The world is being quietly rearranged by people who write very long documents.


The title they went with VERTIGO: Visual Preference Optimization for Cinematic Camera Trajectory Generation Noisy translates that to

AI can now fix its own bad camera shots by watching the footage it creates


Researchers built a system that lets AI camera directors see what they're filming and fix problems in real time. Instead of generating camera motion and hoping it frames characters correctly, the AI now renders what it would shoot, gets scored on whether it matches the director's intent, and adjusts until the framing works.
For years, AI systems that generate camera motion have been trapped in a one-way pipeline: text prompt goes in, camera trajectory comes out, nobody checks if the shot is actually watchable. The result is lots of off-screen characters and compositions that look technically correct but cinematically useless. This system closes that loop by making the AI watch its own work and grade itself. It means the difference between AI that generates plausible motion and AI that actually produces shots a cinematographer would use.
Watch whether film and game studios start using this approach in their production pipelines, and whether character off-screen rates stay near zero when it's used on longer, more complex scenes than the research tested.

If you insist
Read the original →