LoRA for Maka Albarn from Soul Eater.

Capable of nsfw.

I trained it with 252 pics, which was a mix of fan art and screencaps. I trained it at 20 epochs with a dim/alpha of 32/16 and with 1 repeat.

Leans to a semi-real look mostly, but somewhat more anime at higher LoRA weights. Let me know how it goes for you. Read stuff below on guidelines and known issues :)

Triggers:

Input "MakaAlbarn, 1girl" after the LoRA near the start of the prompt.

Input "light brown hair, green eyes" for more consistency if needed, but I didn't find it necessary.

Input "plaid skirt, gloves, necktie" for classic clothes.
Input "sweater vest" for her non-combat clothes.

NB: can place in negative prompts if the clothes are stubbornly appearing.

Notes:

My experience has shown scale 0.9 ~ 1.4 works best; my best results were 0.9 for semi-realistic, but overall favourite was 1.1.

My prompt order shown in the sample pictures worked best for me.

Works on different models though weights may need to change for good results.

I used Anything4.5 model, Anything3.0, anyLora, and Midnight Maple model.

Refer to my pictures for examples and what upscaler I used, keep in mind I have used other textual inversions to help with poses and negatives.

Description

FAQ

Comments (4)

wheelinghubcapApr 9, 2023

CivitAI

Hey, what's your tagging approach when training? Also, what model did you train with?

reevee

Author

Apr 10, 2023

Hi😊 I used this model for training https://huggingface.co/OedoSoldier/animix/resolve/main/animeScreenshotlikeMix_fp16.safetensors , this is my first upload that used it.

My tagging is done with ( https://github.com/toriato/stable-diffusion-webui-wd14-tagger.git ).

Before captioning all pictures, I load and interrogate one picture of the character with a threshold of 0.25 using wd-v1-4-swinv2-tagger-v2. I then find all tags that are going to be essential to the character in all scenarios. Eg: if a girl has orange hair and green eyes then that would be consistent when generating her, so I add those to the list of tags associated with the character.

Then I add the list of tags always associated with the character and put them in the box for "tags to be excluded" essentially, I don't want any of my captions to contain any of those tags. Then add a trigger prompt in a box to append tags to my captions. That way, all those now missing tags in my captions will be entirely represented by that one trigger tag.

Now I batch process all captions for the images at 0.35 threshold and get the txt files.

wheelinghubcapApr 9, 2023

CivitAI

"keep in mind the position of your <LoRA> matters."

How so? The LoRA weight adjustment is either applied or it isn't, and the LoRA text is erased from the prompt.

reevee

Author

Apr 10, 2023

Thanks for pointing that out I appreciate it.

I honestly remember reading that somewhere and when I tried it for myself it seemed to be the case at the same seed. Hence why I believed it.

I tested it out and you're right it doesn't seem to change anything, maybe I had hirez fix on when I tested the LoRA postion thing.

Thanks for pointing it out though, I'm going to remove that statement from my description to avoid confusion :)