奖励强制：基于奖励分布匹配蒸馏的高效流式视频生成
Reward Forcing: Efficient Streaming Video Generation with Rewarded Distribution Matching Distillation
用 EMA 融合被逐出滑动窗口的 token 替代静态 sink token，解决少步蒸馏流式视频生成中初始帧复制、运动退化的问题；再用 VLM 打分构造奖励偏置的分布匹配蒸馏，把教师模型的运动动态蒸馏进学生模型，可直接迁移到任何流式/长视频扩散蒸馏管线。
原文依据 we propose a novel Rewarded Distribution Matching Distillation (Re-DMD)
摘要译文
由模型翻译 · 术语统一 · 请对照原文核对关键结论高效的流式视频生成对于模拟交互式与动态世界至关重要。现有方法通过滑动窗口注意力对少步视频扩散模型进行蒸馏，并以初始帧作为汇 token 来维持注意力性能并减少误差累积。然而，视频帧会过度依赖这些静态 token，导致初始帧被复制、运动动态性下降。为解决这一问题，我们提出奖励强制（Reward Forcing）框架，包含两项关键设计。第一，我们提出 EMA-Sink：它维护固定大小的 token，这些 token 由初始帧初始化，并在被逐出的 token 离开滑动窗口时通过指数移动平均进行融合、持续更新。在不增加额外计算成本的前提下，EMA-Sink token 同时捕获长期上下文与近期动态，从而在保持长时程一致性的同时避免初始帧复制。第二，为更好地从教师模型中蒸馏运动动态，我们提出一种新的奖励分布匹配蒸馏（Re-DMD）。原始分布匹配对所有训练样本一视同仁，限制了模型优先关注动态内容的能力。Re-DMD 则通过优先选择由视觉语言模型评定为动态性更强的样本，将模型输出分布偏向高奖励区域。Re-DMD 在保持数据保真度的同时显著提升运动质量。我们通过定量与定性实验表明，奖励强制在标准基准上达到最优性能，并能在单张 H100 GPU 上以 23.1 FPS 实现高质量流式视频生成。
英文原文摘要 1635 字符
Efficient streaming video generation is critical for simulating interactive and dynamic worlds. Existing methods distill few-step video diffusion models with sliding window attention, using initial frames as sink tokens to maintain attention performance and reduce error accumulation. However, video frames become overly dependent on these static tokens, resulting in copied initial frames and diminished motion dynamics. To address this, we introduce Reward Forcing, a novel framework with two key designs. First, we propose EMA-Sink, which maintains fixed-size tokens initialized from initial frames and continuously updated by fusing evicted tokens via exponential moving average as they exit the sliding window. Without additional computation cost, EMA-Sink tokens capture both long-term context and recent dynamics, preventing initial frame copying while maintaining long-horizon consistency. Second, to better distill motion dynamics from teacher models, we propose a novel Rewarded Distribution Matching Distillation (Re-DMD). Vanilla distribution matching treats every training sample equally, limiting the model's ability to prioritize dynamic content. Instead, Re-DMD biases the model's output distribution toward high-reward regions by prioritizing samples with greater dynamics rated by a vision-language model. Re-DMD significantly enhances motion quality while preserving data fidelity. We include both quantitative and qualitative experiments to show that Reward Forcing achieves state-of-the-art performance on standard benchmarks while enabling high-quality streaming video generation at 23.1 FPS on a single H100 GPU.