Long-horizon video generation faces a growing challenge: as the generation history expands, retaining all previous content becomes increasingly expensive and redundant. Effective selection of historical frames is therefore essential, but existing methods typically judge relevance based on the current content alone.
The authors argue that information relevant to the present is not necessarily useful for future generation, and seemingly less relevant history may become important later. Their key insight is that historical information should be selected according to its relevance to future information needs.
To capture these needs without generating the full future, FrameMorrow predicts a small set of prospective tokens that represent what will become important next. These tokens guide the selection of explicit historical frames rather than model-specific internal states, enabling plug-and-play integration across diverse generators, including closed-source models, with little additional inference cost.
The method was evaluated across five benchmarks and 11 generative models spanning long-video generation, interactive generation, and action-conditioned world models. Experiments reportedly show consistent improvements in long-range consistency, visual quality, and action alignment across diverse generation settings.