Looking Ahead: Modeling Future Evolution in Autoregressive Video Generation
Predicting distant video chunks improved physical plausibility while using less training compute, the authors report.
Their Future Chunk Prediction objective asks the model to predict temporally separated chunks without seeing the visual context between them. It targets a shortcut in teacher-forced training: local visual continuity can help denoise a chunk without learning how a scene evolves. The objective draws supervision from training videos, requires no additional pretrained visual encoder and leaves autoregressive inference unchanged.
HF Daily Papers' note
Their Future Chunk Prediction objective asks the model to predict temporally separated chunks without seeing the visual context between them. It targets a shortcut in teacher-forced training: local visual continuity can help denoise a chunk without learning how a scene evolves. The objective draws supervision from training videos, requires no additional pretrained visual encoder and leaves autoregressive inference unchanged.
HF Daily Papers' note
score 5