Monday, 5 October 2026 · 6 min read
FrameMorrow selects historical video frames by predicting future needs for better long-horizon generation
FrameMorrow improves long-horizon video generation by predicting compact prospective tokens that represent future information needs, guiding selection of relevant past frames. This plug-and-play method enhances consistency, quality, and action alignment across 11 models and 5 benchmarks with minimal overhead.
Based on the paper FrameMorrow: Future-guided Frame Selection with Prospective Tokens for Long-Horizon Video Generation by Bo Yin, Xiaobin Hu, Jiaqi Zhao, Shuicheng Yan
Challenges in long-horizon video generation and historical frame selection
Long-horizon video generation requires models to maintain coherent subjects, objects, and scenes over extended sequences, but the model's limited input window means earlier generated content gradually falls out of scope, risking inconsistencies in appearance and scene revisits.
Retaining all past frames is computationally expensive and redundant, so selectively preserving historical information that is relevant for future generation is essential.
Existing methods typically select historical frames based on their relevance to the current content, but this can miss frames that appear less relevant now but are important for future generation steps, motivating a future-guided selection approach.
Sign in to read the whole paper
You have read the opening. The rest is free with a Bezen account, and no card is needed.
In the rest of this paper
- How the method works, step by step, with a flow chart
- Three key formulas from the paper, explained symbol by symbol
- Two figures from the paper, and how to read them
- How it was tested, what was found, and where it falls short