What's this about?
WorldCrafter combines recent frames with a compact memory of earlier viewpoints to generate continued exploration from one image or text prompt.
What exactly is camera-queryable memory?
The requested viewpoint determines which old visual evidence is compressed for generation. It behaves like asking a map for only the details visible from a chosen direction.
Why does this matter?
Author-reported preprint; 49 HF upvotes at fetch time.
- Long video generation often forgets objects seen earlier.
- Fixed memory tokens control cost as history grows.
- Minute-scale demos do not yet establish persistent world physics.