Introducing: Alani!

I'd give you some detailed explanation of how I was a follower of Hurian myth and her name means death but uh...I just googled, "cool girl names that mean death" and this came up so shurg.

My favorite character I've ever made is also the most ...temperamental. Some tips:

Doing 1girl, Aik-4l4n1 will give you the character pretty much all the time. As soon as you start to add more tags she'll begin to drift. I've trained her about 6 or 7 times now and I'm still working on it (lol just like I'm still working on recaptioning Jacky for her Flux LoRA). But, by and large, she...works-ish.

So what is she? In DAZ3D there is a character called Faceless which is so good and so cool that I wanted to bring it to AI. If you have DAZ3D I will plug this product because it's really amazing : FPE Faceless is the product name go get it. So, this is not REALLY an original character but what can you do.

She has a mannequin like body and a drawn on face. She can be prompted for NSFW if that's your thing but tbh I wasn't thinking about that too much, I just thought she was cute.

Tips

Alani works really well with "emoji" tags like X X, O O, ^ ^
Put "skull" in the negative as the AI continuously thinks she should have a skeleton face.
The more prompts you add the more you'll need to prompt to get her features back, white skin, white hair, drawn on face, should all help.
Putting "scribble" and "symbol" in the negative can help clean up her look if you find her face too messy for your liking.
She has sharp teeth by default as she's a monster. It is a feature, not a bug.

I'm going to do something I don't usually do and I've included the training data for Alani. If anyone wants to have a go at training her better than I did (she's woefully overfit), have a crack at it. Just let me know what you come up with as I really love the character I just have difficulty finding time to run trainings.

Description

Steps taken:

Lowered the learning rate
Increased epochs
Took a more "low and slow" approach
Upped LORA Rank from 32 to ??
Upped LoRA Alpha from 16 to 32

Result:

Default 1girl, Aik-4l4n1 will produce a much cleaner face without the need for negative prompting.
More flexible - particularly in backgrounds
Can use full strength or lower and still maintain her look without need for additional prompting to a point

Remaining Issues:

Based on the training dataset using many shots that focus on the face, the model tends to not do full body shots well and has a desire for portraits. This can be overcome by dropping the strength but then additional prompting is needed to get her look back.
Still somewhat inflexible - I'll have to recheck my original captions. I've never had a model this reticent to pop off a plain background

Tips:

Negative prompts: Simple background
Drop strength to 0.8 and take the hit in tokens to prompt out whatever features disappear for now

Random thoughts

I am trying to remember to document my training steps - feel free to yell at me if I don't. I think the metadata may be included in the LoRA but that may only be true if it was generated on Civit. I suppose I have a billion blue buzz I could give it a whirl.
This is a VERY touchy training I suppose because the output isn't technically a person. The lower epoch save is too "washed out" and the next one up begins to display some heavy artifacting one sees in overtraining. Dragging it out overnight has helped to pinpoint a better save at least.

LORA

Illustrious

by aikage

Download (Beta) View on SeaArt CivitAI Version