And now for something different!

I noticed a few face-only Lora's a while ago, and decided to do an absurd and somewhat dumb experiment: I took 1,000 images (you used to have a max of 2,000 on this site, damn you CivitAI!) of this ladies' cropped 512x512 face, and trained a Lora on it. I was expecting a disaster, but it's actually sort of okay.

Keep in mind a few things:

1) This lora will create her with a realistic and detailed face, but her body will have typical air-brushed Flux qualities, which looks comical. It oddly picked up on the fact that she's petite, but otherwise knows nothing about her body. It also seems to believe she has a really big bust size.

2) This was simply an experiment. I used 1,000 photos with 5 repeats, for a total of 5,000 steps. I'm uploading epoch 1: 1,000 steps because it had diminishing returns on higher steps (though possibly a crisper image). Why 1,0000 images? Because I'm a moron and felt like throwing away money and time.

3) I did not use natural language tagging. I'm forever done using natural language tags, I always get better results with a trigger word only (specifically trigger word + gender).

4) Given the concentration on her face, that includes blemishes and all, so don't expect smooth skin at all times. Real women aren't air-brushed. On the plus side of this, it's unlikely you'll need a face detailer (but consider using one for the body)

5) If she looks too old, simply put the word "young" in the prompt. I used photos of her from her early adult career to present, and Flux seems to understand the age in the photos.

Note: Normally it pisses me off when people merge work without credit, but for this, as it's an experiment, do as you wish (for non-commercial purposes). I'd be really interested if somebody managed to make a separate body lora that goes well with this (I've noticed the petite body type lora works well with this but only if you use a weight of 0.2 or less).