Discussion of training without root information

[yuyujunjun]

Thanks for your brilliant work!

From my perspective, if there is no root-related loss, the generator might ignore the content encoder because there is no constraint to the content encoder.

The training might teach the style encoder to capture all the motion information, including the style and the content, and teach the decoder to reproduce the motion given by the style latent.