This lora can turn any scene into a night scene, indoors and outdoors. Not just darker, but with flashlights, car lights, etc. Some scenes will need a stronger weight than others. So adjust as needed.
This is an experiment in a training process concept I am working on I call seed locking. Initial results looked good, but it still needs work. The basic idea is to generate training images and regularization images with a specific seed, then alter the training images with the desired effect while maintaining image consistency, and then retrain with those specific seeds to target very specific concepts you want to adjust and keeping everything else the same. It is not perfect because of the sampling methods. I hope to find a way to regularize for the sampler in the loss, hopefully soon.
But for now, enjoy the experiment.
Description
No trigger words necessary. Weight of 1 works well for most images. Increase and decrease as needed.
FAQ
Comments (5)
Seed locking! Interesting concept! Did you come up with that yourself?
I believe so, though I wouldn’t be surprised if there is a paper out there describing something similar. I was honestly surprised when I found out that diffusion models didn’t have repeatable noise for training images so every prompt image combo you train on would always have the same seed. I just assumed that is how it worked.
@Ostris Very smart. Keep going!
Does your code is public?
veryy good!