Long-horizon video generation requires keeping track of an ever-expanding history, but conditioning on a growing token sequence is computationally costly. Existing approaches often restrict or compress historical context, which can discard details needed for coherence.
The arXiv paper "Recast the Past: Stable Long Video Generation with Full Context Memory" addresses this by proposing a method that retains full context rather than compressing it. The title suggests a strategy of "recasting" past information to maintain stability over long generations.
While the abstract is brief, the core contribution appears to be a memory mechanism that avoids the trade-off between cost and fidelity. This could improve consistency in long video outputs.