Better Generative AI Video by Shuffling Frames During Training
A new paper out this week at Arxiv addresses an issue which anyone who has adopted the Hunyuan Video or Wan 2.1 AI video generators will have come across by now: temporal aberrations, where the generative process tends to abruptly speed up, conflate, omit, or otherwise mess up crucial moments in a generated video: https://www.unite.ai/wp-content/uploads/2025/03/examples_temporal_issues_bordered_1-1.mp4…