Echo-Forcing: A Scene Memory Framework for Interactive Long Video Generation

Researchers introduced Echo-Forcing, a training-free memory framework for interactive long video generation.

It splits memory into stable, recent, and compressed historical states, while adding structured scene recall frames.

A difference-aware forgetting mechanism helps models handle smooth transitions, hard cuts, and long-range scene recall.

The approach addresses common failures in long-video generation, such as forgetting earlier scenes and reacting poorly to prompt changes, without exceeding a fixed cache budget.