Model introduction

This is the lora model of Sailor Mars from Sailor Moon serie.

There are already a lot of other loras trained for this characters, but i've found that most of them have some inconsistency in the outfit that i wanted to fix, so here we are.

This lora is a composite lora made by merging Sailor Mars lora by Koronen (https://civarchive.com/models/1051701) and my own one trained on 40 images carefully selected and captioned to achieve the best results.

How does it work?

Let's start by saying that illustrious xl already has some understanding of the character, so it was easy to produce a model which would get the job done, but for the same reason we have different keywords to consider when prompting with this lora:

• character:

sailor mars, hino rei        <-- those are the keywords to produce the character

• additional tags for more stability (character):

magical girl,
long hair, bangs, black hair,  
purple eyes, 
breasts, medium breasts

• Outfit:

sailor senshi uniform,

circlet, gold circlet, red gemstone, earrings, star earrings, red earrings,
choker, star choker, star (symbol), red choker,

sailor collar, red sailor collar, leotard, sleeveless leotard, folded sleeves, white leotard,
brooch, heart brooch, bow, purple bow,
gloves, elobow gloves, white gloves,
back bow, red bow,
skirt, miniskirt, pleated skirt, red skirt,

This model support nsfw content (to remove the outfit you can simply use the keyword "completely nude")

• Style:

bishoujo_senshi_sailor_moon_\(style\)
retro artstyle,
takeuchi naoko,
1990s \(style\),
bishoujo senshi sailor moon, 

Eachone of those tags can affect the style of the image and is related to this lora.

The strongest impact comes from "bishoujo senshi sailor moon \(style\)", but due to faulty training it didn't come out as expected so it's strongly recommended to not use this (or even using it in the negative if you want to try some custom style (see gallery for some examples)).

Recommendation

Keep the lora weight between 0.5 ~ 1. (0.7 is the sweet spot, lower to 0.5 to use with other styles)

Somehow it works with realistic models aswell, but the output is not too pleasing so it's not suggested to use it with those.

Highres fix is suggested, and adetailer might be needed in some niche spot.

Conclusion

It has been a while and even though i am happy with the state of this model i might train a v2 in the future to achieve even better control over the output