Clarification on Random Sampling Interval in Section 4.1.1 of HunyuanVideo Paper

[koking0]

Thank you very much, HunyuanVideo, for your excellent work! I am currently studying your paper, but there is one part I don’t quite understand and hope to get clarification on. In Section 4.1.1 "Training," it mentions:

“To improve the reconstruction of high-motion videos, we randomly choose a sampling interval from the range 1 ∼ 8 to sample frames evenly across video clips.”

What does this mean? Does it mean randomly sampling several segments from a long clip and then stitching them together as a sample? If so, wouldn't this lead to inconsistency in the subject?

[KimbingNg]

Thank you for your interest. Please note that the sampling interval is between 1 and 8, whereas the original videos typically have a relatively high fps (e.g., 25). Even in the sparsest sampling scenario (with a frame interval of 8), the adjacent frames are still within a comparatively close time interval, and thus the consistency will not be lost.

[koking0]

Thank you for your interest. Please note that the sampling interval is between 1 and 8, whereas the original videos typically have a relatively high fps (e.g., 25). Even in the sparsest sampling scenario (with a frame interval of 8), the adjacent frames are still within a comparatively close time interval, and thus the consistency will not be lost.

Thank you for reply, I understanding!