提出自采样训练方法，解决自回归视频生成中的暴露偏差问题。
End-to-End Training for Autoregressive Video Diffusion via Self-Resampling
- 训练时模拟推理错误，用自采样修复历史帧偏差
- 支持长视频生成，时序一致性优于蒸馏基线
- 无需教师模型，适合大规模端到端训练
自回归视频扩散模型在世界模拟中具有潜力，但易受训练-测试不匹配引发的暴露偏差影响。现有方法虽通过后训练缓解此问题，通常依赖双向教师模型或判别器。为实现端到端训练，本文提出无教师框架「自采样强制」（Resampling Forcing），可从头训练且支持大规模训练。核心是自采样机制，在训练中模拟推理时的历史帧误差。基于退化历史，稀疏因果掩码保持时间因果性，并支持帧级扩散损失的并行训练。为提升长时序生成效率，进一步引入无需参数的历史路由机制，动态检索每查询最相关的前k个历史帧。实验表明，该方法性能接近蒸馏基线，且在长视频上表现出更优的时序一致性，源于原生长度训练。
原文摘要 · Abstract (English)
Autoregressive video diffusion models hold promise for world simulation but are vulnerable to exposure bias arising from the train-test mismatch. While recent works address this via post-training, they typically rely on a bidirectional teacher model or discriminator. To achieve an end-to-end solution, we introduce Resampling Forcing, a teacher-free framework that enables training autoregressive video models from scratch and at scale. Central to our approach is a self-resampling scheme that simulates inference-time model errors on history frames during training. Conditioned on these degraded histories, a sparse causal mask enforces temporal causality while enabling parallel training with frame-level diffusion loss. To facilitate efficient long-horizon generation, we further introduce history routing, a parameter-free mechanism that dynamically retrieves the top-k most relevant history frames for each query. Experiments demonstrate that our approach achieves performance comparable to distillation-based baselines while exhibiting superior temporal consistency on longer videos owing to native-length training.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。