Future chunk prediction improves video evolution modeling in autoregressive generation
Looking Ahead: Modeling Future Evolution in Autoregressive Video Generation
Computer Vision and Pattern Recognition
Summary
Videos made by computers are often created one small piece at a time, but just making each piece look okay doesn’t mean the video shows realistic change over time. The authors show that predicting a future chunk of video without seeing everything in between helps the model understand how scenes evolve better. Their method works without extra tools and uses less computing power while producing more believable video movements. This approach may help build smarter systems that imagine or simulate how things change in a scene.
What this means in practice
- •For video game developers: Generate more realistic animations by better modeling how scenes and objects evolve over time without additional computational overhead.
- •For animation studios: Produce physically plausible video sequences by improving temporal consistency in computer-generated footage using future chunk prediction.
Authors
Ruopeng Gao, Yukun Huang, Jiarui Hu, Shuai Wang, Wenbo Hu, Ying Shan, Limin Wang
Abstract
Autoregressive video diffusion models generate videos chunk by chunk, yet producing plausible continuations does not guarantee that they capture how scenes evolve. Under teacher forcing, strong temporal redundancy may allow local visual continuity to serve as a shortcut for current-chunk denoising, reducing the pressure to model visual evolution. We introduce Future Chunk Prediction (FCP), an auxiliary objective jointly optimized with teacher-forced denoising that predicts temporally separated chunks without access to the intervening visual context. Drawing supervision directly from training videos, FCP requires no additional pretrained visual encoder and leaves autoregressive inference unchanged. Our method achieves superior physical plausibility over standard teacher-forced autoregressive training with a lower overall training compute budget, highlighting the effectiveness and efficiency of future-chunk supervision for modeling visual evolution. Intriguingly, our visualizations suggest that FCP fosters a more structured representation space, which may be valuable for building autoregressive world models.