This was an interesting LoRA to create. I decided to take my usual realism approach but also played around with the process a bit. An interesting note is that bad_prompt_version2 brings out some interesting styling in the LoRA. It makes artifacts in the image that make it look more like it is a photo and not AI-generated.

Another thing to note is that her hair can be a bit weird. Using hair over one eye in the negative prompt works well for straightening it out in relation to her face if the model you are using is giving you issues with that. I'm pretty sure this has to do with how she does her hair and how it sits in most of the training data. And on that note, short hair helps keep her hair length just above her shoulders. Prompting for longer hair will depend more on the base model you are using.

For the keywords, Michele Alves is good for general images. portrait will give you the type of image you want consistently, though not using it just adds more of her body to the image. freckles will emphasize freckles in the image.

This LoRA works best with about 50 steps with a vertical resolution of 768. It was trained at 1024x1024 with 14200 steps and most images close to that resolution. It can also do more full-body shots with varying degrees of success.

Of the sample images, most were generated using epiCRealism. The more anime-looking one is from dreamshaper_8 and the one with lower-res eyes is AsianRealistic_SDLife_Chiasedamme.