Looking Ahead: Modeling Future Evolution in Autoregressive Video Generation

One-line takeaway

Tencent ARC Lab adds Future Chunk Prediction (FCP), an auxiliary objective that predicts temporally separated chunks during teacher-forced autoregressive video diffusion training, improving physical plausibility at lower training compute with inference unchanged.

腾讯 ARC 实验室提出“未来片段预测”（FCP）辅助目标：自回归视频扩散训练时额外预测隔开若干片段的未来画面，迫使模型学习场景如何演变；物理合理性更好、训练总算力更低，推理流程不变。