A simple LoRA. Trained at 768x768 on high resolution images.

Once in a while you'll encounter floating , or missing limbs. I think you'll get the best generations at 512x768, but there is a lot of potential using the "wide shot" tag with 768x768, 768x896, or 896x896 images!

Caucasian/ and shallow/deep work better in version 2.

As usual, modify the LoRA weight itself if needed.

Note: make sure to add "open mouth" to the negative if needed. Tons of the training images displayed the woman in /surprise, so their mouth was always open.

Check the release notes on each version for example prompts.