通过自重采样实现自回归视频扩散的端到端训练
End-to-End Training for Autoregressive Video Diffusion via Self-Resampling
AI总结：
提出自回归视频扩散模型的端到端训练框架Resampling Forcing，通过自重采样模拟推理错误，结合稀疏因果掩码和动态历史路由，实现长视频生成的时间一致性。
摘要：
自回归视频扩散模型有望用于世界模拟，但容易受到训练-测试不匹配导致的曝光偏差的影响。虽然最近的工作通过后训练来解决这一问题，但它们通常依赖于双向教师模型或判别器。为了实现端到端解决方案，我们引入了Resampling Forcing，一个无教师框架，能够从头开始并大规模训练自回归视频模型。我们方法的核心是一个自重采样方案，该方案在训练期间模拟推理时历史帧上的模型误差。在这些退化历史的条件下，一个稀疏因果掩码强制实现时间因果性，同时通过帧级扩散损失实现并行训练。为了促进高效的长时程生成，我们进一步引入了历史路由，一种无参数机制，为每个查询动态检索最相关的top-k历史帧。实验表明，我们的方法在性能上与基于蒸馏的基线相当，同时由于原生长度训练，在更长视频上展现出优越的时间一致性。
英文摘要：
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.
机构、总结、中文摘要均由AI分析而来，可能存在错误，请以论文正文为准。
