Long-horizon video generation faces a fundamental tension: memory must stay bounded so generation can continue indefinitely, yet the model must still recover what it observed when the camera returns to a previously seen place. A new arXiv paper tackles this by proposing a slot memory architecture addressed by region tokens.
The idea is to store past observations in a fixed set of memory slots, with region tokens acting as the addressing mechanism. When the camera revisits a location, the relevant region tokens retrieve the corresponding slot content, allowing the model to recall what was there without growing memory without limit.
The abstract frames this as a way to satisfy both competing requirements simultaneously. No experimental results are given in the abstract, but the approach offers a concrete direction for making video generation scalable to arbitrarily long sequences with coherent spatial memory.