Future Chunk Prediction trains autoregressive video models on later chunks
Original title: Looking Ahead: Modeling Future Evolution in Autoregressive Video Generation
The auxiliary objective predicts separated future chunks without intervening frames and does not change autoregressive inference.
The paper argues that teacher forcing lets autoregressive video diffusion models use temporal redundancy as a shortcut instead of modeling how scenes evolve. Future Chunk Prediction is an auxiliary loss, trained with teacher-forced denoising, that predicts temporally separated chunks without intervening visual context. Supervision comes from the training videos, so no extra visual encoder is required, and inference stays unchanged. The authors report better physical plausibility at a lower overall training compute budget.
Unconfirmed: Physical-plausibility gains and the structured representation claim come from the authors' experiments and visualizations.
- FCP predicts temporally separated chunks without intervening visual context.
- Supervision comes from training videos, with no extra pretrained visual encoder.
- Authors report better physical plausibility at a lower training compute budget.
- Visualizations suggest a more structured representation space for world models.
From the original article
1 min read at huggingface.co
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.
By Ruopeng Gao, Yukun Huang, Jiarui Hu · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.