How to Use
I specifically trained this to not be overfit. You can mix and match her outfit & appearance, change her body shape or features, and it won't override any styles or other loras you've got on. You can even generate her as a real person in cosplay.
The only things hard-baked into the lora are her face/hair, her headset (optional) and her outfit (optional).
Lora strength: 1.0 - 1.2
Usually 1.0 is fine, but some checkpoints need a little higher
Try at 1.0 first then adjust up a little if necessary
See below for some specific finetunes I've tested
Prompting:
There are three keywords you can use. The lora works with plain language OR tags.
Anby Demara
This is the required phrase you should have in every prompt
It is case sensitive
Use it as normal in a sentence, like "Anby Demara is sitting on a chair" or whatever
special outfit
This gives you her normal outfit from the game
It's case sensitive, all lowercase
It's optional, you must include this term if you want her to wear her game outfit
You can overwrite parts of the outfit, like by specifying she's wearing pants instead of her normal skirt, and it will still work
To use: just put it in a normal sentence, like "she's wearing her special outfit"
Or if you only want part of it, you can write something like "she's wearing the upper half of her special outfit, and on her lower half she's wearing baggy green pants"
headset
This gives you her headphones
It's case sensitive, lowercase
It's optional, you must include this term if you want her to be wearing her headphones
If you're wondering why I trained it with the term "headset" instead of "headphones" it's because "headphones" was conflicting with Krea 2's base training data in weird ways
To use: just put it in a normal sentence, like "she's wearing her headset and casual clothes"
Putting it all together, if you want Anby with all of her normal stuff you'd write something like this:
A medium shot of Anby Demara standing on a street in a casual pose, wearing her headset and special outfit.Or if you're just doing tags, something like this:
Anby Demara, headset, special outfit, standing, street, casual pose, medium shotOther Prompt Tips
You can get her normal neutral/expressionless face by adding something like "she has a calm impassive expression" or "her expression is neutral" to the prompt.
Her canon body shape is "petite" and she has "small breasts". You can nudge her shape in that direction using those terms, or use whatever other terms you want if you want her to have something more specific or non-canon.
Issues / Downsides
It wasn't trained with multiple characters, so her appearance will bleed into other female characters in the scene
You can partially reduce this by being very specific about the other character's appearance
With some prompts/finetunes it can be difficult to separate her jacket sleeves from her outfit, and they may force themselves into scenes annoyingly (thankfully not often)
Her appearance is less stable when you don't include her headset or special outfit (i.e. you only have "Anby Demara" in the prompt)
If it gets her hair color wrong, you can just specify that her hair is white or silver in the prompt and that will fix it
If it messes up the green tip on her braid, you can usually fix it by bumping the strength up by 0.05 or 0.1
It wasn't trained with her sword or backpack - at least, not very well
Her backpack will sometimes appear in shots (badly) when you don't want it to
You can ask her to have a sword, but it won't necessarily look like her in-game sword very well (low training volume)
Specific Finetunes
Below are a few finetunes I've tested this on and the lora strength that seems to work best for them.
Krea 2 base model: 1.0 strength
Muse by stable yogi, version 3.5: 1.0 strength
Muse by stable yogi, version 4.0: 1.1 strength, rarely 1.2 strength is needed
Kroma: 1.1 strength usually, but sometimes 1.2 strength
SeeKrea: 1.0 strength
Background Info
This is the first ever lora I've trained and I wanted to address the overtraining issues I've seen with a lot of other loras before. I also wanted it to be flexible so you could easily generate her without her outfit or headphones, and also not interfere with styles or other loras.
To do that I grabbed a varied set of around ~55 images, with a roughly even split between shots of her with & without her headphones, and with & without her game outfit - so that's four different buckets of images. I also made sure the art styles across the images were varied so it wouldn't bake in any styles. I then manually defined the training prompts for all of them, and ran a relatively long rank 16 training process of around ~65 epochs (roughly 3,500 steps), which was necessary because the training data was more varied than most people do.
I had to play around with it a lot to get to this point, but it worked out really well! It's very flexible, and it doesn't mess with styles or anything else unwanted.
More?
I'd like to make more of these for other ZZZ characters and other games/whatever as well. I'm also open to doing these as commissions for other characters, just PM me if you have a request.



















