[Rafael Bravin]

Hi!

So, this is what I got so far:

Create my own Lora after some trying and then I went and made another one soon after. The first was a Lycoris LoKr and the second was a standard Lora type, and I used different images as dataset on each one. I’m trying to get consistently the same face in every image, basically the same thing the Loras of famous people do, only my character is fictional. So I used the thecniques described in this website to get face consistency and then used the generated images as dataset for both my Loras.
And it worked! I followed something like this tip here and I am now getting the same person in a somewhat consistent matter.

—–/

The problem.
Most of my training images are of close up photos and now when I add my Lora and my Lycoris I lose the freedom of creation the base model grants, for instance if I prompt “beautiful woman sitting on an office desk” SD only generates model like photos without any desk. How can I mix the power of the Loras with the freedom of the model? Did I fail in something when creating the model?

[Andrew]

Hi! This is a common issue. You simply don’t have a training images for small faces.

You will need to paint the face big to get the face. Try turning on Hi-Res fix. Try upscaling the picture 2x, then inpaint the face with the lora.

[Author]

Posts