Training details

I'm trying to reproduce this work, during training, I divided the video into multiple segments, and summed and averaged their losses, but the video generated by the model trained in this way has obvious breaks in the connection of each segment, I would like to know how the author handled it during training, and what details need to be paid attention to?

No file, test, or entry point is named in the issue. Start by locating the training and video-segmentation code, then document how segment boundaries and loss aggregation are handled; done means the training details needed to avoid visible breaks are clearly explained.

Written by the indexing model from the issue text.