Originally published on CivitAI by Stardeaf.
https://civitai.com/models/556691/bettie-page-type-character
Trigger Words: smp_bettie
This is my first exercise in making character Lora. Bettie Page was famous model in 50s and is no longer alive. This one is not supposed to capture real look of Bettie Page, but the fantasy type she created. So it doesn't actually depict a real person. Sort of. It is supposed to catch key attributes, hairstyle, eye color (blue), eyebrow shape, makeup.
SD1.5 and SDXL models are (usually) well aware of the concept. Though in some XL finetuned models it seems to be overwritten with something else. I wanted to see how would it translate to Pony model.
Lora is based on synthetically created images. I rendered in XL, edited, used i2i in Pony or XL... I also used initially trained Lora to create images in Pony and finetune them in XL. In the end I trained about 6 versions until I got to point where it was somewhat satisfying. It looks like despite all, Lora will assume some quality attributes of source images, which makes historic figures a difficult subject.
My set was about 30 images, some in cartoon/comic style, some realistic style, but not quite a photo. Rather quality digital art. I used "cartoon"/"realistic" to tag them. You may try to use those tags.
I used 1600 steps with 0.0006 learn rate, constant. I'm not quite sure where to settle, but it seemed a middle ground in my small trial and error batch. Feel free to school me in comments. I used Invoke training. Both u-net and text encoder training were ticked. I tried a version with text encoder disabled, as recommended in kohya script documentation. But I didn't like the result.
I was able to use batch size of 2, which filled 24GB of VRAM almost entirely. On my pervious card 12GB was not enough for batch size of 1.
It works with some checkpoint better than with others. Generally not so great with "photographic" finetunes. Working weight range 0.6-1.2, depending on circumstances. Have fun.
