FrameMorrow: Future-guided Frame Selection with Prospective Tokens for Long-Horizon Video Generation

The paper presents FrameMorrow, a prospective frame selector that predicts a small set of prospective tokens representing future information needs to guide historical frame selection. By selecting historical frames according to their relevance to future needs rather than just the current content, FrameMorrow enables plug-and-play integration across diverse generative models, including closed-source architectures, with minimal inference overhead. Experiments across multiple benchmarks and generative models demonstrate consistent improvements in long-range consistency and visual quality.