Version 2:

IMPORTANT: The trigger word has changed! kxbrie woman

A bigger dataset, with:

• bigger and more consistently large bust

• better labeling

• more steps in training

• cleaned up dataset, consisting of new images generated from lora v1 including some topless images, and artifacts assiduously removed

• a trigger word that doesn't include numbers or underscores

The good:

• No more jpeg artifacts; images are clean and crisp

• bust is consistently large (though not consistently the same size; remains sensitive to prompting)

• the trigger word isn't printed all over the place

• still good adherence to poses

The bad:

• You might get some nipples you didn't ask for, I can think of a couple places that labeling might have led to this. I'm already considering a v2.1. But this is only happening to me about 1 in 5 images or so.

• If you start to specify anything about her bust, it also grow a whole lot

• It's very picky about what her skin looks like (e.g.: asking for a tan does not work, and the sample photo with the yellow outfit should have ink on her face). I'm sure I can weaken this as well.

• The likeness tends to apply to the image and bleed into other people generated in the same image (this was probably true of v1 as well).

Version 1

This model's data set is smaller and lower quality than Chloe (my first character LoRA). I've noticed some artifacts appearing in this character's images that she does not suffer from. I also suspect using an underscore in the trigger word kx_br13 woman wasn't the best idea, as it tends to print it everywhere.

About Brie

A character for Krea 2 created from a generated dataset.

Consistent dark rimmed glasses are present in almost every training photo. If you prompt for glasses, expect them to look similar.