Trained for 1800 steps on:
eyes--
with https://github.com/ntc-ai/conceptmod
Then reversed with --save_opposite
animations with the lora value -1.0 to 3.0 and the trigger eyes
Description
FAQ
Comments (8)
What is this exactly?
It's a lora trained on the concept of eyes. It doesn't use data, instead it uses stable diffusion's own internal representation of eyes to further train itself.
The '--' trick means that it tries to delete the concept of eyes then reverses the resulting network. This ends up being more reliable than other techniques I've found so far.
The code is open source and there is a way to train easily with runpod (see the fire tutorial) if you'd like to experiment.
So this will just make eyes better in general?
@ntc So I basically use this as a Lora to make the characters eyes come out better?
@9ball I think it will focus on the eyes or add eyes to the scene. I'm not sure that the eyes get better. It depends on the internal representation of eyes in stable diffusion.
It seems to zoom in on the eyes. So if you want an eye closeup, use a stronger value of this.
The eye quality is great.
I usually never comment on anything, but because this method is so absolutely stunning, I had to.
This is what i Called Masterpiece Eyes Ever