Training images suddenly flipped.

[Randy-H0]

Hi, so I was training a stylegan2-ada model on a bunch of Discord images. You'll see below what discord exactly is.

Now it went smoothly until it suddenly just, flipped.

This was before it flipped.

This was after it flipped.
Why did this happen? It never happened on my previous models I was training.

Thanks!

[Randy-H0]

Clarification, all of the images in my dataset are oriented the same way.

[towr]

I think that can happen when the augmentation probability grows to high. There's an example of it in the paper at the top of page 4 ( https://arxiv.org/pdf/2006.06676.pdf )