After the tests, it was decided to return to the 512x512 resolution for training. This gives less distortion in some cases.

The training was conducted in one epoch (5000 steps). Which is equal to 5 epochs of old versions.

This model is suitable for most anime-like models.