>>62757844Flip augmentation is always off when I train vtubers because Asymmetry is like the biggest vtuber feature. Aono Nemo's model almost always covers the right eye so this should as well. Maybe something in the dataset was confusing but the other one also came out undertrained.