HuggingFace Daily Papers·· 1 天前
Future Chunk Prediction：提升自回归视频生成对未来场景演化的建模能力
Looking Ahead: Modeling Future Evolution in Autoregressive Video Generation
AI 导读
研究者提出 Future Chunk Prediction（FCP），一种与 teacher forcing 去噪联合优化的辅助目标，通过预测间隔时间块、不依赖中间视觉上下文来强化模型对场景演化的建模。FCP 直接从训练视频取监督信号，无需额外预训练视觉编码器，且不改变自回归推理流程。在更低训练算力预算下，其物理合理性优于标准 teacher forcing 自回归训练。
正文 · 原文
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.
| Comments: | Project page: this https URL |
| Subjects: | Computer Vision and Pattern Recognition (cs.CV) |
| Cite as: | arXiv:2610.29989 [cs.CV] |
| (or arXiv:2610.29989v1 [cs.CV] for this version) | |
| https://doi.org/10.48550/arXiv.2610.29989 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Ruopeng Gao [view email]
[v1]
Wed, 21 Oct 2026 16:28:46 UTC (30,710 KB)
来源：HuggingFace Daily Papers · arxiv.org