About the number of frames for `future_images`

[bugmakerwww]

Thanks for this excellent work!
I have view the released code and noticed that the number of frames for future_images is always 0;
And it seems that the data loading process configures different history frames, like [-5, 0] or [0];

I was wondering:

1. for the video_gen task, what is the rational behind using only a single future frame? What issues would arise from adding more future frames during training and inference?
2. why are varying numbers of history frames used?

Looking forward to your reply!

[KaiLiu18]

Thank you for your interest in our work and for taking the time to look into the released code.

Regarding the video_gen task, our current released version predicts only a single future frame. The key consideration is computational efficiency.

As for the varying numbers of history frames, these configurations were mainly introduced for ablation purposes. In our experiments, we found that the different history-frame settings, such as [0] and [-5, 0], did not lead to a significant difference in performance. Therefore, we use [-5, 0] as the default setting.