Video adds a dimension stills never had: time. Widen the machine's memory — how many past frames it gets to look at — and the flipping pancake stays a pancake; shrink it to one frame and the subject melts into nonsense.
Live preview · launch for the interactive version
Set how many past frames the model can attend to — its short-term memory.
Past frames ghost behind the current one, like an animator's light-box.
At window 1 the subject morphs; widen it and the motion locks into a coherent flip.
Predict what a one-frame memory does to a moving subject.
Change only the temporal window; keep playing.
How do the onion-skin ghosts line up at each setting?
Say it plainly: coherence over time is attention, not magic.