A self-supervised AI model can now treat the "flow of time" as a visual concept that it can manipulate and control.
April 24, 2026
Original Paper
Seeing Fast and Slow: Learning the Flow of Time in Videos
arXiv · 2604.21931
The Takeaway
Most AI models process video as a static sequence of frames, but this new approach treats time as a distinct variable. The model can detect when a video's playback speed has been changed and can generate new footage at any specific speed. This allows for a level of temporal control in video generation that was previously impossible. By understanding the physics of fast and slow, the model can create more realistic movements in its outputs. This development moves AI from just generating images to truly understanding the dynamics of motion. It provides a powerful new tool for film editors and game developers.
From the abstract
How can we tell whether a video has been sped up or slowed down? How can we generate videos at different speeds? Although videos have been central to modern computer vision research, little attention has been paid to perceiving and controlling the passage of time. In this paper, we study time as a learnable visual concept and develop models for reasoning about and manipulating the flow of time in videos. We first exploit the multimodal cues and temporal structure naturally present in videos to l