FrameMorrow: Future-guided Frame Selection with Prospective Tokens for Long-Horizon Video Generation
FrameMorrow picks past frames by predicting what the generator will need next, not what matches the current frame.
The paper argues that long video generation gets expensive and redundant when models keep too much prior history. Its proposed selector creates a small set of “prospective tokens” to represent future information needs, then uses them to choose explicit historical frames. Because it selects frames rather than internal model states, the authors say it can plug into different generators, including closed-source ones, with little added inference cost. They report gains across five benchmarks and 11 models in consistency, visual quality, and action alignment. HF Daily Papers' note
The paper argues that long video generation gets expensive and redundant when models keep too much prior history. Its proposed selector creates a small set of “prospective tokens” to represent future information needs, then uses them to choose explicit historical frames. Because it selects frames rather than internal model states, the authors say it can plug into different generators, including closed-source ones, with little added inference cost. They report gains across five benchmarks and 11 models in consistency, visual quality, and action alignment. HF Daily Papers' note
score 5