When generating new videos during video detection inference, why are the new video frames black and empty? What could be the reason for this?