FrameMorrow: Future-guided Frame Selection with Prospective Tokens for Long-Horizon Video Generation
Bo Yin, Xiaobin Hu, Jiaqi Zhao, Shuicheng Yan
National University of Singapore · Harbin Institute of Technology (Shenzhen)
Published September 30, 2026 · on the feed October 5, 2026 · ▲ 112 · arXiv 2609.38839
The abstract
Long-horizon video generation requires models to effectively leverage an increasingly long generation history. As the generated history grows, retaining all previous content becomes increasingly expensive and redundant, making effective historical selection essential. Existing approaches often determine historical relevance based on the current content. However, information relevant to the present is not necessarily useful for future generation, while seemingly less relevant history may become important later. Our key insight is that historical information should be selected according to its relevance to future information needs. Capturing these needs does not require generating the full future; instead, a compact representation of what becomes important next is sufficient to guide historical selection. Building on this insight, we propose FrameMorrow, a prospective frame selector that predicts a small set of prospective tokens representing future information needs and uses them to identify relevant information from history. FrameMorrow selects 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. We evaluate FrameMorrow across five benchmarks and 11 generative models spanning long-video generation, interactive generation, and action-conditioned world models. Extensive experiments demonstrate consistent improvements in long-range consistency, visual quality, and action alignment across diverse generation settings.
The Brief
2:29A short podcast summary of the paper. Written and voiced by AI.
Transcript
FrameMorrow. Future-guided Frame Selection with Prospective Tokens for Long-Horizon Video Generation. By Bo Yin and three others, from the National University of Singapore and the Harbin Institute of Technology, Shenzhen. Published on the thirtieth of September. Its code repository is public but slight: two source files.
A model generating long video has to make use of an ever longer history of what it has already generated. Keeping all of it grows expensive and redundant, so some selection from that history is essential. Existing approaches usually decide what is relevant by looking at the current content. The authors see a mismatch in that. Information relevant right now is not necessarily what future generation will need, and history that looks unimportant now may matter later. A selector that only asks what matches the present frame can throw away exactly the piece of the past that the video is about to need again, and keep material that will never be used.
The key insight is to select history by its relevance to future information needs rather than present ones. That raises an obvious worry: knowing the future would seem to require generating it. The authors argue it does not. A compact representation of what is about to become important is enough to guide which parts of the past get kept. The future does not have to be drawn in full; it only has to be described well enough to know what to look for.
FrameMorrow is a prospective frame selector built on that idea. It predicts a small set of prospective tokens, which stand in for future information needs, and uses them to pick out the relevant information from the history.
A design choice makes it portable. FrameMorrow selects explicit historical frames, not internal states specific to one model. Frames are something any video generator works with, so the selector plugs into diverse generators, closed-source models included, with little additional cost at inference.
The evaluation covers five benchmarks and eleven generative models, spanning long-video generation, interactive generation, and action-conditioned world models. That breadth is the test of the plug-and-play claim, since the same selector has to serve many different generators. Across these settings, extensive experiments show consistent improvements in long-range consistency, in visual quality, and in alignment with actions.
Choosing what to remember is reframed as a question about tomorrow rather than today, and the name says as much. A few tokens sketch what the next stretch of video will need, and the past is searched with that sketch in hand. Because what comes back is ordinary frames, the same selector can sit in front of models it was never built around.
Paper Quest · 7 chapters · about 39 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
Long videos forget. The question is what to remember.
You are generating a long video. As it grows, earlier subjects and scenes fall outside the model's limited input window, and keeping the whole history is costly and redundant.
The road ahead
- 1 · The problem ~6 min A long video outgrows the model's input window, and choosing which past frames to keep by matching the current screen often keeps the wrong ones.
- 2 · What came before ~3 min
- 3 · The key idea ~8 min Pick past frames by what the upcoming video will need, using a few predicted tokens as a cheap stand-in for that future.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Predict a few prospective tokens, score each past frame against them, keep the top frames, and pass them to the generator.
- 6 · The experiment ~6 min You’ll guess: Across many generators and benchmarks, which qualities did the future-guided frame selector improve?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.