Reproduce “Looking Ahead: Modeling Future Evolution in Autoregressive Video Generation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Ruopeng Gao, Yukun Huang, Jiarui Hu, Shuai Wang, Wenbo Hu, Ying Shan, Limin Wang
The authors introduce Future Chunk Prediction (FCP) for autoregressive video generation, adding independent predictors that use intermediate features to predict later chunks without intervening visual context. In a controlled Wan2.2-TI2V-5B image-to-video comparison on Physics-IQ Verified, FCP raises the score from 30.4 to 33.8 points versus teacher forcing. This supports improved physical plausibility on evaluated benchmarks, not general scene understanding.
Key insights
- Predicting farther ahead without intermediate visual context encourages representations to capture information about scene evolution, complementing ordinary next-chunk denoising.
- In the reported ablations, combining clean-context and noisy-target features worked better than using either alone; parallel predictors also outperformed chained prediction.
- The method improves selected physical-plausibility results without changing inference, though the evidence does not establish general world-model capabilities or long-horizon behavior.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.