DecMem: Towards Minute-Long Consistent World Generation with Decoupled Memory

Researchers introduced DecMem, a decoupled memory architecture for long-horizon video generation.

It tackles computational inefficiency and attention dispersion with Sparse Global Memory and Anchored Local Memory.

Sparse Global Memory improves efficient access to long-term context, while Anchored Local Memory stabilizes local extrapolation.

Experiments show DecMem outperforms prior methods on consistent, high-fidelity long video generation.