Image Diffusion Models Exhibit Emergent Temporal Propagation in Videos

Here’s the plot twist no one saw coming: models built to make pretty pictures just learned to follow stuff in videos. Researchers say the attention maps inside image generators can “pass the label baton” from frame to frame, meaning zero‑shot (no extra training) object tracking. They even packaged it into a new system called DRIFT, with a helper tool that cleans up outlines, and it’s clocking state‑of‑the‑art scores. The crowd? Split between mind‑blown and what does any of that mean. One top comment begs, “Can someone smarter than me explain?” while another gushes about how hobbyists keep squeezing smarts from 2022-era Stable Diffusion models like they’re toothpaste that never ends. The big drama: Is this genuine “emergent intelligence” or just clever rebranding of math we already had? Hype vs. homework, in classic tech-forum fashion. Privacy nerves flare too—if a photo model can track your dog across a video, what about your face at a protest? Cue jokes: “Finally, my cat’s chaotic zoomies get a highlight reel,” plus memes about Skynet learning to stalk coffee cups. The vibe: unexpected superpower unlocked, with the usual internet cocktail of awe, confusion, and a dash of surveillance side‑eye.