to use sa[Character name here] plus their traits and appearance to trigger, or just "sa" for show style (works with no trigger too, but weaker and Tarte will still occasionally bleed in without negative as the lora is heavily trained on her specifically). add "Y" at the end for their younger versions from flashbacks and start of the show. Trigger for Tarte's spear is satSpear and it seems to work well as "holding satSpear" if you want her to actually have it in her hands.
Check example images for more trigger words and examples.
I will probably make a v2 of this incorporating what I have learned from my first attempt to tweak tagging, training parameters, trigger words etc
Hope you like it!