The motion fluency of the reconstructed video is very poor, similar to the frame skipping effect of the original video.

[hyc9]

Thank you for your work! I tried to use your training code on the UCF dataset. I skipped the training of the image data, only used the video data for training, and kept the other settings the same as the source code. However, I found that the model trained this way had a frame-skipping effect when reconstructing the video. When I slowed down the video, I found that the adjacent 4 frames were basically the same, without smooth motion. However, when I experimented with the publicly available checkpoint model, there was no such effect. I'm very puzzled by the reason for this phenomenon, and I hope the author may help. Thanks for your attention!