Trigger words are "ponytail" and "copper hair color".
Use strength between 0.80 and 1.20
Sampler: Euler
Guidance: 7.5
Description
2nd attempt at putting though LoRA training.
FAQ
Comments (5)
I think your problem is
1: The lack of unique trigger word.
2: Lack of consistent tagging. This looks like a lot of it was just auto-label only. SD needs to be taught the articles of clothing as they apply to the character. Only styles can really get away with pure auto-label, everything else needs some degree of hand tagging.
Thanks for the feedback. I'm just starting out on making prompts and learning. 😊
@dannyboi588922
If you haven't seen it, I have a guide for LoRA creation. I consider it pretty good (naturally).
The good news on the tagging front is you won't need all 88 pictures. Most of what I've found, and I've found correct (all my LoRAs list the training data quantity to help demystify it) , is that for a character with one outfit and one "appearance" (e.g., no alternate Super Saiyan form, post time-skip appearance etc.) in PDXL, 20 is the recommended minimum and past 30 is diminishing returns on value.
It also helps to use existing booru tags in your tagging and only use unique stuff for triggers or when absolutely needed. "long hair in ponytail" isn't really a tag (prompt will pick out "ponytail", which is a tag, but imperfectly).
@dannyboi588922
For examples sake, I've given a quick go at tags since her outfit is simple enough.
Calista_STAS, red hair, ponytail, green ponytail holder, earrings, gold earring, lipstick, pink lips, yellow sweater, miniskirt, green skirt, pleated skirt, ankle socks, white socks, sneakers, green footwear, yellow shoelaces
You would include all the applicable tags on each image, plus whatever else you see in the image and can tag. Alternate outfits would each get their set of tags (I'm not sure how many pics an alternate outfit needs after the main outfit, but it's somewhere less than 10 unless its super complex) .
Calista Ultrabrilliante High School Alumni C/O 2004
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.


